AUC Undergraduate Journal of Liberal Arts & Sciences

Capstone Issue Vol. 12 2019

Published by AUC Undergraduate Journal of Liberal Arts & Sciences Capstone Issue Vol. 12 2019, published by InPrint The AUC Undergraduate Journal of Liberal Arts and Sciences is a biannual, interdisciplinary publication showcasing outstanding undergraduate academic papers. The Journal aims to demonstrate the strength of undergraduate scholarship at AUC, to reﬂect the intellectual diversity of its academic programme, to encourage best research and writing practices, to facilitate collaboration between students and faculty across the curriculum, and to provide students with opportunities to gain experience in academic reviewing, editing and publishing. The Editorial of the Journal is constituted of members of the InPrint board, a registered AUCSA committee.

Editorial board (InPrint)

Capstone Team

Dr. Maurits de Klepper | Capstone co-Coordinator Wade Geary, MA | Capstone co-Coordinator Bob Kardolus | Capstone Administrator

Series Editing

Prof. Dr. Murray Pratt Dr. Belinda Stratton

Copyright: All texts are published here with the full consent of their authors. Every eﬀort has been made to contact the rightful owners of all content with regards to copyrights and permissions. We apologise for any inadvertent errors or omissions. If you wish to use any content please contact the copyright holder directly. For any queries regarding copyright please contact [email protected]. Foreword

Before they graduate, AUC students are required to perform an independent research project within their intended majors of Sciences, Social Sciences or the Humanities: the Capstone. In the semester they spend writing their Cap- stones, students learn to engage and contribute to the academic dialogue occurring within their chosen ﬁeld. With the current Capstone Issue, we publish six Capstones submit- ted to us from across AUC’s three majors by the graduating students of 2019 which demonstrate AUC students’ diverse skills, interests and academic excellence.

All the Capstones published in this issue have undergone a rigorous selection and editing process carried out by our Editorial Board to improve the clarity and accessibility of the selected works. We’d also like to thank Dr. Lotte Tavec- chio and Dean Prof. Dr. Murray Pratt for their support and enthusiasm for our publication. Additionally, a thank you is due for the Academic Writing Skills and Advanced Research Writing teams, who through the support and inclusion of our journal in their classes have helped ensure InPrint’s longevity and presence within the AUC community. Finally, the template for the LATEX 2ε formatting for this issue was graciously provided by former Editor-in-Chief Lanie Preston.

I sincerely hope that in reading this Capstone Issue you, our dear reader, can see the high level of academic achievement of AUC students, their incredible work ethic, and the wonderful variety in interests and topics that is emblematic of the liberal arts and sciences education. And now, without further ado, I welcome you to our 12th Capstone Issue!

Aisha Erenstein, on behalf of InPrint

ii Contents

Sciences Implementing Quantum-Cryptographic Protocols using SimulaQron: A Quantum-Internet Simulation of 1-2 of Oblivious Transfer in the Noisy-Storage Model – Lynn Engelberts (1-32)

Sciences Restricted Boltzmann Machines and the Renormalization Group: Learning Relevant Information in Statistical Physics – Jesse Hoogland (33-68)

Social Sciences Do You Want to Know More? Curiosity for Positive and Negative Life Stories of Ingroup and Outgroup Members – Afke van Egmond (69-86)

Social Sciences The Political Strategy of Prawo i Sprawiedliwość – Christina Keßler (87-108)

Humanities - Social Sciences Between the World and Me(me): Approaching Internet Memes through Digital Ethnography and Berlant’s Aﬀective Present – Cullen Ogden (109-133)

Humanities Pieces of Resistance: Protest Signs as Objects of Dissent – Nina Klaﬀ (134-158)

iii Sciences

Implementing Quantum-Cryptographic Protocols using SimulaQron

A quantum-internet simulation of 1-2 oblivious transfer in the noisy-storage model

Lynn Engelberts

Supervisor Dr. Christian Schaffner (UvA, QuSoft) Reader Dr. Michael Walter (UvA, QuSoft) Capstone Issue Vol. 12, 2019 Lynn Engelberts

Abstract

Research in quantum technology is currently devoted to the development of a large-scale quantum network, also known as a quantum internet. Since a quantum internet is expected to be realised in the near future, a quantum-internet simulator, SimulaQron, has been created to function as a framework for the development of quantum-internet applications [DW18]. One of the expected applications of a quantum internet is the cryptographic primitive secure two-party computation. However, the suitability of SimulaQron for cryptographic applications, such as secure two-party computation, has not yet been examined. In particular, this thesis demonstrates the potential and limitations of SimulaQron as environment for implementing and analysing cryptographic protocols. Furthermore, we evaluate whether the considered protocols are suitable for quantum-internet applications. To the best of our knowledge, this thesis provides the ﬁrst implementations of quantum-cryptographic protocols in SimulaQron. Our implementations are based on the protocols for 1-2 oblivious transfer (1-2 OT) proposed in [Sch10]. 1-2 OT is a simple two-party primitive on which any protocol for secure two-party computation can be based [GV88; Kil88]. Assuming that the laws of quantum mechanics hold, the protocols in [Sch10] are shown to be secure against attacks in the noisy-storage model – a realistic cryptographic model which assumes that the adversary’s quantum storage is noisy. Our work is based on a critical literature analysis of 1-2 OT in the noisy-storage model and quantum internet, followed by an implementation of the protocols provided in [Sch10] using SimulaQron. Our implementations will allow us to evaluate whether these protocols are easy to implement and suitable for quantum internet. This research will thus contribute to both the ﬁeld of quantum cryptography and the area of quantum-internet development by analysing 1-2 OT in the noisy-storage model as well as evaluating the cryptographic applications of a quantum internet and the potential of SimulaQron.

Keywords and phrases: 1-2 oblivious transfer (1-2 OT), error correction, noisy-storage model, quantum internet, SimulaQron

Acknowledgements work is also known as a quantum internet [Cas18; DW18; DLH17]. Despite the fact that many possi- Firstly, I would like to thank my supervisor Chris- ble applications of a quantum internet have been tian Schaffner for his patience and guidance during suggested in research (e.g. see [Weh08]), the real this project. By introducing me to the area of quan- potential of a quantum internet will only be known tum cryptography and information theory, Chris- when it has been realised. Nevertheless, the pre- tian provided me with the background and notions sumed arrival of a quantum internet will require necessary for this Bachelor’s thesis. My grateful software that can be run over the quantum net- thanks are also extended to Axel Dahlberg for tak- work. Researchers have therefore created a quan- ing the time and effort to help me with any ques- tum internet simulator – called SimulaQron – with tions related to SimulaQron. Finally, I would like to the purpose to function as a test-bed for writing thank Michael Walter for taking the time to read my software for quantum internet applications [ALA18; work. DW18]. One of the potential applications of a future quantum internet is two-party cryptography I. Introduction [DW18; Weh08]. Cryptography is concerned with the tasks of achieving communication or computa- By 2020, the Dutch research centre QuTech tion between two or more parties, even if the par- aims to have built a quantum network between four ties do not trust each other [NC10]. A typical ex- cities in the Netherlands: Amsterdam, Delft, Lei- ample of cryptography is that of two parties who den, and The Hague [ALA18; DW18; QuT]. The ul- want to communicate securely over a communica- timate goal is to construct a large-scale quantum tion channel. We often refer to these two parties network so that quantum information can be ex- as ‘Alice’ and ‘Bob’, and to a (third) party attempt- changed between remote quantum processors, al- ing to intercept their messages as the ‘adversary’ lowing for both quantum computation and quantum or ‘attacker’. A cryptographic protocol then de- communication over an arbitrarily long distance. scribes how Alice and Bob can perform their task Such a large-scale quantum-communication net- securely. In classical cryptography, the security of

2 Capstone Issue Vol. 12, 2019 Lynn Engelberts

such cryptographic protocols is often based on the first model gives an explicit upper bound on the ad- hardness of an unproven mathematical conjecture versary’s quantum storage during the protocol, the and on the assumption that the attacker has limited latter assumes that the quantum storage is noisy computational power [DLH17; KWW12]. However, [KWW12; Sch10; STW09; Weh08], meaning that the security of these protocols could be affected if stored (quantum) information goes lost over time. the attacker obtains more (e.g. quantum) compu- The noisy-storage model is perceived as more re- tational power. For instance, the algorithm devel- alistic compared to the bounded-quantum-storage oped by Shor [Sho95] for integer factorisation on model as it reflects more accurately the prevail- a quantum computer could break RSA [RSA78], a ing technical difficulties of storing quantum infor- cryptoscheme widely used for securing the trans- mation [KWW12; Sch10], hence the paper will prefer of (sensitive) data. It is therefore desirable to dominantly focus on the noisy-storage model. aim for protocols that are information-theoretically Several protocols for secure two-party compu- secure, that is, the security of the protocol can be tation have been designed for the noisy-storage proven without restrictions on the attacker’s com- model, among which protocols for the primitive putational power [Sca+09]. called 1-2 oblivious transfer (1-2 OT), such as pro- Whilst the laws of quantum mechanics may vided in [KWW12] and [Sch10]. 1-2 OT is partic- allow for breaking currently used classical cryp- ularly interesting as it has been shown that any toschemes, these laws could also be exploited to protocol for secure two-party computation can be design quantum protocols for cryptographic tasks, based on this primitive [GV88; Kil88]. 1-2 OT is a which is the purpose of quantum cryptography. cryptographic primitive in which Alice has two se- In fact, information-theoretical security can be cret strings (messages, indicated by m) m0, m1 achieved in the case of quantum-key distribution and Bob has a bit c ∈ 0,1 and the goal is that (QKD)1, which enables two parties to establish a Bob receives mc corresponding to his choice bit c, shared secret key. Nevertheless, not all crypto- but Alice does not learn Bob’s choice c; likewise, graphic tasks can be performed with information- Bob does not learn Alice’s remaining string [GV88; theoretical security, even with the help of quan- KWW12]. The protocols for 1-2 OT in the noisy- tum mechanics. This limitation holds, in particular, storage model require Alice and Bob to be con- for secure two-party computation, which involves nected by a quantum channel. Since it is realistic to cryptographic tasks in which two parties want to assume that in any practical implementation of 1-2 exchange information but do not trust each other.2 OT such a quantum channel may be unreliable and More precisely, it has been shown that additional hence induce errors on the transmitted quantum assumptions on the attacker are inevitable in order information, protocols have also been suggested to allow for secure two-party computation [BCS12; that are robust against errors caused by the quan- Lo97; LC97; May97]. These assumptions need to tum channel, such as the one in [Sch10]. In par- be proven as secure and realistic to effectively re- ticular, both the idealised and the robust protocol strict the attacker’s actions. provided in [Sch10] are shown to be information- Consequently, much research has been de- theoretically secure against any noisy-storage at- voted to designing cryptographic models for secure tack, as long as the laws of quantum mechanics two-party computation protocols which introduce hold and the amount of noise in the quantum stor- such additional assumptions on the adversary. Two age channel is above a particular level. cryptographic models that have been proposed Since secure two-party computation is one of are based on physical ( as opposed to computa- the expected applications of a future quantum net- tional) assumptions, mimicking the technical diffi- work [Weh08] and any secure two-party computa- culties that one encounters when storing quantum tion primitive can be based on 1-2 OT, it is signifi- information: the bounded-quantum-storage model cant to evaluate the implementation of a protocol [Dam+05; Dam+07] and the noisy-storage model for 1-2 OT over such a quantum network. In par- [KWW12; Sch10; STW09; Weh08]. Whereas the ticular, since SimulaQron has been designed to de-

1For example, see [TL17] for a security analysis for QKD protocols. 2Note that, contrary to the previous example and QKD, Alice and Bob now do not trust each other, and thus the “adversary” can be either of the two. A concrete application of secure two-party computation is the so-called millionaire’s problem in which Alice and Bob want to ﬁnd out who is the richest of the two without revealing to the other how much money he or she has [Yao82].

3 Capstone Issue Vol. 12, 2019 Lynn Engelberts

velop software for future quantum-internet applica- Knowledge utilisation tions, it is important to verify its suitability for testing cryptographic tasks. Nonetheless, there are, This research provides new insights into the to the best of our knowledge, no implementations area of quantum internet development. Since a of any cryptographic protocols using SimulaQron. quantum internet has not yet been realised, the Therefore, the present paper aims to implement possible applications of a future quantum inter- both the idealised and the robust protocol for 1-2 net are still speculative. Therefore, by implement- OT from [Sch10] in SimulaQron in order to (i) ex- ing and evaluating one of the presumed quantum- amine whether the selected protocols are adequate internet applications (i.e. cryptography) in a sim- for quantum internet applications, and (ii) evaluate ulated quantum network, our research contributes the extent to which SimulaQron is a suitable en- to developing a deeper understanding of the suit- vironment for implementing and analysing crypto- ability of a quantum internet for such cryptographic graphic protocols. applications. Moreover, since there are, to the best of our knowledge, no previous implementations of cryptographic applications using the recently developed Methodology SimulaQron, the results of this research explain the potential of SimulaQron for developing such appli- The present research consists of a critical and cations. In addition, we shed light on the possibil- extended literature review, followed by an imple- ities and current limitations of SimulaQron in gen- mentation of the two protocols from [Sch10] using eral, which is of use for current and future users of SimulaQron [Sim]. SimulaQron and may contribute to a further devel- By critically analysing the existing literature on opment of this quantum-internet simulator. In fact, 1-2 OT in the noisy-storage model, this paper at- we may in this way also contribute to the develop- tempts to justify the choice for the protocols in ment of the quantum network in the Netherlands [Sch10] and to be able to provide a deeper under- that is planned to be realised in 2020, since a ver- standing of the protocol for implementation using sion of the CQC interface3 used by SimulaQron is in- SimulaQron. Moreover, in order to clarify the extra tended to be integrated into this quantum network steps for error-correction for the implementation of as well. the robust protocol in an unreliable quantum chan- Lastly, this research contributes to the ﬁeld of nel, it was necessary to study the error-correction quantum cryptography as there are, according to techniques that are available. An analysis of the the literature, no quantum-network implementa- existing literature on quantum internet and Simu- tions of the protocols in [Sch10]. The only imple- laQron was then used for understanding the work- mentations of these and similar 1-2 OT protocols ing of SimulaQron and to evaluate the potential of for the noisy-storage model are achieved by using SimulaQron. hardware for QKD, such as done in [Erv+14]. Fur- SimulaQron has been chosen as the environ- thermore, since our code is written in the high-level ment for our implementations as it allows for a sim- language Python and publicly available on GitHub, ulation of the quantum network between Alice and our code can be used to provide an illustration of Bob and hence enables (simulated) quantum com- the protocols for 1-2 OT for learning objectives. munication to be performed, which is required for implementing the protocols. In order to carry out Outline of the paper the implementations, we have used several tools provided by the Python Classical-Quantum Com- The remaining part of the paper is organised in biner (CQC) library that comes with SimulaQron the following way. Section 2 describes some termi- and made the implementations feasible. We have nology of quantum mechanics used in this paper. written our code in Python [Pyt] and hence used Section 3 provides an analysis of the noisy-storage several tools provided by Python and its libraries. model and a description of the protocol for 1-2 OT For the decoding part, we use the source code from in the idealised setting. Section 4 addresses how to [Fil], which is also written in Python. correct errors in the setting of an unreliable quan-

3See Section 5.

4 Capstone Issue Vol. 12, 2019 Lynn Engelberts tum channel and explains the resulting robust pro- probability 1. However, it may seem less trivial if tocol for 1-2 OT. Section 5 presents a background a qubit is in superposition: if the qubit is in state on quantum internet and SimulaQron. The results |ψi = α0 |0i + α1 |1i with α0 and α1 both nonzero, 2 of our implementations are provided in Section 6, we either get outcome 0 with probability |α0| or 1 2 which is followed by a discussion. Finally, we con- with probability |α1| . clude our work and provide some potential directions for future research in Section 7. BB84 states

The following four states are often referred to II. Background to quantum theory as the BB84 states, since these states were used in the BB84 protocol ? for QKD (named after the au- In this section, we provide a brief introduction thors Charles Bennett and Gilles Brassard and the to quantum mechanics and focus on the concepts year in which it was proposed). that are relevant for the remaining part of this the- 1 1 sis. We assume that the reader is familiar with the |0i |1i |+i = √ (|0i+|1i) |−i = √ (|0i−|1i) basic notions of linear algebra. We use the deﬁni- 2 2 tions provided in [NC10]. As aforementioned, {|0i , |1i} is called the computational basis. We will refer to {|+i , |−i} as the Qubits Hadamard basis.

Recall that a classical binary bit either has value Operators 0 or 1. The quantum version of the bit is often called a qubit, short for quantum bit. Similar to a classi- An operator acts on a quantum state and en- cal bit, a qubit can be in state |0i or |1i, where |i ables us to change the state of the corresponding is the frequently-used Dirac notation. However, the qubit. An operator is defined as a linear transforma- 2 main difference is that a qubit can be in any linear tion on the Hilbert space C . By the laws of quan- combination of these two states. tum mechanics, an operator U must be unitary, i.e. More precisely, we define a qubit as a vector in U †U = I, where I represents the identity, and U † 2 a two-dimensional complex vector space C with in- means the conjugate transpose of U. Below, we ner product, also called a Hilbert space. The states provide two examples of operators, which will also be used in the implementations of this paper. 1 0 |0i = and |1i = The first operator we consider is denoted X. It 0 1 acts as follows on the BB84 states: form an orthonormal basis for the Hilbert space, X |0i = |1i, X |1i = |0i, X |+i = |+i, and X |−i = often referred to as the computational basis. We − |−i. In other words, if the qubit is in one of the write the most general quantum state as: |ψi = computational basis states, the X operator swaps the states, whereas if the qubit is in one of the α0 |0i + α1 |1i, where α0, α1 ∈ C are normalised so 2 2 Hadamard basis states it remains in the same state that |α0| + |α1| = 1. Such a state is often called a superposition of states |0i and |1i. Moreover, we (with ket− mapped to − |−i). Note the connection to the logical NOT gate. call αi the amplitude of state |ii for i ∈ {0, 1}. An operator can be represented by a matrix. The matrices corresponding to X and H are: Measurement postulate 0 1 1 1 1 Qubits act in a probabilistic manner when mea- X = H = √ . 1 0 2 1 −1 sured. According to the Copenhagen interpretation of quantum mechanics, measuring a quantum state irrevocably changes its state: after measur- III. 1-2 oblivious transfer in the ing an arbitrary quantum state |ψi, if the outcome noisy-storage model is i, then |ψi is collapsed (i.e. changed) to state |ii. Clearly, if the qubit is in state |ψi = |0i (i.e. As mentioned before, 1-2 oblivious transfer (1-2 α0 = 1, α1 = 0), then we get outcome 0 with OT) is a cryptographic task belonging to the field of

5 Capstone Issue Vol. 12, 2019 Lynn Engelberts

secure two-party computation on which any other est party has a quantum memory of size at least n protocol for secure two-party computation can be 2 , where n is the number of qubits transmitted dur- based [GV88; Kil88]. Oblivious transfer was ﬁrst ing the protocol [Dam+05]. Moreover, the ﬁndings introduced by Rabin [Rab81] and has been gener- of [Dam+05] have been extended to the case of alised to 1-2 OT by Even, Goldreich, and Lempel 1-2 OT: 1-2 OT is shown to be secure against a dis- [EGL82]. We recall that in the setting of 1-2 OT, Al- honest player whose quantum memory is of size n ice has two secret strings m0, m1 and Bob chooses at most 4 − 2l [Dam+07]. Here, n amounts again a bit c ∈ {0, 1}. The task that 1-2 OT needs to ac- to the number of qubits that are transmitted dur- complish is that Bob receives Alice’s input string mc ing the protocol, and l (which may be a fraction of corresponding to his choice bit c, without revealing n) is the length of Alice’s input strings m0, m1. In c to Alice and without learning about Alice’s other addition, [Dam+05; Dam+07] point out that their string. Note that here, and in the remaining part protocols also provide security in the case of im- of this thesis, ‘Alice’ and ‘Bob’ refer to the two par- perfect quantum communication, a situation that ties that want to achieve the task of 1-2 OT, and will be covered in Section 4. that the ‘attacker’ can be any of these two parties There are three reasons why we may consider (recall footnote 2). the bounded-quantum-storage model. Firstly, if Alice and Bob merely communicate over a clas- Additional assumptions sical channel—and thus not use quantum communication—secure 1-2 OT can only be achieved We explained in Section 1 that even with the if the adversary’s classical memory size is at possibility of quantum communication, protocols most quadratic in memory size of an honest party for secure two-party computation—and hence 1- [DM04]. However, using quantum communication, 2 OT—must impose additional assumptions on the the protocols for the bounded-quantum-storage adversary in order to achieve a desired level of model only require a bound on the size of the adver- security [BCS12; Lo97; LC97; May97]. It is rea- sary’s quantum memory. In particular, these pro- sonable to aim for protocols for secure two-party tocols achieve a desired level of security regard- computation that are feasible and based on real- less of the size of the adversary’s classical mem- istic assumptions. Firstly, feasibility is important ory, which, consequently, allows for a bigger ra- to ensure that the protocol can be used in prac- tio between the memory size of the honest par- tice. Moreover, the assumptions should be real- ties and that of an adversary than would be pos- istic in the sense that they take into account the sible in the classical setting. In fact, the honest state-of-the-art technologies and their limitations parties do not need any quantum storage at all, [KWW12; Sch10]. Therefore, two cryptographic and the bound on the adversary’s quantum mem- models have been proposed that are based on ory size only depends on the number n of qubits physical (instead of computational) assumptions: that have been sent during the protocol. There- the bounded-quantum-storage model [Dam+05; fore, the bounded-quantum-storage model allows Dam+07] and the noisy-storage model [WST08; us to outdo the optimal bound for which security is STW09; Sch10; KWW12]. obtained in the case that a dishonest player only has access to a classical memory [Dam+05]. Sec- ondly, the limited size of the adversary’s quantum Bounded-quantum-storage model memory ensures that a dishonest party will lose in- The physical assumption in the bounded- formation: being unable to store all qubits, the ad- quantum-storage model is that the adversary’s versary must store some information in his classical quantum memory size is limited to a certain num- memory (i.e. by measuring the qubits) [Dam+05]. ber of qubits. No restriction is imposed on the ad- It is, in fact, this irreversible loss of information that versary’s classical memory or computing power. In allows for 1-2 OT in the bounded-quantum-storage this model, the version of OT that was introduced in model [Dam+05]. The third reason to consider [Rab81] was proven to be information-theoretically the bounded-quantum-storage model is a practical secure [Dam+05]. In particular, if Alice and Bob one: whereas quantum transmission and quantum are honest, they do not need any quantum memory measurements — needed for the honest parties — and the protocols can only be broken if the dishon- are possible with the present state-of-the-art tech-

6 Capstone Issue Vol. 12, 2019 Lynn Engelberts

nology, storing only “a single qubit for more than a reflects the prevailing technical difficulties of stor- fraction of a second” is technologically still very dif- ing quantum information more accurately than the ficult [Dam+05], impeding the storing possibilities bounded-quantum-storage model. First of all, cur- of a dishonest party. rent quantum memories are not completely reli- Despite the aforementioned reasons, it is ques- able. In particular, with current technology, the tionable whether any practical situation will im- time during which quantum information can be pose such a limit on the quantum memory size stored without being completely lost is limited (e.g. as the bounded-quantum-storage model requires see [Zho+12]). The imperfection of current quan- [WST08]. In fact, [Dam+05] argue that it is more tum memories results from the difficulty to keep realistic to consider the case that the adversary’s the stored quantum information unaltered over quantum memory is noisy, instead of bounded, time and the problem that quantum information where noise induces loss of (stored) quantum in- (i.e. the photons in which the quantum information formation. Therefore, we describe the alternative is encoded) is lost during storage [Erv+14]. In fact, model and its advantages below. [KWW12] argues that it is not even known whether it is physically possible to build a fully reliable memory in the future. In addition, noise may arise when Noisy-storage model quantum information is transferred onto a different Whereas the bounded-quantum-storage model physical carrier representing the quantum mem- gives an explicit upper bound on the adversary’s ory; for example, when photonic qubits are trans- quantum memory size during the protocol, the ferred onto an atomic ensemble or an atomic state noisy-storage model4 assumes that the quantum [STW09]. In fact, this transfer onto the quantum storage (i.e. the quantum memory) is noisy memory is often already noisy [KWW12], irrespec- [KWW12; Sch10; STW09; WST08]. More precisely, tive of the (im)perfection of the quantum memory this model allows an adversary to have the best itself [Weh+10]. available quantum memory, but this memory is affected by noise over time: the longer the adversary stores quantum information, the higher the amount of prevailing noise, and hence the more information is lost [Sch10]. Two-party protocols for the noisy- Yet, we should note that security is only guar- storage model (such as the protocols considered in anteed if the aforementioned technological chal- this paper) make use of this time-dependency by lenges prevail [Erv+14]. Nevertheless, the noisy- inducing a waiting time for the participating par- storage model is resistant against technological im- ties during the protocol. After this waiting time, provements in the future (i.e. security can still the higher level of noise in the adversary’s quan- be obtained) as long as the quantum memory size tum memory results in information loss, see Sec- has a finite upper bound [Erv+14]. This resis- tion 3.2. As a consequence, we can evaluate the tance of the model results from the fact that the security of a protocol for this model in terms of level of security depends on the number of qubits the noise level [WST08]. Furthermore, note that that are transmitted during the protocol, and hence the bounded-quantum-storage model can, in fact, can be monitored. In other words, if the dishon- be obtained from the noisy-storage model by sim- est party has access to a larger quantum storage, ply assuming that there is no noise in the quantum the required level of security can still be obtained memory but that the input of the quantum mem- by sending more (quantum) information during the ory (i.e. the number of qubits that can be stored) is protocol [Erv+14]. Hence, as long as an adversary limited [KWW12]. does not have access to an unlimited and noiseless By exploiting the technological difficulty of quantum memory in the future—which can be per- building a quantum computer (i.e. the presence ceived a reasonable assumption – we conclude that of noise) the noisy-storage model has significant the noisy-storage model is a realistic and feasible practical value. In fact, the noisy-storage model model for cryptographic purposes.

4Since the noise occurs in the quantum memory, researchers sometimes also call this model the noisy-quantum-storage model. However, in this thesis we stick to the more common noisy-storage model. In particular, we assume in this thesis that there is no noise in the classical communication channel and classical storage at all.

7 Capstone Issue Vol. 12, 2019 Lynn Engelberts

Security in the noisy-storage model on her quantum memory because the only information she receives from Bob is classical informa- Given the aforementioned reasons why the tion. Therefore, we only consider noisy-quantum- noisy-storage model could be seen as a realistic re- storage attacks by a dishonest Bob. Suppose now flection of the prevailing technical difficulties, we that Bob tries to store an incoming quantum state based our implementations on protocols for 1-2 OT in his quantum memory. Since the quantum mem- 5 within the noisy-storage model . As previously ex- ory is noisy, the state is affected by noise and de- plained, the noisy-storage model allows a dishon- coheres over time [KWW12], i.e. the quantum sys- est party to have access to an unlimited quantum tem interacts with its environment and informa- memory, yet this quantum memory is noisy. This tion is lost. Then, security in this model can be section provides an intuitive description of what achieved by imposing a waiting time during the such a noisy quantum memory entails for a dishon- protocol: the quantum information that a dishonest party and why this may allow us to achieve se- est Bob may store in his quantum memory under- curity. We also describe another tool, privacy am- goes noise during this waiting time, which results in plification, that is used in the considered protocols loss of information. In this way, we may exploit the to provide security in the noisy-storage model. noise in the quantum memory in order to achieve the desired level of security. Nevertheless, we emphasise that a dishonest Bob is still allowed to per- Noisy-quantum-storage attacks form any encoding attack on the received informa- First of all, note that in the considered protocols tion (i.e. the attacker has unlimited computational for 1-2 OT in the noisy-storage model only a dis- power and can do perfect quantum operations) and honest Bob (i.e. the receiver) can take advantage has an unlimited noiseless classical storage. Also, of his quantum memory, since he is the only one after the waiting time, Bob can do any decoding that receives quantum information. In particular, attack. See Figure 1 for an illustration of a noisy- security against a dishonest Alice does not depend quantum-storage attack by dishonest Bob.

Privacy amplification tion. In particular, the probability that Bob guesses Alice’s shorter string is almost uniformly random We provide an intuition about the tool of pri- [Erv+14]. In other words, by shrinking string A in vacy amplification [BBR88], which is applied dur- this way, privacy amplification allows for removing ing the protocol (i.e. Step 5 of Protocol 3.1) to the leaked information and thus increases the level ensure that any leaked information to a dishonest of security. party is removed, and thus security can be guaranteed. Suppose that Alice holds a random bit- Protocol for 1-2 OT string A of length n about which a dishonest Bob has partial information.This partial information may Several protocols for 1-2 OT in the noisy-storage allow Bob to guess A, since A does not look com- model have been suggested and research distin- pletely random to Bob anymore. Then privacy am- guishes two types of attacks from the adversary in plification is performed using some randomly cho- their security proofs: individual and general stor- sen two-universal hash function. In short, a hash age attacks. When the adversary is allowed to only function is a function that maps a set of some size perform measurements on each incoming qubit in- to a smaller set [CW79]. We say that a class F of dividually, we call this an individual attack. How- functions f : {0, 1}n → {0, 1}l. is two-universal if ever, since multi-qubit measurements are possible 0 −l 0 n Pf∈RF [f(x) = f(x )] ≤ s for all x 6= x ∈ {0, 1} with current technology, it is more realistic to as- [CW79]. Note that f ∈R F means that f is ran- sume that the attacker can perform any arbitrary domly chosen from F. Then, by applying a ran- attack, i.e. a general attack [KWW12]. Whereas the domly chosen two-universal hash-function to her n- protocols in [WST08] and [STW09] are proven to be bit string A, Alice obtains a shorter string, about secure only against individual storage attacks, the which Bob has a negligible amount of informa- protocols in [KWW12] and [Sch10] are proven to be

5For a formal deﬁnition of the noisy-storage model, we refer the reader to [Weh+10].

8 Capstone Issue Vol. 12, 2019 Lynn Engelberts

Figure 1: (Source: [Weh+10]) An illustration of an attack by dishonest Bob. Here, F represents the noisy- quantum memory (as formally defined in [Weh+10]) and ∆t the waiting time during the protocol. secure against any attack. The latter two protocols domised 1-2 OT (1-2 ROT), because, as will be fur- are hence much more realistic to consider. ther clarified, 1-2 OT can be easily obtained from In particular, the protocols in [KWW12] and 1-2 ROT [Sch10]. In 1-2 ROT, Alice has no input [Sch10] have been proven to be information- whilst Bob holds his choice bit c ∈ {0, 1}. After com- theoretically secure against general noisy-storage pletion of the protocol, Alice holds two bit-strings attacks as long as the laws of quantum mechan- of the same length, s0 and s1, and Bob holds sc, ics hold and the amount of noise in the quantum i.e. the string received by Alice corresponding to storage channel is above a particular level. Recall his choice bit c. See Figure 2 for an illustration. from Section 1 that an information-theoretically secure protocol implies that there are no restrictions We first explain some notation, for which we use on the attacker’s computational power, which is in some terminology from Section 2. Note that be- contrast to many classical cryptographic protocols fore running Protocol 3.1 and Protocol 3.2, Alice and that are based on computational assumptions and Bob agree on the length ` of the strings s0 and 1. 6 hence are more restricted in their performance. Al- The level of security is also decided beforehand, though the protocol in [KWW12] requires less stor- and determines the number n of qubits that will be age noise for security, the protocol in [Sch10] is transmitted in Step 1. Moreover, note that X ∈R Y more straightforward and, therefore, simpler to im- means that X is randomly picked from Y . We also plement. For this reason, we have decided to im- note that H in Step 1 is the Hadamard operator, A A plement the protocol given in [Sch10], which will and thus if yi = 1, Alice first applies H to bit xi be covered in Section 6. (for i ∈ [n]) before she sends it to Bob. For clari- fication, this implies that in Step 1, Alice randomly sends one of the four BB84 states (|0i , |1i , |+i , |−i) Protocol for 1-2 OT in the noisy-storage to Bob. Finally, in Step 2, when Bob measures his model incoming qubit in the Hadamard basis, we mean Before we provide Protocol 3.2 for 1-2 OT in the that Bob first applies H, and then measures in the noisy-storage model, we provide a protocol for ran- computational basis. 6With the level of security we refer to the security error-probability for an -secure 1-2 ROT protocol as defined in [Sch10].

9 Capstone Issue Vol. 12, 2019 Lynn Engelberts

l Figure 2: In the setting of 1-2 ROT, Bob inputs a bit c ∈ {0, 1}. After completion, Alice holds s0, s1 ∈ {0, 1} and Bob holds sc, which is one of Alice’s output strings. Before running the protocol, Alice and Bob agree on the length ` of the output strings.

A B The Protocol below (Protocol 3.1) is the protocol Note that, after Step 2, whenever yi = yi for A B for 1-2 ROT as provided in [Sch10]. This protocol i ∈ [n], we also have that xi = xi , since Bob mea- originates from [Ben+92b; Dam+09], and security sures his i-th incoming qubit in the same basis as of this protocol in the noisy-storage model is proven in which Alice encoded her i-th bit. in [Sch10]. As pointed out in [Sch10], 1-2 OT can be achieved from 1-2 ROT. Recall that in the setting [Ben+92b; Dam+09] 1-2 ROTl of 1-2 OT, Alice has two input bit-strings of length `, say m0 and m1, and Bob|holding his choice bit A n A 1. Alice randomly picks x ∈R {0, 1} and y ∈R c ∈ {0, 1}|wants to obtain mc after completion of n {0, 1} . the protocol. Below we provide the resulting proto- A A y A yn A At time t = 0, Alice sends H 1 |x1 i, ..., H |xn i col for 1-2 OT. An illustration of the required steps is to Bob. provided in Figure 3. We emphasise that a quantum channel between Alice and Bob is only required for 2. Bob randomly picks yB ∈ {0, 1}n. R achieving 1-2 ROT, i.e. in Step 1. Moreover, we use Bob measures the i-th qubit he receives in B the same notation as for the preceding protocol and the computational basis if yi = 0 and the B ` and n are again determined beforehand. Hadamard basis if yi = 1. He obtains outcome xB ∈ {0, 1}n. Protocol 3.2: 1-2 OTl Both parties wait time ∆t. 1. Alice and Bob run the protocol for 1-2 ROT. 3. Alice sends her basis string yA to Bob. After completion, Alice holds two uniformly random bit-strings s0 and s1 of length `, and A B 4. Bob forms the sets Ic = {i ∈ [n] | yi = yi } Bob holds sc corresponding to his choice bit c. A B and Ic¯ = {i ∈ [n] | yi 6= yi }, where c is his choice bit and c¯ the remaining bit. 2. Alice computes c0 = m0 ⊕ s0 and c1 = m1 ⊕ s1, and sends them to Bob over the classical Bob sends I0 and I1 to Alice. channel. 5. Alice picks two hash functions f0, f1 ∈R F, where F is a class of two-universal hash func- 3. Bob receives c0 and c1. Bob now obtains mc tions f : {0, 1}n → {0, 1}l. by computing sc ⊕ cc = sc ⊕ (mc ⊕ sc) = mc. Alice sends f0 and f1 to Bob and outputs s0 = A A 7 For an explicit analysis of the correctness and a f0(x I0 ) and s0 = f1(x I1 ). full proof of security of these protocol in the noisy- B 6. Bob outputs sc = fc(x Ic ). storage model, we refer the reader to [Sch10].

7 A A If the length of x I0 or x I1 is m < n, then Alice adds n − m 0’s to the string before applying f0 or f1, respectively.

10 Capstone Issue Vol. 12, 2019 Lynn Engelberts

l Figure 3: In the setting of 1-2 OT, Alice inputs two strings m0, m1 ∈ {0, 1} and Bob inputs a bit c ∈ {0, 1}. The three indicated steps are those needed to achieve 1-2 OT from 1-2 ROT. After completion, Bob holds mc, Alice’s input string corresponding to his choice bit c. Contrary to the setting of 1-2 ROT, Alice now does not have an output.

IV. Error correction and robust 1-2 Information theory OT Before continuing, we formalise some notions of information theory that will be used in this section. We use the definitions provided in [DS17]. We assume that the reader is familiar with the basic In the practical implementation of 1-2 OT it is notions of probability theory; however, the relevant realistic to assume that the quantum channel con- definitions are provided in Appendix A. necting Alice and Bob is subject to a particular level of noise. Specifically, imperfections occur when a An important measure of uncertainty about in- quantum channel is implemented in practice, re- formation is the Shannon entropy. We provide three sulting in errors during the transmission of quan- definitions below. tum information over the channel [Ben+92a; BS94; EMMM11; GPCZ18]. Apart from the transmission Definition 1. (Shannon entropy) of the qubits itself, the preparation of the qubits Let X be a random variable on the set X . We define by Alice and the measurements by Bob will also be the (Shannon) entropy of X as affected by noise [KWW12]. Therefore, we should X make the protocol for 1-2 OT robust against noise H(X) := − PX (x) · log2 PX (x). for the honest parties [STW09] so that the protocol x∈X is suitable for real implementations. In other words, In other words, measures the amount of we need to find a way to deal with the errors re- H(X) uncertainty about : the higher the uncertainty sulting from the noisy communication channel and X about , the higher . Note, however, that it the imperfections during the preparation and mea- X H(X) is a function of , not of itself. surement of quantum states. This section, in par- PX X ticular, will describe how information reconciliation – an additional step during which error-correction Definition 2. (Conditional entropy) techniques are applied – can provide a solution, and Let X and Y be two random variables on the sets X subsequently explain the resulting protocol for ro- and Y, respectively. Then, the conditional entropy bust 1-2 OT in the noisy-storage model8. of Y given X is the average uncertainty about Y

8Note that a noisy communication channel should not be confused with a noisy quantum memory (as in the noisy-storage model), as these notions have a completely different meaning. In particular, if Alice and Bob are both honest, they do not even need a (noisy) quantum memory in order to perform 1-2 OT; however, this does not imply that the quantum channel connecting Alice and Bob should be noiseless: errors can still occur and affect the qubits that are sent from Alice to Bob during the protocol.

11 Capstone Issue Vol. 12, 2019 Lynn Engelberts when the outcome of X is known, i.e. Noisy communication channel

X Before we introduce the concept of information H(Y |X) := − PYX (y, x) · log2 PY |X (y|x). y∈Y,x∈X reconciliation and describe the protocol for robust 1-2 OT, we discuss in more detail the concept of Definition 3. (Mutual information) a noisy communication channel. We consider the The mutual information of two random variables X case that Alice sends a classical bit-string, A, to and Y is defined as Bob over a communication channel. It is often assumed that the communication channel is reliable I(Y ; X) := H(Y ) − H(Y |X) = H(X) − H(X|Y ). or noiseless, i.e. A = B, where B is the bit-string received by Bob. However, in real implementa- In words, I(Y ; X) measures the reduction in uncer- tions, Alice’s string may be affected by noise (i.e. tainty about Y when X is known, and vice versa. errors) during transmission so that A 6= B. In the presence of such noise, we say that the commu- To provide some intuition, suppose that X is the nication channel is unreliable or noisy. Clearly, if input of a communication channel and Y the out- Alice and Bob transmit information over an unreli- put. Then the mutual information measures how able communication channel, the correctness and much information output Y reveals about input X. security of the corresponding cryptographic protocol may be affected [GPCZ18]. Definition 4. (Discrete channel) A noisy communication channel can be mod- A (discrete) channel is a triple (X ,PY |X , Y) consist- elled by a binary symmetric channel (BSC) [BS94]. ing of two finite sets X and Y and a conditional A BSC is a communication channel in which each probability distribution PY |X : Y → [0, 1]. Here, X transmitted bit is flipped (i.e. affected by noise) represents the set of possible channel inputs and with a certain probability p [Mac03]. In other words, Y the set of possible channel outputs. PY |X (y|x) in the case that Alice sends her bit-string to Bob, is the probability of receiving output y ∈ Y when each bit is transmitted correctly with probability input x ∈ X is given. 1 − p. Formally, i.e. according to Definition 4.4, we define a BSC with parameter p by X = Y = {0, 1} Definition 5. (Memoryless channel) and by A channel is called memoryless if the probability distribution of the output depends only on the cur- PY |X (0|0) = PY |X (1|1) = 1 − p, rent input. In other words, if the memoryless channel is used PY |X (0|1) = PY |X (1|0) = p, multiple times, previous inputs and outputs do not where p ∈ [0, 1/2] [DS17]. We will use BSC(p) to affect the current output. refer to a BSC with probability parameter p. Note that a BSC is memoryless, as the probability that Definition 6. (Noiseless channel) a transmitted bit flips (i.e. p) does not depend on A channel (X ,PY |X , Y) is noiseless if both any previous bit that has been transmitted over the H(X|Y ) = 0 and H(Y |X) = 0. channel. Moreover, note that the channel is noise- In other words, the output of the channel com- less if p = 0. An illustration of the BSC is provided pletely determines the input, and vice versa. in Figure 4.

Information reconciliation bit c, even if the quantum channel is noisy.

As mentioned before, the presence of noise re- During this additional step, the errors between quires an additional step during the protocol in the transmitted and received information are iden- which Bob can recover the information that Alice tiﬁed and error correction is performed. In the con- has sent so that the correctness of the crypto- text of QKD, this process of ﬁnding the differences graphic protocol can still be achieved. In the setting between the strings and performing error correc- of 1-2 OT, this implies that Bob eventually receives tion is often called information reconciliation or, Alice’s input string mc corresponding to his choice simply, reconciliation (e.g. in [Ben+92a; BS94;

12 Capstone Issue Vol. 12, 2019 Lynn Engelberts

Figure 4: A binary symmetric channel between Alice and Bob, where each bit that Alice sends to Bob is ﬂipped with probability p.

EMMM11]). Information reconciliation occurs via a information leaked to a potential dishonest party. public classical channel connecting the two parties In particular, since there is additional information [Ben+92a; EMMM11]. leaked during the information-reconciliation proto- We provide a simple example of an information- col, we should perform more privacy amplification reconciliation protocol by adjusting the description to achieve a desired level of security than would be in [BS94] to the setting of 1-2 OT. Suppose again the case if the quantum channel is noiseless and that Alice sends a bit-string A to Bob over a noisy we need not perform information reconciliation. For communication channel, and Bob receives the bit- example, suppose that there are m bits of addi- string B. As B is affected by noise, we may as- tional information sent during information reconcil- sume that B ≡ A + N (mod 2), where the bit-string iation, then the hash function should map the initial N represents the noise. In the setting of 1-2 OT, string to a string that is m bits shorter than in the an information-reconciliation protocol aims to gen- case that the channel is noiseless. erate an estimate, Aˆ, of Alice’s string A, given B. ˆ As we will see, this string A is obtained by exchang- Error-correcting codes ing some classical information over a public channel connecting Alice and Bob. At the end of the rec- Shannon [Sha48] described that reliable com- onciliation protocol, Bob holds the string Aˆ. Clearly, munication over a noisy channel can be obtained the protocol should minimise the amount of infor- by means of encoding and decoding. We refer mation revealed to a potential adversary and the to an encoding and decoding system as an error- probability that the protocol fails to produce the fi- correcting code. Intuitively, we may describe en- nal string (i.e. Aˆ) should be low [BS94]. coding and decoding as follows. Instead of immediately sending her message A to Bob, Alice first encodes her message into T (the so-called codeword), Privacy amplification by adding some redundancy to A [Mac03]. Subse- Nevertheless, it is important to note that by quently, T is transmitted over the noisy channel, transmitting some extra information over a pub- which outputs the received codeword R ≡ T + N lic channel, information reconciliation allows a dis- (mod ()2). Here, N is a bit-string representing the honest party to gain additional information. Recall noise induced by the channel. The decoder then the tool of privacy amplification discussed in Sec- uses the (known) redundancy in order to translate tion 3.2, which allows us to reduce the amount of R into the bit-string Aˆ, Bob’s estimate of A. Figure

13 Capstone Issue Vol. 12, 2019 Lynn Engelberts

5 provides an illustration. Definition 7. (Information rate) The information rate R is the number of bits of information that are transmitted per channel use. Shannon’s noisy-channel coding theorem Definition 8. (Channel capacity) t is important to mention that there is an up- Let (X ,PY |X , Y) be a discrete, memoryless chan- per bound on the amount of reliable communica- nel. Then its channel capacity C is defined as tion that can be achieved in the presence of a noisy communication channel. This limit is known as C := max I(X; Y ). PX Shannon’s noisy-channel coding theorem [Sha48], and hence also referred to as the Shannon limit. In other words, the channel capacity measures the Before we state the theorem, we provide two more maximum amount of mutual information between definitions, obtained from [DS17]. the input X and the output Y of a the channel.

Figure 5: An illustration of an error-correcting communication system for the setting that Alice wants to send a message A to Bob over a noisy communication channel.

The theorem below shows that information can should also be possible in practice to achieve any only be transmitted with arbitrarily small error- rate R < C. We therefore aim to find a coding sys- probability over a given channel as long as the tem that approaches the Shannon limit as much as channel capacity is not exceeded. This is a slightly possible. adapted version of the one provided in [DS17], and we refer the reader to this work for the proof. Theorem 1. Noisy-channel coding theorem Consider a discrete memoryless channel with capacity C. Then, (i) for any rate R < C, there is Robust protocol for 1-2 OT a method of encoding such that information can be transmitted over the channel with negligible In order to make our protocol for 1-2 OT robust maximal error-probability (i.e. the maximal error- against noise for the honest parties, we should add probability approaches 0 as the number of chan- the extra step of information reconciliation. For nel uses approaches infinity); and (ii) for any rate our implementation of reconciliation in the robust R > C, it is impossible to find a coding sys- 1-2 OT protocol, we use a class of error-correcting tem that achieves such a negligible average error- codes called Reed-Solomon (RS) codes. Before we probability. explicitly describe how information reconciliation In other words, whilst the channel capacity C works in the setting of 1-2 OT and provide the ro- indicates the maximal information rate that theo- bust protocol, we briefly introduce the class of RS retically can be realised, Theorem 1 shows that it codes.

14 Capstone Issue Vol. 12, 2019 Lynn Engelberts

Reed-Solomon codes affected by noise with the same probability. Recall from Protocol 3.1 for 1-2 ROT that Alice We provide a high-level description of RS codes randomly picks two n-bit strings xA and yA, and obtained from [LC83] and refer the reader to this Bob randomly picks an n-bit string yB. In this sec- work for a more explicit explanation. RS codes are tion, we consider the case that Alice sends each bit a class of codes that were introduced by Reed and A yA A x encoded as H i |x i to Bob over a noisy quan- Solomon in 1960 [RS60]. Each message consists of i i tum channel. Bob measures each received qubit elements from the Galois (i.e. finite) field GF( r), p either in the computational (yB = 0) or Hadamard for prime. If the original message has length i p n (yB = 1) basis . We now explain what may go wrong and the transmitted codeword length , then there i m in the noisy setting. Recall that if the channel was are redundant symbols, and an RS code can m − n noiseless, then for each i such that yA = yB (i.e. correct up to m−n errors. i i 2 Alice sends the qubit in the same basis as in which In the following we briefly explain how encod- A B Bob will measure) we have that xi = xi with prob- ing and decoding of RS codes is established, and ability 1. However, if the channel is noisy, then the use, for clarity, the same notation as in Figure 5. noise in the channel may affect the qubits that Bob Each message A of length n is associated with a has received. Therefore, when Bob measures each polynomial of the same length. Then, the trans- B received qubit in basis corresponding to yi , it is mitted codeword obtained after encoding can be B possible that his outcome xi is no longer such that considered as a polynomial T (x) of some length A B xi = xi . m > n. The noisy communication channel then Now, recall that at a certain point during the adds some noise N(x) (also called the error pat- protocol Bob forms the set Ic containing all indices tern in [LC83]) to T (x). Hence, we can decompose i such that yA = yB, where c is his choice bit. 9 the received message as R(x) = T (x) + N(x). If all i i Thus, in the presence of noise, it is possible that coefficients of N(x) are zero, then there are clearly A B x I 6= x I , whereas equality would necessarily no errors. However, if there are errors, the infor- c c hold if the qubits were transmitted over a noiseless mation from the received message R(x) is recov- channel. Consequently, classical error-correction ered by the decoder. In short, the decoding pro- techniques need to be applied so that Bob obtains cess consists of four steps [LC83]. First, the syn- A an estimate for x I . drome of R(x) is computed, which, intuitively, is an c In order to understand how information recon- indicator of the error pattern (in particular, if the ciliation works in this setting, we explain how we syndrome is zero, we assume that there are no er- can represent the occurring noise. For simplicity, rors). Then, from this syndrome we may determine let A = xA and B = xB . Let n be the length the so-called error-location polynomial. In the third Ic Ic of the bit-strings A and B.10 We assume that B step, the error locations are retrieved by comput- is affected by noise such that A 6= B. Note that ing the roots of the error-location polynomial, and since the noisy quantum channel is i.i.d., each of thus the error pattern (i.e. N(x)) is obtained. In the the qubits is affected by noise with the same prob- last step, we compute R(x) − N(x) to obtain the ability p, and thus (without loss of generality) we transmitted codeword T (x). can assume that each bit of B is affected by noise with probability p. Hence, we may model the situa- Information reconciliation for 1-2 ROT tion by sending A over a BSC with error-probability p. Note, however, that the BSC is not the physi- Having provided some intuition into RS codes, cal channel, but refers to the theoretical channel we can now give a complete description of the set- model. In fact, the BSC represents a noisy classical ting of randomised 1-2 OT when the quantum chan- channel. To make this more precise, note that Alice nel is noisy. Although the channel is noisy, we sends her n qubits to Bob over the physical quan- assume that the noise in the quantum channel is tum channel. Only after Bob performs his mea- independent and identically distributed (i.i.d.), so surements, he obtains classical information. How- each qubit that is transmitted over the channel is ever, if noise has occurred,it is possible that Bob’s

9 We will only describe the situation for Ic, as the situation for Ic¯ is similar (we only need to replace c by c¯). 10 A As in the protocol, for d ∈ {0, 1}, if the size m of Id is smaller than n, we add n − m zeros to x Id to obtain an n-bit string, as this is according to [Sch10].

15 Capstone Issue Vol. 12, 2019 Lynn Engelberts classical bit-string B differs from the correspond- 2. Bob attaches the m − n bits corresponding to A B ing classical bit-string A encoded by Alice, i.e./ that x Ic to his n-bit string x Ic . B A x Ic 6= x Ic . These differences can hence be rep- 3. Bob uses an RS decoder to correct the errors resented by the BSC, even though in practice these in his resulting m-bit string, if possible. classical bits were not transmitted over a classical If the number of errors is greater than m−n , channel. 2 the protocol is aborted. As mentioned before, we use RS codes to correct the errors induced by the channel. We note that our method of information reconciliation is Robust protocol for 1-2 OT slightly different from the one described in the pro- We will now provide the protocol for robust 1-2 tocol from [Sch10]. In particular, [Sch10] considers ROT in the noisy-storage model on which our code the case that Alice sends the syndromes (which are is based. Protocol 4.2 is a slightly adapted from of length n − k) of her two n-bit strings xA and I0 the protocol in [Sch10]. In particular, the protocol xA to Bob. Instead, to keep the implementations I1 in [Sch10] takes into account the possibility that relatively simpler, in our approach, Alice first en- some of the qubits that are transmitted by Alice codes her two n-bit strings into codewords of length will eventually not be detected by Bob, which may m > n, and then sends the last m − n (redundant) result from imperfections in Bob’s detection appa- bits of both of the codewords to Bob. Bob will at- ratus. However, in order to keep the implementa- tach the m − n bits that belong to xA to his n-bit Ic tion simple and understandable, we assume that string xB , and then use RS decoding techniques Ic all n qubits arrive at Bob. Therefore, a few steps to obtain Alice’s string xA . Although our method Ic in the protocol from [Sch10] are omitted. We ac- might be slightly less efficient than the method pro- knowledge, however, that it would be more realistic vided in [Sch10], we note that the correctness and to assume that not all qubits are detected by Bob. security of the protocol from [Sch10] still apply for Nevertheless, we emphasise that the correctness our approach. Furthermore, we note that because and security analysis from [Sch10] still hold in our in our approach m − n additional bits are sent over case. the classical channel, privacy amplification must be We use the notation from Section 3.3 and again adapted accordingly: the final length of ` of the out- assume that Alice and Bob agree on ` and a par- put strings must be shortened by m bits. Neverthe- ticular security error-probability, which determines less, we will assume that when Alice and Bob agree n. Alice and Bob also agree on the value of m for on the values for ` and n they already take into ac- information reconciliation. count this reduction in final string length. The resulting protocol for information reconcili- [Sch10] Robust 1-2 ROTl ation is given below. All communication is achieved A n A over a noiseless classical channel. We assume that 1. Alice randomly picks x ∈R {0, 1} and y ∈R both Alice and Bob are able to perform RS encoding {0, 1}n. and decoding and have agreed on a parameter m, B n which is the length of the encoded list and deter- 2. Bob randomly picks y ∈R {0, 1} . mines the maximal number of errors that can be A 3. For each i ∈ [n], Alice sends her bit xi en- corrected. Recall that given the length n of Alice’s yA A coded as H i |x i to Bob. and Bob’s strings, using RS codes we can correct i If yB = 0, Bob measures the i-th incoming up to m−n errors. i 2 qubit in the computational basis. Otherwise, he measures the qubit in the Hadamard basis. Information reconciliation for 1-2 ROT Both parties wait time ∆t. A We assume that Alice holds two n-bit strings x I0 A B 4. Alice sends her basis string yA to Bob. and x I1 , and Bob holds an n-bit string x Ic , where c is his choice bit. A B 5. Bob forms the sets Ic = {i ∈ [n] | yi = yi } A A A B 1. Alice encodes x I0 and x I1 using an RS en- and Ic¯ = {i ∈ [n] | yi 6= yi }, where c is his coder and sends the last m − n bits of the en- choice bit. coded strings to Bob. Bob sends I0 and I1 to Alice.

16 Capstone Issue Vol. 12, 2019 Lynn Engelberts

6. Alice performs privacy amplification: she Significance and applications of a quan- picks two hash functions f0, f1 ∈R F, where tum internet F is a class of two-universal hash functions n l It is expected that a quantum internet will be f : {0, 1} → {0, 1} . Alice sends sends f0, f1 to Bob. used alongside the classical internet rather than re- Alice performs her part of information recon- placing it. [WEH18]. Therefore, we explain the sig- A A nificance of having quantum internet next to the ciliation: she encodes x I0 and x I1 using an RS encoder and sends the last m − n bits classical internet. Moreover, we explain what the benefits may be of constructing a quantum internet (say r0, r1) of the encoded strings to Bob. A rather than building a large-scale quantum com- Alice outputs s0 = f0(x I0 ) and s1 = A puter. We conclude with some potential applica- f1(x I1 ). tions of a quantum internet.

7. Bob performs his part of information reconciliation: Bob attaches rc to his own string and Comparison with the classical internet and uses an RS decoder to correct the errors on the quantum computer B his bit-string x Ic . He obtains the corrected B A quantum network may allow us to enlarge the bit-string xcor. B state space used for quantum computation. In par- Bob outputs sˆc = fc(xcor). ticular, with a quantum network we may not only surpass the capabilities of a classical network, but We emphasise that robust 1-2 ROT can be con- potentially also those of a single quantum proces- verted into robust 1-2 OT in a similar way as has sor. been done for the idealised setting in Protocol 3.2. Firstly, each node in a quantum network con- We therefore omit the resulting protocol. sists of a quantum processor, which has several advantages over a classical computer. In particular, a quantum computer allows us to exploit the V. Quantum internet and Simu- laws of quantum mechanics—e.g. by making use of quantum entanglement—in order to tackle prob- laQron lems that are hard to solve, or unsolvable, with only classical means. Moreover, the state space A quantum internet is a quantum communi- of a quantum processor (i.e. the number of states cation network that allows for the exchange of on which it operates) scales exponentially with the quantum information over arbitrarily long distances number of qubits it entails as a consequence of [Cas18; DW18; DLH17]. One of the key aspects of the superposition principle [Kim08; CCB18]. Recall a quantum internet is that it will allow for quan- from Section 2 that a single qubit can exist in a su- tum entanglement between any two points in the perposition of two states, and therefore n qubits world [CCB18; Cas18; DLH17; Kim08]. Quantum can exist in a superposition of 2n states. There- entanglement is a type of correlation between two fore, whereas a classical processor consisting of n quantum systems, which has two important fea- classical bits has a computational space of dimen- tures that can be used for several applications. sion n, a quantum processor consisting of n qubits The ﬁrst is that two maximally entangled qubits has a computational space of dimension 2n. Conse- are perfectly correlated: measuring one quantum quently, this larger state space of the quantum pro- state will immediately affect the outcome of mea- cessors in the network may allow us to solve prob- suring the other state as well [DLH17; CCB18; lems exponentially faster than would classically be WEH18].11 The second feature of quantum en- possible [Kim08; WEH18]. tanglement, complete privacy, is particularly in- However, the potential of a quantum internet teresting for security-related tasks: no more than is not only explained by the fact that its individual two qubits can be maximally entangled, and hence components—i.e. the quantum processors—may no third party can participate in the entanglement perform better than if they were substituted by [WEH18]. classical processors, but also since interconnect-

11Note that we consider the Copenhagen interpretation of quantum mechanics.

17 Capstone Issue Vol. 12, 2019 Lynn Engelberts

ing the quantum processors with each other in- tion of a quantum internet is distributed quantum creases the overall computational power of the net- computation [DLH17; Cac+18]: by linking differ- work. This feature is called “quantum connectivity” ent small-sized quantum computers via quantum by [Kim08] and is explained as follows. Suppose channels [DLH17], we can scale up the computa- that we have k quantum processors, consisting of tional space of each individual node, as previously n qubits each (i.e. each has computational space of explained. Therefore, a quantum internet consist- dimension 2n). Then, if these quantum processors ing of multiple quantum processors may allow us to are connected via quantum channels, the overall solve problems (possibly exponentially) faster than computational space of the network is of dimen- would be possible with only a single quantum pro- sion 2kn. In contrast, if these k quantum proces- cessor. Other applications of a quantum internet in- sors (nodes) were connected via classical channels, clude, but are not limited to, clock synchronisation then the computational space is only of dimension [Kóm+14] and improving the achievable resolution k2n. In other words, because of its quantum con- of telescopes [GJC12]. nectivity, a quantum internet potentially allows for an improved performance than each of its quantum Towards the realisation of a large-scale processors is capable of individually. quantum network In particular, some researchers (e.g. [DLH17]) speculate that the realisation of a quantum internet Having discussed the significance and some po- is more feasible than building a large-scale quan- tential applications of a future quantum internet, tum computer. The development of quantum com- we will now provide a high-level description of how puters places a maximum on the attainable size a quantum internet could be constructed and anal- of individual quantum processors [Kim08], despite yse the feasibility of this construction. Before con- the fact that technological progress may increase tinuing, however, we first describe how quantum this maximum over time. In contrast, a quantum information is transmitted. internet is potentially able to exceed this maximum by interconnecting these individual proces- Transmission of quantum information sors, as previously explained. Moreover, many of the potential applications of a quantum network re- The direct transmission of information in a clas- quire only a quantum processor consisting of a sin- sical or quantum communication network is af- gle qubit, and thus not necessarily a large quan- fected by imperfections, i.e. by signal losses or tum computer. We may therefore consider quan- noise. Classical information is often transmitted as tum internet as a possible (temporary) solution for electromagnetic waves over optical or free-space the challenges that prevail in building a large-scale fibres [DLH17]. In the presence of noise or signal quantum computer. losses, the affected classical information can be recovered by reamplifying the signal at a so-called Potential applications classical repeater [DLH17] or simply by resending the information [WEH18]. However, these classi- The possibility of quantum communication over cal solutions of amplification and repetition cannot long distances opens many potential applications be applied when the transmission of quantum infor- for a future quantum internet. One of the main mation is affected by such imperfections. For direct expected applications of a quantum internet is transmission, quantum information is encoded as quantum cryptography, and thus a quantum in- photons that are transmitted over optical or free- ternet can be used to provide secure communica- space fibres, similar to methods used in classical tion [DLH17]. For example, we could implement communication [DLH17]. However, if a transmitted QKD [DLH17; WEH18]12 in order to establish a se- photon is either lost or affected by noise, then the cure secret key between any two points in the net- quantum information it carries is destroyed. More work. Besides QKD, a quantum internet will also precisely, as a consequence of the no-cloning the- allow us to realise protocols for two-party crypto- orem [Die82; WZ82], which states that we cannot graphic tasks [DW18; WEH18]. Another applica- copy the quantum state of a qubit, we cannot re-

12The realisation of these cryptographic protocols obviously depends on the assumptions imposed by the protocol, e.g. that the adversary’s quantum storage is noisy for the realisation of the 1-2 OT protocols in this paper.

18 Capstone Issue Vol. 12, 2019 Lynn Engelberts

cover the lost quantum information by sending it tion, we explain that such distributed entanglement again. As a result, the occurrence of losses and de- can be obtained by including intermediate nodes, phasing errors during transmission only allow for called quantum repeaters, which apply a technique sending photons over a few hundred kilometres called entanglement swapping. [DLH17], and hence we need to find a different way to transmit quantum information over longer dis- Constructing a quantum internet tances in order to establish a large-scale network. Although quantum error correction may provide We have explained how one can transmit quan- a method to overcome the photon losses and er- tum information from one node to another, and will rors, and thus may still enable the direct transmis- now describe how a quantum internet can be con- sion of photons over long distances [DLH17], fur- structed. ther research in this direction is required before The three concrete building blocks for a quan- we can conclude whether this is an acceptable so- tum internet are end nodes, photonic communica- lution. More precisely, this approach is currently tion channels and quantum repeaters, which will still challenging as it requires a relatively low error be described below [WEH18] (see Figure 6 for a rate [DLH17]. Nevertheless, an alternative tech- simplified illustration). The end nodes are quan- nique known as quantum teleportation was pro- tum processors, which may vary from a single qubit posed, which enables the exchange of quantum to multiple qubits forming a large-scale quantum information between two remote nodes [Ben+93]. computer, depending on the task that must be per- Quantum teleportation is a method to transmit an formed.13 At the end nodes, quantum states are unknown quantum state for which the sender and generated, processed and stored [Kim08]. More- receiver are solely required to share two entangled over, these end nodes are connected to the net- qubits and to be connected by a classical commu- work via a photonic channel—i.e. a quantum chan- nication channel. Hence, quantum teleportation nel—to enable the transmission of (quantum) infor- does not require a quantum communication chan- mation and distribute entanglement across the net- nel between the sender and the receiver at the work [Kim08]. Similar to the classical internet, this moment of transmission. In particular, quantum physical connection can be established via fibre- teleportation achieves long-distance transmission based channels (e.g. optical fibres), free-space of quantum information without physically send- channels, or a combination of these, provided that ing the particle that stores it [Cac+18] and with- photon loss and decoherence (i.e. loss of quantum out violating the principles of quantum mechan- information stored in a qubit over time) is minimal ics [NC10]. Since the quantum information is not [WEH18]. transmitted directly (i.e. physically), quantum tele- The third component of a quantum internet is portation prevents the transmission from being af- a device that increases the distance through which fected by noise or photon losses. the quantum information is transmitted: a quan- Although we have explained that quantum tele- tum repeater. More precisely, a quantum repeater portation is a crucial technique to establish a large- enables the distribution of entanglement over a scale quantum network, teleporting quantum in- long distance. Consequently, we can take advan- formation between two remote nodes forces these tage of this established entanglement to ‘teleport’ two nodes to be entangled with each other. Once quantum information over this longer distance. As entangled, the distance between the two nodes can was explained before, quantum teleportation probe arbitrarily long. However, establishing quantum vides a solution to the imperfections that hinder entanglement between the nodes cannot be done the long-distance transmission of quantum infor- over an arbitrarily long distance, since the entan- mation via photons. When the distance between glement distribution rate gradually decreases over the two end nodes is too large to establish the en- distance [Cac+18]. In other words, before we can tanglement required to perform quantum telepor- apply quantum teleportation over arbitrarily long tation, quantum repeaters are positioned along the distances we still need to find a way to provide en- communication channel as intermediate nodes to tanglement over such a distance. In the next sec- bridge this distance. In particular, quantum re-

13As will also be pointed out later on, Wehner, Elkouss and Hanson [WEH18] explain that not all applications of a quantum internet require a full-blown quantum internet; sometimes a slightly less developed “version” is already sufﬁcient to perform a certain task.

19 Capstone Issue Vol. 12, 2019 Lynn Engelberts

peaters are placed in such a way that the distance net. These challenges are imposed by quantum to each end node is small enough to establish en- mechanical features such as quantum entangle- tanglement over the channel [WEH18]. After gen- ment and no-cloning [Cac+18]. Since these fea- erating entanglement between itself and each of tures do not have a classical counterpart, a consid- the end nodes, the quantum repeater implements erable ‘paradigm shift’ is needed in order to coun- entanglement swapping (see Figure 7): it tele- teract these barriers. An example of such a barrier ports the entangled qubit that it shares with one is the occurrence of errors (e.g. due to decoher- of the end nodes to the other end node, such that ence or imperfect operations) when qubits inter- the two end nodes eventually share an entangled act with the environment [Cac+18]. Speciﬁcally, qubit [WEH18]. In this way, entanglement is estab- since quantum information cannot be copied, we lished over a distance that could be longer than the are unable to apply classical error-correction meth- maximum distance possible via direct transmission ods that depend on information cloning, and thus a [WEH18]. new solution is required [Cac+18]. Nevertheless, research is devoted to overcome the challenges Difﬁculties of establishing a quantum inter- posed by the development of a quantum internet net [Kim08] and the realisation seems to be

From the above, it is clear that different issues arise when aiming to establish a quantum inter-

Figure 6: A simpliﬁcation of the components of a quantum internet. The end nodes are the quantum processors where quantum states are generated, processed and stored. The quantum repeaters function as intermediate nodes, bridging the distances between the end nodes. The end nodes and quantum repeaters are connected via photonic channels (lines), over which quantum information is distributed.

technologically possible, even though we are still Secure two-party computation on a quantum at an early stage [DLH17]. Speciﬁcally, Wehner, internet Elkouss and Hanson [WEH18] even predict that a small-scale quantum internet will be established Fortunately, a fully developed world-wide quan- within a few years. Therefore, the current devel- tum internet is not required to implement proto- opments seem to provide an optimistic outlook for cols for secure two-party computation. In partic- the establishment of a quantum internet. ular, the process of establishing a quantum internet is divided into several developmental stages by [WEH18], including a description of the known applications of a quantum internet that can (already)

20 Capstone Issue Vol. 12, 2019 Lynn Engelberts

Figure 7: (Source: [MT13]) In order to increase the distance between two end nodes, one or more quantum repeaters are placed along the quantum channel that connects the end nodes. Here, three quantum repeaters are placed between end node A and end node E. For the end nodes and repeaters, entangled pairs are shared between adjacent nodes. First, repeater B teleports the qubit she shares with A to repeater C using her entanglement with repeater C. After this process of entanglement swapping, node A now shares an entangled pair with repeater C. In a similar way, entanglement swapping between repeater C, repeater D and end node E results in an entangled pair between repeater C and end node E. Then, once more entanglement swapping takes place to ensure that end node A and end node E share an entangled pair. All other qubits are freed. be established at each stage. This work demon- [DW18]. SimulaQron has the purpose to serve as a strates that protocols for secure two-party compu- framework in which software for quantum-internet tation can already, to some extent14, be realised in applications can be written and debugged. To the stage that allows for end-to-end transmission of the best of our knowledge, SimulaQron is the only qubits as well as the preparation and measurement framework for developing software for quantum- of a quantum state at any node [WEH18]. More- internet applications currently available.15 over, experimental results show that the essential elements required for this stage—including quantum repeaters—are within reach [WEH18]. There- Working of SimulaQron fore, we may expect that the realisation of secure We will brieﬂy describe the working of Simu- two-party computation is feasible in a near-future laQron, which is visualised in Figure 8.16 The vir- implementation of a quantum internet. tual simulation network is established by classically connecting so-called virtual nodes with each other. Simulating quantum internet: Simu- A virtual node is labelled by a name (say Alice) and laQron can be viewed as a server program running on a local classical computer that wants to connect to If a quantum internet will be deployed in the fu- the network. The network can be simulated locally ture, we will need adequate software for quantum- by running the different server programs all on one internet applications. For this reason, the quantum- physical computer, but a distributed simulation is internet simulator SimulaQron has been created also possible (i.e. different virtual nodes are con-

14Nevertheless, the realisation of these protocols is only possible if the inaccuracies in transmission and measurement and the probability that the prepared quantum state is lost are below a certain level. It is still an open question, however, what the speciﬁc bounds are for these three parameters are [WEH18]. 15Another quantum-network simulator called NetSquid is currently under development at QuTech in Delft, but not yet publicly available. 16For an extended description of the working of SimulaQron we refer the reader to [DW18].

21 Capstone Issue Vol. 12, 2019 Lynn Engelberts

nected to different computers). In other words, the quantum-internet applications in SimulaQron. virtual nodes corresponding to ‘Alice Computer’, There are two ways to program the quantum ‘Bob Computer’, and ‘Charlie Computer’ in Figure network simulated by SimulaQron. The first 8 could either be run on a single computer, but also method—called programming ‘in native mode’—is on two or three distinct classical computers. Simu- by directly assessing the back end of SimulaQron laQron uses an existing quantum-hardware simula- using the Twisted [Twi] library for Python. How- tor to simulate the quantum processor on each vir- ever, the recommended way is via the provided tual node, enabling a virtual node to simulate and classical-quantum-combiner (CQC) interface. The manipulate qubits. By default, SimulaQron uses the CQC interface could be perceived as an interme- stabilizer formalism17 as quantum-hardware sim- diate point between the application level and the ulator, but it also supports ProjectQ [SHT18] and SimulaQron back end, which allows the user to pro- QuTip [JNN12].18 gram at a higher level and hence facilitates the Besides simulating and manipulating qubits, a writing of code. Moreover, the CQC interface comes virtual node can also connect to other nodes in with both a Python and a C library that consist of the network to enable classical and (simulated) useful methods to perform quantum operations in quantum communication. This communication be- Python and C, respectively. tween virtual nodes is illustrated in Figure 8 by the lines between the different computers. The Simu- Benefits of SimulaQron laQron servers on the distinct virtual nodes allow for connecting the underlying simulated quantum Using SimulaQron has several advantages, as hardware in order to establish (simulated) quantum explained in [DW18]. First of all, SimulaQron pro- communication and quantum entanglement, which vides software developers with a toolkit to write is labelled “SimulaQron internal communication” in software that could later be put into practice on a the figure. The classical-quantum-combiner (CQC) real quantum internet with little or no adjustments. server in each node functions as a link between the In particular, application development is indepen- SimulaQron back end and the level at which appli- dent of the underlying quantum hardware (as a re- cations are written. In other words, applications sult of the CQC interface) and is promoted by pro- on a particular computer (i.e. virtual node) com- viding programming libraries for Python as well as municate with the SimulaQron server via this CQC C. Furthermore, the fact that it is written in Python server. In addition, as quantum-network applica- facilitates a further extension of SimulaQron. An- tions will often require classical communication as other benefit, which is of particular interest to our well, SimulaQron also allows for classical communi- work, is that we may simulate noise in the channel cation between distinct virtual nodes, labelled “ap- by turning the setting noisy-qubits on. Hence, plication communication” in the figure.19 SimulaQron seems to provide the right tools in or- Figure 9 illustrates how we can program der to implement our protocols for 1-2 OT.

Expected limitations fore important to keep in mind that we cannot use SimulaQron to accurately evaluate the robustness of our protocols. Nonetheless, we are still able to Nevertheless, SimulaQron also creates some simulate a noisy quantum channel, and hence can drawbacks, which may become a limitation for our still implement our protocols. Another drawback of implementations. A ﬁrst drawback of SimulaQron SimulaQron is that we cannot rely on its security for that is relevant to our implementations is that time real usage. More precisely, since the entanglement is not modelled accurately, and hence we cannot created with SimulaQron is only a simulation of en- simulate time-dependent noise [DW18]. It is there-

17The stabilizer formalism is a speciﬁc approach to simulate so-called stabilizer circuits on a classical computer, Thesewhich are quantum circuits that are restricted to speciﬁc quantum operations. For example, see [Got97] for more information. 18The underlying back end can be changed to ProjectQ or QuTip in the settings via simulaqron set backend. However, one could also implement another existing simulator as long as it supports working with Python [DW18]. 19More precisely, this classical communication can be established by opening a socket connection between the nodes. We are allowed to program such a client/server setup in Python, but the Python CQC library of SimulaQron also provides a built-in feature to establish classical communication via the methods sendClassical and recvClassical (see Section 6.1).

22 Capstone Issue Vol. 12, 2019 Lynn Engelberts

Figure 8: (Source: [DW18]) An illustration of the communication in a quantum network simulated by Simu- laQron.

Figure 9: (Source: [DW18]) An illustration of programming a quantum network simulated by SimulaQron, which can be done either by directly assessing the SimulaQron back end using Twisted (left) or by running applications via the CQC interface (right). tanglement, the security guarantees of real entan- Python [Pyt]. glement do not hold [DW18]. However, as we do not need to establish entanglement for our protocols, we do not expect this to become a limitation. Before we continue, we provide our reasoning for the choice of the Python environment for our code. Firstly, SimulaQron is itself written in Python VI. Implementation of the proto- and comes with a Python CQC library that provides cols us with the right tools to implement our algorithms in the simulated quantum network. Therefore, our Methods code corresponds to the examples provided in the SimulaQron documentation [Sim], making it under- In this section, we explain how we have imple- standable for current and future SimulaQron users. mented the protocols for 1-2 ROT (and 1-2 OT) in Secondly, Python is a well-known high-level pro- the noisy-storage model using SimulaQron and the gramming language. We therefore expect that our Python CQC library. The version of SimulaQron we code and implementations are clear to other users, work with is version 3.0.3 and our code is written in which may allow for further optimisation and use.

23 Capstone Issue Vol. 12, 2019 Lynn Engelberts

How to get started with SimulaQron 4. recvClassical(), which receives a classical message sent to this node and closes the Section 5.3 explains how SimulaQron estab- socket connection. lishes a simulated quantum network. We will now explain in more detail how SimulaQron allows us The second tool from the Python CQC library to program our protocols and which tools of Sim- that we use is the qubit object, which takes as in- ulaQron we have used for our implementations. put the corresponding CQCConnection (allowing for Before we can run our code we have to config- communication with SimulaQron) and is initialised ure the simulated quantum network. Since we con- to be in state |0i. The useful methods for a qubit sider two parties for our protocols, the setup for our object are: implementations only requires two virtual nodes: Alice and Bob. Although SimulaQron allows for a 1. X(), which applies the X operator to the qubit distributed simulation, we use the default configuration of SimulaQron in which the servers run in a 2. H(), which applies the Hadamard operator to centralised setting (i.e. localhost). In order to start the qubit the SimulaQron back end that consists of the virtual 3. measure(), which measures the qubit and re- nodes we run the command simulaqron start, turns the outcome which will by default start a network with five nodes (labelled Alice, Bob, Charlie, David and Eve). We For example, the following code corresponds to use the default back end (i.e. the stabilizer formal- sending qubit q in state |1i to node ‘Bob’: ism), as this back end is the most efficient in the # initialise the connection number of qubits [Sim]. We need to change two of with CQCConnection( " Alice " ) as Alice : the default settings for our implementations, which # the code should be done before starting the back end. These q = qubit(Alice) two settings are max-qubits, the number of qubits q.X( ) Alice .sendQubit(q, "Bob" ) allowed per quantum register in a virtual node (by default set to 20), and noisy-qubits, the option to For a more explicit description of the possibil- apply noise to the qubits transmitted over the net- ities of SimulaQron and the Python CQC library, work (by default turned off). we refer the reader to the documentation found in Furthermore, for our implementations we make [Sim]. use of the functionalities provided by the Python CQC library SimulaQron. As explained in Section Construction of the code 5.3, this library allows us to program our protocols using the CQC interface without needing to access Our code is based on Protocol 3.1, 3.2, 4.1, and the SimulaQron back end directly. In order to use 4.2. Using the above methods provided by the the library, we need to initialise a CQCConnection Python CQC library, we can implement the required object, which takes as input the name of the node quantum and classical information in the protocol, (say, Alice) to which it corresponds. A CQCConnec- as well as the manipulation of the qubits (i.e. the tion object enables the node to communicate with encoding of Alice’s random string xA as qubits to be SimulaQron and with other virtual nodes. Several sent to Bob). In order to facilitate the computations _ _ _ methods can be applied to a CQCConnection ob- in the code, Alice’s and Bob’s strings (s 0, s 1, m 0, _ ject, of which the following are of use for our work: and m 1) are programmed as Python list objects. Furthermore, during the protocol we print several 1. sendQubit(q, name), which sends qubit q to messages to indicate that communication between node name; Alice and Bob has occurred. For the privacy amplification step we use random binary matrices to 2. recvQubit(), which receives a qubit sent to serve as two-universal hash functions. this node; Additional methods for robust 1-2 OT 3. sendClassical(name, msg), which opens a socket connection and sends msg (integer or As explained in Section 4, in a realistic set- list of integers) to name; ting we may expect that the quantum channel is

24 Capstone Issue Vol. 12, 2019 Lynn Engelberts

(a) Alice’s terminal (b) Bob’s terminal

Figure 10: 1-2 ROT for Alice and Bob. Alice receives two lists, s_0 and s_1. Bob holds choice bit 0 and receives Alice’s output list s_0. noisy. In order to implement Protocol 4.2 for ro- via the setting t1, the resulting noise turned out to 20 1 bust 1-2 ROT we therefore need to find a way to be unrealistic: the error-probability approaches 2 deal with the errors induced by the channel. As as the number of qubits increases, which implies explained, we have decided to work with Reed- that the received information is completely ran- Solomon (RS) codes and use the reedsolo library dom. Therefore, we have made a change to the [Fil]. We are aware that there are more efficient en- underlying SimulaQron code by manually setting 1 coding and decoding techniques available, such as the error-probability to be 10 , as this would be more sum-product decoding for low-parity-density-check plausible in a real setting. (LDPC) codes as introduced by [Gal62]. However, there is to the best of our knowledge no suitable Results open-source implementation of such decoding algorithms that could be easily extended to our set- We have implemented both the idealised and ting. Moreover, our implementations are currently the robust protocols for 1-2 ROT and 1-2 OT in not designed with the purpose to be as efficient as the noisy-storage model using SimulaQron. The possible, and we do not intend to work with very code is written in Python and available on GitHub large values for the length of the encoded word. (https://github.com/lengelberts/simulaqron-qc). Therefore, we do not expect that using the slightly The source code can also be found in Appendix less efficient Reed-Solomon coding techniques will B. We emphasise that in order to run our code both introduce significant problems. simulaqron and cqc need to be installed as speci- Furthermore, in order to simulate the noise fied in the SimulaQron documentation [Sim]. in the quantum channel, we turn the setting noisy-qubits on. However, as also mentioned in Implementation of the idealised protocols [DW18], SimulaQron currently does not allow for an accurate simulation of noise. Although the setting We have implemented Protocols 3.1 and 3.2 for noisy-qubits induces some noise in the channel, 1-2 ROT and 1-2 OT, respectively, in the noisy- and the level of noise can be somewhat monitored storage model using SimulaQron. We have tested

our code both in the setting of a noiseless quan- plementation of the idealised protocol results in an tum channel and in the setting that Alice and Bob incorrect output for Bob. So the idealised protocol are connected by a noisy quantum channel, which breaks down in a noisy setting. We provide a few is simulated by turning the option noisy-qubits examples of running our code. 21 on and manually setting the error-probability to be 1 p = 10 . In the presence of noise, running our im- Example 1. The ﬁrst example is our implemen- 20Note that this error-probability is the probability parameter for the BSC model, as explained in Section 4.1. 21Note that our examples have the purpose to show the working of our implementations. In reality the number n of transmitted qubits is very large, much larger than we will consider for our examples.

25 Capstone Issue Vol. 12, 2019 Lynn Engelberts

tation of Protocol 3.1 for 1-2 ROT in the ide- received Alice’s output list s_0, corresponding to alised setting. 1-2 ROT is realised by the func- his choice bit c = 0. tions called Alice_ROT(l, n, waiting_time) and Bob_ROT(c, l, n). Here, the length of the output Example 2. In our second example, we show our lists is l = 10 and Bob holds choice bit c = 0. We implementation of Protocol 3.2 for 1-2 OT in the ide- also set n = 100 and the waiting time in seconds alised setting. between Step 2 and Step 3 is waiting_time = 2. For 1-2 OT, we wrote two functions called _ _ _ _ Before starting the SimulaQron back, we set the Alice OT(m 0, m 1, l, n, waiting time) and _ maximal number of qubits to 500, as we want to Bob OT(c, l, n). Again, we set l = 10, _ transmit n = 100 qubits (i.e. any other number n = 100, and waiting time = 2. However, greater or equal to 100 would also work). We there- Bob now holds choice bit c = 1 and, since fore run the following commands: we will now run 1-2 OT, Alice holds two 10-bit lists m_0 = [0,1,1,0,0,1,0,1,1,0] and $ simulaqron set max−qubits 500 _ $ simulaqron start m 1 = [0,1,1,1,0,1,1,0,1,1]. We use the same settings as in Example ??. We run the code in two separate Python pro- In Alice’s terminal we now run: cesses,22 one for Alice and one for Bob. In Alice’s terminal we run: >>> Alice_OT (m_0, m_1, l =10, n=100, waiting_time=2) >>> Alice_ROT( l=10, n=100, waiting_time=2) In the second Bob’s terminal we run:

In Bob’s terminal we run: >>> Bob_OT( c=1, l =10, n=100)

>>> Bob_ROT(c=0, l =10, n=100) The resulting prints and outputs are illustrated in The resulting prints and outputs are illustrated in Figure 11. As expected, Bob receives Alice’s input Figure 10. As the ﬁgure illustrates, Bob has indeed list m_1, corresponding to his choice bit c = 1.

(a) Alice’s terminal (b) Bob’s terminal

Figure 11: 1-2 OT for Alice and Bob. Alice inputs m_0 and m_1. Bob holds choice bit 1 and receives m_1.

Example 3. In our third example, we show $ simulaqron stop what happens when we run 1-2 ROT in a noisy- $ simulaqron set noisy−qubits on $ simulaqron start setting. Hence, we use the functions Alice_ROT and Bob_ROT. We therefore ﬁrst have to stop the Note that this will automatically set the value for SimulaQron back end, change the settings, and t1 to be 1. We use the same inputs as for our then start the back end again. We do this by typing ﬁrst example. That is, c = 0, l = 10, n = 100, and the following commands: waiting_time = 2.

22Our code needs to be run in two separate Python processes in order to enable Alice and Bob to exchange classical information via the sendClassical and recvClassical commands, because the sendClassical method will wait until a socket is set up to a remote node and will wait forever if such a socket is never set up. Nevertheless, starting the SimulaQron back end and changing the settings can be done from within one of the two terminals.

26 Capstone Issue Vol. 12, 2019 Lynn Engelberts

(a) Alice’s terminal (b) Bob’s terminal

Figure 12: 1-2 ROT for Alice and Bob in a noisy setting. Alice receives two lists, s_0 and s_1. Bob holds choice bit 0, but does not receive the same output as Alice’s output list s_0.

From Python in Alice’s terminal, we run: a separate code for the part of robust 1-2 ROT dur- >>> Alice_ROT( l=10, n=100, waiting_time=2) ing which information reconciliation (i.e. according to Protocol 4.1) is applied. The example below will In the second terminal we run: illustrate that our code for information reconcilia- >>> Bob_ROT(c=0, l =10, n=100) tion results in the expected outcome and is a work- The resulting prints and outputs are illustrated in ing code. However, we note that we obtain the Figure 12. However, the list received by Bob does same problem as when running our implementation no longer correspond to Alice’s output list s_0. The of Protocol 4.2 if Alice runs her code before Bob (i.e. present noise has affected the output. before we have called the function for Bob). More precisely, the Python terminal again stops responding. Nevertheless, as the code returns the desired Implementation of the robust protocol outcomes in the case that Bob first runs his code, In addition, we have considered the case for the we expect that the problem is not caused by our robust protocol for 1-2 OT in the presence of noise. use of the reedsolo library, but by something else. We have therefore investigated the noise model From further experimentation with the code, we that is used in SimulaQron (when turning the op- speculate that the problem is caused by the way tion noisy-qubits on). As was previously briefly classical communication is achieved: via the meth- outlined, we may manipulate the error-probability ods sendClassical and recvClassical. We recall via the setting t1, which represents the coherence from Section 6.1 that each time sendClassical is time of the qubits: the lower the value, the more called by (say) Alice, a socket connection is opened noise is added to the channel [Sim]. Nevertheless, to Bob, which is only closed after Bob has received as we mentioned before, the noise in the channel is his message by recvClassical. In fact, note that unrealistic, regardless of how we tune t1, and thus the information reconciliation protocol requires Al- we decided to fix manually the error-probability to ice to send two more classical strings to Bob (i.e. 1 be p = 10 . the two m − n strings r0 and r1). Therefore, we However, regardless of whether we allow noise presume that the problem is caused by an imper- in the channel or not, the implementation of the ro- fect opening or closing of the socket connection be- bust protocol for 1-2 ROT turned out to be currently tween Alice and Bob during this additional classical infeasible. More precisely, we wrote the two func- information. However, we emphasise the need to tions Alice_robust_ROT and Bob_robust_ROT ac- investigate further this encountered problem in or- cording to Protocol 4.2. Yet, when we run the code, der to eventually realise our robust 1-2 ROT imple- the Python terminals no longer respond, which oc- mentation. curs more or less halfway through the information Despite the fact that running the code cur- reconciliation part of the code. We have experi- rently does not work, we still provide the code mented with our code in order to find out what may of our implementation of the robust protocol cause the problem. In particular, we have provided for 1-2 ROT in Appendix B. However, as pre-

27 Capstone Issue Vol. 12, 2019 Lynn Engelberts

viously explained, we have been able to pro- Potential of SimulaQron vide a running code for Protocol 4.1 which establishes the information reconciliation part First of all, our implementations elucidate the needed for robust 1-2 ROT. The two implemented potential and some current limitations of Simu- functions are called Alice_reconciliation and laQron. SimulaQron turns out to be an accessible Bob_reconciliation. We provide an example of platform for users to explore the possibilities of a usage below. Recall that for the setting of 1-2 ROT potential future quantum network. SimulaQron is in a noisy quantum channel, when information rec- written in Python and provides a Python CQC li- onciliation is applied, Alice has two strings x0 and brary which consists of an extensive number of x1, and Bob holds a string x˜c for his choice bit c, say useful commands that facilitate the user to simu- 0. In particular, we may assume that the noise in late quantum communication and to apply quan- the quantum channel entails that x˜c 6= x0. Hence, tum operations. In particular, since a version of the task of information reconciliation is to achieve this CQC interface is intended to be available on xc for c = 0. the expected quantum network in the Netherlands in 2020 [DW18], software written for SimulaQron Example 4. We illustrate our implementation of could possibly be extended for use on this real Protocol 4.1. Our implementation only requires quantum network. classical information to be transmitted, and hence In addition, recall that SimulaQron has been we do not need to change any of the default set- designed with the goal to provide a developmen- tings and can start SimulaQron immediately: tal framework for writing and testing software for $ simulaqron start quantum-internet applications, quantum cryptography among these. Despite some limitations, our Alice holds two binary lists of length 10: results demonstrate the feasibility of implementing namely x_0 = [0,1,1,1,0,0,1,1,0,1] and 1-2 OT in SimulaQron, and hence the possibility of x_1 = [0,1,0,0,0,1,0,0,1,0]. Bob holds choice extending it to implementing any other secure two- bit c = 0 and x_c = [0,1,1,0,0,0,1,0,0,1]. party computation protocol. Furthermore, since the Note that the lists x_0 and x_c differ in two ele- quantum part of the considered protocols is sim- ments, since we assume that the quantum channel ilar to that of the BB84 protocol [BB84] for QKD, is noisy. Moreover, Alice and Bob agree on m = 16. our implementations also suggest the possibility Recall that RS codes can correct up to m−n = 3 2 of a QKD implementation in SimulaQron. Hence, errors, i.e. we may expect that the reedsolo codes our results indicate the potential of using Simu- from ? will be able to correct the two errors on laQron as environment for implementing quantum- Bob’s x_c. Indeed, that is what we see below. cryptographic protocols. In particular, our results In Alice’s terminal we run: show that SimulaQron could be used for quantum- >>> Alice_reconciliation(x_0, x_1, m=16, n=10) cryptographic implementations as long as a realistic time-dependent noise model is not desired for In Bob’s terminal we run: testing and the user (i.e. the writer of the code) >>> Bob_reconciliation(c=0, x_c, m=16, n=10) has sufﬁcient programming experience. However, The resulting prints and outputs are illustrated in note that the security of SimulaQron itself is not Figure 13. As the ﬁgure illustrates, Bob returns Al- comparable to the security of a real quantum in- ice’s input list x_0, corresponding to his choice bit ternet. Therefore, we emphasise that SimulaQron c = 0. is intended as a framework to develop software for quantum-internet applications such as quantum cryptography, and not actually to use this software. Nevertheless, we encountered two main prob- Discussion and implications lems for our implementations which explain some limitations with respect to SimulaQron. Firstly, In the previous section we have provided the re- the methods provided by the Python CQC library sults of our implementations and some examples. (sendClassical and recvClassical) to enable We will now discuss what these results entail in classical communication have very limited possi- greater detail. bilities. These methods merely allow us to trans-

28 Capstone Issue Vol. 12, 2019 Lynn Engelberts

mit (lists of) integers and not, for example, arrays Quantum internet and quantum cryptogra- or Python str objects. In addition, we speculate phy that the current method for establishing a socket connection between any two nodes for classical Our implementations provide an illustration of communication is causing a problem and explains performing the task of 1-2 OT over a potential quan- why our implementation of 1-2 ROT did not fully tum internet. Secure two-party computation is one succeed. Although we are aware that SimulaQron of the potential applications of a quantum internet also allows for the implementation of one’s own [WEH18], and this thesis shows how such crypto- socket connection [DW18], we consider a built-in graphic applications may be realised in a quantum feature for classical communication essential, since network. Moreover, we expect that our implemen- many quantum-cryptographic applications will re- tations might be of use in a future quantum inter- quire the use of classical communication [DW18]. net: the protocols on which our code is based are For these reasons, we recommend further updating designed in a relatively simple way and our code the possibilities of the Python CQC library to allow is written in Python, a well-known, high-level lan- for more realistic classical communication. guage. Nevertheless, we speculate whether we would really need a quantum internet to implement A second limitation was that noise is not ac- secure two-party computation. In fact, secure two- curately modelled by SimulaQron, as is also ex- party computation is already shown to be possible plained in [DW18]. Although our results show that with QKD hardware (e.g. see [Erv+14]). Moreover, the noise model allows us to demonstrate that pro- it is questionable whether we really need a large- tocols may break down in the case of a noisy quan- scale network in order to implement the considered tum channel, the level of noise it achieves is un- protocols in practice, as it is reasonable to consider realistic, as previously explained. For this reason, that secure two-party computation is performed at it becomes difficult to demonstrate the working of a short distance [Ben+92b]. protocols in a noisy setting. Nevertheless, by man- In particular, we have not only showed how the ually setting the error-probability to a certain (time- idealised setting of 1-2 OT (for the noisy-storage independent) level, we were able to get a more re- model) can be implemented in a future quantum alistic level of noise, and we expect that this ap- internet, but have also considered a more practical proach may be sufficient for future users as well. setting, namely when the quantum communication is affected by noise. Although we have not been Furthermore, we note that SimulaQron is still at able to run our code for robust 1-2 ROT in Simu- a developmental stage: in the period of this work laQron, we have demonstrated the need for such SimulaQron has been updated from version 1.3 to protocols, since the idealised protocols break down version 3.0.3. The main difficulties we encountered in a noisy setting. Moreover, we may expect that when using SimulaQron generally occurred during it will eventually be feasible to implement (a ver- the installation of SimulaQron and when updating sion of) our code for robust 1-2 ROT in SimulaQron to the newer versions. In particular, we encoun- as we have provided an explicit description of infor- tered problems when installing SimulaQron in Win- mation reconciliation in the context of 1-2 OT and a dows. Moreover, one needs to be careful when us- working code to establish it. ing SimulaQron as even minor errors such as ty- pos could result in SimulaQron malfunctioning. Al- though these difficulties have provided significant Other implications of our work obstacles during the process of writing our code, the updated newer versions of SimulaQron also al- One of the benefits of our work is that our im- lowed for more consistent and clear documenta- plementations are all publicly available on GitHub. tion. In addition, the frequent updates of Sim- Also, the decoding algorithm23 that we have cho- ulaQron forced us to be very aware of any new sen to use for the implementation of the robust changes in SimulaQron, which resulted in a deeper protocol is publicly accessible. This is in contrast understanding of the working of SimulaQron and its to the currently existing implementations of 1-2 possibilities. OT (namely those implemented using QKD hard-

23However, we note that our code is most likely not optimal, since most techniques previously used are not available for public use but are probably “better” in terms of efﬁciency. Therefore, we recommend further improvements be made to our code.

29 Capstone Issue Vol. 12, 2019 Lynn Engelberts

ware), for which the used code is licensed and (to ture quantum internet. On the other hand, we con- our knowledge) not publicly available. Our work is tributed to the field of quantum cryptography by more easily accessible than other implementations providing a quantum-internet simulation of proto- of 1-2 OT, and could therefore be used for future re- cols for 1-2 OT. Additionally, our work shows that search in this direction and to further improve the the protocols provided in [Sch10] are indeed ad- code. equate for quantum-internet applications and are Moreover, as our implementations are simple, feasible to implement. clear, and publicly accessible, our work could be used for learning purposes. More precisely, since Future research we have explicitly described the steps in our work and since our code is properly documented, it may At the same time, our work reveals several po- be a good source for newcomers in the field as it ex- tential areas for future research. plains both the working of SimulaQron (and hence Firstly, we emphasise that our code is not cur- a future quantum internet) and of secure two-party rently designed to be as efficient as possible and computation in the setting of a noisy-quantum could be further optimised. In particular, future re- memory. Previous work, such as in [Sch10] and search may be devoted to improving the decoding [KWW12], provides a complete description of 1-2 part in our implementations of robust 1-2 OT. In ad- OT in the noisy-storage model, yet some steps are dition, it would be valuable to further investigate not included. Although these steps are trivial to ex- why our code for robust 1-2 OT currently does not perts in the field, our work may be a valuable clar- work, which may become beneficial for the devel- ification for newcomers and outsiders. In addition, opment SimulaQron as well. Moreover, as previ- our code could potentially be used to provide more ously explained, the current built-in client/server examples in the SimulaQron documentation, as the setup for classical communication in SimulaQron documentation currently consists of a limited num- is quite restrictive and inefficient, and hence we ber of examples. would advise that this feature be improved in the future. Furthermore, it may be noteworthy to consider VII. Conclusion the simulation of possible adversarial attacks. Al- though the security of the implemented protocols The two primary aims of this study were (i) to is already proven for the given parameters, such ascertain the suitability of the protocols of [Sch10] an imitation of possible attacks may further clarify for quantum-internet applications and (ii) to inves- the setting of secure two-party computation. This tigate the extent to which SimulaQron is an ade- could prove useful for a better understanding of quate environment for implementing and analysing these protocols, but may also explain to outsiders quantum-cryptographic protocols. of the field the significance and relevance of such To the best of our knowledge we have pro- quantum-cryptographic protocols. vided the first implementations of quantum- Another significant research direction may be cryptographic protocols in SimulaQron. As a re- to compare our implementations with existing im- sult, this thesis elucidates the potential of Simu- plementations of 1-2 OT using QKD hardware. Al- laQron and some of its limitations. In particular, though our quantum-cryptographic implementa- our implementations show that SimulaQron may be tions are, to the best of our knowledge, the first a suitable environment for implementing quantum- ones in the setting of a quantum internet, there cryptographic protocols, and indicate some issues are already experimental implementations of 1-2 that may be overcome in the future. In this way OT using QKD hardware (e.g. see [Erv+14]). There- we contribute to a further development of Simu- fore, it would be an interesting direction of future laQron, but also to the current development of a research to compare the performance and feasibil- quantum internet. More precisely, we have illus- ity of the implementations in both settings. How- trated how quantum-cryptography may be used on ever, it should be noted that—as previously ar- a future quantum internet. The analysis in our work gued—SimulaQron does not allow for a realistic may therefore be significant as quantum cryptog- comparison of time performance, and therefore we raphy is one of the expected applications of a fu- recommend using another simulator (potentially

30 Capstone Issue Vol. 12, 2019 Lynn Engelberts

NetSquid) to continue in this direction. Deﬁnition 14. (Conditional probability distri- Lastly, SimulaQron is, to date and to our knowl- bution) edge, the only quantum-internet simulator. How- If A is an even such that P [A] > 0, then we deﬁne ever, as another simulation framework, NetSquid the conditional probability distribution of X given [Net], is under development, it would be interest- A by ing to compare SimulaQron’s working, qualities and P [X = x, A] PX|A(x) := . limitations with those of NetSquid. In particular, P [A] we have explained that although our results show If Y is another random variable and PY (y) > 0, then that SimulaQron provides a suitable developmental we write framework for quantum-internet applications, it is not ideal. Therefore, it may be interesting to inves- PXY (x, y) PX|Y (x|y) := PX|Y =y(x) = tigate whether NetSquid overcomes the limitations PY (y) of SimulaQron. for the conditional distribution of X given Y = y.

Appendix A Appendix B Probability theory We provide the source code for our implemen- We define some basic notions of probability the- tations. The code is also available on GitHub, click ory, using the definitions from [DS17]. here. Definition 9. (Probability measure) A probability measure P is a function P :Ω → R≥0 such that Works Cited X P(ω) = 1. ω∈Ω [ALA18] Ebrahim Ardeshir-Larijani and Farhad Arbab. “Reo coordination model for simu- Definition 10. (Probability space) lation of quantum internet software”. In: A (discrete) probability space is a triple (Ω, F, ), P Software Technologies: Applications and where Ω is a non-empty sample space, F is an Foundations. STAF 2018. Ed. by Manuel event space, and is a probability measure. P Mazzara, Iulian Ober, and Gwen Salaün. Definition 11. (Random variable) Vol. 11176. Lecture Notes in Computer Sci- A discrete random variable X on a discrete prob- ence. Springer, Cham, 2018, pp. 311–319. ability space (Ω, F,P ) is a function X :Ω → X , isbn: 9783030047702. where X is a (discrete) set. [BB84] C. H. Bennett and G. Brassard. “Quantum Definition 12. (Probability distribution) cryptography: Public key distribution and Let X be a random variable on the set X . The prob- coin tossing”. In: Proceedings of IEEE In- ternational Conference on Computers, Sys- ability distribution of X is a function PX : X → [0, 1] defined as tems, and Signal Processing. Bangalore, 1984, pp. 175–179. PX (x) := P[X = x], [BBR88] Charles H. Bennett, Gilles Brassard, and where X = x denotes the event {ω ∈ Ω|X(ω) = x}. Jean-Marc Robert. “Privacy Amplification Note that PX is often called the marginal probabil- by Public Discussion”. In: SIAM Journal on ity distribution of X. Computing 17.2 (1988), pp. 210– 229. doi: Definition 13. (Joint probability distribution) 10.1137/0217014. Let X and Y be two random variables defined on [Ben+92a] Charles H. Bennett et al. “Experimen- the same probability space, with ranges X and Y, tal quantum cryptography”. In: Journal of respectively. The pair XY is a random variable and Cryptology 5.1 (1992), pp. 3–28. issn: has probability distribution PXY : X × Y → [0, 1] 1432-1378. doi: 10.1007/BF00191318. given by url: https://doi.org/10.1007/BF00191318. [Ben+92b] Charles H. Bennett et al. “Practical PXY (x, y) := P[X = x, y = Y]. quantum oblivious transfer”. In: vol. 576.

31 Capstone Issue Vol. 12, 2019 Lynn Engelberts

Springer Verlag, 1992, pp. 351–366. isbn: — FOCS 2005 (2005), pp. 449–458. arXiv: 9783540551881. quant-ph/0508222v2. [Ben+93] Charles H. Bennett et al. “Teleporting [Dam+07] Ivan B. Damgård et al. “A tight high- an unknown quantum state via dual classi- order entropic quantum uncertainty rela- cal and Einstein-Podolsky-Rosen channels”. tion With applications”. In: Advances eng. In: Physical review letters 70.13 in Cryptology — CRYPTO 2007. Vol. (1993). issn: 1079-7114. 4622. Lecture Notes in Computer Sci- [BS94] Gilles Brassard and Louis Salvail. “Secret- ence. Springer, Berlin, Heidelberg, 2007, key reconciliation by public discussion”. In: pp. 360– 378. arXiv: quant-ph/0612014v2. vol. 765. Springer Verlag, 1994, pp. [Die82] Dennis Dieks. “Communication by EPR 410–423. isbn: 9783540576006. devices”. In: Physics Letters A 92.6 [BCS12] Harry Buhrman, Matthias Christandl, and (1982), pp. 271–272. doi: 10.1016/0375- Christian Schaffner. “Complete insecurity 9601(82)90084-6. of quantum protocols for classical two- [DS17] Yfke Dulek and Christian Schaffner. Infor- party computation”. In: Physical Review mation Theory. Lecture Notes. University Letters 109.16 (2012), p. 160501. doi: of Amsterdam, Master of Logic. 2017. 10.1103/PhysRevLett.109.160501. arXiv: [DLH17] Wolfgang Dür, Raphael Lamprecht, and 1201.0849v2. Stefan Heusler. “Towards a quantum inter- [Cac+18] Angela Sara Cacciapuoti et al. “Quan- net”. In: European Journal of Physics 38.4 tum Internet: Networking Challenges in (2017), p. 043001. doi: 10.1088/ 1361- Distributed Quantum Computing”. In: (Oct. 6404/aa6df7. 19, 2018). arXiv: 1810.08421v2 [quant- [DM04] Stefan Dziembowski and Ueli Maurer. “On ph]. Generating the Initial Key in the Bounded- [CCB18] Marcello Calefﬁ, Angela Sara Cacciapuoti, Storage Model”. In: Advances in Cryptol- and Giuseppe Bianchi. “Quantum internet: ogy - EUROCRYPT 2004. Ed. by Chris- from communication to distributed com- tian Cachin and Jan L. Camenisch. Springer puting!” In: (2018). arXiv: 1805. 04360v1. Berlin Heidelberg, 2004, pp. 126–137. [CW79] J.Lawrence Carter and Mark N. Weg- isbn: 1978-3-540-24676-3. man. “Universal classes of hash functions”. [EMMM11] David Elkouss, Jesus Martinez-Mateo, In: Journal of Computer and System Sci- and Vicente Martin. “Information Rec- ences 18.2 (1979), pp. 143–154. doi: onciliation for Quantum Key Distribution”. 10.1016/0022-0000(79)90044-8. In: Quantum Information and Computation, [Cas18] Davide Castelvecchi. “The quantum in- 11.34 (2011). arXiv: 1007.1616v2 [quant- ternet has arrived (and it hasn’t)”. In: ph]. Nature 554 (2018), pp. 289–292. doi: [Erv+14] Christopher Erven et al. “An experimen- 10.1038/d41586-018-01835-3. tal implementation of oblivious transfer in [DW18] Axel Dahlberg and Stephanie Wehner. the noisy storage model”. In: Nature Com- “SimulaQron—a simulator for developing munications 5.1 (2014). doi: 10. 1038/n- quantum internet software”. In: Quan- comms4418. tum Science and Technology 4.1 (2018), p. [EGL82] Shimon Even, Oded Goldreich, and Abra- 015001. doi: 10.1088/2058-9565/aad56e. ham Lempel. “A randomized protocol [Dam+09] Ivan Damgaard et al. “Improving for signing contracts”. In: Advances in the Security of Quantum Protocols via Cryptology-CRYPTO ’82. Plenum, 1982, pp. Commitand-Open”. In: Advances in 205–210. Cryptology - CRYPTO 2009, LNCS 5677, [Fil] Tomer Filiba. reedsolo 0.3. url: pages 408-427 (Feb. 23, 2009). arXiv: https://pypi.org/project/reedsolo/. 0902.3918v4 [quant-ph]. [Gal62] Robert G. Gallager. “Low-density parity- [Dam+05] Ivan B. Damgård et al. “Cryptography check codes”. In: IRE Transactions on Infor- in the bounded quantum-storage model”. mation Theory 8.1 (1962), pp. 21–28. issn: In: Proceedings of the 46th IEEE Sympo- 0096-1000. sium on Foundations of Computer Science [GPCZ18] Ran Gelles, Anat Paskin-Cherniavsky, and

32 Capstone Issue Vol. 12, 2019 Lynn Engelberts

Vassilis Zikas. Secure Two-Party Compu- A 56.2 (1997), pp. 1154–1162. doi: tation over Unreliable Channels. Cryp- 10.1103/PhysRevA.56.1154. tology ePrint Archive, Report 2018/506. [LC97] Hoi-Kwong Lo and Hoi Fung Chau. “Is https://eprint.iacr.org/2018/506. 2018. quantum bit commitment really possible?” [GV88] Oded Goldreich and Ronen Vainish. “How In: Physical Review Letters 78.17 (1997), to solve any protocol problem - an efﬁ- pp. 3410–3413. doi: 10.1103/Phys- ciency improvement (extended abstract)”. RevLett.78.3410. In: Advances in Cryptology — CRYPTO ’87. [Mac03] David J. C. MacKay. Information Theory, Ed. by Carl Pomerance. Vol. 293. Lec- Inference and Learning Algorithms. Cam- ture Notes in Computer Science. Springer, bridge University Pr., Sept. 1, 2003. isbn: Berlin, Heidelberg, 1988, pp. 73–86. doi: 0521642981. 10.1007/3-540-48184-2_6. [May97] Dominic Mayers. “Unconditionally se- [Got97] Daniel Gottesman. “Stabilizer Codes and cure quantum bit commitment is impossi- Quantum Error Correction”. In: (May 28, ble”. In: Physical Review Letters 78.17 1997). arXiv: quant-ph/9705052v1 [quant- (1997), pp. 3414–3417. doi: 10.1103/Phys- ph]. [GJC12] Daniel Gottesman, Thomas RevLett.78.3414. Jennewein, and Sarah Croke. “Longer- [MT13] Rodney Meter and Joe Touch. “Design- Baseline Telescopes Using Quantum Re- ing quantum repeater networks”. In: IEEE peaters”. In: Physical Review Let- Communications Magazine 51.8 (2013), ters 109.7 (2012). doi: 10.1103/phys- pp. 64–71. doi: 10.1109/mcom.2013. revlett.109.070503. 6576340. [JNN12] J.R. Johansson, P.D. Nation, and Franco Nori. [Net] NetSquid. url: https://netsquid.org/. “QuTiP: An open-source Python framework [NC10] Michael A. Nielsen and Isaac L. Chuang. for the dynamics of open quantum sys- Quantum Computation and Quantum Infor- tems”. In: Computer Physics Communica- mation. Cambridge University Pr., Dec. 1, tions 183.8 (2012), pp. 1760–1772. doi: 2010. isbn: 1107002176. 10.1016/j.cpc.2012.02. 021. [Pyt] Python. url: https://www.python.org/. [Kil88] Joe Kilian. “Founding cryptography on oblivi- [QuT] QuTech. Quantum Internet and Networked ous transfer”. In: Proceedings of the Twen- Computing. url: https://qutech.nl/roadmap tieth Annual ACM Symposium on Theory of /quantum-internet/. Computing. STOC ’88. ACM, 1988, pp. [Rab81] Michael O. Rabin. How to exchange secrets 20–31. doi: 10.1145/62212.62215. by oblivious transfer. Technical Report TR- [Kim08] H. J. Kimble. “The quantum inter- 81. Harvard University, 1981. net”. In: Nature 453 (2008), pp. [RS60] I. S. Reed and G. Solomon. “Polynomial 1023–1030. doi: 10.1038/nature07127. Codes Over Certain Finite Fields”. In: Jour- arXiv: 0806.4195v1. nal of the Society for Industrial and Applied [Kóm+14] Peter Kómár et al. “A quantum network Mathematics 8.2 (1960), pp. 300– 304. doi: of clocks”. In: Nature Physics 10.8 (2014), 10.1137/0108018. pp. 582–587. doi: 10.1038/nphys3000. [RSA78] R. L. Rivest, A. Shamir, and L. Adleman. [KWW12] Robert König, Stephanie Wehner, and “A method for obtaining digital signatures Jürg Wullschleger. “Unconditional se- and public-key cryptosystems”. In: Com- curity from noisy quantum storage”. munications of the ACM 21.2 (1978), pp. In: IEEE Transactions on Information 120–126. doi: 10.1145/359340.359342. Theory 58.3 (2012), pp. 1962–1984. [Sca+09] Valerio Scarani et al. “The Security of doi: 10.1109/TIT.2011.2177772. arXiv: Practical Quantum Key Distribution”. In: 0906.1030v4. Reviews of Modern Physics 81.3 (2009), pp. [LC83] Shu Lin and Daniel J. Costello. Error Control 1301–1350. doi: 10 . 1103 / RevMod- Coding: Fundamentals and Applications. Phys.81.1301. arXiv: 0802.4155v3. Prentice Hall, 1983. isbn: 013283796x. [Sch10] Christian Schaffner. “Simple protocols for [Lo97] Hoi-Kwong Lo. “Insecurity of quantum se- oblivious transfer and secure identiﬁcation cure computations”. In: Physical Review in the noisy-quantum-storage model”. In:

33 Capstone Issue Vol. 12, 2019 Lynn Engelberts

Physical Review A 82.3 (2010), p. 032308. Quantum World”. In: (2008). arXiv: doi: 10.1103/PhysRevA.82.032308. arXiv: 0806.3483v1 [quant-ph]. 1002.1495v2. [WEH18] Stephanie Wehner, David Elkouss, and [STW09] Christian Schaffner, Barbara M. Terhal, Ronald Hanson. “Quantum internet: a vi- and Stephanie Wehner. “Robust cryptogra- sion for the road ahead”. In: Science phy in the noisy-quantum-storage model”. 362.6412 (2018). doi: 10.1126/science. In: Quantum Information and Computa- aam9288. tion 9.11-12 (2009), pp. 963–996. arXiv: [WST08] Stephanie Wehner, Christian Schaffner, 0807.1333v3. and Barbara M. Terhal. “Cryptography [Sha48] Claude E. Shannon. “A mathematical the- from noisy storage”. In: Physical Review ory of communication”. eng. In: The Bell Letters 100.22 (2008), p. 220502. doi: System Technical Journal 27.3 (1948), pp. 10.1103/PhysRevLett.100.220502. arXiv: 379–423. issn: 0005-8580. 0711.2895v3. [Sho95] Peter W. Shor. “Polynomial-Time Algo- [Weh+10] Stephanie Wehner et al. “Implemen- rithms for Prime Factorization and Dis- tation of two-party protocols in the noisy- crete Logarithms on a Quantum Com- storage model”. eng. In: Physical puter”. In: SIAM J.Sci.Statist.Comput. Review A: Atomic, Molecular and Opti- 26 (1997) 1484 (Aug. 30, 1995). cal Physics 81 (2010), urn:issn:1050–2947. doi: 10.1137/S0097539795293172. arXiv: issn: 1050-2947. quantph/9508027v2 [quant-ph]. [WZ82] William K. Wootters and Wojciech H. Zurek. [Sim] SimulaQron. url: http://simulaqron.org/. “A single quantum cannot be cloned”. In: [SHT18] Damian S. Steiger, Thomas Häner, and Nature 299.5886 (1982), pp. 802–803. doi: Matthias Troyer. “ProjectQ: an open source 10.1038/299802a0. software framework for quantum comput- [Yao82] Andrew C. Yao. “Protocols for secure ing”. In: Quantum 2 (2018), p. 49. doi: computations”. In: 23rd Annual Sympo- 10.22331/q-2018-01-31-49. sium on Foundations of Computer Science [TL17] Marco Tomamichel and Anthony Leverrier. (sfcs 1982). IEEE, 1982. doi: 10.1109/ “A largely self-contained and complete se- sfcs.1982.38. curity proof for quantum key distribution”. [Zho+12] Zong-Quan Zhou et al. “Realization of In: Quantum 1 (2017), p. 14. Doi: Reliable Solid-State Quantum Memory for 10.22331/q-2017-07-14-14. Photonic Polarization Qubit”. In: Physi- [Twi] Twisted. url: https://twistedmatrix.com/trac/. cal Review Letters 108.19 (2012). doi: [Weh08] Stephanie Wehner. “Cryptography in a 10.1103/physrevlett.108.190505.

34 Sciences

Restricted Boltzmann Machines and the Renormalization Group: Learning Relevant Information in Statistical Physics

Jesse Hoogland Supervisor Dr. P. Marcos Crichigno (UvA) Reader Prof. Dr. Max Welling (UvA) Capstone Issue Vol. 12, 2019 Jesse Hoogland

Abstract

Recent work has drawn attention to the links between statistical physics and machine learning (ML) and, in particular, to comparisons between the renormalization group (RG) and deep neural networks, respectively. These have inspired renewed interest in the information-theoretic framework underpinning these ﬁelds, prompting a better understanding of what RG is. In this capstone, we introduce and expand upon these connections from the ground up. Starting with the basics of ML and RG, we work our way to an algorithm implemented on neural networks that learns optimal, model-independent RG procedures: the real-space mutual information (RSMI) algorithm. In doing so, we review the current state of the literature, clarifying misconceptions in earlier works. With the RSMI algorithm, we review a novel calculation of the Ising model critical exponent and generalize this approach to arbitrary lattice systems. We release an open-source library, rgpy, for implementing these novel procedures, and close with a discussion of the wide-ranging implications.

Keywords and phrases: machine learning, restricted Boltzmann machines, the renormalization group, information theory, mutual information

Acknowledgements searchers in both fields can leverage to solve tough problems. I would like to thank Marcos Crichigno for super- In particular, recent work has drawn attention vising this capstone. I do not think either of us knew to the similarities between ML and a class of tech- what we were getting into when I first approached niques from statistical physics known as the renor- him, and I am stunned at how much I have been able malization group (RG) [4, 5, 6, 2]. Developed in the to learn during this time. Marcos provided guidance, last century, RG has been crucial in making sense of much-needed prodding, and a watchful eye, keeping critical behavior – those phenomena characterizing the focus on the big picture. The topics were new to phase transitions. In 2014, Mehta and Schwab pub- both of us, ML for him and RG for me, and getting to lished a seminal paper describing an exact equiva- integrate these different perspectives was a wonder- lence between a technique from RG and a type of ful experience. I would also like to thank my dad, Jiri neural network (NN) from ML [4]. This, however, Hoogland, for helpful feedback on drafts and, most encountered criticism, and it works only under a nar- of all, for his expertise on on all things Monte Carlo. row set of circumstances. The similarities between After being stuck with implementation details of the ML and RG, then, are still largely qualitative, and RSMI algorithm for two weeks, his words finally let this remains an active area of research. In addition, the penny drop. As a last note, I would like to thank the research landscape maintains lingering miscon- my friends and (other) family for putting up with ceptions about the details of the intersection [6, 2, me during the last weeks (when I was a little less 7, 8]. This warrants further investigation, and in than sociable), in particular, Nick van der Woude, order to facilitate and encourage such research, our for mutually infectious enthusiasm. first contribution is to provide a clarifying overview of the competing views; in doing so we resolve a I. Introduction number of inaccuracies. In 2018, Koch-Janusz and Ringel derived an al- In the age of big data, machine learning (ML), gorithm which uses neural networks to learn RG a subset of artificial intelligence (AI), has become transformations on lattice systems: the real-space more than just another set of data analysis tools mutual information (RSMI) algorithm [5]. Notably, [1]. First, ML’s connections with theoretical physics this method is unsupervised, which is particularly are multivarious and deeply conceptual; the very relevant for research into poorly understood physi- success of ML may, in part, result from physical prin- cal systems; information-theoretic approaches like ciples including symmetry, locality, and hierarchy this algorithm may guide researchers towards the [2]. Furthermore, ML and theoretical physics share locations of critical points and even calculations of a powerful conceptual framework in information the- critical exponents. Furthermore, Koch-Janusz and ory [3]. Beyond data analysis, the intersection of ML Ringel’s derivation is optimal in the rigorous sense and physics contains a unique set of ideas that re- we define in Chapter 5 [5, 9]. This is exciting be-

34 Capstone Issue Vol. 12, 2019 Jesse Hoogland cause many well-established practices in RG lack culation of ν and a generalization of this technique, precise justification. More exact formulations like paving the way for a new class of RG techniques. these may inspire more effective implementations, In Chapter 7, we close with a discussion, reflecting not to mention a better understanding of why these on our comparisons of ML and statistical physics ML and RG techniques work. and emphasizing the wide-ranging impacts of these In our investigation, we will develop a set of ideas. tools for tackling critical phenomena. First, we consider some of the standard techniques of statistical Notation physics [2], building towards an ML-derived implementation of RG [5]. Second, we introduce elements To refer to single microscopic elements (e.g. of ML, emphasizing their utility in a variety of statis- spins in the Ising model or pixels in an image), we tical physics contexts [3]. We anchor this investiga- use lowercase letters with a lower index (xi, yj, etc.). tion around the Ising model, a description of ferro- To refer to collections of microscopic elements, mi- magnetism and one of the most important models crostates or images, we use boldface, lowercase in statistical physics. To compare these various tech- letters (x := {xi}, y := {yj}, etc.). niques, we evaluate their ability in predicting the To refer to collections of microstates, we use up- Ising model’s correlation length critical exponent, ν. percase, cursive letters, X := {x}. We will be inter- To accomplish this, we have built and shared an ested in performing sums and averages over these implementation of the RSMI algorithm [10] in the sets. Rather than introduce an index to keep track open-source Python library rgpy 1 [6]. Hereby, we of each term, we do so implicitly in the sums. For ex- provide a calculation for ν [6]. Then, we describe ample, given some function A(x), the following are P (n) P P a generalization of this algorithm to arbitrary lat- equivalent: n A(x ) ≡ x∈X A(x) ≡ x A(x). tice systems, giving rise to a family of RSMI-inspired Most often, we use the last notation. approaches. If we partition our microstates into subsets (as It is our aim to enable and inspire researchers with block renormalization), we also use boldface, to build further on our results. We accomplish this lowercase letters (v, h, etc.). To distinguish parti- by reviewing the current state of research, sharing tions, we may use an upper index: v(n). an implementation of the RSMI algorithm, and de- For partial derivatives, we typically use the short- : ∂ scribing avenues of future research. Although we hand ∂t = ∂t . focus on the perspective of statistical physicists, For Ising models, we will consider systems with this capstone is accessible for both ML and physics binary units ∈ {−1, 1}, following standard conven- researchers, even at the undergraduate level. tion. For RBMs, we use the standard notation of In Chapter 2, we begin by introducing techniques binary units ∈ {0, 1}. When using RBMs on Ising native to statistical physics. We describe mean-field data, then, we map −1 → 0. theory and its failures, which brings us to the renormalization group. In Chapter 3, we discuss two examples of ML in physical investigations. First, we II. Foundations of Statistical use neural networks to classify Ising model phases. Physics Then, we use the same neural networks to generate new samples of Ising models. These examples Statistical physics emerged in the second half of serve to introduce the basics of ML, assuming no the nineteenth century as an answer to unresolved prior knowledge (except mathematical maturity), questions in thermodynamics, the study of heat and and the same is true for the portion on statistical work. Was heat continuous and wavelike, or might it physics. In Chapter 4, we explore the similarities be something else, discrete and atomic? Founding between ML and RG, and by being more explicit in figures in the field, such as Rudolf Clausius, Ludwig how we define “relevant” information, we achieve a Boltzmann, and James Clerk Maxwell, answered the more precise comparison between the two. In Chap- latter. Introducing the kinetic theory of gases, these ter 5, we explain and justify the RSMI algorithm, scientists posited gases as large collections of tiny following the formulation of Koch-Janusz and Ringel. molecules and heat flow as the net effect of unbal- In Chapter 6, we provide our own results: a recal- anced molecular collisions. By translating these mi-

1A programming library contains a set of functionalities easily accessible to other computer scientists.

35 Capstone Issue Vol. 12, 2019 Jesse Hoogland croscopic descriptions to experimentally-verifiable, know its distribution over microstates s, P (s). The macroscopic predictions, these physicists were able trick is to introduce a reservoir, R, that surrounds to defend this theory of gases. The techniques pi- S. Just like S, the reservoir could be anything: a oneered in performing this translation would give gas, liquid, etc., where we impose two conditions: rise to the field of statistical physics [11]. S and R exchange only energy, and the combined To complicate matters, these scientists lacked system, X = S ∪ R (the universe), is isolated. Then, equipment that could resolve the proposed micro- the total energy, E, is conserved. If we denote the scopic length scales, and a square cubic centimeter energy of microstates, s and r = {rj} as E(s) and of gas can contain upwards of a million million mil- E(r), respectively, energy conservation requires lion molecules [11]. The key insight in statistical E(s) + E(r) = E.3 Furthermore, we treat S as a physics is to focus on the properties of the collec- much smaller fraction of X than R, though S is still tion rather than on the individual components – on macroscopic (1 |s| |r|, where |x| denotes the averages and distributions rather than microscopic number of degrees of freedom of x). details. The translation between microscopic and The fundamental assumption of statistical me- macroscopic is the essence of statistical physics, chanics states that, for an isolated system, each and it is to this task we dedicate our efforts. microstate is equally probable. Then, our probability Our investigation begins by defining a mi- distribution is P (x) = 1/Ω(x), where Ω(x) is the crostate, s: a full description of the microscopic de- number of possible microstates x. Probabilities over grees of freedom of our system, s := {si}, where si subsets of such a system may be more complicated. is the i-th DOF, some fundamental way in which the Consider that the probability of s is proportional to system can vary.2 Our aim, as statistical physicists, the number of ways we can rearrange the reservoir, is to predict the outcomes of macroscopic measure- keeping the energy constant. If we let Ωr(s) de- ments. A key assumption of statistical mechanics note the number of microstates, r, with this energy is that we can express measurement outcomes as E − E(s): averages over all microstates, S (see Appendix A.1). If we are interested in measuring energy, E, our P (s) = cΩr(s), expectation hEi, will be: (2.2) X hEi := P (s)E(s) where c is the constant of proportionality. By the fundamental assumption of statistical mechanics, s∈S the above holds for any choice in microstate, s0: (2.1) 0 0 P (s ) = cΩr(s ). Making macroscopic predictions boils down to evaluating probabilities of microstates and the sums (2.3) thereof. In the following section, we will derive the Although we do not have enough information to probability distribution P (s) and encounter the first evaluate absolute probabilities, we can now com- of the fundamental challenges in statistical physics, pare the relative likelihood of different microstates. following the treatment of Domb [11] and Cardy [12]. P (s) Ωr(s) 0 = 0 . P (s ) Ωr(s ) (2.4) Probabilities and Partition Functions To transform the above into a more manageable Let us consider an example: we begin with some form, we introduce the Boltzmann Entropy: physical system, S. It could be metal, gas, or another material. As we previewed, the microscopic S(S) := k log Ω(S), details are irrelevant to the macroscopic picture. To measure macroscopic properties of S, we need to (2.5)

2For example, the concept of spin, which we will encounter in the next section. 3Let us consider how s and r might exchange energy physically. If s is a metal and r a gas, then the two will transfer energy whenever gas particles collide against the metal, exchanging energy stored in the metal’s vibrations with the kinetic energy of moving gas particles.

36 Capstone Issue Vol. 12, 2019 Jesse Hoogland

where k is Boltzmann’s constant, a scaling factor from the microscopic to macroscopic. Another 1 X crucial definition is that of macrostate: a collection P (s) := e−βE(s) Z := e−E(s) of microstates, S ⊂ S, indistinguishable to the ex- Z perimental observer. Whereas this observer can (2.9) measure differences between different macrostates, the differences within a given macrostate are non- where Z is the normalizing factor, the partition differentiable. For the statistical physicist, trans- function. It holds for any system s so long as s ex- lating microscopic predictions to macroscopic pre- changes only energy with its surroundings. Together, dictions amounts to making statements about how these conditions—fixed temperature, number of par- probable different macrostates are as a function of ticles, and volume—form the canonical ensemble. how probable the underlying microstates are. To Within the canonical esemble we have found a re- facilitate this, we introduce Ω(S), the number of mi- lation between the energy of a microstate and the crostates corresponding to a macrostate, Ω(S) = probability that a system occupies that microstate |S|. We interpret S(S) as a measure of uncertainty: at any given time. If we specify an energy func- higher entropy gives a lower probability of correctly tion we can calculate probabilities and, from those guessing the true microstate the system occupies, probabilities, our desired measurement outcomes. which can complicate the translation between mi- In Chapter 3, we will see this equation again, de- crostates and macrostates. scribing the evolution of artificial neurons in neural Returning to the task at hand, we can plug our networks. With the appropriate choice in energy equation for the entropy of the reservoir into equa- function, equation (2.9) helps characterize certain tion (2.4). Subsequently, we get: neural networks, and much of our subsequent analysis will translate readily. P (s) 1 0 k (Sr(s)−Sr(s )) 0 = e . P (s ) Ising Model

(2.6) An example of a system with a suitable energy function for the Boltzmann distribution is the Ising By our assumption that the reservoir is much model depicted in Figure 2.1. In this model, we re- larger than S, we can approximate the above us- quire that the microscopic degrees of freedom si ing the second law of thermodynamics, T ∆S ≈ ∆E are binary-valued (si ∈ {1, −1}), and we call these (constant volume and number of particles), to get: 4 degrees of freedom spins. The Ising model was conceived as a minimal model for ferromagnetism, the P (s) 0 = e−β(Es(s)−Es(s )), P (s0) phenomenon by which metals form permanent magnets. In nature, spin is an intrinsic property of parti- (2.7) cles that induces and interacts with magnetic ﬁelds. Though it is a quintessentially quantum effect, we where ββ := 1/(kT ), the thermodynamic beta. can approximate spin classically as orienting either This requires that S and R are in thermal equilib- “up” or “down” (+1 and −1 respectively). rium, i.e. their temperatures are the same (for a more thorough justiﬁcation of these steps, see Ap- pendix A.2). By the fact that s and s0 are independent, we can separate (2.7) to get that:

P (s) ∝ e−βE(s).

(2.8) Figure 2.1 – The Ising model is a minimal model for ferromagnetism. Image source [13]. P By normalizing (solving s P (s) = 1), we get an equation for the probability distribution, the so- We write the following Hamiltonian (energy func- called Boltzmann distribution: tion) for the Ising model:

4For other statistical ensembles, we might include other terms, like the number of particles.

37 Capstone Issue Vol. 12, 2019 Jesse Hoogland

results will, in general, differ from the true inﬁnite X X E(s) := −B si − J sisj, sums we would like to calculate. However, as we i hi,ji increase the size of the lattice and the number of samples, our results can get arbitrarily close. The (2.10) key trick in MCMC techniques is that relative prob- 0 where, in the ferromagnet, B is the external abilities, P (s )/P (s), are much easier to evaluate magnetic ﬁeld, J is the interaction energy between than absolute probabilities, P (s), since the partition P functions cancel. Monte Carlo techniques proceed neighboring pairs of spins, and hi,ji denotes a sum over adjacent sites. We see that the system is in a according to some variation of: lower energy state when spins si align* with B and P 1. Begin with a randomly-chosen starting state, their neighbors J j→i sj, where j → i denotes the neighbors of i. s.

2. Consider a small variation to the state, s0 (for Intractable sums example, by ﬂipping spin si).

For the vast majority of conceivable Hamiltoni- 3. Decide whether to accept this variation accord- ans, equation (2.9) is intractable. This stems from ing to P (s0)/P (s). Z, the partition function. For an Ising magnet with N N spins, Z will contain 2 terms. Beyond around 4. Repeat steps 2 and 3 until s0 converges the N = 300, this exceeds the number of atoms in the equilibrium distribution, P (s0). universe [14]. In the standard thermodynamic limit that N goes to infinity, this diverges. Even in ev- Seeing as these methods are quite computation- eryday (finite) life, N is on the order of Avogadro’s ally intensive, earlier generations of physicists de- number so is already incredibly large. How are we veloped other sets of techniques. The first we will to proceed? To complicate matters further, equa- consider is mean-field theory (MFT). tion (2.1) requires another sum of 2n terms. It turns out that with regard to this last quandary, Z will be our saving grace. Z, being more than merely a normalizing factor, contains all the relevant information about our specific system. From Z, we can determine any desired macroscopic parameters of interest by taking derivatives. For example, if we define the free energy, F := −β ln Z, then for the Ising model our expectations for the magnetization will be hMi = ∂BF (see Appendix A.3), where Figure 2.2 – Samples of the 2D Ising model near the : P M(s) = i si, the net orientation of all spins. In fact, critical temperature generated with the Swendsen- with clever tricks, we can actually solve (i.e. add up) Wang Algorithm, implemented rgpy. the infinite sums for the Ising model Hamiltonian in one or two dimensions to derive exact solutions for expectation values. However, this is the exception, Mean-Field Theory not the rule. Most models of interest, we cannot solve exactly. To start, let us consider a simpler problem. If we Then, the best we can do is make approximate so- look at a single spin, si, having specified all other lutions. One class of possibilities is Markov Chain spins, can we calculate its partition function? First, Monte Carlo (MCMC) techniques. In using these we can rewrite equation (2.10) more instructively: techniques, rather than evaluating our sums and averages over all microstates, S, we evaluate these   over a representative, finite set of samples, S X X data E(s) = − Ei(si),Ei(si) := −  Jsj + B si = −(nJhsj→ii+B)si, (see Figure 2.2). Furthermore, rather than evalu- i j→i ating our average over a lattice infinite in extent, we evaluate our average over a finite lattice. The (2.11)

38 Capstone Issue Vol. 12, 2019 Jesse Hoogland

where n is the number of neighbors of si and have to resort to numerical techniques. For intuition : 1 P hsj→ii = n j→i sj is the average spin of si’s neigh- though, we can get far with a graphical approach bors. Then, using Boltzmann’s equation, the proba- (see Figure 2.3). bility over si is:

e−β(nJhsj→ii+B)si P (si) = , e−β(nJhsj→ii+B) + eβ(nJhsj→ii+B)

(2.12)

and Figure 2.3 – Mean-Field Theory predictions for the spontaneous magnetization M|B=0. 2 cosh β(nJhsj→ii + B) hsii = P (si = 1)−P (si = −1) = = tanh β(nJhsj→ii+B). 2 sinh β(nJhsj→ii + B) Restricting to the case that B = 0, let us distinguish two cases: (2.13)

So we see that the orientation of a given spin 1. βJn < 1. There is only one solution: m = 0. depends on the average orientation of its neighbors, 2. βJn > 1. Suddenly, there are two additional as one might expect. solutions. Something interesting seems to If we were given hs i, evaluating this equa- j→i happen at the point βJn = 1; we call this tion is trivial. Just as it is easy to calculate rel- a critical point, and it occurs at the critical ative probabilities, it is easy to calculate condi- temperature, Tc = Jn/k. tional probabilities like P (si|{sj→i}). Here, too, the partition functions cancel: P (si|{sj→i}) = Taylor-expanding the right side of equation 5 P (si, {sj→i})/P ({sj→i}). (2.14), we get that in the vicinity of the critical point: For now though, these observations are not of ( −1/2 much help since we do not know hsj→ii. This brings = 0 T > Tc ∼ ±(3|t|) us to the mean-field approximation. Since we could T < Tc, have chosen any spin si as our starting point (including its neighbors), we expect, hsii ≈ hsj→ii, known (2.16) as the principle of mediocrity. In the mean-field : approximation we assume, more stringently, that where t is the reduced temperature, t = (T − Tc)/Tc. In fact, these equations contain more infor- hsi := hsii = hsj→ii, so: mation about our system than the magnetization hsi = tanh(βnJhsi + B). alone. We can use mean-field theory to derive other parameters like the magnetic susceptibility and spe- (2.14) cific heat (see Table 2.2, some are particular to Ising- P like models, others are more general). These are If we find a solution using hM(s)i = h i sii = P the quantities that we can measure in a laboratory, ihsii, we would have our measurement outcome: so by calculating these figures we can confirm (or X deny) that the Ising model is a suitable model of hM(s)i = hsii = Nhsi, i ferromagnetism. In this light, we see why this translation step is so crucial: it provides the means of (2.15) testing our theories. Until now, our discussion as- and we see that hsi is nothing more than magne- sumed an arbitrary number of neighbors, n, so we tization per site m = hMi/N, where N, as before, is would expect our equations for these parameters to the number of spins. hold for any number of dimensions. It would seem It turns out that we cannot solve equation (2.14) like we have accomplished our goal for this chapter analytically (it is a transcendental equation), so we a full section in advance.

5This avoiding of joint probabilities with clever choices in conditional probabilities will be the basis for efﬁciently training RBMs in the next section.

39 Capstone Issue Vol. 12, 2019 Jesse Hoogland

Unfortunately however, in less than four dimensions, mean-ﬁeld theory provides incorrect predictions: equation (2.16) is wrong. Intuitively, systems with less than four dimensions have too little order for MFT to hold. Introducing more dimensions means there are more paths between any two spins and more correlation between them. Past the critical dimension of four, there is enough order for our MFT approximation to hold. However, for the Ising model in 2 and 3 dimensions (those cases most interesting to our daily lives) we have to resort to a different approximation approach. Figure 2.4 – The qualitative behavior of the correlation length, ξ, and magnetization, M, around the critical point.

Correlation Length

An important quantity in Table 2.2 is the correlation length, ξ. This is the average distance across Critical Phenomena and the Renormaliza- which spins in the system under investigation tend tion Group to fluctuate together. Spins farther apart than ξ are effectively independent of one another, so sev- ering such a connection has no appreciable effect Although the quantitative predictions of the on the macroscopic properties: thus we can think mean-field theory are incorrect, its qualitative pre- of ξ as a measure of how macroscopic our system dictions are instructive, particularly those that pre- is [12]. Mean-field theory predicts that near the dict the existence of a critical point. From equa- critical point the correlation length scales as: tion (2.16), we expect a phase transition between a paramagnetic, disordered phase and a ferromag- ξ ∼ |t|−1/2 netic, ordered phase (in which the system sponta- (2.17) neously magnetizes). This bears out experimentally, although not at the predicted temperature, see Fig- Although the exponent does not line up with ex- ure 2.4. perimental results as the true value is 1, mean-field

40 Capstone Issue Vol. 12, 2019 Jesse Hoogland theory correctly predicts that ξ diverges at the criti- where hhi, jii denotes next-nearest neighbors, cal point (see Figure 2.4). In fact, this is the defining and the continued sum will, in general, contain all characteristic of critical points. When the correlation possible interactions. This equation describes all length diverges the entire system becomes corre- possible physical models and lattice. To recover lated. Any perturbation, no matter how infinitesimal, (1) the Ising Hamiltonian, we would choose Ki = βB, will have macroscopic ramifications. For the statis- (2) ( Kij = βJ, and all K ν)=0forν>2. tical physicist, critical points are excellent places Alone, equation (2.18) is not enough of a require- to test theories as they allow closer access to the ment. A “good” RG transformation satisfies a spe- microscopic realm. cial set of criteria: it should preserve long-distance information while discarding short-distance information. Arbitrary transformations satisfying equation Critical Exponents (2.18) need not extract the information we deem relevant. Formally, we are interested in extracting the Another valuable prediction of mean-field theory relevant operators, those describing macroscopic is that of critical exponents. We see from equation properties, and suppressing the irrelevant opera- (2.17) and equation (2.16) that near the critical point tors, those describing microscopic properties. For the correlation length and the magnetization obey example, we might accomplish this transformation simple power-laws. These are examples of a more by summing over even spins (known as decimation): general trend: near critical points, macroparameters will follow power-law scaling formulas. We call 0 0 X e−H (s ) = e−H(s), the exponents that define these relations critical s2,s4,...,sN exponents (see Table 2.3). We shift our goal to the (correct) calculation of (2.20) these critical exponents. To this end, we turn to the 0 renormalization group, a set of ideas for tackling where now s ranges over the odd spins. In other precisely these critical phenomena. words, we integrate out or marginalize over the short-distance degrees of freedom.

The Renormalization Groups

Instead of trying to compute Z head-on, let us consider a different angle. We will try to re-express Z with a simpler set of parameters while preserving the physical, long-distance information. Repeating Figure 2.5 – Three steps of majority-rule block-spin these transformations, we will discard the irrelevant, renormalization, preceding left to right (block size microscopic ﬂuctuations, keeping only the macro- b = 2) scopic information. Formally, an RG transformation will look something like: Consider, ﬁrst, a descriptive example: (majority- rule) block-spin renormalization, a set of RG tech- X −H0(s0) X −H(s) e = e , niques intended for lattice systems. For a given s0 s microstate, block renormalization proceeds as fol- (2.18) lows, (see Figure 2.5):

constraining, for example, |s0| < |s|. We consider 1. We partition the configuration into non- H0 and H parameterized with sets of couplings {K0} overlapping blocks. For each block, we de- and {K}, for example, termine which spin is in the majority and we assign that value to a new, single spin. These define a new coarse-grained system. X (1) X (2) X (3) H(s) = − Ki si− Kij sisj− Kij sisj ..., i hi,ji hhi,jii 2. We rescale the coarse-grained configuration so that each block takes the size of an original (2.19) spin.

41 Capstone Issue Vol. 12, 2019 Jesse Hoogland

Formally, we can write the block transformation rule as:

−H0(s0) X Y 0 −H(s) e := π(s ; si)e , s blocks

(2.21) Figure 2.6 – Renormalization induces changes in the effective temperature towards fixed points. Images from Wilson [17]. where π is the projection operator implementing the majority rule Let us be more exact and determine how this behavior might arise. We will write the RG transformation rule as a function R of couplings: {K0} = R({K}). With two general assumptions we can ( 0 P 0 1, ifs = sgn i si build a descriptive theory of RG. First, we assume π(s , s1, . . . , s9) := 0 otherwise. that there exists a fixed temperature point (or multiple), as would be the case with any other couplings. These are the specifications of {K∗} that (2.22) are stable under our transformation rule, namely, {K∗} = R({K∗}). Second, we assume we can differentiate this transformation near the fixed point, so we can linearize6: Though we can easily perform this procedure 0 ∗ X ∗ for any individual configuration, equation (2.21) re- Ka ∼ Ka + J ab(Kb − Kb ), quires us to do this for all microstates, and this b remains intractable. Ultimately, for most systems, (2.23) RG will use variational schemes [12]. It is, how- 0 ∂Ka ever, the qualitative insights RG offers, in spite of where J = |K=K∗ . This is the Jacobian, a ∂Kb these approximations, which merit our immediate generalization of the derivative to vector-valued attention. functions of multiple variables. We thus denote J ’s left eigenvectors ei, corresponding to eigenvalues, Consider what happens when we apply block λi so that renormalization to Ising configurations at different temperatures as in Figure 2.6. We see three trajectories of block renormalization for three different initial X i i i 7 eaJ ab = λ eb. temperatures: below the critical point, at the critical a point, and above it. Our first observation is that (2.24) these transformations induce flows away from the critical temperature. There are three fixed points Next, we define scaling variables, ui := P i ∗ (T = 0, T = Tc, and T = ∞) for the Ising model, a ea(Ka − Ka), which are combinations of devi- ∗ and, indeed, macroscopically these correspond to ations Ka − Ka that transform multiplicatively near the three phases (ordered, critical, and disordered). the fixed point:

6Taylor-expansions are one of the ﬁrst steps in any physicists’ toolkit. Here, we perform the equivalent of a Taylor expansion for a transformation of multiple variables. 7There is no reason to assume J to be symmetric or even to have real eigenvalues though we will restrict ourselves to considering real-eigenvalued Jacobians.

42 Capstone Issue Vol. 12, 2019 Jesse Hoogland

us to relate the exponents to one another in scaling relations. These relations had been postulated be- 0 X i 0 ∗ X i ∗ X i i ∗ i ui = ea(Ka − Ka) = eaJ ab(Kb − Kb ) = λ eb(Kforeb − theKb advent) = λ ui of. RG but many only as inequalities. a a,b b RG provided a rigorous means to link these different exponents. (2.25 & 2.26, respectively) For example, we show a derivation in (see Ap- y For later convenience, we introduce λi ≡ b i , pendix B.3) for the correlation length critical expo- where b is the rescaling size (the width of the blocks nent: in block-spin RG). These are the renormalization 1 group eigenvalues and are distinguished in three ν = cases: yt (2.27) • yi > 0: ui is relevant. Repeated RG iterations where y is the thermal RG eigenvalue which is drive ui away from its ﬁxed point value. t related, as its name suggests, to the temperature • yi < 0: ui is irrelevant. Repeated RG iterations of the system. From the exact solution (of the Ising drive ui towards 0. model in 2D), we get that yt = 1. Then, we derive for the correlation length critical exponent: • yi = 0: ui is marginal. The linearized equations are not enough to tell us about ui’s behavior. ν = 1

From this, we see that our ability to distinguish (2.28) between microscopic and macroscopic is a conse- In general, since most of the sums we encounter quence of simple dimensional analysis: there are are not exactly solvable, we do not have access to finitely many relevant eigenvalues. Those are the solutions like these. Therefore, we consider three ones you see when you zoom out far enough. The ir- approximate schemes. relevant eigenvalues span a critical surface of points which, under RG transformations, are attracted to- 4 − Expansion. wards the critical point. Macroscopically, points on this hypersurface are indistinguishable, and their We can rephrase the thermodynamics limit (in behavior is fully characterized by the critical point which we let the number of spins go to infinity) as alone. This brings us to the remarkable principle the limit in which we hold the total size of the sys- known as universality. tem fixed while letting the distance between spins go to zero. In this way, our discrete lattice becomes Universality a continuous field, and we can reformulate the Ising model as a quantum field theory (QFT). Whereas In the last century, experimentalists faced a puz- the above are examples of renormalization in real- zling situation. In their efforts to measure more space, in QFT, we typically perform renormalization precise critical exponents for all manner of systems, in momentum-space. Here, we can use Wilson’s they discovered that their experimental set-ups did 4 − expansion, treating the dimensionality of our not matter: all ferromagnets for a given number of system as a perturbation [18]. Then, with perturba- dimensions possessed the same critical exponents tion theory and diagrammatics, we approximate the and so, too, for all superfluids [11]. The exponents values of our scaling variables. are universal. The Ising model, then, is not only a minimal model for ferromagnetism but can describe Monte Carlo Simulation fluids, neural networks, metal alloys, and more. Another key result of the RG formalism is that RG allows us to take advantage of the finite-size we can express all of our critical exponents in terms effects that dominate Monte Carlo techniques, and of the relevant RG eigenvalues. First, we use RG we can predict how our results will deviate from the to derive a scaling rule for the free energy (see infinite-size limits (see Appendix B.1). We can com- Appendix B.2). Then, from its derivatives, we can bine the two approaches, performing the RG sums determine the critical exponents which even allows (like equation (2.21)) over Monte Carlo samples.

43 Capstone Issue Vol. 12, 2019 Jesse Hoogland

Kardanoff’s Variational Technique revealing universality, simple scaling laws, critical exponents, and the relations between them. In this Another approximation scheme is that of chapter, we turn to machine learning. Like RG, ML Kadanoff. He writes the renormalization transfor- is a blanket term that refers to a wide range of mation as: techniques. Both involve the iterative manipula- 0 0 X 0 tion of information with the goal of extracting “rele- e−H (s ) = eT λ(s ,s)−H(s), vant” information. However, we will see that ML is s more flexible in its definition of relevant. Whereas (2.29) in statistical physics relevant is synonymous with long-distance, in ML, relevance will depend on the with T serves as a function that couples the orig- λ problem at hand. We will also see that the level at inal and coarse-grained systems. Kadanoff derives which RG and ML manipulate information is different. upper- and lower-bound estimates of the free energy Where statistical physics works with partition func- that depend on T . By choosing λ to tighten the λ tions, ML works with probabilities. In practice, many bounds, Kadanoff variationally minimizes the differ- of the challenges (namely, intractable sums) and ence in free energy between the initial and coarse- the solutions are the same. In spirit, however, this grained systems, ∆F = F (H0 )−F (H) (knowing the distinction reveals something fundamental about margin of error). However, this technique does not the types of challenges that characterize statistical guarantee reasonable estimates of macroparame- physics and ML. ters. As Kadanoff himself observed: In physics as a whole, our investigations are largely reductionist: we begin with a Hamiltonian Hopefully, one might obtain good re- which, in turn, defines a partition function, and our sults for physical quantities by choosing aim is to predict something about large sets of mi- the upper (lower) bound recursions that crostates. To this end, computational techniques, give the smallest error in the free en- like MCMC algorithms, may use P (s) (more pre- ergy.. . We say “hopefully” because usu- cisely, the relative probability) to generate a finite ally one is not interested in the free en- set of samples S that, we hope, is representa- ergy itself. Rather its derivatives are of data tive of S, the true, complete set of microstates. By the major physical interest. Since the this, we mean that the statistical properties of S variational principles pertain to the free data and S should converge as we increase the number energy, there is no guarantee that the of samples. In ML, the investigation often works in derivatives will be accurate [19]. the opposite direction: we are given some Sdata and This reflects one of the major challenges with RG. assume it is representative of S. Then, our goal is It is by no means “easy” to construct adequate RG to learn something about P (s). While physics con- schemes, and findings are rarely backed with rigor- cerns properties of S, ML concerns properties of s. ous justification. The details of appropriate transfor- These are not absolute distinctions, but it will be mations depend on the systems under investigation valuable to bear them in mind. and require an amount of intuition on the part of the Let us introduce some of the essentials of ML and physicists [5]. We will see that, ultimately, being show its value for statistical physics through a practi- more precise in defining physically-relevant informa- cal example. We follow the treatment of Hinton [20] tion will allow us to circumvent this problem. We will as well as Mehta et al.’s A high-bias, low-variance be able to formulate a system-independent criterion introduction to Machine Learning for physicists [21]. of quality for RG transformations. We refer the interested reader to these sources for elaboration where our analysis goes quickly.

III. Machine Learning in Physical Phase Classification Investigations Our first question, as physicists, is the following: In the previous chapter, we began with the goal can we train a neural network to classify the likely of translating microscopics to macroscopics. Where phase of some Ising configuration, x? This immedi- MFT failed, RG made sense of critical phenomena, ately reflects the different spirits of ML and physics

44 Capstone Issue Vol. 12, 2019 Jesse Hoogland

we discussed above: our goal is to learn something about the microstate s, whereas in the previous chapter we cared only for S. Formally, our aim is to learn the conditional probability distribution P (y|x), where y encodes the phase.

Figure 3.1 – Restricted Boltzmann machines are a binary-valued two-layer neural network. RBMs trace their origins to the Hopfield model, an early model for associative memory which itself was inspired by We start with a dataset. In this case, we as- the Ising model [22]. sume that we have access to a large collection of Ising configurations at different temperatures and phases, Xdata := {x}, as well as their phase labels A full review of the techniques available in ML Ydata := {y}. We often distinguish two kinds of ML: is beyond the scope of this capstone. As such, we 8 supervised and unsupervised learning. Our access will focus our attention on one particular class of to labels, Y, means this example of classification algorithms, those of restricted Boltzmann machines 9 falls under supervised learning. In contrast, unsu- (RBMs): bidirectional artificial neural networks for pervised learning works without labels, typically at modeling probability distributions. Depicted in Fig- the level of P (x), detecting patterns in the raw data ure 3.1, RBMs consist of two layers of binary units: itself. : N a visible layer, v = {vi}i=0, and a hidden layer, N h := {hj}j=0, where vi, hj ∈ {0, 1}. The visible layer will serve as the input for our data, x → v, and the hidden layer as our prediction, h → y, the label. For the purposes of classifying the phases of the 2D Ising model, it will suffice to consider an RBM with Before doing anything else, we randomly divide one hidden unit, h. If h = 1, we predict that the configuration is in the ordered phase, and for h = 0, Xdata’s elements into a training set, Xtrain, from we predict it is disordered. From our exploration in which our network learns, and a testing set, Xtest, on which we test the trained model’s results. This divi- the previous chapter, we could probably come up sion is crucial because, in ML, we want our results to with some function that performs this prediction. For generalize to new samples that may not have been example, we might sum all of the spins (compute the magnetization) and if the result is close to 0 we in our dataset Xdata but that could have come from know the system is disordered: h := 0. Otherwise, Xtrue. A central problem in ML is the bias-variance tradeoff. An algorithm may overfit the training set we would output h := 0. The power of ML will be to which compromises its ability to generalize to new extract rules like these without explicit instructions. data, or it might fail to learn enough detail, per- Instead, we teach our model implicitly through forming poorly on both training and testing data the cost function, C(Pθ(y|x), {Xdata, Ydata}). This (low-bias/high-variance and high-bias/low-variance, acts as a moderator, and it tells the model how respectively). By splitting the dataset into a test- “incorrect” its predictions (labels) are. Formally, the ing and training set, we get a good impression of ML goal becomes to find the parameters θ that mini- the bias and variance for the models we trained. mize C. Experience from theoretical physics tells us Often, we will further split the training set into cross- that finding the ground state (global minima) can validation sets, trying out different kinds of models be really, really hard, so instead, we use stochastic and ultimately keeping those that best accomplish gradient descent (SGD). For each θ, we calculate a balance of low bias and variance. ∂θC, with which we implement the update rule: 8A third axis of differentiation, reinforcement learning is outside the scope of this capstone. 9We could, however, also formulate classification of phases as an unsupervised learning problem. We might try “clustering,” where we specify a number of clusters, k, and try to group training examples into as many clusters, according to some notion of similarity. Then, we hope that the clusters our model learns correspond to the phases. We could also use this procedure with different choices of k to determine a likely number of phases.

45 Capstone Issue Vol. 12, 2019 Jesse Hoogland

(3.3) θ → θ − η∂θC, We note that this is the exact same equation (3.1) as our Ising model (2.9 and 2.10). Here, {ai} and {b } take the role of the external magnetic ﬁeld B, where η is the learning rate. From calculus, we j which now varies from site to site (hence the indices know that the negative gradient of a function is i andj). Then {w } takes the role of J varying from the direction in which that function decreases most ij pair to pair. rapidly. This update procedure, then, adjusts our Naturally, we run into the same intractability parameters so as to reduce the cost. Iterating this issues. For large enough networks we cannot eval- procedure, we end up in a local minimum of the cost uate Z. Instead of calculating joint distributions, function (see Figure 3.2). In practice, we calculate we, similar to MFT and MCMC, consider instead con- these gradients not over the entire data set, but ditional and marginal distributions. Due to RBM’s over subsets – minibatches. This means the gra- bipartite structure these factors are easy to evalu- dients will vary from iteration to iteration (hence, ate. Explicitly, in the case of a single hidden spin h , stochastic). This is both computationally faster and j we get (with a bit of algebra): introduces noise (like temperature) that improves the chances of escaping poor local minima.

P (h , v) (e−E(v,hj ))/Z P (h |v) = j = j P −E(v,h0) P (v) ( h0 e )/Z −h (P w v +b ) e j i ij i j = . (P w v +b ) 1 + e i ij i j

(3.4 & 3.5, respectively)

This conditional probability is itself a Boltzmann distribution, but over only two states (tractable!). Figure 3.2 – Stochastic Gradient Descent (SGD). If We can use this to express to the full system as: we view the cost function as defining an “energy” M landscape, SGD allows us to find local minima (sta- Y 1 P (h|v) = θ −h (P w v +b ) ble or metastable) energy states. Image taken from 1 + e j i ij i j j=1 [23]. (3.6) Now, if we can formulate an adequate prediction function, P (h|v), for the RBM, we can improve it and similarly: with SGD. The crucial step is to define an energy N function: Y 1 Pθ(v|h) = P −vi( j wij hj +ai) i=1 1 + e X X X E (v, h) := − a v − b h − w v h θ i i j j ij i j (3.7) i j ij All that remains is for us to choose an adequate (3.2) cost-function, and we can start training our RBMs. For the example of binary-classification, an appro- where θ := {{ai}, {bj}, {wij}}. Then, we can model the system’s joint probability with a Boltz- priate choice is the cross-entropy loss: mann distribution: X X C(Pθ(y|x), {Xdata, Ydata}) = Ptrue(y|x) log Pθ(y|x), 1 X 0 0 x∈Xdata y∈{0,1} P (v, h) := e−Eθ(v,h),Z := eEθ(v ,h ) θ Z v0,h0 (3.8)

46 Capstone Issue Vol. 12, 2019 Jesse Hoogland

see Appendix C.1 for elaboration. Differentiating (see Appendix D.1): with respect to θ and with some simple algebra, we ( !)! get the training rule: X X X Eθ(v) = − aivi − log 1 + exp − bj + viwij , X i j i ∂bj C = Ptrue(y|x) 1 − Pθ(y|x) (hj) , x∈Xbatch X (3.12) ∂wij C = Ptrue(y|x) 1 − Pθ(y|x) (vihj) . x∈Xbatch and we can perform an analogous computation for the marginal distribution over h to get P (h) as a (3.9 & 3.10, respectively) θ Boltzmann distribution in terms of energy, Eθ(h). As in the classiﬁcation example, we require a cost func- We have all the necessary elements of a phase tion – in this case, the Kullback-Leibler divergence classiﬁer: a dataset, a model of P (y|x), and a (KLD): means to train this model. In the next section, we will consider an example more relevant to the statistical physicist: generating samples. X Ptrue(x) C(Pθ(x), x) = DKL(Pdata(x)||Pθ(x)) := Ptrue(x) ln . Pθ(x) x∈Xdata

Generative Modelings: Gibbs Sampling (3.13)

In Section 2.3, we considered how to use MCMC This is closely related to the cross-entropy and, sampling to estimate macroparameters and even similarly, it is an important information-theoretic critical exponents. Here we will consider an MCMC quantity (see Appendix C.2) that provides a notion technique called Gibbs Sampling which will use of similarity for probability distributions. It is 0 if and RBMs to generate samples of P (x). only if the two distributions are equal: Suppose we are given an already trained RBM. Gibbs sampling, like other MCMC techniques, consists of a series of update steps. In one step, we DKL(Ptrue(x)||Pθ(x)) = 0 ⇐⇒ Ptrue(x) = Pθ(x). input some initial state, transform it into a hidden (3.14) state using Pθ(h|v), and transform it back to a new visible state using Pθ(v|h). This process is imperfect, We claimed that this is unsupervised, but in com- so the output of one step will differ from the input, paring C in equation (3.13) with equation (3.8), we and this difference will become increasingly smaller might interpret x as a kind of label for itself. A more as the network learns. If we repeat this process appropriate description is self-supervised. We em- many times, then the distribution of the outputs will phasize this point because it has been a source of converge to the equilibrium distribution Pθ(v).. misunderstanding. ML always requires the speciﬁ- This ﬁrst requires Pθ(v) to be an appropriate cation of a cost function, and even in the absence model of Ptrue(v). To get to this point, we need to of labels, the cost function guides how the model derive a marginal distribution over v from Pθ(v, h). learns which information is relevant and irrelevant. Similar to the conditional probabilities, the architec- If we differentiate the KLD with respect to θ, then ture of the RBM allows us to perform the marginal- after a bit of algebra, we get: P ization Pθ(v) = h Pθ(v, h) explicitly. If we write

Pθ(v) as a Boltzmann distribution with its own en- ∂ai DKL = hviitrue − hviiθ ergy Eθ(v), ∂bj DKL = hhjitrue − hhjiθ ∂ D = hv h i − hv h i X −Eθ(v) wij KL i j true i j θ Pθ(v) = Pθ(v, h) ∝ e , h (3.15, 3.16 & 3.17, respectively)

(3.11) where h...itrue is an average over the true distribution Ptrue(v) and h...iθ is an average over the then we can express Eθ(v) in terms of Eθ(v, h) model distribution Pθ(v).

47 Capstone Issue Vol. 12, 2019 Jesse Hoogland

Naturally, we approximate. Wherever hj shows up, we use P (hj|v), and for the expectations over Ptrue, we calculate a Monte Carlo average over our dataset, Xdata. The expectations over Pθ are trickier. Fortunately, we can approximate them with Gibbs sampling, initializing with samples from Xdata. Together, these approximations constitute the contrastive-divergence algorithm, see [20]. Hav- Figure 4.1 – Two iterations of block renormalization ing trained our RBM, we can generate new samples and a deep Boltzmann machine of three layers. and start calculating critical exponents.

RBMs trained in this way have applications other However, questions like this are too general. We than generation. Consider training these on black have already seen two examples of RBMs used for and white images: if we constrain the number of different goals, and what constituted relevant infor- hidden nodes to be less than the number of visi- mation differed in either context. In RG, relevant ble nodes, then the hidden layer will learn a com- information is understood more narrowly: it is the pact representation of the input. We can even stack long-distance information. If we are to make com- multiple RBMs on top of each other to form deep parisons between RG and RBMs, we must be more Boltzmann machines (DBMs). In DBMs, the hidden specific, and this begins by being more precise in layer of one RBM becomes the visible layer of the our definition of ‘relevant’ information. Although next. Trained with contrastive divergence, each Claude Shannon, the founder of information theory, layer learns a progressively more compact version avoided this topic explicitly, Tishby et al. showed of the input. This should remind you of RG. In the that information theory provides a natural and exact next section, we will make these similarities more formulation of relevant information: relevant infor- exact. Ultimately, this allows us to machine learn mation is simply the information contained in one RG transformations. signal, x, about another y [3]. For our phase classifier, the relevant information is the information contained in the Ising samples x about the labels y. Once we have trained RBMs for compression, the relevant information is the information contained in IV. Machine Learning and the the hidden layer h about the visible layer v. Infor- mation theory quantifies how much information is Renormalization Group shared between two signals with the mutual information:

In the previous chapter, we considered two examples of ML in physical investigations: neural net- X P (x, y) I(x; y) := P (x, y) log , works used as a phase classifier and MCMC sampler. P (x)P (y) x,y Consider the superficial similarities between these neural networks and our treatment of RG in Chap- (4.1) ter 2. The classification example calls to mind the infinite RG limit in which all disordered phases flow We see this quantity is minimized (I(x; y) = to one fixed point and all ordered phases to another. 0) when the random variables are independent Here, these fixed points would correspond to the (P (x, y) = P (x)P (y)). It is maximized when the values of our label being either 0 or 1. The genera- variables are maximally dependent. We have a tive example is more immediately similar, and the quantitative basis for extracting relevant informa- very language is analogous: we speak of hidden lay- tion: maximizing mutual information between ap- ers that hierarchically decompose the input visible propriately chosen signals. layer to coarse-grained hidden layers. We might be In the next chapter, we will derive the correct tempted to ask: does a DBM of, say, five layers learn choice for x and y on lattice systems that recovers to implement four iterations of some RG procedure physically-relevant (i.e. long-distance) information. (see Figure 4.1)? We then derive a variational proxy to the mutual in-

48 Capstone Issue Vol. 12, 2019 Jesse Hoogland

formation that we can differentiate and, thus, learn inquiry, Mehta and Schwab train RBMs as described in a neural network. in Chapter 3.2. They ﬁnd that the RBMs learn to With the knowledge of an exact formulation of couple hidden units to local blocks of visible units. relevant information, we can start investigating links Crucially, this was not possible without regulariza- between particular RG implementations and RBMs. tion, a technique that promotes sparse connections, Such comparisons will revolve around the question between hidden and visible units see [21]. of relevance. What information does a given cost- function deem relevant? Are there examples where this overlaps with our notions of physical relevance? Let us consider a seminal example: the equivalence posed by Mehta and Schwab.

A Comment on Mehta and Schwab’s Equivalence

As a first attempt, we will consider Mehta and Schwab’s correspondence between Kadanoff’s variational procedure (Chapter 2.3) with RBMs trained by contrastive divergence (Chapter 3.2) [4]. Specifi- cally, we consider the narrower case, of “exact” RG Figure 4.2 – Visualization of the receptive filters of and “exact” RBMs. Exact RG means that Kadanoff’s the hidden units in one of the layers of the DBM transformation preserves all the information con- trained by Mehta and Schwab [4]. Each plot cor- tained in the hidden system; i.e. the free energy responds to one hidden unit, and each pixel cor- of the input and coarse-grained systems is exactly responds to one spin in the initial layer. Higher the same. Similarly, an exact RBM has learnt to intensities mean the hidden unit is more coupled perfectly recreate its inputs Ptrue(x) = Pθ(x) in to that spin. Indeed, we see that the RBM learns to equation 3.13; the RBM will have reached a global couple local transformations. However, the transfor- minimum of the cost function. mations do not preserve nearest neighbor relations. First, Mehta and Schwab show that, under the This does not matter for the total amount of infor- above conditions, the two transformations are equiv- mation stored in coarse-grained layers. In RG, these alent under: concerns do matter because they affect the practicality and interpretability of the result [5]. T θ(v, h) = −Eθ(v, h) + H(v),

(4.2) Mehta and Schwab write:

and we provide the derivation in Appendix D.2. Surprisingly, this local block spin struc- However, the exact case is not generally pos- ture emerges from the training process, sible, and when it is, it is often not particularly in- suggesting the [deep neural network] is teresting. Exact RBM transformations likely mean self-organizing to implement block spin overﬁtting which is often opposed to the aims of renormalization [4]. the ML investigation. Exact RG transformations are few and far between, so this correspondence would These RBMs may indeed learn a kind of block-spin apply only under narrow circumstances. More in- transformation. However, this block-spin transfor- teresting then, would be to consider a non-exact mation need not be block-spin renormalization. As comparison. Koch-Janusz and Ringel put it: Here, the two approaches will vary: Kadanoff’s variational method operates at the level of free en- [T]he usefulness (and practicality) of ergies while contrastive divergence works at the the RG procedure depends on choosing level of probabilities. Optimizations at these differ- [the transformation] . . . such that the ef- ent levels will not necessarily coincide. Still, there fective Hamiltonian. . . remains as short may be qualitative similarities. Pursuing this line of range as possible [5].

49 Capstone Issue Vol. 12, 2019 Jesse Hoogland

More precisely, in the Taylor expansion of our coarse- Compression and generation are invariant under 10 grain Hamiltonians, the shortest range terms should partial flips hj : (0, 1) → (1, 0) for some fixed j. dominate: Blocks of spins which were perfectly correlated in the visible layer may be perfectly anticorrelated in the hidden layer. Although the information content X (1) X (2) X (3) H(s) = − Ki si− Kij sisj− Kij sisj ..., is the same, the hidden layer Hamiltonian will take i hi,ji hhi,jii a more complicated form than is necessary, going against notions of what is a good and practical RG (4.3) procedure. (n) where K are exponentially suppressed in n. Mehta and Schwab fail to mention these condi- In RG, we place additional constraints on how to or- tions, so their suggestion that RBMs learn block-spin ganize the information in subsequent layers. Block renormalization falls short. However, if we make renormalization procedures, including Kadanoff’s these conditions explicit, we can easily recover a technique, respect the symmetries and topology of stronger version of their claim. To do this, we in- the system under consideration. The locality of introduce convolutional neural networks (CNNs) as teractions means neighboring blocks of visible spins depicted in Figure 4.3. These are networks with a become neighboring individual hidden spins. Trans- special architecture: instead of all-to-all couplings, lational symmetry means the same transformation we explicitly couple spins in the hidden layer to is applied to each block. Mehta and Schwab’s RBMs blocks in the visible layer. Additionally, CNNs per- will not, in general, satisfy these two conditions. form the same translation for each block and re- Let us be more exact by rephrasing renormal- spect the topology of the input system. Had they ization in probabilistic terms: we parametrize our required a convolutional architecture upfront, Mehta RG transformation as the conditional probability and Schwab may, indeed, have successfully trained distribution distribution P (h|x), where x is our ini- RBMs to learn even block renormalization. tial configuration and h is the hidden or coarse- grained configuration. Locality and translational invariance of the Ising model mean we can factor the RG transformation into a product of local single- block transformations. We divide the system into M blocks, x → {v(1), v(2), . . . , v(M)}, with corresponding hidden units {h1, h2, . . . , hM }. Then we can factor P (h|x) as: Figure 4.3 – This is LeNet-5, an example of a convolutional neural network used for digit recognition M and a seminal architecture [24]. Y (j) P (h|x) = P (hj|v ). j=1 From these observations, we can make the link (4.4) between majority-rule block-spin RG and ML fully Consider what happens, when you permute the precise. In Appendix D.3, we derive an equiv- units in the hidden layer; i.e. use a different trans- alence between RBMs and RG, describing a pa- (j) 0 formation transformation P (hj0 |v ) where j 6= j . rameterization which implements the majority-rule: (j) P (j) As long as our reverse rule, P (x|h) also takes this P (hj|v ) = sgn i vi . into account, such a permutation has no impact on the performance of Mehta and Schwab’s contrastive Relevant Information divergence trained RBMs. This is not the case in RG. There is a unique permutation of hidden layer de- For the exact case discussed by Mehta and grees of freedom which maximizes the above short Schwab, we preserve the full probability measure range condition, and acceptable RG transformations which necessarily preserves the long-distance in- identify this permutation. formation. However, in the non-exact case, the Furthermore, the block transformations that the KLD has no preference for long-distance information RBM learns may not be the same for each block. over short-distance: all information is valued equally.

10RG transformations are invariant under a collective ﬂip of h.

50 Capstone Issue Vol. 12, 2019 Jesse Hoogland

This is acceptable in the case of compression and model’s locality and translational invariance condi- generation, as we have no a priori knowledge of tions reduce the number of acceptable RG proce- which features in the data are more important. dures. Rather than devise an update rule for how In RG, we have stronger requirements: transfor- an entire configuration should transform under RG, mations must favor long-distance information over we can start with equation (4.4). Then, we only short-distance information. We can say resolutely: need to learn the transformation for a single block, (j) RBMs trained with the KLD (even with the appro- P (hj|v ). We already know that we can model con- priate convolutional architecture) need not learn ditional probabilities like these with RBMs. This is acceptable RG transformations. Perhaps in the case the first insight of Koch-Janusz and Ringel, and in this of the simple Ising model, the KLD is an adequate chapter we will follow their information-theoretic heuristic. However, we have no guarantee that treatment [5]. this extends to novel systems. In fact, Koch-Janusz and Ringel showed that RBMs trained with the KLD An Information-Theoretic Formulation of on the dimer model, another model from statisti- Renormalization Group cal physics, couple to local fluctuations rather than to the correct hidden variables [5]. We recall that We first partition a full lattice configuration x appropriate transformations for one system need into four areas: a visible block, v, that is surrounded not translate to others (see Chapter 2.3). Critique by, in order, a buffer, b, an environment, e, and of Mehta and Schwab’s paper revolves primarily an outer area, o (see Figure 5.1). In RG, we intro- around the KLD being an inappropriate criterion for duce a hidden area, h := {hi} which is a function of RG transformations [2, 6]. the degrees of freedom in v. Our aim is to choose In reply, Mehta and Schwab point out that this coupling, P (h|v), so that h encodes the long- Kadanoff’s method also offers no guarantee of ex- distance degrees of freedom about our system at tracting the correct physical information. We come large, x11. By our assumption that o is farther than back to the difficulty of devising appropriate RG the correlation length, v contains no information techniques, and the need for creativity. In fact, this about o. We can ignore o in devising P (h|v). Fur- need for intuition in RG is mirrored quite clearly in thermore, in coming up with a function for h, we ML. Decisions regarding model architectures and can reasonably ignore b. The information contained cost functions require creativity on the part of the in v about b is likely to be short-range. By elimi- researcher. Both fields would benefit from a clearer nating b, we effectively remove the shortest-range understanding of when techniques will and will not fluctuations. work. In the next chapter, we will see that information theory may provide the framework to answer questions like these, and we will encounter a system- independent formulation of RG. This might mean better, more efficient models in both disciplines.

V. Renormalization in Information Theory

In the previous chapter, we used information theory to formalize the notion of relevant information. Figure 5.1 – The RSMI algorithm partitions conﬁgura- In this chapter, we will adapt this to our physical tions into four regions, a visible block v, buffer zone context of long-distance as relevant information. Ul- b, environment e and outer zone, o. We introduce a timately, this allows us to devise a cost-function that coarse-grained block of spins h. measures the physical information. From this, we can derive a learning rule to train RBMs to perform That which remains is the physically relevant optimal transformations. information; we see it is the information shared be- In the previous chapter, we saw that the Ising tween v and e. Our goal is to extract this information

11We have dropped the indices for convenience, and now we let block h consist of multiple coarse-grained spins.

51 Capstone Issue Vol. 12, 2019 Jesse Hoogland and encode in h. We choose the parameters, Λ, of and [9] for these results. In the remainder of this our RBM that models P (h|v) to maximize the mutual capstone, we discuss our own implementation and information between h and e: generalization.

X P (h, e) I(h; e) = P (h, e) log , P (h)P (e) Measuring Critical Exponents h,e

(5.1) Having trained an RBM according to the RSMI algorithm, we can use finite-size scaling (see Ap- where these probabilities are defined as pendix B.1) to estimate, finally, critical exponents. marginalizations over P (x): An important detail will be to devise a “thermometer” that can measure the effective temperature X at successive RG-iterations. We discuss several op- P (h, e) = P (h|v)P (v, e) tions in Appendix E.2. These options, moreover, are v X intrinsic and do not require explicit knowledge of the P (h) = P (h|v)P (v, e) temperature. This means the RSMI algorithm is fully v,e unsupervised. If we encounter new systems where X P (v, e) = P (x) we do not know the proper RG transformations (or b,o even how to measure “temperature”), then unsu- X P (e) = P (x). pervised approaches might save us considerable headache. Rather than guess and check possible b,v,o transformations, we would first train a neural net- (5.2, 5.3, 5.4 & 5.5, respectively) work to learn how to perform a renormalization procedure on the system, then observe and interpret, From previous chapters, we know that the prob- only later attempting more calculation-heavy ap- ability measure P (x) and its marginalizations are proaches. As Koch-Janusz and Ringel put it, the RSMI generally intractable. The key insight of Koch-Janusz algorithm could form the basis of the “physical rea- and Ringel is that we can use RBMs not only to soning process” itself kjr]. Instead of retroactively model P (h|v) but also to model these other distribu- explaining trends in our data, ML would proactively tions. Koch-Janusz and Ringel ultimately use three guide our exploration. RBMs: one P (v, e), another for P (v), and finally the P (h|v) RBM already mentioned. The other distributions are calculated as MC-averages over the dataset. Hereby, Koch-Janusz and Ringel derive a VI. Results: Machine Learning Crit- proxy to the mutual information (see Appendix E.1) which they can optimize with stochastic gradient ical Exponents descent. To validate their ideas, they provide both experi- In [10], we release rgpy, an open-source library mental and theoretical justification. For the 1D Ising for implementing ML-based renormalization group model, these marginalizations can be calculated techniques. This is still in its infancy, and develop- exactly, and they show that maximizing the RSMI ment is ongoing. As of now, it contains a full-stack rederives decimation, an RG procedure known to be realization of the RSMI algorithm implemented in optimal for the 1D Ising model in that the procedure Tensorflow: i.e., the package includes implemen- does not increase the range of the coarse-grained tations of various MCMC techniques (Metropolis- Hamiltonian. In follow-up research, Lenggenhager Hastings, Swendsen-Wang, and Wolff algorithms), et al. (along with the aforementioned authors) show the RBMs necessary for the RSMI algorithm, and an that RSMI coarse-graining is more generally opti- implementation of standard block-spin renormaliza- mal, maintaining short-distance in any number of tion. We also provide a host of already-generated dimensions. They further show this holds even in samples at various lattice sizes and temperatures. some cases that the mutual information is not fully We invite the reader to explore and try out these saturated. We direct the interested reader to [5] tools.

52 Capstone Issue Vol. 12, 2019 Jesse Hoogland

A Novel Calculation of ν for the 2D Ising rent formulation works only for systems with binary Model degrees of freedom: spin-1/2 Ising models. The authors avoid explicitly generalizing these results. Our exploration of statistical physics and ML was Though the generalization is straightforward, it is centered around the Ising model and its macropa- important enough to warrant elaboration. rameters and critical exponents. The calculation of Let us consider systems with either n spins or these exponents served as unifying thread, and as n components of spin, then, the visible units of our a validation of the RSMI algorithm we provide the RBM should be able to take n values. Accomplishing following approximation of : ν this is quite simple: we let our visible units become n vectors with n components: vi → {vid} . We say ν ≈ 0.79 ± 0.39 d=1 that vi takes the value of spin d when: (6.1) ( v = 1 For how we calculated this, see Appendix E.3. v ≡ d ⇐⇒ id i 0 This is not, by any means, an improved calculation. vid0 = 0, d 6= d. However, it is a highly promising ﬁrst result. The (6.2) quality of this calculation was ultimately limited by both time and hardware constraints. With more In fact, we will use the RBM to model a slightly time and resources, the prediction should further different vector: converge to the right quantity. vid = P (d), (6.3) P where the vector is normalized d vid = 1. Such a vector that contains probabilities of classes is called a one-hot encoding, and with these probabilities, we can randomly sample values of spin to generate n-valued spins. For n-vector models, we leave the probabilities as they are. To allow our RBMs to produce this encoding, we have to introduce the index d in the parameters of our model:

ai → aid

wij → wijd.

Figure 6.1 – The ﬁnite-size scaling curve for the cor- (6.4 & 6.5, respectively) relation length critical exponent. To derive a one-hot encoding, we change the RBM’s reverse update rule, P (v|h) (3.7). The condi- In fact, this result is just the start. Our inves- tional probabilities will still factor, but normalization tigation revealed a large number of possible im- over the scalar values vi = 0, 1, becomes normal- provements and generalizations. Due to the time ization over the components of the vector, d. If we constraints of this capstone, we have not yet imple- : P denote i = −( j wijhj + ai), then, whereas before, mented these ideas, but all the same, they merit we had: attention. The generalization we discuss gives rise − v to a family of RSMI-inspired techniques, together e i i P (vi|h) = 0 , P −iv constituting a new branch of RG methods. 0 e i vi∈{0,1} (6.6) A Generalization to n-Spin and O(n) Sys- : P tems now we have, id = −( j wijdhj + aid), so:

Koch-Janusz and Ringel claim the RSMI algorithm e−idvid P (vid|h) = 0 . works for general lattice systems. However, the cur- Pn −idv d=1 e id

53 Capstone Issue Vol. 12, 2019 Jesse Hoogland

(6.7) For all the exact results of Koch-Janusz and Ringel, their practical implementation uses a proxy For n = 2, we re-derive the previous rule with which does not fully capture the above mutual infor- a suitable choice in wid, aid. For reasonable values mation [5]. Here, we see two possibilities. First, we of d (anything we might encounter in practice), we can improve Koch-Janusz and Ringel’s approxima- can perform these calculations explicitly. If we ap- tion. The authors use a cumulant expansion ply this change to each of the RBMs in the RSMI P∞ 1 C algorithm, we will be able to model n-spin and O(n) hexp{K(x)}i = e κ=0 κ! κ , models. Note that since we did nothing to the hidden variables, these will still be either binary or (6.9) Bernoulli-valued. We may require that the hidden with cumulants expressed in terms of the mo- spins take the same form as the input variables, in ments which case we also need to apply a completely anal- (j) ogous change to the hidden unit of the P (hj|v ) C1 = hKi, RBM. This is often the case in RG implementations, 2 2 and it allows us to apply the RSMI algorithm to C2 = hK i − hKi , 3 2 3 generic lattice systems. C3 = hK i − 3hK ihKi + 2hKi , There is, in fact, more room for generalization. Consider an arbitrary Hamiltonian for a system with binary-valued data H(x). If we Taylor expand this (6.10, 6.11 & 6.12, respectively) function, we would get: and so on. In deriving the proxy, only the first X X term is kept (see Appendix E.1). For better results, H(x) = a + bixi + cijxixj. i ij a first improvement could be made by introducing additional terms.13 Second, we can consider other (6.8) approximations of the mutual information, and we draw the reader’s attention to a sampling of options 12 These are all the terms in the expansion. Then, in the literature: contrastive predictive coding [26], barring the bipartite structure, the RBM energy func- Deep InfoMax [27], and Mutual Information Neural tion already has the most general form we could Estimation [28]. All three examples use mutual in- possibly consider. However, when we make the formation as the basis for powerful unsupervised switch to continuous or n-ary valued data, this is techniques. Though we avoid details, similarities no longer true: the Taylor-expansions become in- (and differences) with the RSMI algorithm warrant finite. In these cases, we might consider energy further investigation. functions with different, possibly higher-order inter- Perhaps the most exciting future course of action actions. As long as we maintain the bipartite struc- is extending these ideas to momentum-space renor- ture, we can ensure that conditional probabilities malization. Rather than focus on lattice models, factor. Then, we can use variants of contrastive- our goal would be to solve quantum field theories. divergence to efficiently train these models. We A first attempt might use the quantum-to-classical need not even require that P (v|h) and P (h|v) take mapping of the path integral formalism [5]. More the same form. The first might be Gaussian and powerful however would be to use the same physi- the second a gamma distribution. Formally, we can cal and information-theoretic arguments to devise consider the extension of RBMs into the exponential conditions equivalent to equation (5.1) from first family, and for a full discussion of the possibilities, fundamentals. we refer the reader to [25]. This extension offers more powerful options for modeling probability distributions. Though the choices implemented in this VII. Discussion and Conclusions capstone suffice for the spin-1/2 Ising model, it may be necessary to extend the work to the exponential Let us summarize what we have accomplished. family to investigating more complicated systems. We began with an overview of the basic elements

12 p For p > 1, xi = 0 or 1. 13We may even be able to avoid this expansion altogether, see Appendix E.1.

54 Capstone Issue Vol. 12, 2019 Jesse Hoogland

of statistical physics, which has the goal of turn- tional invariance, we managed to peek deeper into ing microscopic theories into concrete macroscopic ‘black box’ of neural networks than may have oth- observables. One quickly ran into a problem: in- erwise been possible. We quote Koch-Janusz and tractable probabilities. To avoid this fundamental Ringel, “the internal data representations discov- problem, we sampled a host of techniques. MCMC ered by suitably designed algorithms are not just techniques avoid absolute probabilities with itera- technical means to an end, but instead are a clear tions of relative probabilities. Mean field theory reflection of the underlying structure of the physical uses a clever constraint on conditional distributions. system” [5]. Integrating physical and ML perspec- RG creates new probability distributions, iteratively tives proves a promising basis for a better insight simpler, keeping the relevant long-distance infor- into these algorithms. Not to mention, these tech- mation. After an introduction to ML, we noticed a niques are unsupervised, so they should translate resemblance between neural networks and RG. We readily to problems other than the Ising model. identified a need for a more precise notion of ‘rele- Comparisons and integrations of ML and RG are vant’ information, which one formalizes with infor- at an early phase, and there remains much to be mation theory. At the level of ‘relevant’ information, uncovered. It seems that information theory is the ML and RG appeared to behave similarly, extract- appropriate framework with which to lay these links. ing relevant information and suppressing irrelevant Within physics, a better understanding of this area information. However, we saw that RG contained may guide research into new systems (on disor- a more narrow conception of relevance: limited to dered, glassy systems and quantum field theories), long-distance. though the ramifications extend much further [5]. We evaluated Mehta and Schwab’s seminal com- Ultimately, the synthesis of ML, physics, and infor- parison of Kadanoff’s RG and CD-trained RBMs and mation theory may teach us a thing or two about identified a flaw in one of their claims: a ‘good’ RG why these techniques work as well as they do. Until should uncover compact, short-range hidden repre- then, the links should inspire powerful new tech- sentations and respect the original system’s symme- niques in both of these disciplines. tries. Evidently, RBMs do not obey these constraints these unless explicitly instructed to do so. By being more precise in formulating conditions, requiring Appendix A convolutional architectures, we derived an exact correspondence between majority-rule block-spin Basics of Statistical Physics renormalization and restricted Boltzmann machines. Using the information-theoretic formulation of The proceeding derivations follow Cardy [12] and relevance, one can derive a system-independent RG Domb [11]. criterion that optimally satisfies the short-distance condition. We implemented these ideas in rgpy, A.1 Measurements as Averages performing a recalculation of the critical length correlation exponent. This library is publically available, Suppose we want to test whether our theory, and we will continue developing it into the future. the Ising model, lines up with experimental results: We encourage the reader to try this out, and we what are the kinds of predictions we can make? Our hope to enable many more calculations of critical first attempt might be to measure the system’s mag- exponents. To kickstart this project, we provided an netization, rather appropriate as this is the defining overview of possible improvements and generaliza- characteristic of ‘magnets.’ Before we do, however, tions: the next likely targets are the XY-model and let us be more precise. What is a measurement, the spin-1 Ising model. specifically for magnetization? The magnetization, We place special emphasis on the role played M, of any one microstate, s is simply the sum of the by physical reasoning throughout our exploration. orientations of all spins, si ∈ s in a given state: Though information theory presented us a formal notion of relevant information, it was physical rea- X M(s) = si, soning that led us to an appropriate RG condition. i By explicitly thinking about the properties physical systems possess, such as locality and transla- (A.1)

55 Capstone Issue Vol. 12, 2019 Jesse Hoogland

Generally, the microscopic state of a system is From equation (2.6), we have ∆S = Sr(s) − 0 changing rapidly, and the system will explore many Sr(s ). If the reservoir is much larger than s, this different microstates over human time-scales. Our is a tiny difference, so ∆S ≈ dS. Then, we get 1 measuring devices have finite resolution in time, ∆S ≈ T (dE). Note that this requires the reser- so, in a laboratory, we cannot perform sums over voir and system to have the same temperature individual microstates. A measurement becomes (i.e., to be in thermal equilibrium, Tr = Ts). We the average over the microstates visited during the can use the same approximation to expand this out: 0 measurement period, weighted by the time spent in dE ≈ Er(s) − Er(s ) = ∆E. We end up with equa- each state. In statistical physics, we often work in tion (2.7). the limit that the update time goes to zero compared Note that in the thermodynamic limit, |s|, |br| → to the time of measurement. In this light, the notion ∞ while maintaining the inequality 1 |s| |r|, of a probability of a microstate is also the fraction of these statements are precise (this is simply the cen- time the system spends in that microstate over an tral limit theorem). infinite duration. Our measurement outcome, M, is an expectation value, hM(s)i, the average over the set of all microstates, S = {s}, weighted by the amount of time spent in each state, P (s): A.3 Observables as Derivatives of the Parti- tion Function X M = hM(s)i = P (s)M(s). Let us define the free energy, 1 , s∈{s} F = − β ln Z where β = kT , the so-called thermodynamic beta. (A.2) Consider an energy function with a dependence on parameter A according to: The same holds for any measurable quantity. Given some function of a microstate, A(s), the E(s) = E0(s) + λA(s) macroscopic value it corresponds to is its expectation value over all microstates, hA(s)i. Common examples (not restricted to the Ising model) include (A.5) E(s), the energy, P (s), the pressure, and V (s), the volume. If we can find a probability measure over Then, hAi = ∂λF . The derivation proceeds as: microstates, we can use the above to calculate any macroparameter we choose. 1 hA(s)i = ∂ − ln Z λ β A.2 Thermal Equilibrium and the Second Law 1 of Thermodynamics = − ∂ Z βZ λ Consider the first law of thermodynamics (differ- 1 X = − e−βE(s)∂ (−βE(s)) ential form with only one kind of particle): βZ λ s 1 X −βE(s) dE = T dS − P dV + µdN, = e (βA(s)) βZ s (A.3) X e−βE(s) = A(s) Z where E is the energy, T the temperature, P the s X pressure, V the volume, µ the chemical potential, = P (s)A(s). and N the number of particles. s For the canonical ensemble (our example), dV and dN are 0. Then, this reduces to: (A.6, A.7, A.8, A.9, A.10 & A.11, respectively) 1 dE = dS. T Consider an example, taking the Ising Hamilto- (A.4) nian (2.10):

56 Capstone Issue Vol. 12, 2019 Jesse Hoogland

crossing-over behavior. By controlling some relevant parameter we change the effective dimension- 1 X −B P s −J P s s ∂ (− log Z) = − ∂ e i i hi,ji i j ality of our system and navigate between different B Z B s critical points. X e−E(s) X An example of this finite-size behavior is that of = − (− s ) Z i the correlation length, ξ. It will no longer diverge s i X and depending on the boundary conditions will ex- = P (s)M(s) perience a peak at either above or below the infinite s critical point. For cyclical boundary conditions, the = hM(s)i. effective critical temperature will increase, as the system is more ordered with more paths between spins. For zero-field boundary conditions, the tem- (A.12, A.13, A.14 & A.15, respectively) perature will decrease as there less paths between spins. Two others: We have seen from Table 2.3, that diverging −1 −ζ 2 quantities A(t, N = 0) scale as |t − tc| for some χ = ∂BM = ∂ F B critical exponent ζ. In the large N-limit, we expect E = ∂βF. that the system will continue to behave as such, provided that N is much greater than the system’s characteristic length scale, the correlation length, (A.16 & A.17, respectively) −ν ξ ∼ |t − tc| . This amounts to:

−1 −ζ ζ/ν Appendix B A(t, N ) ∼ |t − tc| ∼ ξ (N ξ, t → tc).

Scaling and Renormalization (B.1)

B.1 Finite-Size Scaling Analysis The system’s geometry will act as a cutoff: rather than diverge, now, as ξ → N, A will behave This section follows [29]. In our theoretical dis- as cussion, we have typically considered infinite lattices: for a fixed lattice size L, we have let the mi- −1 ζ/ν croscopic size a (the lattice width) go to zero. Then, A(t, N ) ∼ N (N ξ, t → tc). the dimensionless N = L/a (the number of lattice sites in a given dimension) diverges; this is the clas- (B.2) sical thermodynamic limit. Of course, truly physical Together, these considerations give rise to the systems are finite with correspondingly non-singular finite-scaling ansatz: partition functions. For any such systems, there can be no divergences. An important concern is, then, −1 ζ/ν −1 −1 determining under which conditions we can treat A(t, N ) ∼ ξ φ((Nξ) )(N → 0, t → tc), our systems as infinite and under which conditions finite-size effects have a non-negligible effect. (B.3) An important tool we use to study critical sys- where tems is MCMC simulations: here, we quickly run into computational limits on the sizes of lattices we can ( = const. for |x| 1 simulate. Especially for MCMC techniques, the role φ(x){ . ζ/ν of finite-size scaling is crucial. ∼ x for |x| → 0 Consider a d-dimensional Ising model infinite in (B.4) all directions. If we were to decrease N of one dimension to some finite value, the system’s criti- phi(x) is a scaling function that controls the cal behavior begins to be dominated by a d − 1 finite-size effects. By convention, we choose the dimensional critical point. This is an example of scaling function φ˜(x) = x−ζ φ(xν). Then, we get that

57 Capstone Issue Vol. 12, 2019 Jesse Hoogland

The free energy of new system is equal to that of the original system with some rescaling, plus the ζ/ν ˜ 1/ν A(t, L) = L φ L (t − tc) , (L → ∞, t → tc), addition of a constant term. Note that this is an inho- mogeneous transformation. However, for singular (B.5) behavior (near the critical point), we only care about the second term. g() originates from summing over with short degrees of freedom, and it should be an analytic (non-divergent) function of Ka even at critical ( = const. for x → 0 (L ξ), point. φ˜(x) −ζ/ν −ζ Considering just the transformation rule for the ∼ L (% − %c) for |x| 1 (L ξ). singular part: . −d 0 fs({K}) = b fs({K }). (B.6) (B.9) This is valuable because it allows us to plot experimental data aexp.(t, L) collected at some tem- We use our knowledge that there are only two perature and lattice size, on a single curve: relevant scaling variables, ut and uh (for the interested reader, we refer to Cardy [12]). These possess 1/ν ˜ −ζ/ν φ(L (t − tc)) = L A(t, L) important symmetries: ut corresponds to the even subspace of couplings (invariant a collective sign- . ﬂip s → −s), and uh to the odd subspace (equivari- (B.7) ant with the collective sign-ﬂip). Let us re-express the above equation in terms of our scaling variables. 1/ν 14 Plotting Plotting L (t − tc) against First, we Taylor-expand the scaling variables . By −ζ/ν L A(t, L)|aexp.(t,L), our experimental data will these symmetry arguments, odd terms vanish from ‘collapse’ onto a single line. In practice, it is enough ut and even terms from uh. Looking at the formula to consider only the effective critical temperature for ui, we see they must vanish at t = h = 0. at different length scales. Plotting this against log L, t 2 2 the points should fall onto a single line with slope ν. ut = + O(t , h ) t0 (B.10) B.2 Scaling Rules for the Free-Energy

In this section, we will derive a transformation h rule for the free energy. From Appendix A.3, we uh = + O(th) h0 know that we can use this transformation rule to (B.11) determine our critical exponents of interest. Starting with equation (2.18), we can intuit that Near the critical point, we take these scaling the free energy will take a form similar to: variables to be proportional to h and t15. Combining our equations for the scaling variables, (B.10) and f({K}) = g({K}) + b−df({K0}). (B.11), with the singular free energy transformation:

−d yt yh (B.8) fs(ut, uh) = b fs(b ut, b uh). we expect a function similar to our starting point (B.12) but of the updated coupling f({K0}) and where we If we repeat this transformation n times, we get: have rescaled the system by (for example, a block- size) b. For d dimensions, we rescale in each direc- −nd nyt yh fs(ut, uh) = b fs(b ut, b uh). tion. Furthermore, we could have some (analytic) function of our coupling at a given instance, g({K}). (B.13)

14This is valid since we have already limited ourselves to the immediate vicinity of the critical point. 15In more complicated systems, and those with tricritical points, our relevant variables will not necessarily be directly proportional to the experimental variables (the knobs) we control.

58 Capstone Issue Vol. 12, 2019 Jesse Hoogland

nyt We now choose an arbitrary point ut0 = |b ut|. and differentiate the free energy (f = ln Z) This constrains our n to not be too large that the twice: linear approximation breaks down. Solving for n we get: ∂2 G(r − r ,H) = ln Z({H})| −1 1 2 ∂h(r )∂h(r ) h(r)=0 1 ut 1 2 n = logb ( ). ! yt ut0 ∂ 1 X H(s)−P h(r)s(r) = − s(r )e r | ∂h(r ) Z 2 h=0 (B.14) 1 s 1 X 0 H(s0)−P h(r)s0(r) X H(s)−P h(r)s(r) = − s (r )e r s(r )e r Plugging equation (B.14) into equation (B.13): Z2 1 2 s0 s 1 X P y H(s)− r h(r)s(r) d − h − s(r1)s(r2)e |h=0 yt yt Z ut ut s fs(ut, uh) = fs(±ut0, uh ). ut0 ut0 ! ! 1 X 0 H(s0) 1 X H(s) 1 X H(s) = − s (r1)e s(r1)e + s(r1)s(r2)e (B.15) Z Z Z s0 s s Rewriting in terms of t and h, where we incorpo- =hs(r1)s(r2)iH − hs(r1)iH hs(r2)iH . t t rate ut0 into a new scale factor t0 (i.e., ← ): t0 ut0t0 (B.19, B.20, B.21, B.22, B.23 & B.24, respectively)

d/yt h t h0 We suppose that h(r) varies over scales much fs(t, h) = Φ( ) yh/yt t0 |t/t0| larger than block size ba. Applying block renormalization we can effectively ignore the varying of h(r) (B.16) within a given block. Then, for a speciﬁed block, it

ut0 has allowed us to express the (approximate) transforms as a uniform ﬁeld would. The renormal- function of two variables, fs, in terms of just one ization Hamiltonian should be of the same form. variable in the scaling function Φ. 0 0 X 0 0 0 0 This scaling function may appear to include a H (s ) − h (r )s (r ) r0 ut0 dependency, but since the l.h.s. does not, this will cancel. These scaling functions turn out to be (B.25) universal, it only depends on the system through t0 Since RG preserves the partition function, we and h0. can write:

B.3 The Correlation Length Critical Exponent ∂2 ln Z0(h0) ∂2 ln Z(h) 0 0 0 0 = 0 0 0 0 As an example of using free energy scaling to ∂h (r1)∂h (r2) ∂h (r1)∂h (r2) solve critical exponents, let us attempt correlation (B.26) length critical exponent. Consider the two-point correlation function: Now the l.h.s. is just the correlation function of the RG system with the new Hamiltonian. In terms of our original correlation function, we have to rescale G(r1 − r2,H) ≡ hs(r1)s(r2iH − hs(r1)iH hs(r2)iH the distance by a factor b.

0 (B.17) G((r1 − r2)/b, H )

Our ﬁrst step will be to express this in terms (B.27) of (derivatives of) the free energy. We introduce a 0 0 non-uniform magnetic ﬁeld to our Hamiltonian: Onto the r.h.s. If we make a change in h (r ), we will change all the spins contained in that block: X H → H − h(r)s(r) −yh 0 0 r δh(ri) = b δh (r1), (B.18) (B.28)

59 Capstone Issue Vol. 12, 2019 Jesse Hoogland

so this reduces to: Appendix C

Basics of Information Theory −2yh (1) (1) (2) (2) b h(s1 + s2 + ... )(s1 + s2 + ... )iH . We begin with a random variable x ∈ X , with the (B.29) probability distribution P (x). The information in x is: Each block contains bd spins, so expanding this 2d gives us b 2-point correlations. For our assump- I(x) = − log2 P (x). tion that r =|r − r | is large, each will give about 1 2 (C.1) the same result. For isotropic systems (respecting rotational symmetries of the lattice), the correla- To understand why this is more useful than P (x), tion function will only depend on distance between consider another random variable y ∈ Y with distri- points, not on orientation. bution P (y). If the two variables are independent We end up with the transformation rule: (P (x, y) = P (x)P (y)), we get that their information adds:

0 2(d−yh) G((r1 − r2)/b, H ) = b G(r1 − r2,H). I(x, y) = I(x) + I(y). (C.2) (B.30) The logarithm is a powerful tool that converts For simplicity, we set h = 0 and iterate n times: multiplication (hard) to addition (easy). Intuitively, it makes sense. An event which is 100% likely P (x) = 1, is guaranteed to happen, so when it hap- G(r, t) = b−2(d−yh)G(r/b, byt t) = b−2n(d−yh)G(r/bn, bnyt t), pens, it carries no information. An event which is inﬁnitely unlikely, P (x) = 0 should never happen. (B.31) If it does, somehow, that event carries an inﬁnite

nyt amount of information. Precisely, information is stopping at some ﬁxed point where b (t/t0) = 1 (as we did for the free energy in Appendix B.2). measured in bits a 0 or 1. Consider trying to encode Then, samples of P (x) as effectively as possible. When P (x) = 1/2, we need only one bit to encode x. For 1 P (x) = 1/4, two bits, and so on. n = − logb(t/t0), yt In this light, entropy is the expected value of information: were we to sample P (x), then what is the (B.32) average information we glean from any one sample? In this limiting case of inﬁnite samples, we perform and plugging this in: a sum of the information of each state weighted by the probability of that state: ! r 2(d−yh)/yt X G(r, t) = |t/t0| Ψ . −1/yt S(X ) = hI(x)i = P (x)I(x). |t/t0| x∈X (B.33) (C.3)

By deﬁnition, −r/ξ for , near the G(r) ∝ e r ξ C.1 Cross-Entropy critical point (see argument in, for example, [12]). From the argument of the scaling function, P si, we Assume we have one set of events, X , but two can identify that: ξ ∝ |t|−1/yt . distributions P (x) and Q(x). We have devised an op- Then, timal encoding of the events in X using Q(x). This means we need the least amount of bits that is pos-

ν = 1/yt sible to identify events x from Q(x). If instead, the events suddenly come from P (x) and P (x) 6= Q(x), (B.34) we will need, on average, more bits to identify x.

60 Capstone Issue Vol. 12, 2019 Jesse Hoogland

The average number of bits we need is the cross- entropy, H, identiﬁed as the expectation over P of P P P X −E (v,h) X − aivi− bj hj − viwij hj the information over Q: e θ = e i j ij h h P P P − aivi X − (bj + viwij )hj X X = e i e j i . H(P,Q) = hIQ(x)iP = P (x)IQ(x) = − P (x) log Q(x). h x∈X x

(C.4) (D.3, D.4 & D.5) Choosing Q so as to minimize this quantity, we derive an encoding closer to P (x).

Furthermore, C.2 Kullback-Leibler Divergence

If we rewrite the KLD (3.13) as X − P b +P v w h X Y − b +P v w h e j ( j i i ij ) j = e ( j i i ij ) j h h j DKL(P ||Q) = H(P,Q) − S(P ), Y X − b +P v w h = e ( j i i ij ) j (C.5) j hj ∈{0,1} Y − b +P v w P and Q minus the entropy over P . In other = 1 + e ( j i i ij ) . words this is the average number of additional bits j we need to identify x from P when using the im- proper coding Q. This amount is 0 if and only if H(P,Q) = S(P ). From equation (C.4) and equation (D.6, D.7 & D.8, respectively) (C.3), we see, by the positive semi-deﬁniteness of entropy, this immediately implies that P =Q.

Plugging in, we see, Appendix D   Restricted Boltzmann Machines P P − aivi Y −(bj + viwij ) Eθ(v) = − log e i 1 + e i  D.1 Factoring of the Marginal Distribution j P X X −(bj + viwij ) Our aim is to solve for Eθ(v), the energy function = aivi − log 1 + e i . describing the marginalized system of visibles of an i j RBM, equation (3.12), in terms of the joint energy function of the full system, equation (2.10). Combin- ing these two equations, we get the following. (D.9 & D.10, respectively)

X e−Eθ(v) = e−Eθ(v,h) h ! X −Eθ(v,h) ⇐⇒ Eθ(v) = − log e , D.2 Correspondence between Kadanoff’s h Variational RG and RBMs

(D.1 & D.2, respectively) When we form a comparison with restricted Boltzmann machines, we will consider the limited where the partition functions have cancelled out. case of exact transformations (i.e., ∆F = 0 ⇐⇒ We can rewrite the right hand side: F (v) = Fθ(h)). This implies that:

61 Capstone Issue Vol. 12, 2019 Jesse Hoogland

T (v,h) −E(v,h)+H(v) F (v) = Fλ(h) e = e !   X −E(v,h) −H(v) T λ(v,h)−H(v) e X −E(v,h) P (v) −H(v) − ln e = − ln Trhj ,vi e = · e · · e P −E(v,h)   v v ,h e P (v) i j vi,hj X −H(v) T λ(v,h)−H(v) e = Trhj ,vi e P −E(v,h) P (v, h) v ,h e RBM v i j −H (v)−H(v) = · RBM e X X T (,h) −H(v) P (v) P e−H (v) = e λ e vi RBM v h∈H P −H (v) v e RBM X X = P (h|v) · i e−H (v)−H(v) = e−H(v) eT λ(v,h), P e−HRBM (v) vi v h∈H RBM = P (h|v) · e−H (v)−H(v).

(D.11, D.12, D.13, D.14 & D.15, respectively) (D.21, D.22, D.23, D.24 ˆ D.25)

which is true if and only if: This lends Kadanoff’s visible-hidden coupling T an interpretation as a kind of variational conditional T λ(v,h) probability distribution. Using the exact case, equa- Trhj e = 1. tion (D.16), we show:

(D.16) X 1 = eT λ(v,h) (RBM) We know that P (h) is a marginalization hj over the energy function E(v, h). Combining this X P (v) = P (v|h) with equation (D.6): Ptrue(v) hj P (v) = . −HRBM (h) −E(v,h) e X e Ptrue(v) P (h) = = Zh Z vi (D.26, D.27 & D.28, respectively) X eT (v,h)−H(v) = . Z v i Therefore, P (v) = Ptrue(v). This derivation triv- RG e−H (h) ially goes both ways. = . Zh D.3 An Exact Correspondence between Majority-Rule RG and Convolutional RBMs (D.17, D.18 & D.19, respectively) In this section, we will see that it is possible to We identify: implement a majority-rule block update with an RBM. Note that this does not concern whether or not any given RBM will actually learn this rule. For argu- HRBM (h) = HRG(h) ments considered in Chapter 4, we need only con- j sider the update function for a single block, P (hj|v ). (D.20) Ignoring the index on v(j), let us remind ourselves of the RBM’s conditional distribution:

The Hamiltonian describing our hidden spins af- 1 P (h |v) = ter block-spin renormalization also describes the j −h (P w v +b ) 1 + e j i ij i j . conﬁguration of hidden spins in our RBM. Further- more, we can expand: (D.29)

62 Capstone Issue Vol. 12, 2019 Jesse Hoogland

For block renormalization, we want each spin v i X to contribute equally, so we can consider the simpler P (h, e) = P (h|v)P (v, e). wij → wj. Let, v X (E.1) −hj(wj vi + bj). i In Chapter 6, we saw that we can interpret an (D.30) RG transformation as a conditional probability distribution P (h|v), equation (4.4) (dropping the index j If we have N, visible spins (i.e. i ∈ {1 ...N}), and allowing multiple hidden units per block). As we then we can rewrite the sum over visibles in terms discussed in Chapter 5, we model this distribution of their average hvii as with an RBM, with parameters Λ, such that:

−hjwj (Nhvii + bj) . e−EΛ(h,v) (D.31) PΛ(h, v) = 0 0 , P −EΛ(h ,v ) h0,v0 e If we choose bj = −Nwj/2, we get:   X (Λ) X (Λ) X (Λ) 1 EΛ(h, v) = −  w vihj + a vi + b hj . −h w N hv i − . ij i j j j i 2 ij i j (D.32) 1 Restricting our attention to hvii − 2 , we see this (E.2 & E.3, respectively) is positive if and only if more than half of the spins As we saw, we can use the above to derive easy in vi are up, 1, and negative if and only if half the spins are down, 0. Narrowing our attention to to evaluate equations for PΛ(h|v), PΛ(h), and PΛ(v). In fact, we can also derive an equation for P (v|h). hhji = P (hj|v), the trick to recover the majority rule Λ However, for RG, we only care about transforma- transformation will be to let wj → −∞. tions in one direction, v → h. Therefore, for sim- (Λ) 1 plicity, we can set a = 0 (this term factors out in hhji = lim i w →∞ −w N hv i− 1 j 1 + e j ( i 2 ) PΛ(h|v)). We get that: (D.33) We distinguish three cases: X e−EΛ(v) PΛ(v) = = P ve−EΛ(v)  hPΛ(h,v) 0 ⇐⇒ hvii < 1/2  P X −(bj + i viwij ) hhji = 0.5 ⇐⇒ hvii = 1/2 EΛ(v) = − log 1 + e ,  j 1 ⇐⇒ hvii > 1/2.

(D.34) (E.4 & E.5, respectively) This is exactly the majority-rule, including even the probabilistic update rule for blocks of even num- and bers of spins. P (h, v) Λ −EΛ(h,v)+EΛ(v) PΛ(h|v) = = e . Appendix E PΛ(v)

The Real-Space Mutual Information Maxi- (E.6) mization Algorithm Combining with our mutual information condition, Equation 5.1, we get: E.1 A Proxy for the Mutual Information

As a ﬁrst step, let us factor the joint distribu- P ! tion in equation (5.1) (h on e is mediated entirely X v P (v, e)PΛ(h|v) AΛ(h : e) = PΛ(h|v)P (v, e) log P 0 0 0 P (v , e)PΛ(h|v ) through v): h,v,e v ,e

63 Capstone Issue Vol. 12, 2019 Jesse Hoogland

(E.7) K(x) P∞ 1 C he i = e κ=0 κ! κ , We assume that all of these distributions (not only those deﬁned by the Λ-RBM) are of Boltzmann (E.13) form. Then, we can factor the partition functions, keeping an equation of Boltzmann form: Since these with the cumulants expressed in terms of the probabilities are all of Boltzmann form, we can factor moments. The ﬁrst three terms are: out partition functions, leaving behind an equation with Boltzmann terms: C1 = hKi 2 2 C2 = hK i − hKi P P −E(v,e)−EΛ(h,v)+EΛ(v) v P (v, e)PΛ(h|v) v e 3 2 3 → 0 0 0 . C3 = hK i − 3hK ihKi + 2hKi . P 0 0 P −E(v )−EΛ(h,v )+EΛ(v ) v0 P (v )PΛ(h|v ) v0 e

(E.8) (E.14, E.15 & E.16, respectively) We can reexpress the argument of the logarithm as follows: This extends to distributions that include a dependence on other variables (i.e., e, h). Then, we can approximate: P e−E(v,e)−EΛ(h,v)+EΛ(v) P e−E(v)−EΛ(h,v)+EΛ(v)e−E(v,e)+E(v) P e−EΛ(h,v,e)e−∆E(v,e) v = v . = v , P −E(v0)−E (h,v0)+E (v0) P −E(v0)−E (h,v0)+E (v0) P −E (h,v,e) 0 e Λ Λ 0 e Λ Λ 0 e Λ v v X v AΛ(h : e) ≈ PΛ(h|v)P (v, e)h−∆E(v, e, h)i[e, h], (E.1) h,v,e

where (E.17)

EΛ(v, e, h) = E(v, e) + EΛ(h, v) − EΛ(v) where now ∆E(v, e, h) = E(v, e) − E(v).

P (∆E (v, e, h)) e−EΛ(v,h) h∆E (v, e, h)i[e, h] ≡ v Λ . Λ P −E (v0,h) 0 e Λ (E.10 & E.11, respectively) v

This is the expectation of e−∆E(v,h,e) over the (E.18) Boltzmann distribution with energy EΛ(h, v, e), In general, we will not have access to the en- where h and e are clamped. We write ergy functions E(v) and E(v, e). Even if we were −∆E(v,h,e) he iΛ[e, h], where [e, h] denotes the depen- to have access to P (x = {v, b, e, o}), the neces- dence of this expectation value on the clamped P sary marginalizations (P (v, e) = b,o P (x) and values. Although ∆E has no dependence on Λ, P P (v) = b,e,o P (x)) are not necessarily tractable its expectation value gains a dependence through calculations. Therefore, we approximate E(v, e) and PΛ(h, v, e). We get the following expression for the E(v) with two other RBMs with parameters Θ and mutual information proxy: Ψ, respectively:

X −∆E(v,e,h) AΛ(h : e) = PΛ(h|v)P (v, e) log he iΛ[e, h] , E(v, e) ≈ EΘ(v, e) h,v,e E(v) ≈ EΨ(v). (E.12)

Here we see that the clamped values of e and h (E.19 & E.20, respectively) are given by the outside sum, so we write [e, h] after the expectation value to denote its dependence on Note that (v, e) and (v) are the inputs, the visible these variables. layers of these two RBMs. We introduce a second, To further simplify this expression, we use a cu- hidden layer, which we marginalize over to get the mulant expansion: above quantities.

64 Capstone Issue Vol. 12, 2019 Jesse Hoogland

Plugging in our learned distributions, we get: E.2 Intrinsic Thermometer

In order to measure critical exponents, we need X a means of measuring the values of macrovalues AΛ,Θ,Ψ(h : e) ≈ PΛ(h|v)P (v, e)h−∆EΘ,Ψ(v, e, h)iΛ,Θ,Ψ[e, h], h,v,e through our iterations. If our aim is to measure temperature, we may, however, not have access (E.21) to an explicit thermometer. To further complicate matters, as we saw previously, RG transformations We have derived an expression which we evalu- generally introduce higher order correlations. From ate using Monte Carlo averages. data alone, it is not necessarily possible to identify the contributions from different terms in the Hamil- tonian. For our approach to be general, then, we 1 X A (h : e) ≈ P (h|v)h−∆E (v, e, h)i need[e, h a], means of implicit macroparameter calcula- Λ,Θ,Ψ N (v,e) Λ Θ,Ψ Λ,Θ,Ψ h,v,e tions. With regard to temperature, we have several options.’ (E.22) 1. First they looked at using the MC configura- First, we generate samples of e and v simply tions at given iterations. We can compute from partitions of our dataset x16. Then, we use expectations of functions like the nearest- our Λ-RBM to translate samples of v into samples neighbor and next-nearest-neighbor correla- of h. For each combination of e and h, we generate tion. Since these depend monotonically on the wholly new samples, v0, with the energy function temperature near the critical point, we can use these to recreate an effective temperature at EΛ,Ψ(v, e), over which we perform the internal MC- average. successive length scales. In fact, we are not interested in the quantity, 2. Next, they used the mutual information as a AΛ,Θ,Ψ(h : e) as much as we are in its derivative. proxy to the temperature. This, too, depends With some simple (though exceedingly tedious)17 monotonically on the temperature near the Algebra, we can evaluate an expression for the critical point. derivative with respect to Λof our mutual informa- 3. Finally, they mentioned the possibility of using tion proxy. It is crucial we calculate this explicitly, be- the - and -RBMs. These we can also use cause through the stochastic nature of MC-sampling, Λ Ψ to measure expectations of correlations, and our proxy AΛ will often have a zero-gradient. We we can even intrinsically evaluate effective would not be able to use stochastic gradient de- temperatures. scent. *Comment on the Cumulant Expansion. It is not There are yet other options, not considered by entirely clear why the authors felt this expansion Koch-Janusz and Ringel. We can formulate the task was necessary. The expectations are ultimately cal- of measuring the effective temperature of a set of culated over Monte Carlo samples, and hK(x)i is samples as a supervised learning problem. Then, we not much of an improvement over heK(x)i in compu- can further leverage the power of neural networks tational complexity. Furthermore, it disregards infor- to act as our thermometers18 So far, we have only mation about higher order terms. It may be that this considered neural networks which take a fixed num- step manages to suppress irrelevant fluctuations, ber of inputs. Our aim for a NN thermometer would but a more rigorous understanding is needed. In the be that it could accurately measure the temperature future, we may implement the algorithm, including of systems with different numbers of spins (at dif- additional terms in this expansion, comparing the ferent RG steps). Two immediate solutions come to ultimate performance against an un-expanded base- mind. First, we could train the networks on subsets line. Then, we can more rigorously justify or critique of the x samples. Then, we would train a separate this assumption. network for each successive step. Second, we could

16We could have generated these samples de novo using the Θ-RBM, but that would introduce needless time-complexity. Instead, we use that data we already have access to 17I mean whiteboards and whiteboards of it. 18Accomplished for example by Iso et al. [6]).

65 Capstone Issue Vol. 12, 2019 Jesse Hoogland

make use of recursive neural networks19. These recursive neural networks employ, as their name suggests, recursive weight-sharing, which explicitly allows for variable-sized inputs. However, this would reduce the extent to which the RSMI algorithm is truly unsupervised. In practice, this need not be an issue: we assume that we have access to some data, and more often than not, this will be given to us by MCMC techniques. Then, we necessarily have a means of measuring parameters like these, at least in the non-coarse-grained systems. Figure E.1 – The Magnetic Susceptiblity, χ, at different iterations of the RSMI algorithm. Though we did not have the time to implement these ideas, we will continue developing rgpy, and Works Cited we intend to introduce these features in later re- leases. [1] Yann LeCun, Yoshua Bengio, and Geoffrey Hin- For our implementation, we used the next- ton. Deep learning. Nature, 521(7553):436, nearest correlation function. For each value of this 2015. function we derive in successive iterations, we as- [2] Henry W. Lin, Max Tegmark, and David Rol- sign the temperature that it is closest too in our nick. Why does deep and cheap learning sample set. In the future, we will use more compli- work so well? Journal of Statistical Physics, cated functions: this ‘nearest-neighbors’ approach 168:1223–1247, Sep 2017. introduces significant error margins. [3] Naftali Tishby, Fernando C Pereira, and William Bialek. The information bottleneck method. arXiv preprint physics/0004057, 2000. [4] Pankaj Mehta and David J. Schwab. An exact mapping between the variational renormal- E.3 Experimental Realization ization group and deep learning. arXiv e- prints, page arXiv:1410.3831, Oct 2014. Ultimately, our calculation of the correlation [5] Maciej Koch-Janusz and Zohar Ringel. Mu- length critical exponent, ν, proceeded as follows. tual information, neural networks and the We generated samples (1000 per temperature) of renormalization group. Nature Physics, Ising configurations of various lattice widths (8, 16, 14:578–582, Jun 2018. 32, 64). From values for the next-nearest neighbor [6] Satoshi Iso, Shotaro Shiba, and Sumito Yokoo. correlation function, we devised a “thermometer”, Scale-invariant feature extraction of neu- see preceding section, for each length-scale. We ral network and renormalization group flow. trained the RSMI algorithm on samples of 64-by-64 arXiv e-prints, 97:053304, May 2018. spins for a total of 3 RG iterations. Each RBM we [7] David J. Schwab and Pankaj Mehta. Comment on trained for 30 epochs. For the rest, our experimen- ”why does deep and cheap learning work tal set-up mirrored that of Koch-Janusz and Ringel ?. so well?” [arxiv:1608.08225]. arXiv e-prints, Having generated results, and measurements of the page arXiv:1609.03541, Sep 2016. magnetic susceptibility, χ, for each step in each [8] Jaco ter Hoeve. Renormalization group con- temperature sequence. From the peaks in suscep- nected to neural networks, 2018. tibility, we identified the effective critical tempera- [9] Patrick M. Lenggenhager, Zohar Ringel, Sebas- ture. Plotting this on one curve against 1/L, using tian D. Huber, and Maciej Koch-Janusz. Op- arguments from finite-size scaling (see section ??), timal renormalization group transformation we collapsed the data onto a single curve whose from information theory. arXiv e-prints, slope equals ν. page arXiv:1809.09632, Sep 2018.

19Not to be confused with recurrent neural networks, a particular kind of recursive neural network used for processing 1-dimensional (typically temporal) data.

66 Capstone Issue Vol. 12, 2019 Jesse Hoogland

[10] Jesse Hoogland. rgpy. https://github.com/jqhoogland/rgpy, page arXiv:1803.08823, Mar 2018. 2019. [21] John J Hopfield. Neural networks and physical [11] Cyril Domb. The critical point: a historical in- systems with emergent collective computa- troduction to the modern theory of critical tional abilities. Proceedings of the national phenomena. CRC Press, 1996. academy of sciences, 79(8):2554–2558, [12] John Cardy. Scaling and Renormaliztion in Sta- 1982. tistical Physics. Cambridge University Press, [22] Mukul Rathi. Learning through gradient descent, Cambridge, United Kingdom, 1996. Aug 2018. [13] JGTechSol. Optical computing, Feb 2017. [23] Yann LeCun, Léon Bottou, Yoshua Bengio, [14] Alan H Guth. Time since the beginning. arXiv Patrick Haffner, et al. Gradient-based learn- preprint astro-ph/0301199, 2003. ing applied to document recognition. Pro- [15] Christian Robert and George Casella. Monte ceedings of the IEEE, 86(11):2278–2324, Carlo statistical methods. Springer Science 1998. Business Media, 2013. [24] Max Welling, Michal Rosen-Zvi, and Geoffrey [16] Jian-Sheng Wang and Robert H Swendsen. Clus- E Hinton. Exponential family harmoniums ter monte carlo algorithms. Physica A: with an application to information retrieval. Statistical Mechanics and its Applications, In Advances in neural information process- 167(3):565–579, 1990. ing systems, pages 1481–1488, 2005. [17] Kenneth G Wilson. Problems in physics with [25] Aaron van den Oord, Yazhe Li, and Oriol many scales of length. Scientific American, Vinyals. Representation learning with con- 241(2):158–179, 1979. trastive predictive coding. arXiv preprint [18] Michael E Peskin. An introduction to quantum arXiv:1807.03748, 2018. field theory. CRC Press, 2018. [19] Leo P Kadanoff, Anthony Houghton, and [26] R Devon Hjelm, Alex Fedorov, Samuel Lavoie- Mehmet C Yalabik. Variational approxi- Marchildon, Karan Grewal, Adam Trischler, mations for renormalization group trans- and Yoshua Bengio. Learning deep rep- formations. Journal of Statistical Physics, resentations by mutual information esti- 14(2):171–203, 1976. mation and maximization. arXiv preprint [19] Geoffrey E Hinton. A practical guide to train- arXiv:1808.06670, 2018. ing restricted Boltzmann machines. In Neu- [27] Mohamed Ishmael Belghazi, Aristide Baratin, ral networks: Tricks of the trade, pages Sai Rajeswar, Sherjil Ozair, Yoshua Bengio, 599–619. Springer, 2012. Aaron Courville, and R Devon Hjelm. Mine: [20] Pankaj Mehta, Marin Bukov, Ching-Hao Wang, mutual information neural estimation. arXiv Alexandre G. R. Day, Clint Richardson, preprint arXiv:1801.04062, 2018. Charles K. Fisher, and David J. Schwab. A [28] Andreas Sorge. (2015). pyfssa 0.7.6. Zenodo. high-bias, low-variance introduction to ma- DOI: https://pyfssa.readthedocs.io/en/stable/fss- chine learning for physicists. arXiv e-prints, theory.html

67 Social Sciences

Do You Want to Know More? Curiosity for Positive and Negative Life Stories of Ingroup and Outgroup Members

Afke van Egmond

Supervisor Dr. Suzanne Oosterwijk (UvA) Reader Dr. Sennay Ghebreab (AUC) Capstone Issue Vol. 12, 2019 Afke van Egmond

Abstract

Current research on information seeking, an important aspect of curiosity, has shown that it can depend on context; for example, information seeking is different for positive as opposed to negative information, as well as about ingroup as opposed to outgroup members. The present study examines these relationships experimentally, using a paradigm developed for the purpose of this study. In this paradigm, participants were provided with pictures of ingroup and outgroup members, each with an emotional facial expression, along with a short description of this person’s situation. These pictures and stories were either of positive or negative valence. Participants were constantly asked whether they wanted to know more about the person in the picture. The data was analysed using a generalised estimating equations (GEE) model. The results revealed that information seeking was higher for negative than for positive information, which can be explained by morbid curiosity. Overall, people did not seek information differently about ingroup than outgroup members. However, this changed when valence was taken into account: participants sought the least information about ingroup members in trials containing positive emotional information. This could be because this type of information is the least novel and uncertain. Future research should examine how the way people seek information inﬂuences what people learn about others.

Keywords and phrases: information seeking, curiosity, outgroup bias, valence, empathy

I. Introduction empathising with others is also driven by intrinsic motivation. We may empathise when we learn about others’ life events, and it has been shown In today’s society we are continuously con- that empathy can help in reducing outgroup bias. fronted with the life stories of other people. We The present study incorporates all these ideas and listen to the life events shared by family, friends, aims to answer the question: How is curiosity for neighbours, and colleagues, whether they are part the positive and negative life stories of other peo- of the social we belong to, the ingroup, or the group ple inﬂuenced by social group status? we do not belong to, the outgroup We learn about other people’s lives through news, books, movies, Deﬁned as the drive to learn and explore, to and shows. When we learn about life stories of seek new information, with the goal to expand others we are seeking information, which is an im- knowledge, curiosity is what intrinsically motivates portant behaviour resulting from a state of curios- people to seek information about their (social) envi- ity. Research indicates that this information seek- ronment (Kashdan & Silvia, 2009). One of its coming behaviour seems to differ depending on the so- ponents is social curiosity, which explains that peo- cial context and the valence of the information that ple acquire information about other people in order is sought; for example, we might learn differently to gain a better understanding of their social en- about people that are part of our own social group vironment (Kashdan et al., 2018; Renner, 2006). than people who are not. Since information seeking Gottlieb, Oudeyer, Lopes and Baranes (2013) ar- helps us keep track of our social environment, it is gue that since human social and cultural structures important that we learn about people in our own so- change so quickly, understanding this environment cial group in the same way as people in other social by seeking information in this context is evolution- groups, in order to keep a realistic view of the so- arily adaptive. They argue that the motivation for cial environment. However, it seems that this does information seeking is independent of a reward, not always happen; outgroup bias, or generalised and that learning alone is enough of a reinforcer negative beliefs about outgroup members, exists to engage in the process of seeking information. in virtually all diverse, modern societies. There- Therefore, they conclude that information seeking fore, it is important to examine where this outgroup is associated with a high degree of intrinsic motiva- bias comes from, and whether it is related to infor- tion. Loewenstein (1994) explains this relationship mation seeking and curiosity, because this knowl- between motivation and curiosity with his informa- edge can help in targeting problems such as the tion gap hypothesis, which states that people are racist and discriminatory behaviour that may result motivated to seek information when there is a dis- from it. Information seeking is driven by intrinsic crepancy between what they know and what they motivation; similarly, new research suggests that want to know. They then seek out information in

70 Capstone Issue Vol. 12, 2019 Afke van Egmond

order to move from a state of uncertainty to one ple have different expectations of and assumptions of certainty. People are intrinsically motivated to about ingroup and outgroup members, which often seek out information about others in order to learn persist because of confirmation bias, or the ten- more about their social environment and to resolve dency to seek out information that confirms one’s uncertainty. own beliefs (Nelson, 2014). Aboud (1976) con- When people are learning about their social en- ducted an experiment to show that people indeed vironment, for example when they want to know seek out different kinds of information about in- more about the life stories of others, the way they group than outgroup members. She offers differ- seek information can differ depending on the con- ent explanations for why these discrepancies exist, text. For example, Baumeister, Bratslavsky, Finke- the most significant one being that when people nauer and Vohs (2001), as well as Taylor (1991), are communicating with members from other social have shown that the processing of negative infor- groups, they seek information that affirms their ex- mation is more ‘costly’ than positive information, pectations that these people are somehow different meaning that it requires more cognitive resources. from themselves - effectively performing according Based on this idea, it can be hypothesised that peo- to confirmation bias. Similarly, when people are ex- ple are less likely to seek out negative informa- posed to information that does not agree with their tion than positive information and would, therefore, beliefs, they disregard it. Uncertainty could be an- prefer to learn more about positive life stories than other reason why people might seek information negative ones. A contradicting idea comes from differently for outgroup than ingroup members. As the concept of morbid curiosity, which describes previously discussed, uncertainty drives curiosity a curiosity for intense, extremely negative infor- (Gottlieb et al., 2013). It can be hypothesised that mation (Litman, 2005; Oosterwijk, 2017). Zucker- people are more uncertain about individuals from man (1990) argues that people seek out this type different social groups than their own, and thereof negative information because they expect that fore are more driven to learn about these people it will induce a feeling of sensation. Baumeister et (Weary & Jacobson, 1997). In conclusion, seeking al. (2001) hypothesise that being sensitive to neg- information about a person seems to depend on ative information is evolutionarily adaptive, which what social group that person belongs to in relation might also explain why people tend to seek out to one’s own. negative information (Oosterwijk, 2017). People re- Seeking out information about others, such as spond to positive and negative information differ- learning more about their life stories, might nat- ently, and therefore information seeking in the con- urally activate empathy. Empathy is the cogni- text of learning about other’s positive and negative tive process that helps people understand each life stories may also differ, although the exact di- other’s mental and emotional state, and in partic- rection of this difference is not clear yet and should ular enables them to share this state (Avenanti, in the future be further investigated. Bueti, Galati, & Aglioti, 2005; Stephan & Finlay, Information seeking may depend on the social 1999). Until recently, the consensus was that context, in particular the interaction with differ- empathising happened mostly automatically; evi- ent social groups. The social group that one be- dence from multiple fields supported this claim. For longs to is referred to as the ‘ingroup’, and the instance, early research done by developmental group that one does not belong to is called the psychologists Haviland and Lelwica (1987) showed ‘outgroup’ (Fiske, 1998; Tajfel, 1970). Tajfel and that infants as young as ten weeks old mimicked Turner (1979) argue that the sense of belonging to their mothers’ emotions, and emotional mimicry a group helps to build not only someone’s social is thought to be a predictor of empathy (Harri- identity, but also their self-image and self-esteem. son, Morgan, & Critchley, 2010). Congruently, re- For that reason, people tend to enhance the sta- search from social psychology has robustly shown tus of their ingroup in order to reinforce their self- that emotional contagion, also an important build- image. Similarly, evolutionary psychologists speak ing block of empathy, happens mostly automati- of ingroup favouritism and outgroup hostility as cally (Hatfield, Carpenter, & Rapson, 2014; Hat- adaptive features that promote survival (Masuda & field, Rapson, & Le, 2013). Neuroscientific research Fu, 2015; Neuberg, Smith, & Asher, 2000; Van Vugt supports this idea on a neural level, by showing & Schaller, 2008). These ideas illustrate why peo- that performing an action is supported by simi-

71 Capstone Issue Vol. 12, 2019 Afke van Egmond

lar neural systems as observing that same action Since processing negative emotions or information (Gallese & Goldman, 2002). Zaki (2014) calls this is more costly than positive information (Baumeis- parallel activation ‘neural resonance’, and he ar- ter et al., 2001; Taylor, 1991), it is inherently im- gues that this happens reflexively and even in the plied that empathy can be more costly depending absence of attention. on the emotional context, and therefore might be Even though the previously discussed studies avoided in some cases. Indeed, Cameron et al. all seem to indicate that empathy is automatic, (2016) as well as Zaki (2014) argue that empathy recent evidence has shown that empathy is also is an effortful process with a high cognitive cost. context-dependent. First of all, Decety, Yang and Based on this idea, Cameron, Cunningham, Saun- Cheng (2010) showed that in some cases, neural ders and Inzlicht (2018) state that empathising is a resonance is not completely automatic. They found choice, based on the cost and benefits of its conse- that physicians, who are trained to ‘turn off’ em- quences. In short, we cannot take the automaticity pathy for their patients, had different early-latency of empathy for granted; the present study assumes event-related potentials than control participants that people do have some agency in empathising, when witnessing others being exposed to painful in the form of motivation to engage in the resource- stimuli. This early latency indicates that the dif- demanding process of empathy. ference in the responses of physicians and con- Based on these ideas, it seems that empathis- trol participants was rooted mostly in automatic- ing with is similar to seeking information about ity. In other words, physicians had less of an au- someone in the sense that they both are driven by tomatic empathy response; after years of down- motivation, and can depend on context. Empathy regulating empathy for pain, this reaction simply is relevant for the process of learning about oth- did not unfold. Similarly, Sheng and Han (2012) ers’ life stories; research on motivated empathy, measured participants’ neural activity when ob- as well as research on information seeking, indi- serving ingroup and outgroup members experienc- cate that learning about others’ life stories might ing pain, and found that this activity significantly be different depending on the social group this differed in these two conditions. This effect again other belongs to. Additionally, research on infor- happened at such an early latency that it can be mation seeking indicates that there might be a dif- concluded that people empathise less automati- ference in positive and negative information seek- cally with outgroup members than ingroup mem- ing; this preference for either positive or nega- bers. Other studies using several different meth- tive information might even depend on the social ods have supported the idea that people empathise group of the person someone is seeking informa- less automatically with outgroup members than in- tion about. People might be more motivated to group members (Cikara & Fiske, 2011; Cikara & Van seek negative information about ingroup members Bavel, 2014; de Waal, 2008; Harris & Fiske, 2006). than outgroup members, since the former are more The contextual influence of empathy was also ex- relevant to them than the latter (Fiske, 1998; Tajfel, amined by Meffert, Gazzola, Den Boer, Bartels and 1970), and processing negative information is more Keysers (2013), who found that psychopaths have costly than positive information. Therefore, peo- been shown to have abnormal neural activation in ple might be less motivated to spend these ex- situations where empathy is expected to be sponta- tra resources necessary for negative information neously elicited, but also that this difference is re- on the less relevant (the outgroup) than on what duced when they are explicitly asked to empathise. directly affects them (the ingroup). Conversely, Based on the evidence found by Meffert et al. based on the idea of confirmation bias, it is possible (2013) regarding empathy in psychopaths, Keysers that people seek more positive information about and Gazzola (2014) make a distinction between the ingroup members and more negative information capacity and the willingness to empathise. This about outgroup members, since this information suggests that empathy is not something that can agrees with people’s expectations. The present only be activated automatically, but that willing- study will examine these relationships and will aim ness or motivation can play a role in it. Zaki to answer the question: How is curiosity for positive (2014) also claims that since empathy is relatively and negative life stories of other people influenced resource-demanding, motivation is necessary to ei- by social group status? ther approach or avoid empathising with others. In order to measure how curiosity for life sto-

72 Capstone Issue Vol. 12, 2019 Afke van Egmond

ries is influenced by the social context, the present Ward, & Prkachin, 2011; Inzlicht, Gutsell, & Legault, study carried out an experiment, for the purpose 2012; Sheng & Han, 2012). If it is indeed true that of which a new paradigm was developed, which people are less motivated to learn more about out- simulated real-life experiences of learning about group members than ingroup members, it can be others’ life stories. In this paradigm, participants concluded that motivation is what should be tar- were shown images of ingroup members and out- geted in stimulating learning more about and thus group members expressing positive (joy, surprise) sharing experiences of others, which can reduce and negative (sadness, fear) emotions, that were outgroup bias. Furthermore, the academic rele- taken from the Amsterdam Dynamic Facial Expres- vance of the study is expressed in multiple ways. sions Set (ADFES; van der Schalk, Hawk, Fischer, First of all, it is an exploratory study aiming to & Doosje, 2011). This dataset contains pictures of establish a relationship between concepts of cu- North-European and Mediterranean models, which riosity, social identity, motivation, and empathy. were validated by van der Schalk et al. (2011) In order to establish these links, validated ques- to be the ingroup and outgroup, respectively, for tionnaires measuring concepts of curiosity (social Dutch participants. Along with these pictures, par- as well as morbid) and empathy are included, so ticipants were given a one- sentence description of that correlations between these concepts and pos- an emotional event, corresponding to the facial ex- itive and negative information seeking can be expression, that the person in the picture had sup- amined. This is helpful in validating the present posedly experienced. Participants were then asked paradigm and will shed more light on how seeking whether they wanted to know more about this per- information about other people’s life stories is re- son. Information seeking was quantified in this lated to trait social curiosity and different aspects way, representing the number of times participants of trait empathy. Moreover, by putting these con- asked for more information in each trial. cepts in the same context for the first time in an The present study expected first of all that in- innovative manner, the present study could make formation seeking would be lower for negative than use of multiple insights from research about each for positive information, based on the idea that proof these concepts and link them to build different cessing negative information is more costly than hypotheses. These hypotheses were tested using processing positive information (Baumeister et al., a paradigm that was newly developed for the pur- 2001; Taylor, 1991). Additionally, it was expected pose of the experiment. This paradigm mimics real- that information seeking would be lower for out- life situations, which increases the external validity group members than for ingroup members, since of the present study; in addition, it can be used in outgroup members are less relevant, and stud- future research, perhaps instead of paradigms that ies have shown that people are have negative are currently used to measure concepts of informa- generalised assumptions about outgroup members tion seeking or curiosity. (Fiske, 1998; Tajfel, 1970). Last of all, it was expected that the difference between information seeking for outgroup and ingroup members was bigger for negative than for positive information, II. Methodology because of the higher relevance of ingroup members as well as the cost of negative information processing. Participants The present study is relevant both in a societal as well as in an academic way. The societal rele- Participants were recruited from the participant vance is evident: Prejudice exists as a result of (im- pool of the Psychology department at the Univer- plicit) bias against outgroup members, and gaining siteit van Amsterdam (UvA; N = 56). The sample a better understanding of where this bias comes consisted of 12 males and 42 females; the average from is key in trying to reduce its negative effects age of the sample was 19.3. Participants received such as discriminatory and racist behaviours that credit for participating in the study. The Ethics Re- manifest in diverse societies worldwide. Research view Board of the Faculty of Social and Behavioural has shown that empathy, or sharing others’ expe- Sciences at the UvA reviewed the proposed study riences, can reduce racial bias (Drwecki, Moore, and gave permission to conduct the experiment.

73 Capstone Issue Vol. 12, 2019 Afke van Egmond

Design crash.”. For an example trial of the paradigm, see Appendix 1. All stories that were used (all in Dutch) The present study has a 2 (social are provided in Appendix 2. group: ingroup(0)/outgroup(1)) by 2 (valence: positive(0)/negative(1)) design. Social group and valence are both within-subject variables. The de- Procedure pendent variable is the number of times partici- Before the experiment started, participants pants asked for more information about the person read an information letter and gave informed con- in each picture. This variable is referred to as in- sent. They were given a short instruction, in which formation seeking, and represents the number of they were told that the aim of the study was to times a subject asked for more information in one examine how people responded to emotional in- trial. formation and pictures of certain emotional facial expressions. They received a short explanation of Materials the paradigm and were told that after the paradigm they would have to fill in some questionnaires, fol- The experiment was designed using the online lowing which the paradigm started. Participants survey tool Qualtrics. The pictures of ingroup and were shown a picture of a person displaying a cer- outgroup members displaying emotional facial ex- tain emotional facial expression, along with a short, pressions were selected from the validated ADFES one-sentence description of an emotionally evoca- (van der Schalk et al., 2011). This set contains tive event that this person experienced. Some ex- stills from videos of both male and female models, amples of events that were used are losing a fam- showing one of nine emotions: joy, surprise, pride, ily member, being promoted, getting into an acci- anger, fear, sadness, disgust, contempt, and em- dent, and having a child. The event was related barrassment. In a validation experiment, joy and to the emotion the model expressed in the pic- surprise were rated as positive emotions, and fear ture. After viewing this picture and reading the and sadness were rated as the most negative (van description, participants were asked whether they der Schalk et al., 2011). Therefore, the present wanted to know more about this person’s situa- study chose to use these four emotions in order tion or not. If they did, they were given another to avoid a confounding influence of valence. As sentence describing the person’s situation. Partici- previously mentioned, the models in the images of pants were given this choice eight times; after ev- the ADFES are either North-European or Mediter- ery new sentence they could choose to stop. When ranean; these groups are respectively defined as they stopped, or when they had been given eight ingroup and outgroup for Dutch participants, as val- new additions to the story, participants answered idated by van der Schalk et al. (2011). questions about the emotions the person in the pic- The stories that were presented along with ture was feeling. These questions consisted of a these pictures were situational and emotional (cor- list of nine emotions, for each of which participants responding to the emotion on the picture), and rated how strongly they felt that the model was ex- were written for the purpose of this study. The top- periencing that emotion, on a Likert scale from 1 ics of these stories were counterbalanced across in- (not at all) to 10 (very strong). This marked the end group and outgroup members of the same gender, of one trial; the full paradigm consisted of 16 trials, to ensure that effects could not be driven by par- showing a new picture with a new story each time. ticular combinations between stories and models. Out of the 16 models, eight were ingroup Stereotype-evoking information, such as names or (North-European) and eight were outgroup gendered pronouns, was avoided as much as pos- (Mediterranean). In each set of eight models, four sible during writing. The stories were split into were male and four were female. Because it is pos- nine parts that all carried equal amounts of emo- sible that ingroup/outgroup status interacted with tional information. For example, the following two the gender of the model, in terms of chosen infor- sentences, each carrying a substantial amount of mation, gender was taken into account during the emotional information, were subsequently used: “A data analysis. Each of the four models in all gen- few months ago, my sister suddenly passed away.”; der groups displayed one of the four emotions that “She was travelling, and she got into a big bus were chosen to be used; two with positive valence

74 Capstone Issue Vol. 12, 2019 Afke van Egmond

(joy, surprise), and two with a negative valence the middle value of 4. This already suggested non- (fear, sadness). Subjects were randomly assigned normality. Additionally, normality tests, normal Q-Q to one of two counterbalance conditions so that the plots, as well as a histogram showing the frequency same stories were paired with pictures of ingroup of information seeking, showed that the variable members in one condition and outgroup members information seeking was not normally distributed. in the other, and vice versa. The order in which the Interestingly, the histogram showed peaks at both sixteen trials were presented to participants was extremes (0 and 8). Therefore, it was concluded fully randomised. that GEE was indeed a good fit for this data. After the main experiment was over, partici- In GEE, there are different types of models that pants were asked to report some demographic in- can be used; the present study used a repeated formation such as their age and gender. In addi- measures Poisson log-linear model. This model was tion, they were asked whether they identified as chosen since a Poisson distribution with a log-linear Dutch or non-Dutch, in order to define their ingroup link function is known to best approach the distri- and outgroup positioning. This was done because bution of count data (Garson, 2013). The type of even though the survey was conducted in Dutch, correlation matrix that was assumed was chosen it could not be assumed that all participants iden- by trying out different types of correlation matrices tified as being Dutch. Lastly, they were asked to that SPSS offers in GEE and checking the goodness- fill in the Five-Dimensional Curiosity Scale (5-DC; of-fit of each model. In GEE, the goodness-of-fit Kashdan et al., 2018), the Interpersonal Reactivity is approached by the quasi-likelihood under inde- Index (IRI; Davis, 1983), and the Morbid Curiosity in pendence criterion (QIC). The correlation matrix as- Daily Life scale (Oosterwijk, 2017). These question- sumed in the model with the best fit, and thus naires were included in order to obtain data on the lowest QIC, was used in the final model. When effectiveness and validity of the paradigm itself, by testing these correlation matrices, a robust estima- finding correlations between each of the subscales tor was used, since this is robust against choos- of curiosity and empathy with information seeking, ing the wrong correlation matrix and is generally both positive and negative. used when the correct correlation matrix has not yet been selected. However, after the correct correlation matrix had been identified, a model-based Data Analysis estimator was used. This estimator does assume The data analysis of the experiment was done a correct selection of the correlation matrix, and using IBM’s SPSS statistical analysis programme. works better for smaller sample sizes. Because the dependent variable, information seek- Two models were tested regarding social group: ing, measures count data, normality cannot be as- first of all, all participants were included (N = 56), sumed. The most appropriate analysis technique and secondly, the participants who indicated that for count data in combination with a repeated mea- they identified as non-Dutch were filtered out so sures design is a generalised estimating equations that N = 46, in order to examine whether these (GEE) model. GEE is an extension of generalized participants being included in the first model influ- linear models (GLM) and is used to model the av- enced the results. In the second model, the sig- erage response over a (sub)population or cluster, nificance of all three effects that were tested for as opposed to subject-specific responses (Garson, did not change; excluding participants who identi- 2013). It allows for non-normality of the dependent fied as non-Dutch did not change the results signifi- variable, as well as correlated data. In addition, cantly. Since van der Schalk et al. (2011) used only GEE allows for more complex data analysis than participants that identified as Dutch to validate the GLM, such as repeated measures. In order to verify difference between ingroup and outgroup members whether GEE was indeed the best model to use for in the pictures in the ADFES, the present study de- this data, the normality of the dependent variable cided to use this same assumption and exclude all (information seeking) was tested. Measurements participants that identified as non-Dutch. existed over the full range of information seeking, This final model was used to test for the main its minimum and maximum at the most extreme effects of both social group and valence, and an in- values of 0 and 8 respectively. Its overall mean teraction effect of social group and valence on in- was revealed to be 2.74, which is relatively far from formation seeking. In GEE, Wald chi-square (W 2) is

75 Capstone Issue Vol. 12, 2019 Afke van Egmond

used as a test statistic; the bigger Wald chi-square, group and outgroup members was bigger for nega- the stronger the effect. For significance, an alpha tive than for positive information (H3). Lastly, the of 0.05 was used. present study expects correlations (positive, nega- Besides analysing data obtained by the experi- tive, or both) between information seeking, social ment, the questionnaires that were included were curiosity, and morbid curiosity. Additionally, a cor- used to examine correlations between different relation (again positive, negative, or both) of sub- personality traits and positive and negative infor- scales of empathy with information seeking, is ex- mation seeking. The 5-DC calculates the five di- pected. mensions of curiosity: joyous exploration, depriva- tion sensitivity, stress tolerance, social curiosity, and thrill seeking. In the present paper, only cor- III. Results relations with social curiosity are reported; Cron- bach’s alpha for the social curiosity scale was α = Social Group 0.787. Correlations with the IRI are reported for all Contrary to H1, the analysis showed that social its four subscales of empathy: fantasy (α = 0.789), group did not have a significant effect on informa- empathic concern (α = 0.702), perspective taking tion seeking (W 2 = 1.837; p = 0.175). (α = 0.556), and personal distress (α = 0.777). The Morbid Curiosity in Daily Life is a single scale, which Valence had a Cronbach’s alpha of α = 0.601; this scale was also correlated with positive and negative informa- As predicted, the analysis demonstrated a sig- tion seeking. Lastly, in order to establish the ef- nificant effect of valence (W 2 = 13.107; p = 0.000). fectiveness of the paradigm, some statistics about However, in contrast to H2, participants wanted to the way participants responded across trials were know more about negative (M = 2.99; SE = 1.181) retrieved. than positive (M = 2.64; SE = 0.170) life stories.

Hypotheses Social Group*Valence

The present study expected that the main effect Congruent with H3, an interaction between va- of social group on information seeking would be lence and social group on information seeking was significant and negative, meaning that participants found (W 2 = 6.800; p = 0.009). Pairwise com- would want to know more about ingroup members parisons, shown in Table 1, revealed that for pos- than outgroup members (H1). The main effect of itive information about ingroup members, informa- valence on information seeking was expected to be tion seeking was significantly lower than all other significant and positive. In other words, the present combinations of social group and valence. The study expected that participants would want more other groups did not significantly differ among each information about positive than negative stories other. These significant relationships are presented (H2). Additionally, a significant interaction between in Figure 1 by use of asterisks. In conclusion, like social group and valence on information seeking Hypothesis 3 predicted, social group and valence was expected; the present study hypothesised that do interact, but not in the way that was hypothe- the difference between information seeking for in- sised.

Correlations but there was a significant positive correlation between negative information seeking and social curiosity (r = 0.278; p = 0.040). In other words, In order to find out whether the paradigm is the higher the participants scored on social curios- a good measure for social curiosity, participants’ ity, the higher their scores for negative information scores on the 5-DC social curiosity subscale (Kash- seeking were. dan et al., 2018) were correlated with the positive and negative information seeking score. Pos- In exploratory fashion, the four subscales of the itive information seeking was not significantly cor- IRI were correlated with positive and negative infor- related with social curiosity (r = 0.161; p = 0.240), mation seeking. The IRI’s subscale of perspective

76 Capstone Issue Vol. 12, 2019 Afke van Egmond

Table 1.

Figure 1. Information seeking for Valence*Social group.

taking was positively correlated with positive infor- Paradigm mation seeking (r = 0.362, p = 0.035), but not with negative information seeking (r = 0.199, p = 0.154). To establish the effectiveness of the paradigm The three other subscales did not signiﬁcantly cor- that the present study used, the way participants relate with either positive or negative information responded across trials was examined. All partici- seeking (all p > 0.05). pants (N = 56) were taken into account during this analysis. There were three participants who chose Lastly, the Morbid Curiosity in Daily Life scale the maximum available information for all sixteen was correlated with positive and negative informa- trials; four participants never chose to get more in- tion seeking, but no signiﬁcant correlations were formation in any of the sixteen trials. The rest of found (r = 0.064, p = 0.642; r = 0.055, p = 0.696, the participants’ (N = 49) scores for information respectively). seeking varied across trials. From this it seems that

77 Capstone Issue Vol. 12, 2019 Afke van Egmond

the majority of the participants reacted differently cause of its higher cost. In fact, the paradigm cre- to different pictures and stories, which is congruent ated for the present study may be better at measur- with the intention of the paradigm. ing morbid curiosity than paradigm currently used by Oosterwijk (2017). In that paradigm, participants are provided with a description of a situation IV. Discussion and Conclusion and are asked whether they want to see the picture that belongs to the description. In this case, they The results of the statistical analyses were com- have to choose between looking at a picture or a pared to the hypotheses. First of all, H1 stated white screen, both of which take the same amount that a negative effect of social group on informa- of time. In the case of negative emotional pic- tion seeking was expected. This effect was not tures, it makes sense that participants sometimes found, meaning that there was no significant dif- choose not to look at this picture, and instead look ference in the number of times participants asked at a white screen. However, in the case of positive for more information about ingroup members than emotional pictures, participants may prefer looking outgroup members. This might be explained by at a picture rather than at a white screen for sev- demand characteristics playing a role in the cur- eral seconds, and thus their default response may rent experiment; at least some participants might be to say yes. In other words, when participants have guessed the true purpose of the study, which have to invest the same amount of time into either might have affected their responses. Additionally, choice, even when they are not particularly inter- gender might have been a confounding variable ested in the picture, it may still be preferable to see in this effect. The present study studied ingroup the picture over an empty screen, but only when and outgroup in terms of (appeared) ethnicity; how- the information is positive. As a result, research ever, one’s ingroup and outgroup may also be de- with this paradigm has found that people choose fined by gender. Since the present sample was 78% positive information more often than negative in- female, but the trials were 50% female, the two- formation (Oosterwijk et al., in prep.). In the cur- dimensionality of social group and the fact that the rent paradigm, however, choosing not to get more sample was predominantly female together may information saves participants time. Therefore, the have confounded the results and lead to a dimin- default choice may be not to get additional informa- ished effect of social group on information seek- tion. Increased information seeking for negative in- ing. Lastly, since the present sample consisted of formation as opposed to positive information in the psychology students, and thus they had all stud- present study happened against the cost of being ied psychological concepts such as stereotypes and emotionally disturbed by the negative information, outgroup bias, they might have been more aware and against the cost of time investment necessary of their bias. Awareness of outgroup bias has been if participants wanted to know more. In this sense, shown to reduce the bias itself (Kawakami, Dion, & the presently developed paradigm may be a bet- Dovidio, 1998; Pope, Price, & Wolfers, 2018); there- ter measure of morbid curiosity than the currently fore, the results of the present study might be bi- used paradigm. ased by these characteristics of the sample. However, it should be noted that no correlations A significant effect of valence on social group were found between information seeking, positive was found, however, not in the form that H2 pre- or negative, and the Morbid Curiosity in Daily Life dicted. Even though processing negative infor- scale. This might be related to the scale’s relatively mation is more costly, participants in the present low Cronbach’s alpha (α = 0.601); the scale may study chose to get more information in trials con- not have properly measured the concept of mor- taining negative emotional pictures and stories bid curiosity in the present study. It is also good than in positive trials. This effect may be explained to note that the stimuli used in both the present by morbid curiosity - the seeking of negative in- paradigm and the paradigm for morbid curiosity formation because of its associated feeling of sen- used by Oosterwijk (2017) are very different. The sation (Zuckerman, 1990) or higher informational present study used negative life stories about peo- value (Baumeister et al., 2001; Oosterwijk, 2017). ple, which contained different information than the This curiosity seems to be stronger than the ten- pictures shown in the paradigm used by Oosterwijk dency to avoid processing negative information be- (2017), which portray scenes involving violence

78 Capstone Issue Vol. 12, 2019 Afke van Egmond

and death. Therefore, the tendency to seek neg- learn about outgroup members, and they wanted ative information found in the present study may to know more about them regardless of the valence not reflect the same concept as measured in the of the information they received. This might mean morbid curiosity paradigm. Although the construct that they are not biased against outgroup mem- that the present study measured (which we may bers. However, this conclusion cannot be drawn call negative social curiosity) is likely to be related with certainty, since the present study did not mea- to morbid curiosity, it is also different, since the sure participants’ bias. negativity of visual scenes involving violence and Future research should include a measure of death cannot be compared to the negative valence outgroup bias in order to establish whether learn- of the life stories of the present study. Therefore ing about outgroup members as much as ingroup it is concluded that negative social curiosity should members is correlated with a reduced outgroup not be measured using a morbid curiosity scale. It bias. In addition, studies have shown that empa- is good to take this distinction into account in case thy can reduce racial bias (Cameron et al., 2016; the present paradigm is used in future research. Drwecki et al., 2011; Sheng & Han, 2012); future Even though no main effect of social group in in- research should try to determine whether learn- formation seeking was found, there was an interac- ing more about an outgroup member also changes tion effect between social group and valence on in- one’s view of that person and possibly reduces formation seeking, which agrees with H3. However, bias. For example, the paradigm could be adapted the specific interaction that was found was differ- to include personal judgement scales of the per- ent than the one H3 had predicted; participants son whose picture was seen. Then, data could be wanted to know the least in trials that contained analysed to find out whether learning more about positive information about ingroup members, than someone changes one’s judgement of that person. in all other trials with other combinations of social Another interesting way to look at the effect of group and valence. This might be explained by the learning more about ingroup and outgroup mem- fact that curiosity is often elicited by novelty and bers is to utilize an implicit bias task about people uncertainty (Kashdan & Silvia, 2009; Kashdan et that participants did and did not learn about. Al- al., 2018; Litman, 2005). Ingroup members are who ternatively, a memory task could be used to find we know most about and thus are least uncertain out how information learned about ingroup and out- about, since they are part of our own social group group members is remembered. If participants re- (Zebrowitz, Bronstad, & Lee, 2007). Conversely, we member information about outgroup members bet- are more uncertain about outgroup members, and ter, this may indicate that they do process this in- specifically positive information about them may formation differently, which could give an insight be the most novel, since we are often biased to into the underlying mechanism of outgroup bias. think negatively about outgroup members (Masuda Adding to knowledge about what causes and what & Fu, 2015; Neuberg et al., 2000; Tajfel & Turner, reduces outgroup bias is of tremendous importance 1979; Van Vugt & Schaller, 2008). This uncertainty in attempting to change (often implicit) racist per- and novelty regarding outgroup members and es- spectives adopted by people in many diverse soci- pecially positive information about them may ex- eties of today. plain the present findings. Besides examining the effects of social group The lack of a main effect of social group on and valence on information seeking, several scales information seeking does not explain where out- were included in order to examine whether the con- group bias comes from; the expected finding that cepts that these scales measured were correlated participants wanted to know less about outgroup with information seeking in this paradigm. The so- members than ingroup members could have sug- cial curiosity subscale of the 5-DC measures the gested that people are biased against learning new tendency to acquire information about the social information about people in different social groups environment; people who score high on this scale than our own. On the contrary, the present study have a strong interest in other people’s thoughts found no difference in information seeking about in- and actions. It was expected that social curiosity group and outgroup members. This finding poses was correlated with both positive and negative in- a positive outlook on problems of outgroup bias: formation seeking; people who score high on the The participants of the present study did want to personality trait social curiosity, should choose to

79 Capstone Issue Vol. 12, 2019 Afke van Egmond

know more about other people’s life stories. How- and personal distress) did not significantly corre- ever, there was only a significant correlation be- late with information seeking, positive or negative, tween the social curiosity subscale of the 5-DC and which suggests that participants’ behaviour in the negative information seeking; the higher partici- present paradigm does not reflect these aspects of pants scored on social curiosity, the more often empathy. they wanted to know more about other people’s In conclusion, the present study is an ex- negative life events, but this relationship was not ploratory one, and developed an innovative found for positive information. It is worth mention- paradigm in order to establish a relationship being that Kashdan et al., (2018), in their study val- tween information seeking, social group, valence idating the 5-DC, found that their subscale of so- and empathy. It was found that participants pre- cial curiosity correlated strongly with a tendency to ferred negative emotional information over posi- gossip. Hartung and Renner (2013) state that so- tive, and that they were least curious of positive cial curiosity and gossip are related concepts, but information about ingroup members. Because of that social curiosity serves a function of gathering the exploratory nature of the present study, it is information about the social environment, whereas important that it is replicated in order to verify that gossiping is more sensational, and done for enter- similar results are found across studies. If they tainment purposes. Since Kashdan et al. (2018) are, these results greatly add to current knowledge found a correlation between their social curiosity in the field of each of these concepts and open subscale and a tendency to gossip, it might be that up many new directions for future research. Fu- this subscale is in fact a better measure of the con- ture studies can elaborate on the results found in struct of gossip than of social curiosity. The fact the present study, in order to more strongly es- that gossip often involves discussing negative in- tablish how people seek information and how their formation about others (Feinberg, Willer, Stellar, & way of seeking information changes their knowl- Keltner, 2012) could then explain why the present edge about the world and opinions about others. study found a correlation between the social curios- Relating this knowledge to outgroup bias will be in- ity subscale of the 5-DC and negative information credibly important in reducing this bias and thus seeking. reduce the racism and discrimination that come as Alongside the 5-DC, the IRI was included in its result. the present experiment. The IRI consists of items Works Cited measuring four distinct aspects of empathy: fantasy, perspective taking, empathic concern, and Aboud, F. E. (1976). Self-evaluation: Informa- personal distress. A correlation was found with tion seeking strategies for interethnic so- the subscale of perspective taking and positive in- cial comparisons. Journal of Cross-Cultural formation seeking, but not with negative informa- Psychology, 7(3), 289–300. tion seeking. This means that people who score Avenanti, A., Bueti, D., Galati, G., & Aglioti, S. M. high on perspective taking, or in other words often (2005). Transcranial magnetic stimulation spontaneously adopt other people’s point of view highlights the sensorimotor side of empa- (Davis, 1983), had a tendency to ask for more pos- thy for pain. Nature Neuroscience, 8(7), itive information, but not necessarily negative in- 955– 960. https://doi.org/10.1038/nn1481 formation. It can be said that because these peo- Batson, C. D. (2013). These things called empathy: ple spontaneously experience someone else’s per- Eight related but distinct phenomena. In J. spective, they enjoy being provided with positive Decety & W. Ickes (Eds.), The Social Neu- information, since they then empathise with and roscience of Empathy. Cambridge, MA, US: therefore experience the positive emotions asso- MIT Press. ciated with the information (Batson, 2013). How- Baumeister, R. F., Bratslavsky, E., Finkenauer, C., ever, one has to be careful with drawing this con- & Vohs, K. D. (2001). Bad is stronger than clusion, since Cronbach’s alpha of the perspective good. Review of General Psychology, 5(4), taking subscale was low (α = 0.556), and therefore 323–370. https://doi.org/10.1037/1089- this subscale may not have accurately measured 2680.5.4.323 this construct in the present study. The other three Cameron, C. D., Cunningham, W., Saunders, B., & subscales of empathy (fantasy, emotional concern Inzlicht, M. (2018). The ends of empathy:

80 Capstone Issue Vol. 12, 2019 Afke van Egmond

Constructing empathy from value-based chology (pp. 357–411). choice. Psyarxiv.Com. Retrieved from Gallese, V., & Goldman, A. (2002). Mirror neurons https://psyarxiv.com/d99bp/download/?format=pdf and the simulation theory of mind-reading. Cameron, C. D., Hutcherson, C. A., Ferguson, Trends in Cognitive Sciences, 2(12), A. M., Scheffer, J. A., Hadjiandreou, E., 493–501. https://doi.org/10.1016/s1364- & Inzlicht, M. (2016). Empathy is hard 6613(98)01262-5 work: People choose to avoid empathy be- Garson, G. D. (2013). Generalized linear mod- cause of its cognitive costs. Psyarxiv.Com. els/generalized estimating equations. Sta- https://doi.org/10.31234/osf.io/jkc4n tistical Associates Publishing. Cikara, M., & Fiske, S. T. (2011). Bounded Gottlieb, J., Oudeyer, P. Y., Lopes, M., & Baranes, empathy: Neural responses to outgroup A. (2013). Information-seeking, cu- targets’ (mis)fortunes. Journal of Cog- riosity, and attention: Computational nitive Neuroscience, 23(12), 3791–3803. and neural mechanisms. Trends in https://doi.org/10.1162/jocn_a_00069 Cognitive Sciences, 17(11), 585–593. Cikara, M., & Van Bavel, J. J. (2014). The neu- https://doi.org/10.1016/j.tics.2013.09.001 roscience of intergroup relations: An Harris, L. T., & Fiske, S. T. (2006). Dehumanizing the integrative review. Perspectives on lowest of the low: Neuroimaging responses Psychological Science, 9(3), 245–274. to extreme out-groups. Psychological Sci- https://doi.org/10.1177/1745691614527 ence, 17(10), 847–853. Retrieved from 464 https://journals.sagepub.com/doi/abs/10.1111/j.14 Davis, M. H. (1983). Measuring individual differ- 67-9280.2006.01793.x?casa_token=qnpQXh6am_ ences in empathy: Evidence for a mul- 0AAAAA:8zXXQQzT9r0A81-fin9BoocTdOMu- tidimensional approach. Journal of Per- LBTP6j c7dziyXZ33Bajb5sUEOyNxH59OF0CCc5C8rGiPOoQ sonality and Social Psychology, 44(1), bw 113–126. https://doi.org/10.1037/0022- Harrison, N. A., Morgan, R., & Critchley, H. D. 3514.44.1.113 (2010). From facial mimicry to emo- de Waal, F. B. M. (2008). Putting the altru- tional empathy: A role for norepinephrine? ism back into altruism: The evolution Social Neuroscience, 5(4), 393–400. of empathy. Annual Review of Psy- https://doi.org/10.1080/17470911003656330 chology, 59, 279–300. Retrieved from Hartung, F. M., & Renner, B. (2013). Social curiosity https://www.annualreviews.org/doi/abs/10.1146/an and gossip: Related but different drives nurev.psych.59.103006.093625 of social functioning. PLoS ONE, 8(7). Decety, J., Yang, C. Y., & Cheng, Y. (2010). Physi- https://doi.org/10.1371/journal.pone.0069996 cians down-regulate their pain empathy Hatfield, E., Carpenter, M., & Rapson, R. L. response: An event-related brain poten- (2014). Emotional contagion as a tial study. NeuroImage, 50(4), 1676–1682. precursor to collective emotions. In https://doi.org/10.1016/j.neuroimage.2010.01.025 Collective Emotions (pp. 108–122). Drwecki, B. B., Moore, C. F., Ward, S. E., https://doi.org/10.1093/acprof:oso/978019965918 & Prkachin, K. M. (2011). Reduc- 0.003.0008 ing racial disparities in pain treatment: Hatfield, E., Rapson, R. L., & Le, Y.-C. L. The role of empathy and perspective- (2013). Emotional contagion and em- taking. Pain, 152(5), 1001–1006. pathy. In J. Decety & W. Ickes (Eds.), https://doi.org/10.1016/j.pain.2010.12.005 The Social Neuroscience of Empa- Feinberg, M., Willer, R., Stellar, J., & Keltner, thy. Cambridge, MA, US: MIT Press. D. (2012). The virtues of gossip: Rep- https://doi.org/10.7551/mitpress/9780262012973. utational information sharing as proso- 003.0003 cial behavior. Journal of Personality and Haviland, J. M., & Lelwica, M. (1987). The in- Social Psychology, 102(5), 1015–1030. duced affect response: 10-week-old in- https://doi.org/10.1037/a0026650 fants’ responses to three emotion expres- Fiske, S. T. (1998). Stereotyping, prejudice and dis- sions. Developmental Psychology, 23(1), crimination. In The handbook of social psy- 97–104. https://doi.org/10.1037/0012-

81 Capstone Issue Vol. 12, 2019 Afke van Egmond

1649.23.1.97 Nelson, J. A. (2014). The power of stereotyp- Inzlicht, M., Gutsell, J. N., & Legault, L. ing and confirmation bias to overwhelm (2012). Mimicry reduces racial prej- accurate assessment: the case of eco- udice. Journal of Experimental So- nomics, gender, and risk aversion. Journal cial Psychology, 48(1), 361–365. of Economic Methodology, 21(3), 211–231. https://doi.org/10.1016/j.jesp.2011.06.007 https://doi.org/10.1080/1350178X.2014.939691 Kashdan, T. B., & Silvia, P. J. (2009). Curiosity and Neuberg, S. L., Smith, D. M., & Asher, T. (2000). interest: The benefits of thriving on nov- Why people stigmatize: Toward a biocul- elty and challenge. In S. J. Lopez & C. tural framework. In The social psychology R. Snyder (Eds.), The Oxford Handbook of stigma. of Positive Psychology (pp. 367–374). Oosterwijk, S. (2017). Choosing the nega- New York: Oxford University Press. tive: A behavioral demonstration of mor- https://doi.org/10.1093/oxfordhb/9780195187243. bid curiosity. PLoS ONE, 12(7), 1–20. 013.0034 https://doi.org/10.1371/journal.pone.0178399 Kashdan, T. B., Stiksma, M. C., Disabato, D. J., Oosterwijk, S., Snoek, L., Te Koppele, J., Engelbert, McKnight, P. E., Bekier, J., Kaji, J., & L., & Scholte, H. S. (in prep.). Choosing Lazarus, R. (2018). The five-dimensional to view morbid stimuli involves reward cir- curiosity scale: Capturing the bandwidth cuitry. of curiosity and identifying four unique Pope, D. G., Price, J., & Wolfers, J. (2018). Aware- subgroups of curious people. Journal ness reduces racial bias. Management Sci- of Research in Personality, 73, 130–149. ence, 64(11), 4988–4995. https://doi.org/10.1016/j.jrp.2017.11.011 Renner, B. (2006). Curiosity about people: Kawakami, K., Dion, K. L., & Dovidio, J. F. The development of a social curios- (1998). Racial prejudice and stereo- ity measure in adults. Journal of Per- type activation. Personality and So- sonality Assessment, 87(3), 305–316. cial Psychology Bulletin, 24(4), 407–416. https://doi.org/10.1207/s15327752jpa8703 https://doi.org/10.1177/0146167298244007 Sheng, F., & Han, S. (2012). Manipulations Keysers, C., & Gazzola, V. (2014). Dissociating the of cognitive strategies and intergroup ability and propensity for empathy. Trends relationships reduce the racial bias in in Cognitive Sciences, 18(4), 163–166. empathic neural responses. NeuroIm- https://doi.org/10.1016/j.tics.2013.12.011 age, 61, 786–797. Retrieved from Litman, J. A. (2005). Curiosity and the https://www.sciencedirect.com/science/article/pii/S1 pleasures of learning: Wanting and 053811912004247 liking new information. Cogni- Stephan, W. G., & Finlay, K. (1999). The role of tion and Emotion , 19(6), 793–814. empathy in improving intergroup relations. https://doi.org/10.1080/02699930541000101 Journal of Social Issues, 55(4), 729–743. Loewenstein, G. (1994). The psychol- https://doi.org/10.1111/0022-4537.00144 ogy of curiosity: A review and Tajfel, H. (1970). Experiments in in- reinterpretation. Psychological Bul- tergroup discrimination. Scien- letin. https://doi.org/10.1037//0033- tific American, 223(5), 96–102. 2909.116.1.75 https://doi.org/10.1038/scientificamerican1170- Masuda, N., & Fu, F. (2015). Evolution- 96 ary models of in-group favoritism. Tajfel, H., & Turner, J. C. (1979). An in- F1000Prime Reports, 7(March), 1–12. tegrative theory of intergroup conflict. https://doi.org/10.12703/P7-27 In The Social Psychology of Intergroup Meffert, H., Gazzola, V., Den Boer, J. A., Bar- Relations. https://doi.org/10.1016/S0065- tels, A. A. J., & Keysers, C. (2013). Re- 2601(05)37005-5 duced spontaneous but relatively normal Taylor, S. (1991). Asymmetrical effects of pos- deliberate vicarious representations in psy- itive and negative events. Psycholog- chopathy. Brain, 136(8), 2550–2562. ical Bulletin, 110(l), 67–85. Retrieved https://doi.org/10.1093/brain/awt190 from https://taylorlab.psych.ucla.edu/wp-

82 Capstone Issue Vol. 12, 2019 Afke van Egmond

content/uploads/sites/5/2014/10/1991_Asymmetric sonality and Social Psychology, 73(4), al-Effects_Positive«_Negative-Events_Mobilization- 839–848. https://doi.org/10.1037/0022- Minimization-Hypothesis.pdf 3514.73.4.839 van der Schalk, J., Hawk, S. T., Fischer, A. Zaki, J. (2014). Empathy: A motivated account. H., & Doosje, B. (2011). Moving Psychological Bulletin, 140(6), 1608–1647. faces, looking places: Validation of the https://doi.org/10.1037/a0037679 Amsterdam Dynamic Facial Expression Set (ADFES). Emotion, 11(4), 907–920. Zebrowitz, L. A., Bronstad, P. M., & Lee, H. K. https://doi.org/10.1037/a0023853 (2007). The contribution of face famil- Van Vugt, M., & Schaller, M. (2008). Evolution- iarity to ingroup favoritism and stereo- ary approaches to group dynamics: An in- typing. Social Cognition, 25(2), 306–338. troduction. Group Dynamics, 12(1), 1–6. https://doi.org/10.1521/soco.2007.25.2.306 https://doi.org/10.1037/1089-2699.12.1.1 Zuckerman, M. (1990). The psychophysiology Weary, G., & Jacobson, J. A. (1997). Causal of sensation seeking. Journal of Per- uncertainty beliefs and diagnostic in- sonality. https://doi.org/10.1111/j.1467- formation seeking. Journal of Per- 6494.1990.tb00918.x

Appendix A

Example of the stimuli material used.

Een half jaar geleden ben ik ernstig gewond geraakt in een verkeersongeluk. Wil je meer weten?

• Ja

• Nee

Additional lines:

1. Het was ochtend, het regende, en ik zat op de ﬁets onderweg naar mijn werk.

2. Op een kruising heeft een auto die door rood reed me geschept.

3. Ik kan me nog goed herinneren hoe de auto eruit zag: het was een blauwe Opel.

4. Het ging heel erg snel, ik realiseerde het me pas toen ik op de grond lag.

83 Capstone Issue Vol. 12, 2019 Afke van Egmond

5. Er stond meteen een hele groep mensen om me heen die heel bezorgd keken.

6. Ik voelde eerst even helemaal niks, maar na een paar minuten kwam de pijn.

7. Ik kon niet meer praten of bewegen, ik was totaal in shock.

8. Uiteindelijk heeft een ambulance me meegenomen naar het ziekenhuis.

84 Capstone Issue Vol. 12, 2019 Afke van Egmond

Appendix B

Emotional stories written for and used in the present study (in Dutch only).

85 Capstone Issue Vol. 12, 2019 Afke van Egmond

86 Social Sciences

The Political Strategy of Prawo i Sprawiedliwos´c´

Christina Keßler

Supervisor Dr. Carlos Reijnen (UvA) Reader Dr. Christian Noack (UvA) Capstone Issue Vol. 12, 2019 Christina Keßler

Abstract

This capstone analyses the political strategy and party behaviour of Prawo i Sprawiedliwosć´ (PiS), the current governing party of Poland. It aims to explain how PiS managed to gain electoral support in the parliamentary elections of 2015 and how PiS continues to appeal to the public despite various acts which are deemed highly controversial. This work draws from scholarly theory concerning both the topics of political strategies of parties and the ongoing process of so-called democratic backsliding in Poland. Inspired by Strömbäck’s (2007) four arena model, this capstone is divided into two sections. In the first part, this research project will use textual analysis of a press conference given by leading PiS politicians in Septem- ber 2017 to illustrate the narrative the party tries to convey to its electorate. The second part analyses concrete acts of the party in order to shed light on PiS’s political strategy. The findings demonstrate that PiS successfully employs a double strategy of reaching out to new voters by softening their image whilst at the same time satisfying its more radical core electorate. Furthermore, findings show that the party makes use of pre-existing narratives concerning ‘political issues of the day’, most notably during the 2015 refugee crisis.

Keywords and phrases: Poland, illiberalism, party strategy, democratic backsliding, PiS

I. Introduction deed been nominated illegally, but also clearly stated that three out of the five appointments were Scholars are of the opinion that over the last constitutional (Szczerbiak, 2017). Nevertheless, few years, countries in Central and Eastern Eu- the new government swore in five different judges rope (CEE), most notably Poland and Hungary, have who had been nominated by the new parliament been transforming into so-called ‘illiberal democra- (Szczerbiak, 2017). However, the then-President cies’ (Cianetti et al., 2018; Rupnik, 2018). Research of the Constitutional Tribunal, Andrzej Rzeplinski,´ seeking to explain this phenomenon has largely only allowed two of the new judges to assume their focused on structural explanations, such as the duties (Szczerbiak, 2017). Trying to find a way larger historical and sociological context in which around this impasse, the PiS government passed this process, oftentimes referred to as ‘democratic a law which both increased the number of judges backsliding’, occurs (Rupnik, 2018). This research required to make rulings and made it obligatory for project takes a different approach, focusing not the Tribunal to consider cases not in order of discre- on structural factors, but rather aiming to gain a tion but in the order in which those cases have been comprehensive understanding of the political strat- received (Szczerbiak, 2017). When the Constitu- egy employed by the Polish political party Prawo tional Tribunal declared this law unconstitutional, i Sprawiedliwosć´ (Law and Justice; PiS). PiS, which the PiS government refused to publish the judge- has been governing Poland since 2015, is the driv- ment in the official journal, which is a necessary ing force behind the country’s transition into an il- step for rulings to become binding, and declared liberal democracy (Cianetti et al., 2018; Grzymala- that the Tribunal had no constitutional power to re- Busse, 2018; Rupnik, 2018). It is therefore ques- view the law (Szczerbiak, 2017). This drew sharp tionable how PiS continues to appeal to the public criticism from various actors, such as the Venice despite a variety of acts that undermine institutions Commission1 which concluded that the actions of of liberal democracy and are perceived as highly the PiS government endangered the rule of law as controversial. well as the democratic system of Poland (Rupnik, A prime example of such actions is the PiS 2018; Szczerbiak, 2017). party’s restrictions of the Constitutional Tribunal, When analysing the ongoing events in Poland, Poland’s constitutional court. After taking office in many scholars use the aforementioned term demo- November 2015, the new PiS government annulled cratic backsliding (Cianetti et al., 2018; Przybylski, the appointment of five judges to the Tribunal who 2018). This phenomenon is not unique to Poland - had been nominated by the previous parliament indeed, various academics are of the opinion that (Szczerbiak, 2017). The Constitutional Tribunal we are currently witnessing an erosion of democ- subsequently ruled that two of the judges had in- racy all over the Central and Eastern Europe region

1The Venice Commission is an advisory body of the Council of Europe that deals with matters of constitutional law.

88 Capstone Issue Vol. 12, 2019 Christina Keßler

(Cianetti et al., 2018). and Hungary can be seen as illiberal democra- The goal of this research project is to examine cies remains largely undisputed and is even em- the situation of Poland in the context of democratic braced by leading political figures such as Hungar- backsliding, by focusing on the political strategy ian Prime Minister Victor Orbán, who publicly de- employed by PiS. Specifically, this research aims to clared himself in favour of an ‘illiberal state’ (Rup- answer the question of how PiS managed to gain nik, 2018). The concept of illiberal democracy can electoral support and what strategies it employs to be defined as a regime which falls somewhere be- keep this support steady. Most importantly, this tween democratic and nondemocratic, and which capstone questions the kind of narrative PiS aims combines both democratic and illiberal institutions to create about its role in contemporary Poland. and practices (O’Neil, 2015). In illiberal regimes, Further questions address the concrete actions of the rule of law might be in place, but institutions the party, such as policies implemented by PiS. resting on the rule of law (as well as the rule of This capstone contributes to the academic discus- law itself) are weakly institutionalized and not well- sion concerning rising illiberalism in the Central and respected (O’Neil, 2015). The Polish turn towards Eastern European region and offers a new perspec- illiberalism is a very recent development and repre- tive by focusing less on structural explanations and sents one of the most extreme cases of deteriorat- more on contemporary political party strategies. ing democracy in the region (Cianetti et al., 2018). Apart from its academic pertinence, the implica- In order to fully comprehend this, one must be fa- tions of this research are of societal and political miliar with the historical background of the country relevance as they contribute to our conception of and of Central and Eastern Europe. ongoing political processes in Poland by furthering Poland is the largest state in the region and our understanding of the success of PiS. Moreover, is considered to be especially important from a while the rise of illiberal democracy is not a dis- geopolitical perspective (Cianetti et al., 2018). For- tinctly European phenomenon, the research of its merly partitioned by Germany, Austria, and Russia, European instances can contribute to a general un- Poland was occupied by Nazi Germany and Soviet derstanding of this phenomenon. Russia during the Second World War and belonged to the Soviet Bloc until 1989 (Mucha, 2006). The agricultural country has been linked to Eastern Eu- II. Literature Review rope through its internal social structure, its economy, and its politics, while it is connected to West- The notion of democratic backsliding and what ern Europe by its Latin Christianity, its alphabet, it constitutes is a question of ongoing debate (Wald- and its cultural aspirations (Mucha, 2006). Poland ner & Lust, 2018). The term implies incremental in- proved more resistant to communism than other stead of radical change; “a process related to yet countries in the region, a fact that has been at- still distinct from reversion to autocracy” (Waldner tributed to its culture, which Rupnik calls a “com- & Lust, 2018, p. 94). According to Waldner and bination of nationalism and Catholicism” (Rupnik, Lust (2018), backsliding always refers to a dete- 2018, p. 28). The country’s democratic transition rioration of some sort, while democratic backslid- after the collapse of the communist regime in 1989 ing specifically is defined as “a decline in the qual- proved a challenging task (Przybylski, 2018). Infla- ity of democracy” (p. 94). Over recent years, tion rose to 640 percent and the newly introduced the scholarly consensus seems to support the no- reforms left many people losing their wealth, their tion that we are currently witnessing a process of social status, and their trust in the new institutional democratic backsliding in the CEE region (Cianetti arrangements (Kowalewski & Rybinski, 2011; Przy- et al., 2018; Grzymala-Busse, 2018; Rupnik, 2018). bylski, 2018). In the following years, Poland un- Cianetti et al. (2018), however, problematized this derwent ambitious reforms (Przybylski, 2018), and view, bringing forth various points of criticism, key was the first among the countries in the CEE re- among them being that calling the currently on- gion to recover from the sharp economic decline going process ‘democratic backsliding’ implies a following the transition of 1989 (Kowalewski & Ry- preceding successful democratisation. While the binski, 2011), experiencing a period of continuous concept of democratic backsliding in CEE remains economic growth from 1992 onwards (Kowalewski controversial, the notion that contemporary Poland & Rybinski, 2011; Mucha, 2006; Przybylski, 2018).

89 Capstone Issue Vol. 12, 2019 Christina Keßler

In 1997, Poland became a member of NATO and the paign 2015, such as weakness of the opposition, country joined the European Union in 2004 (Mucha, the ‘Poland in ruins’ narrative employed by PiS, the 2006; Przybylski, 2018). political strategy of softening the party’s image, Due to Poland’s economic development as well campaign promises, and a focus on identity poli- as the fact that it was the only EU member state tics during a time in which the topic of migration to survive the 2008 financial crisis without a re- dominated public discourse (Fomina & Kucharczyk, cession, Poland has been heralded as an exam- 2016). In combination, these factors contributed ple for successful post-communist democratic re- to the electoral success of PiS in the parliamentary forms (Przybylski, 2018; Rupnik, 2018). However, election in 2015. the recent changes (both legislative and institu- Party behaviour and electoral strategy have tional) brought about by PiS have evoked fears long been a focus of scholarly research (Brysk, that the accomplishments of the successful demo- 1995; De Vries & Hobolt, 2012; De Sio et al., 2018; cratic reform might be reversed (Przybylski, 2018). Somer-Topcu, 2015). While the research presented Poland’s path to illiberalism and the role of PiS in the following paragraphs is not directly linked to in this process is traced by Anna Grzymala-Busse the Polish context, it provides significant informa- (2018), who claims that contemporary Poland must tion about political strategy on a general level. In be defined as an illiberal democracy, because it this research, these theories on political strategy remains a democracy in which the critical institu- will be synthesized and applied to the Polish con- tions of liberal democracy such as the courts and text in order to further understand the party be- the public media are no longer able to fulfil their haviour of PiS. When thinking of a political strat- purpose. This argument aligns with the definition egy, we oftentimes think first and foremost of cam- of illiberal regimes as outlined by O’Neil (2015). paigning and elections. While elections are an im- Seeking to understand the surge of illiberalism in portant part of party competition, they are by no the CEE region, Rupnik (2018) points out the de- means the only one. Strömbäck (2007) uses the coupling of liberalism and its focus on individual ‘four arena model’ in order to conceptualize the dif- freedom from democracy. While democracy has ferent areas in which political parties compete: an historical antecedents in Central and Eastern Eu- electoral arena, a parliamentary arena, a media rope, liberalism is understood as a foreign concept arena, and an internal arena (Strömbäck, 2007). imported to the region after 1989 (Rupnik, 2018). Depending on the larger context in which they are Rupnik (2018) argues that Poland and Hungary competing, political parties adopt different strate- view liberalism as a model of society which is not gies for each of the respective areas. (Strömbäck, only imposed but also failing in critical situations 2007). A strategy commonly employed by politi- like the 2015 refugee crisis. Additionally, he points cal parties in the electoral arena is elaborated on towards the loss of confidence in the institutions of by Somer-Topcu (2015): many parties try to appeal liberal democracy and the fears of identity loss in to broad audiences in order to gain more votes. Polish society (Rupnik, 2018). In the context of ris- Political parties can achieve this goal in a variety ing illiberalism, Przybylski (2018) introduces differ- of ways, for instance by taking distinct positions ent national narratives. Those include a narrative on various issues from different sides of the polit- of sovereign democracy (according to which demo- ical spectrum; by using candidates with particular cratic legitimacy places the actions of a govern- profiles to strengthen or soften their positions; or ment above criticism) and the ‘antemurale myth’ by beclouding their own policy positions (Somer- (according to which refugees and migrants mobilize Topcu, 2015). Such strategies help parties win on a march through Europe and must be stopped votes (Somer-Topcu, 2015). De Vries and Hobolt by countries such as Poland). Fomina & Kucharczyk (2012) link the study of political strategies to theo- (2016) further point towards Norris’ theory of “con- ries of issue evolution which emphasize the impor- temporary authoritarian populism” (p. 58) accord- tance of a changing issue-agenda (2012). ‘Political ing to which we can understand the current devel- losers’ can advance their position in a party sys- opments in Poland as a backlash against long-term, tem by taking up an issue entrepreneurial strategy, ongoing social change. They also name various PiS thus introducing new issue dimensions (De Vries & campaign strategies as well as circumstantial fac- Hobolt, 2012, p. 246). Research shows that vot- tors which were present during the electoral cam- ers are responsive to such strategies and reward

90 Capstone Issue Vol. 12, 2019 Christina Keßler issue entrepreneurs (De Vries & Hobolt, 2012). The III. Methodology importance of specific issues is also the focus of the research of De Sio et al. (2018), which intro- This interdisciplinary capstone combines the duces the issue yield model, according to which po- fields of political sciences and communication stud- litical parties select issues based on the considera- ies in order to gain a comprehensive understand- tion of whether they are positively associated with ing of the political strategy employed by PiS. In or- the party and whether a stance is widely shared by der to contextualize the political behaviour of PiS, the electorate. Reeves et al. (2006) highlight the scholarly literature on the current political situation importance of an increasing focus on the demands and the rise of illiberalism in Poland is examined. of the electorate: they establish that British politi- In addition to this, this capstone draws from schol- cal parties have become more marketing-oriented arly theory concerning political strategy and party and that the importance of ideology has declined. behaviour. Its structure is inspired by Strömbäck’s This, however, results in tension between making (2007) four arena model. According to this model, responsible long-term decisions and being exclu- political parties acting in multiparty systems com- sively voter-driven (Reeves et al., 2006). Johnston pete in four areas: the electoral arena, the media et al. (2018) also focus their research on politi- arena, the internal arena, and the parliamentary cal parties in Britain and conclude that voters often arena. Analyzing merely the electoral arena is lim- base their decisions about a political party on their iting and does not adequately reflect the aims of a feelings towards a party leader instead of specific political party, as a party’s aims go beyond elec- policies. If party leaders change, voters might very toral victory. Each arena has respective strategic well change how they feel about a party (Johnston goals, primary actors and decision concerns asso- et al., 2018). While both the research of Reeves et ciated with it (Strömbäck, 2007). It is important to al. (2006) and Johnston et al. (2018) focus on a note that the strategies pursued in different are- British context, they can give us indicators on how nas can very well contradict each other: the posi- political parties might behave in general. tions a party takes on the campaign trail are not always reflected in the parliamentary behaviour of that very same party (Strömbäck, 2007). It would However, changes in an electorate’s perception also be misguided to conceptualize the four arenas of a political party cannot always be attributed to as strictly separate from each other; in practice, concrete policies or changes in the party structure. the arenas overlap. Strategic decisions regarding Brysk (1995), in contrast to the previous theorists, a party will often affect more than one of the are- highlights the importance of stories and the role of nas, and the arenas influence each other as well narrative in politics. In order to understand politi- (Strömbäck, 2007). Keeping in mind that the are- cal change, we must comprehend that the decision nas are not isolated entities but rather deeply inter- of the electorate is not only based on fixed needs twined, the four arena model can serve as a start- but that the adoption of new narratives concerning point to gain a comprehensive view of party being interests and identities can lead to paradigm haviour. Inspired by the four arena model, this cap- changes which influence voting behaviour (Brysk, stone is divided into two sections: one examining 1995). While the term narrative is oftentimes as- the narrative employed by PiS, the other analyzing sociated with fiction rather than with political sci- the party’s concrete actions. ence, it has over time become an invaluable tool in The first section (Narrative) is influenced by social science methodology (Patterson & Monroe, Strömbäck’s conception of the electoral and the 1998). In the field of political science, the narrative media arena. In those arenas, it is of the utmost plays an essential role in one’s perception of one’s importance what kind of story a political party tells political reality and hence greatly influences politi- to its electorate. In order to gain a comprehen- cal behaviour (Patterson & Monroe, 1998). Theories sive understanding of this story, the following re- of narrative construction can, therefore, be under- search makes use of content analysis. The narra- stood as yet another set of theories further explain- tive was derived from a textual analysis of an offi- ing political strategy and party behaviour: the nar- cial press release by PiS which covers a joint press rative PiS constructs and conveys to its electorate conference by the then Prime Minister Beata Szydło is an essential part of its political strategy. and PiS party leader Jarosław Kaczynski´ in Septem-

91 Capstone Issue Vol. 12, 2019 Christina Keßler

ber 2017 (Chancellery of the Prime Minister, 2017). political actors can create new political opportuni- This press conference was chosen as it was given ties by revealing, challenging, and changing nar- by the then most important political leaders of PiS, ratives about interests and identities” (p. 561). and clearly frames the policies of the party. In order Therefore, the ability to create a compelling narra- to determine the narrative from the official press tive or to insert oneself in an already existing one release, an analysis guided by emergent codes an essential part of bringing about political and col- was carried out (Chancellery of the Prime Minis- lective change and influences the success of polit- ter, 2017). Afterwards, the content of the press ical actors (Brysk, 1995). The term narrative can release was organized in a matrix in which every be defined as a story describing the way in which paragraph of the text was thematically sorted into one constructs facts into a coherent manner in or- at least one of the categories that the coding pro- der to make sense of reality (Patterson & Monroe, cess had brought forth (Appendix A). Then, the find- 1998). When looking at Strömbäck’s four arena ings for each of the emergent codes were gath- model, narrative is most closely related to the elec- ered and used to describe how a certain topic area toral and the media arena. Strömbäck (2007) dis- was framed in the press conference. Subsequently, tinguishes different strategic goals for the different an analysis was carried out that examined which arenas in which political parties compete: in the topics had been emphasized and which topics had electoral arena, the goal is “to maximise electoral been barely mentioned. The analysis also put the and voter support”; in the media arena, the goal findings in context with current scholarly literature of a political party is “to maximise positive public- concerning the party strategy of PiS. ity” (p. 59). Both of these strategic goals require The second section (Action) is inspired by that the political party engage in an act of story- Strömbäck’s conception of the parliamentary and telling by conveying a new, compelling narrative to the internal arena. Through an analysis of con- the public. As Brysk (1995) points out: “We think crete policies and actions of PiS, this research seeks about politics in stories, and our consciousness is a holistic understanding of PiS’s political strategy, changed when new stories persuade us to adopt and limits itself to three concrete actions, prioritiz- a new paradigm” (p. 561). Subsequently, under- ing those which are widely regarded as the most standing the narrative PiS tries to convey to the prominent by current literature on the PiS govern- Polish electorate is an essential part of analysing ment. Concretely, these actions are social spend- the party’s political strategy. Narrative, as under- ing pledges made by the party, decisions taken stood in this research project, is based on Somer regarding leadership positions in the current Pol- and Gibson’s definition of narrativity (Patterson & ish government, and PiS’s migration policy (Fomina Monroe, 1998). Somer and Gibson’s definition, & Kucharczyk, 2016; Przybylski, 2018). Those ac- which is particularly relevant to social sciences re- tions are thoroughly described and contextualized search, contains four features (Patterson & Monroe, in a findings section, followed by an analysis of the 1998): 1) relationality of parts (events need to be actions supported by scholarly literature regarding placed in relation to other events), 2) causal em- political strategy and party behaviour. plotment (events in the narrative have a causal relationship with each other), 3) selective appropriation (some elements are incorporated while oth- IV. Narrative ers are omitted), and, taken together as a fourth feature, 4) temporality, sequence, and place (re- When trying to understand political outcomes ferring to the way in which the elements are lo- the assumption of a rational choice model, in which cated vis-à-vis each other) (Patterson & Monroe, actors base their decision on fixed needs, falls short 1998). Based on this definition, this section will an- (Brysk, 1995). Rather, one must take the impor- alyze the narrative brought forth by PiS and explain tance of ideas and normative beliefs into consider- how it contributes to the party’s popularity. Firstly, ation, because they have the power to shape agen- the narrative will be derived through the analysis das, build identities and shift paradigms (Brysk, of an official press release covering a joint press 1995). According to Brysk (1995), “interests are conference in November 2017 by then-Prime Min- not fixed needs, but rather deeply subsumed sto- ister Beata Szydło and the leader of PiS, Jarosław ries about needs, [...] [and] symbolically mobilized Kaczynski.´ After the analysis of the aforementioned

92 Capstone Issue Vol. 12, 2019 Christina Keßler

primary source, the findings will be put into context is set in November 2017, two years after PiS’s elec- with scholarly literature concerning PiS. Finally, the toral victory in 2015. It is an ideal source from chapter will answer the question of how the narra- which to derive PiS’ narrative considering that it tive employed by the party contributes to its ap- was given at the occasion of the two-year hallmark peal. of PiS government rule, and frames the actions of In order to contextualize the press conference, the party during the past two years. The press con- some information on the political setting in which ference was also held by two of the most prominent it takes place is needed. While the current PiS gov- PiS politicians at the time, respectively occupying ernment is the first instance of a single-party major- two of the most important positions: the head of ity government in post-communist Poland’s history, the party (Kaczynski)´ and the head of government PiS has previously been part of the Polish govern- (Szydło). ment from 2005 to 2007 (Jaskiernia, 2017). How- ever, the coalition led by PiS fell apart in 2007, triggering early elections (Fomina & Kucharczyk, V. Joint press conference by Szydło 2016). During these, a centrist government un- and Kaczynski´ der then Prime Minister Donald Tusk and his party Platforma Obywatelska (Civic Platform; PO) was The analysis of the press conference was guided elected and subsequently “moved Poland back to- by emergent codes. The codes were the follow- ward the mainstream of European Politics” (Fom- ing: development of the Polish economy, reclaim- ina & Kucharczyk, 2016, p. 59). Between 2006 ing the country and feeling at home in Poland, out- and 2014, the Polish electorate chose PO over PiS side pressures, ordinary people and families, equal eight consecutive times in elections at the local, na- opportunities and dignified life, failures of the past, tional, and European level (Fomina & Kucharczyk, judiciary reform, security, and success. The con- 2016; Jaskiernia, 2017). As Fomina and Kuchar- tent of the press release (Chancellery of the Prime czyk (2016) describe, an electoral victory of PiS had Minister, 2017) was analyzed using a matrix which been deemed increasingly unlikely by experts: “An- thematically sorted every part of the press release alysts had begun to regard PiS as unelectable, and into at least one of the aforementioned categories dismissed its authoritarian longings and conserva- (Appendix A). The following analysis clarifies how tive social ideology as lacking appeal outside older, the press conference frames the different topic ar- less educated, and poorer sections of Polish soci- eas mentioned above. Hence, this analysis does ety” (p. 59). However, in 2015, PiS managed to not only show which topics PiS party leaders choose win the Sejm (the lower house of Polish parliament), to talk about, but it also analyses how these differ- the Senate and the presidency, and became the ent issues are framed by the party. first electoral victor in the country’s democratic history to form a government without needing coalition partners (Fomina & Kucharczyk, 2016; Jask- Findings iernia, 2017). While many political scientists con- Development of the Polish economy. sider PiS to be a ‘far-right’ party, delineating the party with a clear ideological label remains a chal- The press conference spends large parts fo- lenge (Jaskiernia, 2017). Positions of the party en- cusing on the development of the Polish economy. tail an increase in social spending, increased tax- Prime Minister Beata Szydło, for example, empha- ation of the wealthy, and the re-nationalization of sizes that strengthening the Polish economy is one key economic sectors (Jaskiernia, 2017). Kaczynski,´ of the primary goals of the PiS government, stating: the party’s current leader, stated that PiS is op- “Family, development and security are the three posed to “immigrants, gays, feminists, liberals, and pillars of our governance” (Chancellery of the Prime most foreigners” (Jaskiernia, 2017, p. 238). Fur- Minister, 2017). Various PiS government initiatives ther, PiS makes a harsh distinction between who with economic dimensions are mentioned, such does and does not belong to the Polish people and as lowering the retirement age and “the flagship explicitly opposes ideas of multiculturalism (Fom- PiS government project, namely Family 500+ pro- ina & Kucharczyk, 2016; Jaskiernia, 2017). The gramme” (Chancellery of the Prime Minister, 2017), press conference analyzed in the following section which provides child benefits to Polish citizens. It is

93 Capstone Issue Vol. 12, 2019 Christina Keßler

emphasized that “Putting Polish economy back on be able to implement social policy that a sound footing” (Chancellery of the Prime Minis- would be beneficial for groups that had ter, 2017) is one of the main priorities of the PiS benefited little, if at all, from changes government. Regarding the state of the economy taking place in Poland. This diagnosis in previous years, Beata Szydło states that : “PiS has turned out to be true” (Chancellery is freeing Polish economy and Polish entrepreneurs of the Prime Minister, 2017) from the shackles of inability, bureaucracy, all that had blocked the energy of Polish economy” (Chan- Judiciary reform. cellery of the Prime Minister, 2017). While the judicial reform sought by the PiS gov- Equal opportunities and dignified life. ernment is mentioned, the idea is not developed at length in the press release. The controversy sur- The notion of bringing about “equal opportuni- rounding judicial reform is not brought up at all, in- ties” and a “dignified life” (Chancellery of the Prime stead, the reform of the judiciary wing is presented Minister, 2017) for the Polish people is mentioned as something positive: several times in the press release, albeit not as often as, for example, the development of the Polish “I hope very much that we will com- economy. According to the text, “The fundamen- plete the process of judiciary reform, - tal assumption of the programme proposed by the Prime Minister Beata Szydło declared. Law and Justice was to ensure that all Poles - re- She highlighted that the judiciary regardless of their place of residence or their occupa- form makes no sense unless it is radical tion - have equal opportunities” (Chancellery of the and gives Polish people the feeling that Prime Minister, 2017). Notably, the press release courts are given back to them and that does not only mention the goal of giving people a citizens are treated fairly” (Chancellery life of dignity but explicitly uses the expression of of the Prime Minister, 2017). “restoring” a life of dignity twice, therefore imply- By framing completing the reform of the judiciary ing that over the previous years, Polish citizens had wing as something necessary for Polish citizens, been deprived of the very dignity the PiS govern- PiS is able to relate their agenda directly to a so- ment is now giving back to them (Chancellery of called “giving back” of the courts and fair treat- the Prime Minister, 2017). ment (Chancellery of the Prime Minister, 2017).

Failures of the past. Ordinary people and families. While the press release gives few concrete ex- No other topic is as frequently mentioned as amples of past failures, it generally conveys the the topic of ordinary Polish people and families; PiS impression that the living conditions for Polish citi- presents itself as a representative of the “ordinary zens have been improving under PiS rule. There is people” (Chancellery of the Prime Minister, 2017), a one example of past failures which is further elab- term which is used repeatedly in the press release.. orated: it is mentioned that PiS is pursuing policies The press release emphasizes that the PiS pro- which make it possible for “a majority of Polish fam- gramme is directly built on the needs of “ordinary ilies, which earlier had not benefited from economic people” (Chancellery of the Prime Minister, 2017) development in Poland” (Chancellery of the Prime and was created in dialogue with them. Every pol- Minister, 2017), to finally do so. In the press re- icy area is directly related to improved living con- lease, this issue is directly related to corruption and ditions for ordinary people and Polish families, who fraud by Jarosław Kaczynski:´ benefit from programmes such as 500+. Repeat- “He then pointed out that two years edly, a notion of restoring dignified living conditions ago, representatives of PiS had found for ordinary Poles and Polish families is brought up. out that the level of all kinds of corrup- When Prime Minister Szydło states that, ”We wish tion and fraud in Poland is high. We to cordially thank all Polish people who allow for the had concluded that if we harness, even good change to happen and who help us to put it in partially, these pathologies, we would effect” (Chancellery of the Prime Minister, 2017),

94 Capstone Issue Vol. 12, 2019 Christina Keßler

she notably explicitly talks to the Polish people and people” (Chancellery of the Prime Minister, 2017). expresses her gratitude for their support. Thirdly, Beata Szydło notes that “Poland is perceived as a terrorism-free country” (Chancellery of the Prime Minister, 2017); this notion is directly fol- Outside pressures. lowed by commenting on the migration policy of Outside pressures on the PiS government are the EU and the former PO government, implying only mentioned twice throughout the entire press that Poland’s status as a terrorism-free country is conference. The first time, they are merely hinted due to the harsh migration policy of the PiS gov- at, when Kaczynski´ thanks Prime Minister Szydło ernment. However, in connection with the migra- for “never failing us in difficult moments, both at tion crisis, Beata Szydło also points out that Poland the beginning of her term and later, despite the would be delivering humanitarian help. different pressures put on her” (Chancellery of the Prime Minister, 2017). Later, the press release gets Success. more specific and mentions the dispute over migration with the EU, stating that: “We have won The press conference frames the two years of the migration dispute with the European Union”, PiS government as a success story. Several gov- and effectively framing the situation as a victory ernment projects are praised and are said to have for Poland and Polish influence over the discourse fulfilled their purpose. Kaczynski´ goes as far as (Chancellery of the Prime Minister, 2017). to claim “that all the undertakings of PiS government have been successful” (Chancellery of the Prime Minister, 2017). This framing makes sense, Reclaiming the country and feeling at home as the press conference took place two years af- in Poland. ter the PiS government started their work. The press conference conveys the notion that the PiS Repeatedly the press release expresses a no- government is a successful administration, which tion of ‘reclaiming’ the country. This notion is con- keeps its promises and betters the living conditions nected to several policy fields, such as economic of ordinary people. The work of the administra- development, security issues, and judicial reform. tion is by no means finished though: “As Jarosław What exactly Poland is being ‘reclaimed’ from re- Kaczynski´ said, changes require longer time than mains unsaid. However, the notion of ‘reclaiming’ just one term of office and thus it is our policy, the the country is connected to the living conditions of policy of PiS, to govern for a longer time, to be able ordinary Polish people and families, for whom the to change Poland completely” (Chancellery of the country must be ‘reclaimed’. The press release em- Prime Minister, 2017). It is made clear that the phasizes that significant changes are needed in or- “good change”(Chancellery of the Prime Minister, der to ensure that “Poland is safe and is growing, 2017), the slogan of PiS’s 2015 election campaign, and all Polish families feel that they are rulers of is not completed yet. their own country” (Chancellery of the Prime Minis- ter, 2017), implying that this is not the case under the current circumstances. Analysis

Joint press conference by Szydło and Security. Kaczynski´

Security, according to the press conference, is The matrix shows that three issues dominate one of the three pillars of PiS governance (together the press release: ordinary people and families, the with family and development), and is brought up successes of the PiS government, and the develop- in relation to three speciﬁc issues. One of them ment of the Polish economy. According to the press is the Warsaw NATO summit, which Beata Szydło conference, these issues are closely related. The described as a “breakthrough” (Chancellery of the PiS government is presented as a government that Prime Minister, 2017). The second one is gov- puts improving the living conditions of the so-called ernment support for the armed forces as well as “ordinary people” (Chancellery of the Prime Minis- the police forces, which are framed as having con- ter, 2017) at the front and centre of its programme. tributed to the “everyday sense of security of Polish PiS’s claim to only represent ‘the real people’ in

95 Capstone Issue Vol. 12, 2019 Christina Keßler

contrast to an elite establishment is a textbook def- on reproductive rights. The topic of judicial reform inition feature of populism (Grzymala-Busse, 2018). is only brought up twice in the press conference: However, while PiS has denounced the Polish “post- once presented as a promise to the Polish people 1989 elite establishment” as a “[corrupt] cartel” that must be fulfilled in the two years ahead, and (Grzymala-Busse, 2018, p. 96) in the past, this once as a reform that “makes no sense unless it press release stays vague and does not specify is radical and gives Polish people the feeling that from whom the country must be reclaimed. courts are given back to them and that citizens are Large parts of the press conference are de- treated fairly” (Chancellery of the Prime Minister, voted to the development of the Polish economy 2017). The approach of PiS to the reform of the ju- - more specifically focusing on various economic diciary wing could be considered radical - shortly and social policies introduced by the PiS govern- after being elected, the PiS government started at- ment, such as the 500+ child benefit programme tacking judicial independence and Poland’s Con- and the lowering of the retirement age. This is institutional Tribunal through legislation, while us- sofar striking as it is expected, as it allows the PiS ing procedural obstructions to stall the announce- to join in a historic tradition: “From the very start of ment of court vacancies (Przybylski, 2018; Rup- the transition era in Poland, political narratives and nik, 2018; Fomina & Kucharczyk, 2016). These ac- policy decisions have rested to a large extent on tions provoked nation-wide protests, within Poland an economic rationale, with the promise of prosper- most notably from a grassroots-initiative called ity at its center” (Rupnik, 2018, p. 54). While the Komitet Obrony Demokracji (Committee for the De- Polish economy has been performing well over the fense of Democracy; KOD), lawyer’s groups, and last few years (Przybylski, 2018; Rupnik, 2018), the other Non-Governmental Organisations (Fomina & press conference argues that the “ordinary people” Kucharczyk, 2016; Grzymala-Busse, 2018). Inter- (Chancellery of the Prime Minister, 2017) have not nationally, the efforts of the PiS government drew sufficiently benefited from this development, an is- sharp criticism from the European Union: sue that the PiS government now aims to tackle. In “Late in December 2015, Poland be- fact, the claim that “other parties have been hoard- came the first EU member state to be a ing the fruits of the transition, forsaking social soli- subject of the Union’s new “pre-Article darity with those who had reason to resent the new 7 procedure,” adopted just the year be- system” (Przybylski, 2018, p. 56) is by no means a fore, for looking into possible breaches novelty, but an argument Kaczynski´ has been mak- of EU standards, including harms to the ing since the 1990s (Przybylski, 2018). The PiS gov- rule of law. In January 2016, the Eu- ernment, led by Kaczynski,´ thus argues that its var- ropean Commission launched an official ious initiatives have already benefited the Polish probe.” (Fomina & Kucharczyk, 2016, p. people economically and will continue to do so in 64). the future. The press conference does not tell a story of While the press conference brushes over the topic struggle, but a story of triumph. The two years of judicial reform rather quickly, it entirely fails to of PiS governance are framed as a success story mention the controversy concerning the PiS gov- that is set to continue. Szydło and Kaczynski´ fo- ernment’s restrictions on reproductive rights. A cit- cus on the accomplishments of the PiS government, izen’s bill of March 2016 which would have enacted placing a strong emphasis on economic and social a full ban on abortion, but was ultimately not voted pledges the party has managed to fulfill. In or- on in parliament after mass protests by Polish cit- der to derive the distinct frame which was chosen izens (Fomina & Kucharczyk, 2016). To claim the for this press conference, it is important to distin- problem was no longer relevant however would be guish not only which political issues were empha- incorrect, as “issues remain subjects of heated pub- sized, but also which issues were just briefly men- lic controversy, with attendant demonstrations and tioned or entirely ignored in the telling of this suc- petition drive” (Fomina & Kucharczyk, 2016, p. 65). cess story. Two topics that provoked controversy This is also due to the Health Ministry’s public an- both within Poland and abroad during the first two nouncement to consider restricting access to con- years of PiS governance were the state of the rule traception (Fomina & Kucharczyk, 2016). Likely the of law and the restrictions the government imposed two aforementioned topics were not elaborated on

96 Capstone Issue Vol. 12, 2019 Christina Keßler

in the press conference due to the domestic crit- years in opposition came from its strong icism they provoked, which, in the case of both stance on issues connected to national judicial reform and of abortion rights, led to mass identity and sovereignty and from its protests by Polish citizens. bond with the Catholic Church.” (p. 66). Apart from these two issues, there is another However, the Church was not brought up even once topic which strikingly is not elaborated on in the during the press conference, nor were left-liberal press conference: migration. Migration is only ex- values (such as secularism, feminism, LGBTQ plicitly mentioned twice throughout the entire con- rights, multiculturalism) towards which PiS often ference: first, when Szydło and Kaczynski´ empha- presents itself in opposition (Rupnik, 2018). size that the PiS government will not take part in The fact that the press release fails to elaborate “the misguided EU migration policy” (Chancellery on issues which are generally perceived to be cru- of the Prime Minister, 2017) while pointing towards cial to the party’s identity and of high importance the fact that terrorism does not occur in Poland; to its core electorate, suggests PiS trying to ap- secondly, when they declare that the PiS govern- peal to new parts of the electorate - people who ment has “won the migration dispute with the Eu- traditionally would not have voted for PiS and are ropean Union” (Chancellery of the Prime Minister, not necessarily swayed by the ideology that PiS 2017) and changed the discourse concerning mi- puts forth. Indeed, the narrative presented in this gration. While the issue of migration, therefore, press conference emphasises not ideological, but is certainly mentioned, it is not discussed, espe- rather socioeconomic considerations for support- cially compared to topics such as economic devel- ing the PiS government. Controversial topics are opment. This is remarkable, as the party’s posi- largely avoided or brushed over. For a party which tion on migration is generally perceived as one of often verges on xenophobia (Fomina & Kucharczyk, PiS’s political assets (Narkowicz, 2018). When try- 2016), the little stab that Szydło and Kaczynski´ take ing to understand the electoral victory of PiS in at EU migration policy appears almost negligible. 2015, scholarly literature regularly points to the The strategy of focusing not on overtly ideological party’s position on migration as one of the reasons issues but rather appealing to more moderate parts for its appeal, as its stance resonates with the Pol- of the electorate has also proven successful for PiS ish public (Fomina & Kucharczyk, 2016; Narkow- in the past: the conscious deradicalization of the icz, 2018; Przyblyski, 2018; Rupnik, 2018). Rup- party image has been on the agenda of the party nik (2018) attributes the appeal of PiS largely to since the 2015 elections. Back then, “Running on the fact that “Central European countries perceive the slogan “Good Change”, PiS leaders called for the redistribution of migrants across national bor- compassionate conservatism, and sought to offer ders [...] as an attempt to impose on them a mul- undecided voters an alternative to the “boring” PO” ticultural model of society that they consider a fail- (Fomina & Kucharczyk, 2016, p. 61-62). This strat- ure” (p. 33). According to Fomina and Kucharczyk egy was rewarded by the electorate: in 2015, PiS (2016), “[Kaczynski’s]´ hard line on refugees, verg- “reaped the benefits of a more moderate appeal ing on xenophobia, won over people who normally to the general public” (Przyblyski, 2018, p. 56). would never have voted for PiS” (p. 62) in light of PiS’s strategy to deradicalize its message as much the 2015 migration crisis. The omission of an elab- as possible in order to appeal to a broader public orate discussion on the topic of migration during also makes sense regarding the setting of the press the press conference is worthy of scrutiny, espe- conference, considering that it was an official event cially considering the importance of the issue for with the then two most prominent PiS party politi- PiS (Narkowicz, 2018). There are other issues gen- cians, and therefore likely to draw the attention of erally perceived as PiS’s core issues that were ig- a broad audience. nored during the press conference. According to Whilst PiS’s ideology is somewhat softened, rad- Fomina and Kucharczyk (2016), ical features of the party’s politics keep shining “Although Law and Justice was elected through if one knows what to look out for. This on a platform of generous socioeco- becomes especially apparent in the repeated no- nomic promises, its dominant position tion that the country needs to be ‘reclaimed’ and on the righthand side of the politi- ‘restored’, insinuating that the country must have cal spectrum and its resilience despite been at least somewhat destroyed by previous gov-

97 Capstone Issue Vol. 12, 2019 Christina Keßler

ernments. Fomina and Kucharczyk (2016) do in- tory, suggesting a double strategy. The traditional deed attribute the success of PiS to the fact that ideological core of PiS with its authoritarian and the party managed to successfully convey the nar- xenophobic longings still very much exists, as cer- rative of a “Poland in ruins” (p. 60) to the Polish tain comments and allusions in the press release electorate. According to this narrative - which ig- demonstrate. However, the party aspires to send a nores the economic progress Poland experienced message that does not only appeal to its traditional after it had recovered from transition - Poland has electorate and subsequently softens its message. to be restored and rebuilt “after the devastation allegedly wrought by PO’s eight-year rule, or even the quarter-century of Poland’s democratic trans- VI. Action formation” (Fomina & Kucharczyk, 2016, p. 60). The success of this narrative can be partially cred- Understanding the narrative a political party ited to the reforms of the 1990s. The privatization employs is an integral but ultimately insufficient of whole economies over a relatively short period of step towards understanding its strategy. In order time, “often with a contempt for the rule of law and to draw a holistic picture of PiS’ political strategy, societal sentiments about equity” (Rupnik, 2018, one must also examine the party’s concrete ac- p. 34), resulted in a general distrust of the politi- tions in and outside of parliament. This dimen- cal elite (Kowalewski & Rybinski, 2011; Przybylski, sion is reflected in Strömbäck’s four arena model 2018). While deradicalization may take place con- through the parliamentary and the internal arena cerning the party’s explicit ideology, PiS does not (Strömbäck, 2007). According to the model, the hide the fact that its goal is “to change Poland com- goal of a political actor in the parliamentary arena pletely” (Chancellery of the Prime Minister, 2017). is “to maximize parliamentary influence” (Ström- The narrative presented here seamlessly fits bäck, 2007, p. 59), while the strategic goal in the definition of narrativity by Somers and Gibson, the internal arena is “to maximise internal cohe- which includes relationality of parts, causal emplot- sion” (Strömbäck, 2007, p. 59). These two are- ment, selective appropriation, and temporality, se- nas make apparent that, in addition to an analy- quence and place. (Patterson & Monroe, 1998). The sis of PiS’s political message, an analysis of PiS’s press conference places events in a causal rela- concrete actions is necessary. Again, it must be tionship to other events and therefore meets the noted that while narrative and concrete party ac- first two features: relationality of parts and causal tions are treated as two separate dimensions in this emplotment. During the press conference, for ex- capstone, the two areas are intertwined, and over- ample, Szydło and Kaczynski´ not only suggest that lap and influence each other in a variety of ways. Poland is being reclaimed by ‘ordinary citizens’, but In the following section, three concrete actions they also place this ‘reclamation’ in a causal rela- will be examined: 1) social spending policies which tionship to the efforts of the PiS government. The have been implemented by PiS, 2) the decision third feature of the definition, selective appropria- to select Beata Szydło as Prime Minister instead tion, is likewise apparent in the conference. Issue of party leader Jarosław Kaczynski,´ and 3) the re- areas, such as the controversy regarding abortion fusal to take in refugees and the following migra- laws that have been omitted in the press confer- tion policy dispute with the European Union. These ence, as explained above. The fourth requirement, three actions have been selected as that they all the elements of temporality, sequence, and place, occurred within the two years after the PiS govern- is also met. Exemplar for this is the story that ment came into power in 2015, and are regarded previous (temporality) governments have brought to be cornerstones of PiS popularity with the Polish Poland (place) in a position where it now needs to electorate (Fomina & Kucharczyk, 2016; Narkow- be reclaimed by its citizenship, a task that the cur- icz, 2018; Przybylski, 2018; Rupnik, 2018). Each rent PiS government is now successfully taking on action will first be contextualized and elaborated (sequence). on, answering the question of what exactly the PiS When examining the narrative presented in the government’s actions in specific issue areas were. press conference, one notices that the radical ele- Then, the strategic benefit of each concrete action ments of PiS’s ideology, while hidden, are still no- will be analysed in connection to the relevant schol- ticeable for anyone familiar with the party’s his- arly literature. To conclude the chapter, how the

98 Capstone Issue Vol. 12, 2019 Christina Keßler

aforementioned decisions by PiS contributed to the core electorate and thus understandably improve party’s appeal will be summarized. the party’s standing. As the previous analysis of the narrative put forward by the PiS government shows, the party VII. Concrete actions of the PiS places a lot of emphasis on its social spendings and government chooses to convey a message emphasising these spendings over one portraying a more ideological Findings narrative. We can connect these actions to efforts to deradicalize the image of the party and appeal to Social Spending. a broader electorate. Furthermore, we should note When Beata Szydło ran for Prime Minister in that there is evidence of political parties becom- 2015, her platform included generous social spending less ideologically driven and moving towards ing pledges (Fomina & Kucharczyk, 2016). The two a more ‘market-oriented’ strategy in recent years most prominent examples of these are a child ben- (Reeves et al., 2006). For example. according efit programme and the promise to lower the re- to Reeves et al. (2006), political parties in Great tirement age after it had been raised in previous Britain increasingly attempt a so-called consumer- years (Fomina & Kucharczyk, 2016). Once elected, driven or market-oriented strategy instead of fo- the government maintained the promises they had cusing on party ideology. The concept of political made in the election campaign. The programme marketing is not a new one (Strömbäck, 2007). Ac- 500+ was implemented, financially benefiting fam- cording to Strömbäck (2007), the defining features ilies with children (Grzymala-Busse, 2018; Szczer- of the concept are: biak, 2017). The unpopular 2011 decision by Don- “[First,] [...] that political marketing ald Tusks’ government to gradually raise the pen- is the application of marketing princi- sion age to 67 was reversed as the age of retire- ples and procedures–not just market- ment was moved to 60 for women and 65 for men ing techniques and activities–and sec- (Grzymala-Busse, 2018; Szczerbiak, 2017). While ond, that the processes should be “in re- both of these social spending pledges were rather sponse to the needs and wants” of peo- expensive promises to fulfill, one must note that ple or groups targeted by the produc- the Polish economy has been performing well re- ers–the organizations or campaigns.” cently, with unemployment rates dropping and in- (p. 56). vestments increasing (Szczerbiak, 2017). As Anna Grzymala-Busse (2018) points out: “Claims from The focus of the PiS government on social spending the opposition that such measures could ruin the pledges is a prime example for successful political public finances ring hollow against a background of marketing, as the 500+ programme, as well as the steady economic growth and the lowest unemploy- lowering of the retirement age, can be understood ment rate in decades” (p. 100). as a direct response to the needs of the Polish elec- When talking about the electoral success of PiS torate. By following through on campaign promises in connection to social spending pledges made by that are widely popular and financially feasible (for the party, it should be mentioned that in recent now), PiS positions itself as an honest advocate for years, PO and PiS have appealed to an increasingly the Polish lower and middle class. split electorate whose living conditions differ from each other (Przybylski, 2018). According to Przy- Party Leadership. bylski (2018), “While PO developed its constituen- cies around major urban areas, mostly in the north- Jarosław Kaczynski,´ the head of PiS, has a long west regions of the country, PiS turned to small and track record in Polish political history. In 2001, he medium-sized towns, generally in the southeast” founded the party together with his brother Lech (p. 55). This split encompasses different socioeco- Kaczynski,´ and served as Prime Minister in the nomic realities, as economically lower performing PiS-led coalition government from 2005 until 2007 regions frequently vote for PiS (Przybylski, 2018). (Grzymala-Busse, 2018). Currently, however, he The social spendings made by the PiS government does not hold any official government positions. In therefore directly affect the living conditions of its fact, in 2015 it was Kaczynski´ who selected Andrzej

99 Capstone Issue Vol. 12, 2019 Christina Keßler

Duda as the party’s presidential candidate and de- on leading party candidates instead of a party’s cided that Beata Szydło would serve as Prime Minis- policies. According to Johnston et al. (2018), re- ter of the PiS government (Grzymala-Busse, 2018). search conducted in Great Britain shows that the Emphasizing a message of “compassionate conser- electoral success of a political party can at least vatism” (Fomina & Kucharczyk, 2016, p. 61) under partially be attributed to the electorate’s feelings the slogan of ‘Good Change’, Jarosław Kaczynski´ about leading candidates: was too controversial a personality, too much associated with the authoritarian leanings of PiS, to “Many voters [...] avoid at least some occupy an official government position (Fomina & of the effort involved in assembling Kucharczyk, 2016). This leads to a somewhat odd and assessing information about par- power structure, with Kaczynski´ being acknowl- ties’ policies and instead use heuris- edged, as Anna Grzymala-Busse phrases it, as the tics such as their feelings about the “power behind the throne” (2018, p. 96). To Przy- party leaders as major determinants of bylski (2018), “hiding its most extreme politicians their decisions. When party leaders are from the public eye” (p. 56) is part of the deradi- changed, therefore, differences in vot- calization PiS tries to convey to its electorate. ers’ feelings about predecessor and suc- cessor could lead to changes in party The efforts of PiS to deradicalize its message in choice.” (p. 2). order to become more broadly acceptable have already been touched upon earlier. In fact, the strat- The strategic benefit of selecting Beata Szydło as a egy of appealing to different subgroups is often pur- candidate for the office of Prime Minister is appar- sued as well as rewarded by the electorate (Somer- ent. Her persona, unlike that of Jarosław Kaczynski,´ Topcu, 2015). According to Somer-Topcu (2015), is not as firmly linked to the more controversial there are several ways through which a political leanings of PiS (Fomina & Kucharczyk, 2016; Przy- party can broaden its outreach to different groups bylski, 2018). Through her, PiS manages to ap- in society. One of them is to have a variety of posi- peal to a more moderate part of the electorate tions which represent policy issues from both the who would hesitate to vote for Kaczynski´ (Fomina left and the right ends of the political spectrum & Kucharczyk, 2016; Przybylski, 2018). We can, (Somer-Topcu, 2015). Another possible strategy therefore, understand the decision to select Beata is to not clarify one’s political position but rather Szydło and not Jarosław Kaczynski´ for the position do the opposite by engaging in a strategy of ‘be- of the prime minister as another part of a larger clouding’ (Somer-Topcu, 2015). However, a political strategy to appeal to the general public. party that tries to reach out to a broader electorate does not necessarily need to change its positions: it might be sufficient to change the phrasing of these Migration Policy. positions: “[while] the term party strategy is used to refer to actual party behavior, it is not the actual In order to comprehend the importance of PiS’s party positions, but voters’ perceptions of these po- stance on issues of migration, it is important to sitions that influence their vote choice and deter- note that the 2015 election campaign which re- mine the electoral consequences of a party strat- sulted in PiS’s victory was dominated by the Euro- egy” (Somer-Topcu, 2015, p. 845). It is hence pos- pean refugee crisis, which was then at its height. sible to pursue the strategy of appealing to a broad (Narkowicz, 2018). Prior to the elections, PO, then electorate without changing one’s political position. in power, had accepted the EU proposition of quo- Somer-Topcu (2015) points towards the selection of tas for a relocation scheme and pledged to ac- a specific candidate as another strategy to appear commodate 7000 refugees (Narkowicz, 2018). On broadly acceptable. In the case of Poland, PiS has the campaign trail, PiS took a particularly tough managed to do so through the selection of Beata stance on the refugee question and strongly op- Szydło as their candidate for Prime Minister, who is posed the idea of the EU relocation scheme, point- perceived as more moderate and not as controver- ing towards Poland’s national security and terror- sial than Jarosław Kaczynski´ (Fomina & Kucharczyk, ism threats (Narkowicz, 2018). The electorate re- 2016; Przybylski, 2018). Indeed, there is evidence warded the tough stance on the refugee question suggesting that voters make their decisions based at the voting booth (Narkowicz, 2018).

100 Capstone Issue Vol. 12, 2019 Christina Keßler

Shortly after the 2016 Brussels terrorist at- a political party’s position on a singular issue can tacks, the new Prime Minister Beata Szydło an- indeed inﬂuence voters’ decisions. As De Sio et nounced that Poland would retract from the EU al. (2018) elaborate: “A number of studies have relocation scheme and not accommodate any mi- shown the increasing importance of the political is- grants (Narkowicz, 2018). The Polish discourse con- sues of the day for voting behavior, on both sides cerning refugees was dominated by the “imagined of the Atlantic” (p. 1209). De Vries and Hobolt threat of a Muslim terrorist arriving from Syria dis- (2012) demonstrate how political parties can ac- guised as a refugee” (Narkowicz, 2018, p. 358), tively make use of the importance of singular is- as the Polish public started to fear terrorist attacks sues and attract new voters by adopting an ‘issue in Poland could be commited by migrants. In ad- entrepreneurial strategy’ (therefore, introducing a dition to the linkage of the refugee question with new issue dimension). In 2015, PiS was able to issues of national security, the discourse was input forward a compelling policy standpoint on the creasingly racialized, with PiS party leader Jarosław refugee crisis: that refugees would not be accepted Kaczynski´ suggesting refugees could bring para- in Poland. While other parties also adopted posi- sites to Poland, a statement that was criticized tions that reﬂected the public’s anxiety regarding as being “reminiscent of Nazi rhetoric” (Narkowicz, the refugee question, the stance of PiS was the 2018, p. 366). As shown in the previous chapter, most credible due to the party’s history (Narkow- the refugee crisis is used by the PiS government to icz, 2018). It is therefore indeed possible that a lot criticise both the EU and the previous government of voters who would not have traditionally consid- (Fomina & Kucharczyk, 2016). ered voting for PiS changed their minds because of The hostile stance of PiS towards accepting the refugee crisis (Fomina & Kucharczyk, 2016). refugees is supported by large parts of the Polish In its dispute with the European Commission electorate. According to a May 2017 CBOS (Cen- over the question of whether to accept refugees, trum Badania Opinii Społecznej; Centre for Public the PiS government managed to position itself as Opinion Research) survey, only 25 percent of Poles a force protecting Poland from foreign interference were in favour of accepting refugees, while 70 per- (Narkowicz, 2018). While Poland’s citizenry is over- cent were against it (Szczerbiak, 2017). In fact, in whelmingly pro-European (Fomina & Kucharczyk, a country that as of today is “one of the most reli- 2016), many citizens agreed with the PiS govern- giously and ethnically homogenous countries in Eu- ment instead of the European Commission when rope with around 90% of the population declaring it came to a EU relocation scheme for refugees themselves as Catholics”, the 2015 refugees crisis (Narkowicz, 2018). PiS politicians openly accuse has heightened Islamophobic sentiments and fears the EU of trying to impose on Poland, with a concerning national identity (Narkowicz, 2018, p. representative of PiS in the European Parliament, 359). Issues related to national identity, Catholi- Legutko, “accus[ing] the EU of promoting in this cism, and conservative values have been empha- area a “left-liberal” agenda (feminism, LGBT rights, sized by PiS for years and are at the core of the gay marriage, multiculturalism) that must be re- party’s identity (Fomina & Kucharczyk, 2016). In sisted” (Rupnik, 2018, p. 33). fact, Fomina and Kucharczyk (2016) attribute PiS’s Furthermore, the dispute with the EU over “dominant position on the right hand side of the refugee policy enabled the PiS government to po- political spectrum and its resilience despite years sition itself as a victim of unfounded criticism by in opposition” (p. 66) to its strong positions in the EU (Szczerbiak, 2017). This transcends onto those areas as well as its connection to the Catholic other issue areas. When faced with criticism by Church. PiS’s tough stance on the question of the Venice Commission concerning the state of the accepting refugees was therefore credible to the rule of law in Poland, the PiS government invoked Polish electorate, and especially appealing in a a narrative of sovereign democracy in which the time when concerns about accepting refugees were EU is yet again presented as a foreign force trying heightened and considered more pressing than to impose on Polish sovereignty (Przybylski, 2018; usual. Fomina & Kucharczyk (2016) argue that Rupnik, 2018). Przybylski (2018) elaborates: “Ac- PiS’s “hard line on refugees, verging on xenopho- cording to this argument, a party that has won bia, won over people who normally would never the majority of seats in Parliament represents the have voted for PiS” (p. 62). Research shows that sovereign will of Poland. Once the election results

101 Capstone Issue Vol. 12, 2019 Christina Keßler

are in, the new government’s democratic legiti- In addition, voters who in previous years might macy places its actions above question” (p. 59). have been put off by the authoritarian leanings The refugee crisis has therefore not only very likely of PiS did not have to fear that the leading can- benefitted PiS during the 2015 elections, it also has didate of PiS would be too radical. Instead of handed the party a template for dealing with criti- party leader Jarosław Kaczynski,´ Beata Szydło was cism of its actions by the European Union. chosen as Prime Minister, effectively broadening the party’s appeal by deradicalizing its message (Somer-Topcu, 2015). While Jarosław Kaczynski´ has Analysis long been associated with PiS and continues to lead Concrete actions of the PiS government the party, Szydło was perceived as a less controversial, more moderate candidate (Fomina & Kuchar- In order to determine what makes a political czyk, 2016; Przybylski, 2018). Especially consid- party appealing to the electorate, it is important to ering the fact that many voters do not necessarily analyse not only what a party says, but also what research a political party’s platform but are rather a party does (Strömbäck, 2007). The actions of PiS guided by their feelings about leading party fig- have contributed in various ways to the party’s ap- ures as a heuristic (Johnston et al., 2018), select- peal and its domestic support. ing Szydło as Prime Minister was a decision which First and foremost, PiS’s popularity stems from contributed to broadening PiS’s appeal. the fact that it delivered on various social spend- Finally, one must point out the importance of ing pledges made in the 2015 campaign. As men- the 2015 refugee crisis in order to understand the tioned before, these promises included a generous 2015 electoral victory of PiS. Despite the fact that child benefit programme as well as the lowering the country was barely affected by the crisis, the of the retirement age after it had been raised by 2015 election was dominated by the topic (Narkow- the previous administration (Fomina & Kucharczyk, icz, 2018). On this issue, many Poles agreed with 2016). As Poland is currently experiencing a pe- the position of PiS which rejected the EU reloca- riod of economic growth, the PiS government can tion scheme and the notion of accepting Muslim afford these costly projects, and criticism pointing refugees into the country (Szczerbiak, 2017). Re- towards the state of the public finances is largely search shows that the ‘political issue of the day’ ignored (Grzymala-Busse, 2018; Szczerbiak, 2017). might indeed be decisive when it comes to a voter’s PiS’s focus on social spending can be understood decision (De Sio et al., 2018). It is especially ef- as a form of political marketing, a strategy accord- fective when parties manage to explore new ising to which, simply put, the needs of the market sue dimensions and thus explore an issue en- (the electorate) determine the final product (poli- trepreneurial strategy in order to win votes (De cies) (Strömbäck, 2007). Ideology is no longer the Vries and Hobolt, 2012). After the new PiS gov- leading factor which determines a voter’s decision ernment retracted from the EU relocation scheme, at the ballot box (Reeves et al., 2006); instead PiS’ PiS managed to frame itself as the victim of anti- voters might be persuaded by the successful so- Polish sentiments in an EU that tried to impose a cioeconomic policies the party has initiated (Fom- liberal model of society upon Poland (Przybylski, ina & Kucharczyk, 2016). 2018; Rupnik, 2018).

VIII. Conclusion special emphasis on the fact that PiS has managed to improve the living conditions of the so-called ‘ordinary people’. The conference further highlights In order to understand the strategies PiS uses the importance of the development of the Polish to appeal to the public, it is necessary to analyze economy and socioeconomic policies which were both the narrative the party tries to convey, as particularly popular with the electorate, such as well as its concrete actions. A thorough analysis the 500+ programme and the lowering of the pen- of a 2018 joint press conference by Prime Minis- sion age to 60 and 65 for women and men, respec- ter Szydło and PiS President Kaczynski´ reﬂects the tively. Government initiatives that were regarded story the PiS government tells about its own role as controversial and/or triggered mass protests in in Poland. The (until then) two years of PiS gover- Poland (judicial reform and restrictions on abortion nance are represented as a success story, with a

102 Capstone Issue Vol. 12, 2019 Christina Keßler

rights) were only briefly mentioned or entirely ig- been located at the far-right of the Polish politi- nored in the press conference. Another topic which cal spectrum, and its traditional core issues consist is barely mentioned in the press conference is mi- of national identity, Catholicism, and conservative gration. This is surprising given that large parts of values (Fomina & Kucharczyk, 2016). Through its the Polish citizenry agree with PiS on its stances re- tough stance on migration PiS has kept its cam- garding refugees and migrants. The omission of paign pledge to this part of the electorate. One the topic might be due to the fact that questions might be tempted to claim that PiS is not engag- regarding migration are too closely related to PiS’s ing in a double strategy, but that its position on more radical ideology while the press conference the political spectrum has simply shifted to the cen- in general attempted to bypass topics traditionally tre. Is it possible that it was not the message that associated with PiS. Another noteworthy aspect is has been deradicalized, but rather the party? In- that a narrative of ‘reclaiming’ and/or ‘restoring’ deed one must be aware that the character of a po- the Polish country is repeatedly pushed, insinuating litical party necessarily changes to at least some a counterfactual ‘Poland in ruins’ narrative accord- extent with the political message it sends. After ing to which previous administrations left Poland in all, winning over new parts of the electorate sub- shambles (Fomina & Kucharczyk, 2016). While de- sequently influences PiS’s decisions regarding its radicalization of the party’s message seems to be future political strategies and campaign messages. a key concern of the official press conference, PiS PiS’s success in the electoral arena and its changed leaves no doubt that its goal is to change Poland behaviour in the media arena are bound to influ- drastically. ence the party’s behaviour in the parliamentary In addition to analyzing the official narrative and the internal arena because these arenas over- that PiS tries to convey to the Polish citizenry, three lap (Strömbäck, 2007). However, there is evidence significant actions of the PiS government were con- that the radical elements of the party have not textualized and further examined. The party en- changed so much as they have been hidden. PiS’s gages in a classic example of political marketing, ideology has not changed as much as its political which becomes especially apparent when analyz- message has; its authoritarian leanings and xeno- ing its social spending pledges. Programmes like phobic character are still very much existent (Fom- the 500+ child benefit programme are highly pop- ina & Kucharczyk, 2016); Kaczynski’s´ position in the ular with the Polish people and add to the party’s party is a prime example of this. In order to appeal appeal. Not only does PiS employ a strategy of to a more moderate public, the key controversial softening its image in order to appeal to a broader figure does not occupy any governmental positions audience through its narrative, but it also does so (Fomina & Kucharczyk, 2016). However, he is still through its actions. An additional softening of PiS’s universally acknowledged as the de facto leader of image was achieved by selecting Beata Szydło for Poland (Fomina & Kucharczyk, 2016; Rupnik, 2018). the position of Prime Minister, who is perceived as A substantive change in PiS’s ideological makeup less controversial than Jarosław Kaczynski,´ the PiS would require Kaczynski´ not to step out of the spot- party leader (Fomina & Kucharczyk, 2016; Przybyl- light, but to step down as the party’s leader. An ski, 2018). Finally, the rejection of the PiS govern- analysis of the party’s concrete actions, therefore, ment to accept refugees added to the party’s ap- results in the conclusion that the party is engaging peal, as large parts of the Polish citizenry agree in a double strategy which is changing its message with this position and the discourse concerning in order to appeal to a mode moderate public, while refugees dominated the 2015 elections. Here, PiS still not losing its core electorate, by maintaining a successfully used a ‘political issue of the day’ to its more radical image for those ‘in the know’. advantage and subsequently reaped the benefits A greater variety of sources would enable this at the voting booth. research project to have a more holistic analysis. It The findings of this research concerning both is presumed that PiS deradicalizes its message de- narrative and concrete actions suggest that PiS has pending on the audience it targets. The press con- successfully adopted a double strategy in which ference from which the narrative was derived was a the party deradicalized its image in order to ap- highly official event (both Prime Minister and party peal to new voters while not neglecting its core leader participated, its occasion was the celebra- electorate. The electorate of PiS has historically tion of two years of successful governmental rule,

103 Capstone Issue Vol. 12, 2019 Christina Keßler

and it was translated into English for international Works Cited observants) from which one could expect a softening of PiS’s message. It is thinkable that the double Brysk, A. (1995). " Hearts and minds": bring- strategy in which PiS is currently engaging would ing symbolic politics back in. Polity, 27(4), not have been recognized if this research had an- 559-585. alyzed content solely from a different, e.g. more Chancellery of the Prime Minister (2017). Prime local, context. To do so is a task for future research Minister Beata Szydło: family, devel- in the topic area. opment and security are the three pillars of our governance [Press re- While PiS has managed to convey a softer, lease]. Retrieved April 21, 2019, from less radical message to voters, its ideological core https://www.premier.gov.pl/mobile/en/news/news/pr with its authoritarian and xenophobic leanings is ime-minister-beata-szydlo-family-development- still very much existent, as Kaczynski’s´ position as and-security-are-the-three-pillars-of- party leader suggests. However, the very fact that our.html PiS has engaged in an act of deradicalizing its mes- Cianetti, L., Dawson, J., & Hanley, S. (2018). sage shows that the party is changing. By attract- Rethinking “democratic backsliding” in ing new subgroups of voters, the identity of the Central and Eastern Europe – looking party is not drastically altered, but certainly has be- beyond Hungary and Poland. East come less clear-cut. Today, not everyone who iden- European Politics, 34(3), 243–256. tifies with PiS necessarily identifies with the tradi- https://doi.org/10.1080/21599165.2018.1491401 tional notion of conservative values the party was De Sio, L., De Angelis, A., & Emanuele, V. (2018). built on. For Poland, this situation could pose a Issue yield and party strategy in multiparty chance as well as a threat. If PiS were to aban- competition. Comparative Political Studies, don its authoritarian aspirations, its double strat- 51(9), 1208-1238. egy could further a transformation of the PiS into a De Vries, C. E., & Hobolt, S. B. (2012). When di- Christian conservative people’s party without ex- mensions collide: The electoral success of tremist tendencies, for example. Such a party issue entrepreneurs. European Union Poli- could possibly fill the space in the Polish political tics, 13(2), 246-268. spectrum formerly occupied by PO. However, the Fomina, J., & Kucharczyk, J. (2016). Pop- current behaviour of PiS, especially in the realm ulism and Protest in Poland. Jour- of judicial reform, makes this scenario improbable. nal of Democracy, 27(4), 58–68. It seems more likely that PiS will continue to sys- https://doi.org/10.1353/jod.2016.0062 tematically undermine institutions of liberal democ- Grzymala-Busse, A. (2018). Poland’s Path to Il- racy, entailing the courts and the media. Poland’s liberlism. Current History, 117(797), 96- transformation into an illiberal democracy will con- 101. tinue until a substantive part of the Polish citizen- Jaskiernia, J. (2017). The Development of the Polish ship recognizes this transition as a threat. Party System. Polish Political Science Year- On the European level, this means that illiberal book, 46(2), 227-246. movements across the continent will be strength- Johnston, R., Hartman, T., & Pattie, C. (2018). ened. This is especially true in, but not limited to, Feelings about party leaders as a voter’s the CEE region and Orbán’s Fidesz party in Hun- heuristic–What happens when the leaders gary. Future research on the political strategies of change? A note. Electoral Studies. parties comparable to PiS should determine to what Kowalewski, O., & Rybinski, K. (2011). The hidden extent the success of PiS can be attributed to struc- transformation: the changing role of the tures that are specific to the Polish context. Political state after the collapse of communism in movements similar to PiS are found across national Central and Eastern Europe. Oxford Review and even continental borders (Rupnik, 2018), and of Economic Policy, 27(4), 634-657. in order to fully comprehend their root causes, it is Przybylski, W. (2018). Can Poland’s of utmost importance to understand what makes Backsliding Be Stopped? Jour- them more successful in some countries than in nal of Democracy, 29(3), 52–64. others. https://doi.org/10.1353/jod.2018.0044

104 Capstone Issue Vol. 12, 2019 Christina Keßler

Mucha, J. (2006). Poland in Central and Eastern Rupnik, J. (2018). The Crisis of Liberal- Europe, Polish Sociology within the Central ism. Journal of Democracy, 29(3), 24–38. European Context: Introduction. Journal of https://doi.org/10.1353/jod.2018.0042 Classical Sociology, 6(3), 251-256. Somer-Topcu, Z. (2015). Everything to everyone: Narkowicz, K. (2018). ‘Refugees Not Welcome The electoral consequences of the broad- Here’: State, Church and Civil Society Re- appeal strategy in Europe. American Jour- sponses to the Refugee Crisis in Poland. In- nal of Political Science, 59(4), 841-854. ternational Journal of Politics, Culture, and Strömbäck, J. (2007). Political marketing and pro- Society, 31(4), 357-373. fessionalized campaigning: A conceptual O’Neil, P. H. (2015). Essentials of comparative analysis. Journal of political marketing, politics: Fifth international student edition. 6(2-3), 49-67. WW Norton & Company. Szczerbiak, A. (2017, November 01). Ex- Patterson, M., & Monroe, K. R. (1998). Narrative plaining the popularity of Poland’s Law in political science. Annual review of po- and Justice government [web log com- litical science, 1(1), 315-331. Przybylski, ment]. Retrieved May 5, 2019, from W. (2018). Can Poland’s Backsliding Be https://blogs.lse.ac.uk/europpblog/2017/10/26/expl Stopped?. Journal of Democracy, 29(3), 52- aining-the-popularity-of-polands-law-and- 64. justice-government/ Reeves, P., de Chernatony, L., & Carrigan, M. Waldner, D., & Lust, E. (2018). Unwelcome change: (2006). Building a political brand: Ideology Coming to terms with democratic backslid- or voter-driven strategy. Journal of Brand ing. Annual Review of Political Science, 21, Management, 13(6), 418-428. 93-113.

105 Capstone Issue Vol. 12, 2019 Christina Keßler

Appendix A

106 Capstone Issue Vol. 12, 2019 Christina Keßler

107 Capstone Issue Vol. 12, 2019 Christina Keßler

108 Humanities - Social Sciences

Between the World and Me(me): Approaching Internet Memes through Digital Ethnography and Berlant’s Affective Present

Cullen Ogden

Supervisor Dr. Alexandra Brown (AUC, UvA) Reader Dr. Pedram Dibazar (AUC) Capstone Issue Vol. 12, 2019 Cullen Ogden

Abstract

This research addresses the massive number of internet memes encountered daily and over time to consider how the technological capabilities of the internet have been able to open new avenues for relationality and subjectivity, fundamentally shaping how we perceive what’s true, what’s right, and what’s possible. By combining a two-step methodology of gathering auto-ethnographic, qualitative research and then analyzing this data through Lauren Berlant’s theory of the contemporary affective present, this project argues that memes constitute a genre of the contemporary present, a subgenre of comedy made possible by digital technology. From this claim of the meme as a genre, this research extends Berlant’s theory by developing the terms self-connection and virtual intuition to speak to the respective (re)shaping of subjectivity and intuition today.

Keywords and phrases: everyday life, memes, Lauren Berlant, affective present, genre

Acknowledgements Introduction:

Encountering Berlant Thank you to Dr. Robert Y. Chang, my ﬁrst anthropology professor who taught me the magic of Lauren Berlant’s Cruel Optimism is a ground- questioning the world around us, and who showed breaking demonstration of the ambitious and pow- me that a study of culture can encompass any in- erful work that can be done through considering terest you can think of. In addition, thank you to the role of affect in the shaping and apprehension Dr. Pedram Dibazar, who helped me discover the of the world today (2011, p. 4). Berlant asserts beauty of work done between anthropology and that, “the present is what makes itself present to cultural analysis, who has been helpful and sup- us before it becomes anything else, such as an or- portive not only as a professor but as a friend. I chestrated collective event or an epoch on which would also like to extend an enormous thank you we can look back” (p. 4). This is to say that al- to Dr. Alexandra Brown, my supervisor, whose sup- though the present has substance which we ﬁrstly port through this entire process has been so forma- feel, this present is molded by the forms we give to tive that I would have never considered a project of it — forms both temporal and aesthetic which can this nature without her guidance, and to whom I’m carry a profoundly political purpose. By assigning thankful for introducing me to the academic work the present such forms as a crisis, an interruption, that continues to grab my attention. a situation, or a habit, these forms each shape and alter “what forces should be considered responsible and what crises urgent in our adjudication of sur- In addition, to my parents, Carolyne and Chan- vival strategies and conception of a better life than dler Ogden, thank you for constantly supporting what the metric of survival can supply” (Berlant, me in everything I do and encouraging me in all of 2011, p. 4). my oddity. To Manar and the Cult of friends (who feel like family) that I’ve made in life who have Forms matter because today, life has changed fearlessly supported me, who have been there for and feels particularly precarious and turbulent. me at every step of this process, and who have This sense of precarity is not a new condition in heard me attempt to explain that wild and wonder- the world, but its growing presence in the lives of ful thing that is “affect” too many times to count – those previously unaffected has generalized this Thank you. condition to be perceived as an overarching tone of current conditions at large, whose effects and And last, but certainly not least, thank you distribution remains unequal (Berlant, as found in to Lauren Berlant, whose work has been a guid- Puar, 2012). To consider what this precarious world ing force in this project, and who taught me the of today may feel like, Berlant theorizes the present power and potential of articulating that something as punctuated by a sense of crisis ordinariness in to which we all strive for. which the world is disorganized by capitalism and other forces that demand the subject to scramble

110 Capstone Issue Vol. 12, 2019 Cullen Ogden

for new ways of living, “rhythms that could, at any coincidental, and conditions how we adapt and ad- time, congeal into norms, forms, and institutions” just to these forms of the present. By exploring (2011, p. 9). Part of what reinforces Berlant’s claim the many different forms and genres which can to the disorganization of the world is the fraying conflict, contest, and challenge each other, we can of what she sees as the “moral-intimate-economic ground this study of affect in material narratives thing called ‘the good life,’” or the conventional and stories which emerge as genres. Focusing on fantasies of what it means to have a life, what this these newly emerging genres gives us a way to survival is for, and how the world “add[s] up to consider how the contemporary present is under- something” (2011, p. 2). As the system appears stood and encountered as it unfolds. not to have radically changed but rather increasingly fails to reciprocate the promises of life before So what constitutes cruel optimism in the in its unchanging nature, these fantasies are fray- present? What everyday actions, genres, and prac- ing. Such ‘good life’ fantasies include: tices enact the affective structure Berlant identifies? This capstone explores the relevance of Berlant’s theory in examining one of the most si- “upward mobility, job security, political multaneously prolific and quotidian elements of the and social equality.. . lively, durable in- contemporary present, which I will argue provides timacy.. . meritocracy, the sense that a new genre for apprehending and experiencing liberal-capitalist society will reliably pro- the present today: the meme 1. vide opportunities to carve out relations of reciprocity that seem fair and that foster life as a project of adding up “I feel all sorts of ways – angry at a to something and constructing cushions Trump tweet, laughing at a meme about for enjoyment” anxiety but also feeling shame, I find another account which enjoys astrology (Berlant, 2011, p. 3). and I follow them which is exciting, and then I laugh but I’m sad and worried The title of her work, Cruel Optimism, thus reflects about the environment. My thoughts a general affective structure found throughout life aren’t everywhere, exactly, but they’re today; we, as people in this precarious world, re- many places which I’m entering and main attached to objects which we believe build moving between within a single screen. towards this ‘good life’ even as these objects no But within this chaos, I’m still learning longer aide us in achieving these fantasies and in and laughing and imagining absurd sit- fact inhibit our achievement in the process of at- uations and a lot of funny things, and tachment (Berlant, 2011, p. 1). Specifically, Cruel I’m learning new slang and new jokes Optimism is about how ordinary life is home to and new ways of relating in themselves. these conflicting forces of old fantasies and new Maybe I retweet some of the ones I find disappointing realities, and how in renegotiating really funny, I definitely like the ones I what it means to have and live a life, we find that find funny even if I don’t retweet them, adapting to and surviving life today may be as but many I may just scroll past.” close to this good life as we can hope for (Berlant, 2011, p. 3). While to speak of cruel optimism (observation, March 11th, 2019) and more generally of affective structure is not to speak directly to the experience of the structure, The contemporary circulation of and engagement but to recognize how the act of giving shape to an with memes, just as Berlant’s affective present, in- amorphous ongoing present has effects and conse- volves: (i) the setting of expectations, (ii) the shap- quences. Description then also becomes prescrip- ing and reshaping of the present, and in particular, tive: the difference, for example, between ‘climate (iii) an identifiable affective structure which reflects change’ and ‘climate crisis’ is neither neutral nor particular facets of the contemporary present that

1This research focuses on the Internet Meme speciﬁcally, so the use of the word meme is with the assumption that we are discussing memes as they occur online. See Chapter 1 for a working deﬁnition of meme.

111 Capstone Issue Vol. 12, 2019 Cullen Ogden

Berlant describes. While Cruel Optimism takes digital anthropology as an evidential basis for in- a variety of cases from between 1990 and 2010 terpreting memes through the theoretical frame- which capture this titular affective structure, this work offered by Berlant’s concept of the affective project extends Berlant’s affective present to in- present. The digital ethnographic work produces clude social media and digital technology in the an extensive archive of instances from which to emergence of new genres. Considering the meme deduce genres without requiring a classificatory in its context of social media and everyday life definition of the meme to haunt over the fieldwork requires new language which can capture the dy- process. namism and flux behind the term ‘meme.’ Thus, this interdisciplinary research project To think of memes within the context of the af- firstly collected qualitative, auto-ethnographic data fective present is to allow the meme to speak to from engaged participant observation on social and from everyday experience, by identifying pat- media networks to survey memes which were cir- terns across a variety of memes. What is a meme culated between March and May of 2019. Taking at its simplest? In considering how the present the everyday act of scrolling on one’s phone and is the space where we encounter and gather vari- looking at memes as an (auto-)ethnographic field- ous forms of knowledge and develop intuition about site, my fieldwork consisted of three visits of two what’s happening and what might follow from this, hours each via laptop computer and mobile phone what kind of knowledge and intuition do memes to the social platforms of Twitter and Instagram, require? How does the sheer volume of memes which took place on March 11th, April 3rd, and May under continual production reflect the contempo- 4th, 2019. Through a trace ethnographic approach rary present, as identified by Berlant? An interdis- to engaging with memes as digital ‘documents’ ciplinary approach that combines digital ethnogra- within the Internet ‘network,’ this process involved phy, autoethnography, affect theory, and literary deep dives online, wherein I would emulate the criticism will be used to consider what and how we everyday practices of browsing and scrolling to en- can learn about today’s world through memes. counter and engage with memes in a variety of forms and digital locations (Geiger Ribes, 2011). Although it would seem digital worlds would require Project Methodology new ways of understanding what it means to participate in comparison to classical anthropology, Tom The aim of this project is to understand the Boellstorff’s work Ethnography and Virtual Worlds meme by considering what an affective approach provides insight into the validity of participant ob- to the meme can begin to tell us about memes, our- servation both on- and offline, or wherever social- selves, and our world. To accomplish this, Berlant’s ity may be occurring (Boellstorff, 2012, p. 65). formalist approach to locating affect within cultural The work of Larissa Hjorth, John Postill, and Sarah objects and discerning from them a sense of the Pink has also been impactful in demonstrating in- historical present lets us place memes at the inter- novative approaches to considering digital cultures section of the everyday and the larger contempo- through ethnographic engagement. rary present. Through the emulation, reproduction, and potential habituation of the meme form(s), During the ethnographic work, memes were genres are shared and negotiated as active agents encountered not only via my own networks, but in shaping our perception of the present, which es- also by utilizing tags and searching features to see tablishes the gravity of these funny little things. what the social media algorithms of Twitter and Berlant’s theory is productive for an analysis of Instagram might produce – behaviors potentially memes, especially in considering how they func- exhibited by those who have yet to discover what tion and move within everyday life (which will be a ‘meme’ is. Other collection methods included further established in Chapter 1). However, her acts such as writing posts that asked explicitly for own methodology of generalization through read- memes, messaging a number of friends asking for ing cultural artifacts proves difficult to apply to a memes they recently came across, and predomi- single meme or to the blurry definition of memes nantly autonomous navigating of these networks. as a whole. Thus this paper employs methods of By using my own social media accounts, my own

112 Capstone Issue Vol. 12, 2019 Cullen Ogden

experience of these encounters was already fil- affective present. Chapter 3 considers how the tered through the networks of followers and fol- meme affects subjectivity, while Chapter 4 will ad- lowing that I had developed on each platform re- dress how the meme effects intuition. Finally, the spectively. To prevent reliance on my self-curated conclusion will reflect on this research and the work timelines, I utilized the interactive aspects of new of theorizing comedy. media to follow the trails of sociality through the interface instead. Additionally, documenting the entire process through fieldnotes has provided an archive to revisit in further considering the devel- I. Memes Among Us: Memes opment of this argument. and/in Everyday Life

In the second step of this interdisciplinary Limor Shifman, a predominant contemporary project, I incorporated Berlant’s theory of genre scholar on memes, defines an internet meme as, and the affective present to code and analyze my “(a) a group of digital items sharing common char- findings from my digital ethnographic exploration. acteristics of content, form, and/or stance; (b) that In addition to concepts developed in Cruel Opti- were created with awareness of each other; and mism, Berlant’s additional work on precarity and (c) were circulated, imitated, and/or transformed comedy provide useful supplements to this frame- via the Internet by many users” (Shifman, 2014, work of genre. Berlant develops a broad notion p. 8). While this definition draws on the use of the of genre to describe any affective structure which term ‘internet’ to develop a definition of the object trains our expectations of watching something un- in relation to its medium (the Internet network), fold, a useful conceptual tool for an affective anal- this does not consider the ambiguity of the term ysis which will be further examined in Chapter 2 as such and the confusion in defining its multiplic- (2011, p. 7). As the first chapter will describe, ity. To answer what the ‘meme’ at the heart of the Berlant’s work provides a relevant and provocative ‘internet meme’ is requires demystifying the term framework for thinking about everyday life in the ‘meme’ and its double meaning within colloquial contemporary present by highlighting how genre common knowledge and academic use. works to (re)frame and (re)shape the present as and after being perceived.

Memes Ultimately, this research argues that memes constitute a genre of the contemporary present, Meme was first coined by Richard Dawkins in operating within temporal genres which have been 1976 as a biological analogy applied to culture, but enabled and heightened through digital technolo- only started to gain pop-culture traction after Mike gies. As a genre, memes therefore function in mul- Godwin published an article about Nazi memes in tiple ways: first, by developing new modes of sub- Wired magazine in 1993 (Shifman, 2014, p. 1; jectivity through engagement with memes via so- Godwin, 1993). Defined loosely here as ‘units of cial media; and secondly, by heightening our sense culture,’ memes arose as a tool for exploring how of intuition whereby we come to relate in new ways the transmission of culture and particular cultural to the world and its forces. ideas may be similar to that of genes, or at least involve similar processes such as transmission, du- In Chapter 1, this research reviews how memes plication, and mutation. Some eagerly jumped on and everyday life (via anthropology and everyday board to this pursuit, with even the field of ‘memet- life theory) have been studied in the past to contex- ics’ briefly developing in hopes of establishing the tualize Berlant’s and my own work. In Chapter 2, I study of memes in academia. However, critics of define and elaborate on Berlant’s concept of genre this discipline have been unable to reconcile the and introduce the genre of comedy to consider how flawed origins of the cultural-biological metaphor, memes extend a particular form of comedy into the with worries over agency and transposing scientific digital. From this definition, the last two chapters theory onto culture (Shifman, 2014, p. 10; Jenkins take up the meme as genre to consider the effects et al., 2009; Conte, 2000). Many who have worked and implications of this structure in shaping this on memes and confronted these issues have cho-

113 Capstone Issue Vol. 12, 2019 Cullen Ogden

sen to simplify the unruly meme by considering not start out with predesignated outcomes” (High- it within a specific disciplinary context. This has more, 2008, p. 3). While the study of everyday life illuminated only a small piece of the dynamic pro- became more explicit in the latter half of the 20th cesses involved in the production of even a single century, Highmore claims that the origins of turn- meme, and isolated the meme from its context in ing to everyday life first appeared in the works of circulation. Sigmund Freud and Karl Marx (2014, p. 6). Through a framework of skepticism, Freud and Marx planted This can be seen in a variety of fields as memes the seeds to challenge the regularizing and totaliz- are conceptualized within a specific discipline, and ing power of theory in producing notions of agency, have been considered in the domains of philoso- determinism, and rationality that had been present phy (Sterelny, 2006; Holdcraft Lewis, 2000), psy- during the 19th century, and in fostering critical chology (Blackmore, 1999; Blackmore, 2001), an- perspectives on everyday life (Highmore, 2014, p. thropology (Marwick, 2013; Shifman, 2014; Nis- 6). senbaum Shifman, 2015; Banet-Weiser Miltner, 2015; Saito, 2017), linguistics (Zenner Geeraerts, One of the first scholars to directly engage with 2018), politics (Borenstein, 2004; Coker, 2008; the conceptual realm of the everyday and attempt Dagta¸s,2016),ˇ as well as communication studies to theorize it was Henri Lefebvre. In his Critique and aesthetics (Dennett, 1990; Rossolatos, 2015; of Everyday Life, he claimed that everyday life is Gal et al., 2016). My concern with these previous defined by contradictions and stands at the mag- approaches is that investigating a specific meme ical intersection of “illusion and truth, power and from a specific context requires assuming a stance helplessness; the intersection of the sector man toward memes that can distort or generalize par- controls and the sector he does not control” (Lefeb- ticular cases in efforts to fulfill the aim of the disci- vre, 1992, p. 21). His critique of everyday life took pline at large. A solution to this could be to compile issue with the different rhythms that all come to- these different investigations to produce a fuller gether in the everyday, and he argued that by rec- image of what an internet meme is. On the con- ognizing these conflicting forces we would come trary, one could simply claim that all studies of to recognize that the everyday is where capital- memes are destined to fail due to the impossibil- ism reproduces itself (Lefebvre, 1992). It is from ity to ever include all memes and this potential for this capitalist colonization of the everyday that we a meme to contradict any findings. Alternatively, I come to normalize different conflicting narratives would like to ask: what if we instead framed memes as they occur in our lives, and as such the every- as a part of the “residual of life,” that excess which day is where we can investigate the rhythms that cannot be captured by simply one single approach people embody and habituate in the present. but instead requires new paradigms for consideration (Highmore, 2008, p. 3)? What if we position Although Lefebvre provides a dynamic image of ourselves to academically approach memes in the everyday life, it is Michel de Certeau’s The Practice same way we personally encounter them, not in of Everyday Life (1974) that is widely considered isolation but instead in the chaos within which we one of the key works in the study of the every- happen upon them – within the everyday? day. De Certeau’s theory recognized the conflicting forces at work in everyday life by distinguish- ing between strategies and tactics. Institutions and macro-level shaping of society by capitalism occurs Everyday Life through strategies, which people develop tactics To turn to the everyday as a means of inter- for living in, through and against. Berlant believes preting the meme is not an easy task, but cul- that the problem with theory such as de Certeau’s tural studies and anthropology offer two differ- is that it no longer accurately accounts for the situ- ent yet complementary trajectories for theorizing ation most people are living in today (2011, p. 8). this. Considering first the cultural studies approach to everyday life upon which everyday life theory It must be noted that alongside the develop- has been constructed, the everyday develops “a ments in cultural theory from Marx to Lefebvre, kind of heuristic approach to social life that does a correlated trajectory in anthropology beginning

114 Capstone Issue Vol. 12, 2019 Cullen Ogden

with Ruth Benedict and Bronisłav Malinowski has affect, and perceptions online through ethnogra- been concerned with the act of studying culture it- phy has been extremely impactful in illustrating self through developing a methodology for access- how cultural analysis and anthropology can be ex- ing everyday life (Goldenweiser Benedict, 1937; tended to digital culture through new considera- Malinowski, 1922). In order to develop academic tions of how deeply ingrained digital technology validity to the claims and observations made by an has become in everyday life (Hjorth Pink, 2014). anthropologist as to how life is lived somewhere, Additionally, Ralph Schroeder’s work recognizes participant observation emerged as the best so- the need for social theory to account for the im- lution, requiring that “the researcher participates pact of digital technology, and proposes new the- with people in commonplace situations and every- oretical conceptions of the everyday in doing so day life settings while observing and otherwise (Schroeder, 2018). collecting information” (Jorgensen, 2015, p. 1). Instead of generalizing to speak of all people or Another emerging approach that anthropo- all cultures, the early twentieth century with the logical and cultural theory both took inspiration work of both Franz Boas in the United States and from and were influenced by is the affective turn Bronisłav Malinowski in the United Kingdom and Eu- (Clough, 2008). From the mid-1990s onwards, the rope set the standard for participant observation as Humanities and Social Sciences have seen a surge the best way to discuss a specific culture’s customs in attention to the affective which has been at- within their own terms, in hopes of breaking from tending to “a dynamism immanent to bodily mat- the imperialistic past of anthropology (Jorgensen, ter and matter generally – matter’s capacity for 2015). While not overtly operating under the rubric self-organization in being in-formational” (Clough, of everyday life theory, participant observation’s 2008, p. 1). This can be described as theorizing and extended exposure and participation in everyday focusing on impact or qualitative changes, through life has enabled scholars to understand how the ev- focusing on a more dynamic understanding of the eryday may be radically different around the world, way things impact each other in the world. By but also allowed them to discover commonalities developing new methods and theories for affect, between these disparate processes. These meth- scholars have been able to discuss ephemeral real- ods have remained at the heart of anthropology, ities in a newly concrete and tangible way. with ethnography standing for an active participation in one’s hybrid position within and outside of To position this turn within academia, the theo- a culture while exposing that which may be hidden retical basis of the affective turn has roots in a long to those who are too close (Jorgensen, 2015). history of queer and feminist consideration of emotion (Ahmed, 2015, p. 9). In anthropology, emotion The colonial legacy that anthropology carries has emerged as a point of attention from the early has left the political and ethical lines of ethnogra- work of Lutz Abu-Lughod (1990), through more re- phy contested, but has also encouraged scholars to cent diagnoses of anxiety and fear as the experien- consider cultures which may have been historically tial dimensions of globalization (cf. Bauman, 2018, ignored or excluded from studies. One example of Comaroff Comaroff, 2003). Similarly, the work these new fields is the work of digital anthropol- of experimental anthropologists such as Michael ogists, particularly those interested in the cultural Taussig’s The Nervous System and Kathleen Stew- and experiential functions of digital technology and art’s Ordinary Affects have challenged how ethno- the users of such (Hjorth Pink, 2014; Postill Pink, graphic writing grapples with this dynamism and 2012; Hjorth Cumiskey, 2018). Through focus- attends to the something which has force in this ing on aspects such as routine, movement, and world but may not be easily apprehended in dis- sociality in digital spaces and via digital technol- course (Taussig, 2016; Stewart, 2001). In the Hu- ogy, digital anthropologists have attempted to ad- manities, this affective turn inspired work which dress the significance of life online and the par- focuses on affective structure and the formaliza- ticularly apparent gap between the adoption of tion of it in producing cultural objects such as in digital technology and its subsequent academic Eugenie Brinkema’s The Forms of the Affects and interrogation (Murthy, 2008). Within this domain Sinnae Ngai’s work on tone (Brinkema, 2014; Ngai, of study, Larissa Hjorth’s focus on the everyday, 2007). This consideration of affect’s form is also

115 Capstone Issue Vol. 12, 2019 Cullen Ogden

what inspires Lauren Berlant’s research, working this present moment is like. She mirrors Jameson’s to read genres of the affective present as affective structure but denotes this period instead as “the structures (see Chapter 2), which provides a rig- waning of genre,” wherein the traditional realist orous tool for approaching a variety of forms that genres of the past no longer hold under the current may appear. This approach provides the basis for conditions of what it means to have and build a life considering memes from within the everyday. (emphasis my own, Berlant, 2011, p. 6). Gener- ally, genre can be considered “a classification of To address the dynamism of memes from the type or kind. . . [which] regulates the narrative proposition of the everyday, Lauren Berlant’s theory cess producing coherence and credibility through of the affective present returns to Lefebvre’s ap- patterns of similarity and difference”, but Berlant proach to rhythms and patterns in opposition to stretches this term to encompass both lived and de Certeau’s approach. As discussed in the intro- depicted processes which we experience and repli- duction, Berlant proposes that capitalism and other cate (“Genre”, 2004). This chapter takes up this forces of today disorganize the everyday, leading curious concept of genre in order to to disentangle to a sense of crisis ordinariness wherein the failure the multiple meanings and uses that genre has and of past fantasies about having and building a life to develop an argument in support of the meme leaves us searching for new ones. Although this as a genre in itself. This process will incorporate process of finding new genres of living and adapt- Berlant’s more recent work with Sianne Ngai on ing to this crisis ordinariness in some ways ad- comedy to consider how the meme functions as a dresses this disorganization and attempts to over- new genre of comedy in society. This chapter will come it, it does not change the enduring need and conclude with an ethnographic example of a meme, desire for a good life (Berlant, 2011, p. 7-8). In to consider some of the implications of claiming the this chaotic present, Berlant develops a formalist meme as a genre. approach to affect which focuses on the dramas of adjustment and produced narratives of the new pressures of building a life. Within this, she fo- Genre cuses on the place of fantasy and the promises of the future via affective structures to explore how The term genre is traditionally used in classify- these train our intuition of the present in their con- ing literary and artistic objects, but Berlant trans- sumption and replication. Using Berlant’s criteria poses it onto the realm of affect theory to attend for genre to consider the meme, the next chapter to the particular structures and conventions (aes- will define genre and develop an argument for the thetic, temporal, and ways of living) which make meme as genre. the contemporary present apprehensible, tangible, and set expectations for future intuiting in the process. For example, Berlant considers such diverse genres in her work as “the situation, the episode, II. On Genre - The Meme the interruption, the aside, the conversation, the travelogue, and the happening” to explore how the In 1984, Frederic Jameson wrote his now cele- allocation of agency and structuring of temporality brated essay ‘Postmodernism, or, the Cultural Logic capture a particular perception of the world and its of Late Capitalism’, in which he discussed a partic- possibilities (2011, p. 5). Genre can thus be seen ular shift in both art and life that he named the as the mediator which transforms the amorphous ‘waning of affect’, stemming from his perception of present into a mediated affect, sensed and under increasingly depthless artwork (Jameson in Norton, constant revision as each iteration of a genre fur- 2018). The legacy of this argument remains, but ther establishes how the forces may play out in this Berlant conjures up a different and contrary ap- particular structure. This desire for new genres in proach to periodizing this contemporary present by the failure of the old ones show the “wearing out leaning further into affect. Berlant’s theory shifts [of] the power of the good life’s traditional fantasy the logic of Jameson to embrace a non-essentialist bribe without wearing out the need for a good life” approach, focusing on the means through which we (Berlant, 2011, p. 7). By turning towards new gen- build intuitions and expectations of what being in res and their new fantasies, we can consider how

116 Capstone Issue Vol. 12, 2019 Cullen Ogden

the contemporary present when framed in different mism as an affective structure involves “a sustain- ways “provide[s] an affective expectation of the ex- ing inclination to return to the scene of fantasy that perience of watching something unfold, whether it enables you to expect that this time, nearness to is in life or art” (Berlant, 2011, p. 6). These new this thing will help you or a world to become differ- genres emerge as new styles of managing “the ent in just the right way” (Berlant, 2011, p. 2). An- narratives of what’s going on and what seems pos- other example can be seen in describing the genre sible or blocked in personal/collective life” (Berlant, of the situation: “a situation is a state of things in 2011, p. 4). The question is, then: in what partic- which something that will perhaps matter is unfold- ular ways do memes as a genre set expectations ing amid the usual activity of life” (Berlant, 2011, and give shape to the present? p. 5). By establishing a form which can take on a variety of objects within it, affective structures and The need for new genres such as the meme is genres allow for specificity of identification without not simply an organic emergence at the hands of closing off subjectivity and situational specificity. digital technology, but emerges from a particular In this case, the ability of the ‘situation’ to apply historical present in contemporary society wherein to both the situation comedy and the police inter- our conditions are no longer represented by the rogation (“We have a situation”) emphasizes the genres of the past. The titular concept of Berlant’s absorptive nature of genre in Berlant’s use of the theory (cruel optimism) describes a structural re- term (2011, p. 5). lation in which. As discussed in the introduction, Berlant claims that the affective structure of cruel As such, paying attention to affective structures optimism as a sustained attachment to “some- allows for a formally grounded consideration of thing you desire [which] is actually an obstacle how relating to the world is not a neutral act but to your flourishing” (2011, p. 1) is reflective of is rather a process which involves the participation the structure of our contemporary present as a in and formation of new genres as they emerge. To whole, touching upon the anxiety surrounding fail- pursue a cohesive definition of the meme, the next ure in reciprocity from larger societal systems. Al- section will present Berlant’s work on comedy to though the factors which compose our world such position the meme as a new genre emerging within as institutions, the state, and our environment con- comedy, and extending the author’s work to the tinue to fail and disappoint us in returning the work particular context of digital technology. we put towards them, we can still recognize our strong, continued attachment to these factors and our continued investment. To handle this sense Comedy of extended crisis, we do what we have to; we all participate today in the improvisation of gen- Berlant’s account of comedy is useful for clarify- res (simply put, we try out different ways of being ing the complex and ambiguous place of comedy in and living), where the emergence of new genres life today. Berlant describes comedy as “always a today reflects our striving for stability and ways of pleasure-spectacle of form’s self-violation” (Berlant being in a world that feels without any. These gen- and Ngai, 2017, p. 243). It involves the watching of res, which are made physical in a variety of lived something which somehow defies or violates itself, actions and artistic depictions through replicating creating a spectacle from which to derive pleasure. affective structures, have the potential to reconfig- This means that comedy itself is a particularly am- ure our relations to our fantasies of the good life biguous and contextually determined form where and what can be a good life for us. deception and inverted logics provide pleasure. The comedic can involve questions of subjective It is important to specify what I mean when perception, such as “what’s living, what’s mechan- speaking of affective structure: although the expe- ical, and who needs to know”, which also estab- rience of an affective structure may feel any variety lishes comedy as the scene of aesthetic, moral, of ways, the affective structure itself is one which and political judgement (Berlant Ngai, 2017, p. can be deduced through formal consideration of 234). At its core, comedy is based on intuitive in- subjectivity, fantasy, and form coming together in tersubjectivity, whereby we test out what is shared something (Berlant, 2011, p. 2). For example, opti- and not shared in a more practical and discreet way

117 Capstone Issue Vol. 12, 2019 Cullen Ogden

than any traditional theory of comedy claiming to Figure 1 – Ron Swanson Meme (taken from be objective could (Berlant Ngai, 2017, p. 235-6). observations, March 11th, 2019) It is difficult to create a theory of comedy because we feel out what is comedic, with the formal lines of Case Study: A Meme about Memes humor as experienced contingent upon the events and contexts of the world filtered through a subjec- “The meme in question was sent to me tive lens. As a result, we can view comedy as the on my first field site visit, during which pleasure of this reevaluation where we renegotiate I messaged a close group of my friends what in the past had been perceived as harsh lines in a group chat where we frequently ex- (Berlant Ngai, 2017, p. 235). change memes. Asking for any memes that they had seen recently, many of One of Berlant and Ngai’s assertions about com- the memes I was sent in this way were 2 edy which I find particularly productive for this dis- static images , as they were able to cussion is the suggestion that the ‘comedification’ be saved from social media and directly of everyday life has placed comedy as an “over- sent to other users through personal arching tone of late capitalist sociability, affecting messaging. This was the more tradi- how people self-consciously play as well as work tional form of the meme in the past, altogether and the spaces where they do so (in- though the development and changes cluding Twitter, Facebook, Snapchat, Instagram, in social media affordances have al- and YouTube)” (Berlant Ngai, 2017, p. 237). As lowed for the inclusion of GIFs which this is one of the only mentions of social media loop a drama infinitely, or a video which explicitly in Berlant’s work, the inclusion of com- adds an audio-temporal dimension to edy in the discussion of memes as genre is not the drama at hand. The top text in this only warranted, but imperative to understanding meme (see Fig. 1) reads “when the re- the centrality of comedy to this realm of the ev- latable meme is funny but hits one of eryday. Therefore, considering memes as a new your biggest insecurities,” and the ma- genre of comedy within the digital shows how the jority of the image has a picture of an contingency of the meme’s humor is safeguarded angry or disappointed man (the char- by increasing the number of memes one may en- acter Ron Swanson from Parks Recre- counter, suffusing much of life with comedy. To ation), with subtitle text on the bottom start outlining some of these dynamics involved in that indicates laughter in the original the work of a meme and its relation to comedy, the scene.” rest of this chapter will use an ethnographic case (observation, March 11th, 2019) study to explore the general structure of a meme and its relation to comedy as a whole. As can be seen from this ethnographic attempt at describing the meme, memes generally have an ambiguous appearance. While the meme itself is singular, it is never in isolation but rather defined in relation to the variety of memes which come to be absorbed within the genre of meme and further reproduced under this guise (i.e. Ron Swanson’s character, “When you. . . ” memes structure, using a cropped Netflix screenshot). Additionally, the combination of a framing phrase and a reaction image shows how a meme involves the work of context and content in their loosest interpretation. In a meme, context(s) (the form of a setting) and content(s) (the things which are held within this) can layer, mix, persuade, and combine to develop

2In order to disseminate across different platforms, many memes are either directly exchanged as images or are screenshotted and then disseminated through the system. Some of the memes I received directly as images are included in the conclusion.

118 Capstone Issue Vol. 12, 2019 Cullen Ogden

new forms of content which are further contextu- representation, however flawed or misdirected the alized to develop more comedy. In this meme, we recognition in itself may be. This problem of (in- see the context establishing a time but the place is ter)subjectivity will be further explored in the next not necessary. The language of this context estab- chapter 4. lishes a disjunction between the subject and object in both the act of reacting to memes and in the Memes may be spoken of in passing as simply subject who comes to ‘view’ it. By including an im- the source of a small chuckle, but it is this invest- age of a man looking angry but with the subtitle ment of the self into the meme through recognizing of laughter, this disconnect is even more comical and living (in this case, playing out one’s biggest as we identify the meme’s ability to capture what insecurities) that we can begin to see the serious feels like a disjuncture of our experience through dimension of memes. The affective structure, the the content of this context. genre, or set of formal expectations which memes produce indicates how the pleasure and recognition Turning to the ambiguity of the context, we of projecting the self (towards the meme and back come to see how the instability of language to the self) attaches one to fantasies which may emerges. Rather than this destabilizing of memes play out in the meme via this particular object. leading to a rupturing in meaning, meaning multi- In this example (Figure 1), my biggest insecurity plies as the meme absorbs and comes to stand for becomes generalized to part of the affective expe- an entire complex of objects, experiences, mem- rience of viewing memes as a whole while I interact ories, dreams, and forces brought to life through with the meme, where my own experiences cannot one’s ability to ‘feel out’ the scenario. Consider- be excluded in considering these forms. This meme ing my own experience of this meme, I know my (see Figure 1) depicts the doubling effect of think- biggest insecurity, and this meme may suddenly ing about memes through a meme, showing how lead me to consider not only this insecurity but its the process of viewing memes is made tangible in derivative pleasure as this meme transforms this itself as an object of consideration by the meme. vulnerability into an object of humor for myself. By making explicit the conflicting forces involved in What’s important to recognize is that the place- my own extraction of humor from a joke about my holder of ‘my biggest insecurity’ provides access to insecurity, the affective structures of memes as a a deeper awareness of the self through considering whole is made apparent. This is done by character- at the time of viewing what one’s biggest insecu- izing the scene of viewing memes as a framework rity is; but this knowledge is only brought up for for making these connections and renegotiating in- one’s own interpretation and does not travel with tersubjectivity. The meme draws in the viewer to the meme – making space for vulnerability without invest themselves in the scene, allowing them to necessitating disclosure. As for what we bring to feel out new affective relations to their own expe- memes, this insertion of the self into a hollow con- riences via this content by, in this case, engaging text acted out by a relatable representation as in with their own insecurities in the past. Thus, it is this meme is the work which invests the self into the potential for new and different modes of inhab- the digital while still cultivating some sort of fur- iting the self to emerge which furthers the self’s ther knowledge of the self in the process. Contexts quest for development by viewing and interacting come to build, multiply, and challenge meaning as with a meme. The collective body of memes as the interface allows more engagement to layer nor- this growing space for new and complex ways of mative modes of behavior in these different con- considering how we relate to the world and what flicting realms. Thus, we, too, can re-contextualize, it means to have and to build a life at large, is as decontextualize, and find ourselves in an array of inseparable from the meme as its status is contin- situations made tangible by the meme. For exam- gent within the context of memes macroscopically. ple, the vignette featured in the introduction 3 highlights the rapid movement of the subject between The meme considered within its inextricably different scenes within a single screen. But still, a interlinked relation to the entire body of memes meme only works when we are able to inhabit that provides the space for developing new, less rigid

3See end of the Introduction 4See Chapter 3

119 Capstone Issue Vol. 12, 2019 Cullen Ogden

forms of relationality which the world increasingly recently posted to me. Within an in- requires of us in the hopes of deriving pleasure stant, I’m confronted by an acquain- from recognition itself. The viewing of memes is tance from New York announcing that a practice of relief (as I will further elaborate be- he’d just finished some important ex- low), but it is, nonetheless, not devoid of or divided ams, a video of FKA Twigs’ makeup that from feeling out the practical necessities of the was retweeted by someone I don’t represent. It involves the assessment and judgement member which someone posted saying of different means of managing a life, and involves how she has no flaws, and a post from a the celebration of both success and failure as it is drag queen which was liked by someone reproduced beyond our own precarious inconsis- I follow named ‘b’ with a video of some- tency. Thinking of internet memes as a subgenre of one putting on a clown costume com- comedy provides a new lexicon for attending to the menting on how she would be wearing affective work in how memes come to mediate an that outfit tomorrow (“me tomorrow”). otherwise confusing and overwhelming world with- I start to go through them one by one, out collapsing into another discussion of comedy. In working my way down to the line break categorizing and setting expectations of what view- which starts a new flow of posts under ing memes may be like, such that we expect and the heading of ‘In case you missed it’. thus demand comedy from the meme, our own set (Usually, it’s not in case but rather what expectations are troubled in the process. The rest I did miss at some point, what have I of this research will take up the aspects of subjec- missed?) tivity (Chapter 3) and intuition (Chapter 4), which are both important in considering how the meme I stumble across another post on my as interlocutor between the affective present and feed (see Figure 2) and I look at it. First, ourselves is (re)negotiated from both top-down and I see someone I follow who retweeted bottom-up means of control and engagement. the post and captioned it as, “wait they ATE,” which I look up on Urban Dictio- nary to see means that ‘they’ performed something well. I open up the post (see III. Making Me(me)s – Subjectivity, Fig. 3), intrigued by a high-speed video Composure, and Self-Connection of an animated sequence of people who have been personalized to appear to be “I picked up my phone, saw some notifi- Homer and Marge Simpson, dancing for cations of a new text, some news, and 40 seconds to a recently viralized por- that app I kept forgetting to delete. I tion of Nicki Minaj’s Roman Holiday that ignored them and unlocked my home has been edited over many videos. The screen, pushing upward and away what original post says, “Nobody: | My brain at the time was a richly-saturated sur- at 3am:,” with a video completing this realist body painting, painted on and rebuttal to no one’s action. They per- by Dain Yoon’s face with a distorting ef- form even faster than the edited song, fect which makes her ripple in front of a showing that this sequence must have red background. Tapping on the twitter been performed in some context, and icon and entering into the light blue and then found and added to this. I laugh, white interface, returning instantly to a because between the speed, the seem- post which I had last seen earlier in the ingly illogical organization, and yet still day, I find myself back to a space which skilled performance, something in the is familiar to me. meme really does capture what it feels like my brain is doing at 3am. I don’t A quick press of the top bar, and I find watch a lot of the Simpsons anymore, myself whisked abruptly past an ever- but still, I relate.” growing number of tweets to whatever post happens to have been the most (observations, May 4th, 2019)

120 Capstone Issue Vol. 12, 2019 Cullen Ogden

ment of ﬁeldwork, is like an inside joke. That is, to see a meme is like being inside of the joke, or to complete the circuit of the meme and to engage a moment of recognition through inhabiting and feeling out a space, time, or moment such as my brain being represented by this video. It is not the fact that I know who the Simpsons are, or because of a particular afﬁnity with Nicki Minaj, but rather the combination of disparate objects into a context which I am attached to (these things allegedly representing my brain) that brings these objects closer in proximity to my life and compels me to relate to them. Most jokes, as Ngai and Berlant (2017) argue, work to draw on ideas of “what it means when we say ‘us,’” as we navigate how we relate to others and particularly the many identities we may be taking on (p. 235). Our sense of self and our knowledge also affects our sense of literacy, a teaching by the world of meanings and messages from which one can extend to digital spaces and make further connections.

There are few requirements for how a meme Figure 2 – “wait they ATE” (observation, May 4th, must look, be, or express, but memes always in- 2019) volve a ‘me’. To consider the meme without the place of the viewing subject would be to assume that the meme is neutral, that the meme does not affect me but merely represents me (or something to which in some way I relate – such as the video above). The subject is the one who gives life to a meme, not only in activating its logic but also in the potential for its distribution and further uptake as a carrier, standing between the meme and one’s own network. Memes involve the democratic navigation of what we like, and what it’s like to be us – but to see the memes without considering how subjectivity may be shaped by these objects is to disregard how the creation of these bites of bytes involves the materializing of this affective present which I am attempting to tap into.

Thus, this chapter will mobilize Berlant’s conceptual framework to understand subjectivity - or the relation to and sense of one’s self - within the meme. This will begin with an overview of subjectivity within Berlant’s work and then discuss the modes of subjectivity which memes re- Figure 3– Nobody | My Brain @ 3am (observation, inforce through comedy, drawn from this ethno- May 4th, 2019) graphic fragment above. Finally, this section will consider how memes impose and reinforce new A meme, like the one described in this seg- modes of subjectivity which blur the lines between

121 Capstone Issue Vol. 12, 2019 Cullen Ogden

self-continuity, self-expansion, and self-interruption rhythms comprise and produce subjectivity in it- in the cultivation of subjectivity online. self (Lefebvre in Berlant, 2014, p. 198). Berlant and Lefebvre provide a perspective that is fruitful in attending to patterns of meme consumption and construction, to consider the force of these pro- Berlant, Lefebvre, and Sedgwick on Sub- cesses as self-making practices. jectivity

Firstly, a consideration of the present is insepa- Lefebvre is not the only source of inspiration for rable from a consideration of subjectivity, because Berlant’s investigation into subjectivity within the subjectivity is the condition of being a person as contemporary present. Drawing on Eve Sedgwick’s well as the process of becoming one within this work into alternative understandings of desire and present (“Subjectivity,” 2004). As this research the self, Berlant departs from the dialectically tra- focuses on the present moment which Berlant de- ditional formation of neoliberalism’s force as im- scribes as disorganized, unstable, and particularly posed on the subject yet subjectively autonomous, uneasy, subjectivity stands as the tool through and instead looks for something much messier which we simultaneously navigate our sense of self (Berlant, 2014, p. 123-4). Sedgwick makes space and the world. Ergo, our own subjectivity shapes for us to think through “a practice of meticulous how we see the world, while this subjectivity is curiosity” as Berlant calls it, for the potential of a trained and reshaped through the work of genres reparative reading of agency and desire that rec- by setting expectations to reconﬁgure this sense of ognizes where those practices of therapy and relief self. For Lefebvre, it is society’s normative conven- are worthy of study, too (2014, p. 123-122). Al- tions imposed on us which make us see ourselves though there’s been great work done trying to high- as individuals (and this individuality as a reality of light memes as a serious matter, this research fol- existence), while what we claim to be this individ- lows Sedgwick’s mission and instead focuses on the ual subject – what one could call a personality – joy and pleasure that can arise from people think- does not simply reveal or express an innate inner ing and living together in digital space. As much essence. Instead, it is our actions of individual- as this chapter criticizes the way in which memes ity which produce the potential for habits, secur- shape and alter subjectivity, it is also possible to ing physical safety while allowing for intellectual embrace the reparative potential of tracking at- freedom, a securing of life’s continuity to allow for tachments in hoping to place them “back into play life’s betterment in the now (Berlant, 2014, p. 63). and into pleasure, into knowledge, into worlds” One’s personality forms a history brought forward, (Berlant, 2014, p. 123). Between Sedgwick’s ac- and imposed unto the now as a framework through knowledgement of the complexity of attachments, which to see the world. Although we have this and Lefebvre’s dressage to explore the impact of history, the present is full of potential and such a practical rhythms in shaping these attachments, sense of agency that we feel as though we could Berlant’s approach to subjectivity provides a way truly act in any direction we may choose. To resist to consider the shaping work of memes upon the this nihilistic sense of potential and inﬁnite possi- subject and the subject’s agency in concurrently bility, we lean into this personality as a sense of shaping the meme as well. grounding and come to conceptualize our positioning as not the result of society’s active shaping but simply as who we are – our sense of an innate self. Composure and Self-Interruption To unpack the shaping of such subjectivity by social forces, Berlant draws on Lefebvre’s model of dres- It is fair to claim that not everything we do in sage to speak of subjectivity as a process whereby this world is in line with the idea of a self and a we train ourselves and are trained to behave in par- personality which we have or hope to live so far. ticular accepted and expected ways, not through Personality provides comfort, but it also sets pa- active cognition of such a process. Rather, by em- rameters for how we must behave based on the phasizing the importance of social practices and history which we experienced and how we proceed rhythms Lefebvre turns our attention to rhythms, from this. When someone acts ‘out of character’, both biological and imposed, to consider how these what character is it that they are required to live

122 Capstone Issue Vol. 12, 2019 Cullen Ogden

within? Can being someone else - or being just a interruption, self-extension, and self-suspension, little less trapped in the self you have been told you the meme requires new ways of thinking about life- are - even for a moment provide space for relief? bettering practices online.

The conceptual bridge between a subject’s per- Turning to the ethnographic fragment at the sonality and this subjective escape from the self beginning of this chapter, the internet emerges which I will henceforth call impersonality is what as a space in which one faces the pressures of Andrew Phillips calls composure, or the relation- socially required personality imposed upon an in- ship one has in relating to their own mind (Phillips trinsically impersonal interface. For example, tech- in Berlant, 2014, p. 144). Berlant, in her own nology such as my iPhone arrived to me in default words, calls composure “the safeguard of fantasy,” factory settings, but I have continually made my as it allows you to “set the scene for your entrance phone and the content of my applications into my and make the world come to you when you want own image since. This personalizing of impersonal it” (Berlant, 2014, p. 144) acting as a buffer to technology has the double-edged effect of also pro- the world. Through composure, the mind can make ducing a digital body in the form of a profile - one space for feeling out a different sense of self that which the viewer inhabits first handedly while cre- is not based in the practical continuity of the world ating and producing a public facing portion of a and self-continuity of action, but rather operating profile such as those I repeatedly stumbled across perpendicularly to this. Composure makes enrich- in my research. On social media, through the acts ment possible through an understanding of the of scrolling, following, liking, blocking, and messag- mind’s enrichment as something worth pursuing, ing with others, this pushing and pulling of content seeing how acts of relief and enjoyment enrich our around the user makes space for their connection lives while not necessarily aiding in their continuity. to a gargantuan network of connections, while also elaborating on their own self through their connec- Composure allows for the interruption of the tion with others. However, this separation between self, emerging in situations of perceived safety that physical self-continuity and digital self-continuity gives the mind room to think of itself and the world (which is not in itself compulsory in the same ways) in new ways; composure protects the self while is what has allowed for the inspirational and opti- also allowing for new forms of recognition that mistic scene of the online. This mix of personalities are not caught up in the ongoing logic of things and impersonality is fundamental to the internet as (Berlant, 2014, p. 144). Berlant proposes a model a technology which is being made personal. With which claims that “the body and a life are not only their ambiguous appearance, memes further carry projects, but also sites of episodic intermission from this oscillating flow. personality, the burden of whose reproduction is part of the drag of practical sovereignty, of the Where the sense of crisis ordinariness that obligation to be reliable” (2014, p. 116). These Berlant established can leave life feeling “more like intermissions can take form in activities such as desperate doggy paddling than a magnificent swim eating, having sex, and looking at memes like in into the horizon,” the internet and social media the ethnographic fragment above. This involves have provided the means by which the social can “inhabiting agency differently in small vacations be accessed impersonally and texititremotely while from the will itself,” drawing on many kinds of self- requiring the creation of a profile for one’s own par- understanding and often works towards “making ticipation (Berlant, 2014, p. 117). Therefore, al- a less bad experience” (Berlant, 2014, p. 116; p. though the digital is separate from our embodied 117). For example, in the case of eating, Berlant lives, there is still an alternative space for a differ- describes this as a mode of self-extension whereby ent sort of continuity if one chooses. In the case the subject feels out something about the moment of the meme, we are neither forced nor required to (a taste, flavor, sensation), to cultivate “a sense participate in a particular way but rather are con- of well-being that spreads out for a moment, not fronted with memes based on our networks and are a projection toward the future” (Berlant, 2014, p. subsequently able to make judgements on a piece 117). While composure enables a variety of modes of content and act accordingly. In this process, we for subjectivity which Berlant explores, such as self- are taking in new ways of self-understanding and

123 Capstone Issue Vol. 12, 2019 Cullen Ogden

developing new conceptions of how we fit within Nicki Minaj, an amateur animation of the Simp- this larger virtual community made legible and log- sons, and a classic “Nobody | Me” meme form), but ical to us in many small chunks which we can find the recontextualizing of adding the comment “wait and track. Such as in the vignette at the begin- they ATE” by another user redirects the attention ning of this chapter, it is the formal and affective from the relation of the content to your own expe- structure which provokes these connections, as it rience. This then is considered in relation to their is the structuring of a relation mediated by a re- technical skill, which could be extending one’s self- lation which compels the subject to turn towards awareness through a consideration of gracefulness these connections more than to the content of the in performing this dance. Through the rabbit hole meme. What the meme does is present a relational of considerations that multiply from ambiguity, the structure, and allow the subject to gauge their own diverse ways in which a meme can be redistributed connection or disconnection with this relationship. and remixed can draw on and extend one’s own existing sense of self-connection to the networks Thus, I propose that the meme provides a time- online in the process. less (yet contingent) scene to which the self can return and experience a mode of self-connection. Building from self-extension, self-connection involves an extension within a moment rather than towards a future, but one whose main focus involves the navigating of intersubjectivity on many The meme provides the pleasure of recogniz- levels. Recalling my argument in Chapter 2 that ing new forms of intersubjectivity and laughing at memes constitute a genre related to comedy, I turn what arises. In doing so, the subject embraces an here to the intersubjective function of comedy to attitude where any form of knowledge that which consider how the meme generates self-connection, can be logically made apparent (even if the logic is and to further explain this case (Figure 2). visual, incomplete and shaky at best) can provide a punchline of closure to an otherwise constantly Comedy and Intersubjectivity unfolding and unpunctuated present. The product of a meme is pleasure: often in the form of laughter Comedy, as Berlant and Ngai suggest, relates from simplifying the world, but most broadly from heavily to aesthetic judgement, as they both re- recognition. The ability to cultivate an alternative mind us of “forms of intersubjectivity we usually self(s), and these small moments of affinity found don’t think about but that we rediscover as pre- through memes may motivate us to carry them supposed by our very compulsions to make jokes with us as the digital re-enters our navigation of the and judgements in the first place” (2017, p. 235). real. The meme also reveals through its replication The foreshortening of comedy is made even closer that others feel something like ‘my experience’, through social media at the touch of a finger, and such as in Figure 3 which shows that I can feel less the curation of posts into a news feed creates the alone in being awake and thinking at 3am. Even endless stream of content expansively relating it considering how I have not had this feeling for a other content, leading to this layering of contexts long time, the meme still takes me back there, and developed in Chapter 2. As memes involve making in that moment, I find a newly mediated space of claims about and judgements of these relations, affinity. My anxiety is no longer a violent force they extend and grow through further layering of working against me, but through composure trans- new meanings, not only as the self cultivates new forms into the source for my derivative pleasure. and interesting ideas but as you see others see- In some ways, memes draw on the already estab- ing it, when they post it, and who else may have lished genres of knowledge dissemination amongst liked it. For example, the chronological distribution visual culture to consider the kinds of pleasure we of posts by the users one follows into a personal- can generate from recognition through the imper- ized feed can happen to arrange content into new sonal, using the internet as a means for connecting formations that the meme could not predict. Not to many others and drawing out new selves in the only can memes such as the one above combine process. a confusing combination of objects (in this case

124 Capstone Issue Vol. 12, 2019 Cullen Ogden

Recognition as Form from the flurry of life taking place online, basking in the errors we would normally ignore or disregard While memes contain absurd and entertaining while considering how taking things seriously only content, this content only matters in appealing to sets up future disappointments. To create a meme the subjective taste of the user and their poten- is to make a claim to one’s way of adapting to tial literacies. The meme as genre creates a space the world, allowing others to momentarily assume of affinity, wherein composure allows for the plea- this position by animating their own experience for sure of recognition, with the potential for various comedic effect. The challenge, to be taken up in forms of further recognition. Memes, as well as the final chapter, is to consider how this new mode comedy, are incredibly sensitive to context, so it of subjectivity requires new ways of paying atten- is this constant (re)negotiation of intersubjectivity tion. that generates comedic pleasure while exposing and developing new ways of relating and connecting in the process. We focus on the meme as a source of pleasure, but the pleasure comes from IV. Feeling Out and Through – Ide- recognizing your own self, recognizing your rela- ology and (Virtual) Intuition tion to the poster’s self, and considering the wider connection to a topic, an opinion, or even the world “Starting my very first field visit on In- (or how only in the meme an idea may be made stagram, I am unsure as to whether I possible). Memes can break through our stance of should start with my feed as it is in front impersonality to collapse the space between the of me, or if I should start on a meme personal and impersonal, as we are both absorbed page instead. How would I be able to into these “me”s while absorbing further knowl- discern if something is or is not a meme? edge about myself in the process. I decided to start with the meme page, @walkingpathogen, since then I could In terms of the expectations of the genre estab- be able to get an idea and make my lished by our interaction with the meme, the desire own discretional choices of what to in- to find pleasure within the meme leads to what clude and what not to include. The first ultimately has been an impersonalization of the meme I saw on this page (see Figure 4), personal. Therefore, representation of the past is however, already brings these concerns no longer required as I can project myself into a va- back to my attention. The image shows riety of contexts and find recognition in objects as a close up picture of a man’s goatee, diverse as in Figure 2, where I come to recognize with some golden foil object superim- my brain as a crudely animated pastiche of car- posed over his mouth. On top of this im- toon nostalgia converted to artistic performance. age is the text, “Goatee Goat Cheese,” It is the development of self-connection and open- with the first half in black text and the ness to comedic relations which builds the momen- second half in white text. The typed tum of meme consumption, and ultimately forms a caption of the post says: “Exactly how constantly-in-flux domain of relations. I like my men.” While I know that it is a meme for the fact that this meme ac- Thus, memes are like inside jokes because not count posted it, or at least I would as- only are we positioned inside the joke, but the joke sume this is the case, it doesn’t feel to places the user inside of their own reality. One me like a meme. It’s about goatees and may also find that this new meme has something goat cheese, and I’m guessing it’s mak- familiar which brings the pleasure of a past expe- ing fun of these objects as certain as- rience into the present through this new piece in pects (in some way) of who people are, unanticipated directions, allowing one to appreci- but if I were asked what this image is, I ate the stability that perfect reproduction allows would probably call it a meme from the for as its connections can move in unexpected and context more than from the image it- new ways. Encountering memes can be likened self. I’m not sure if it’s because of my to warming your cold hands on the residual heat own taste in memes, or because I’m not

125 Capstone Issue Vol. 12, 2019 Cullen Ogden

accustomed to this one, but I wonder negotiation of the present which forms collective whose meme is it, then?” ideologies. Then, I will propose this concept of virtual intuition and discuss its particularities in (observations, March 11th, 2019) relation to my ﬁeldwork. Finally, this chapter will end by considering how comedy relates to self- connection and virtual intuition, speaking to the effects of the genre as a whole in shaping the contemporary present.

Deﬁning Ideology and Intuition

Ideology, as defined by Louis Althusser, is the, “imaginary relationship of individuals to their real conditions of existence” (Althusser in Felluga, 2015). This definition draws on Jacques Lacan’s use of the term, imaginary, to identify ideology’s grounding in the similarities between people’s perceptions which enable us to give meaning and make sense of the world (Felluga, 2015). There can be many different kinds of ideologies, such as religious, political, and moral, which operate in many different ways, but ideology is ultimately a material condition which cannot be separated from the apparatuses of society which reinforce particular ideologies. Interpellation, or being hailed by an ideology, is a constant process of indoctrination which converts concrete individuals into subjects Figure 4 – Goatee Goat Cheese (observation, - a process that is already underway but is still March 11th, 2019) susceptible to radical change. Ideology, therefore, Subjectivity speaks to the place of the individ- cannot be separated from its practices, as these ual within larger social dynamics, but this chapter practices are “our performance of our relation to considers the training of intuition as the mediator others and to social institutions that continually between subjective experience and collective bi- instantiates us as subjects” (Felluga, 2015). By ography (Berlant, 2014, p. 53). Building from the considering genres as material objects which me- account of subjectivity developed in the previous diate the affective present, Berlant suggests that chapter, we can consider how the mode of self- we must turn to the topic of intuition to address connection, which is enabled through the meme, ideological concerns (Berlant, 2014, p. 53). Thus, draws on a form of virtual intuition to access and to consider genre as the normative repetition and respond to the complex, performative space of the development of a collective relation to the world digital. To access this space requires new modes which builds expectation for the future, is to make of being a subject (i.e. “I’m baby”) and new ways space for the importance of intuition in genre as a of paying attention to the world through this sub- tool for renegotiating ideology. jecthood (i.e. “That’s a Big Mood”). As the vignette above shows, it is in the pursuit of drawing mean- Berlant defines intuition as “the process of dy- ing and pleasure from the meme where intuition namic sensual data-gathering through which affect is stretched and warped to encompass a variety takes shape in forms whose job it is to make re- of possible meanings that are made available and liable sense of life” (2014, p. 53). This is not to possible by virtual intuition. First, this chapter will say that intuition is simply the process by which discuss ideology in relation to intuition in order to people interpret and respond to the world in time, ground this discussion of intuition in the collective but rather sets up the claim that visceral responses

126 Capstone Issue Vol. 12, 2019 Cullen Ogden

are trained as we gather genres, set expectations, virtual intuition also involves the consideration of and make tangible the emerging event (Berlant, where relations can be left open, and the possibility 2014, p. 52). As Berlant aptly describes, “intuition to revel in humor that, if it existed outside of the is where affect meets history, in all of its chaos, internet, may not actually be considered as all that normative ideology, and embodied practices of dis- funny. cipline and invention” (2014, p. 52). Although ideology is important to consider in this context, it Ironically enough, as the content of memes is intuition which provides an avenue for consider- crosses lines and makes new comedic connections, ing what feels ‘normal’, what is required of us both the structure of scrolling through memes involves macroscopically and within the everyday, and how encountering many memes to absorb new forms this may be changing. of content constantly, sensing out how things are happening now. If a meme is not funny, then it may be too ambiguous or not provide enough of a re- Virtual Intuition action to the context. Yet through seeing more examples of the meme elements in varying contexts, Intuition provides a way of adjusting to the these elements may become freed up enough for world and, as genres solidify into a habitual form, them to make sense and to provide pleasure for the a way of managing life according to the assump- user, such as Roman Holiday’s necessitated accom- tions we hold of how to manage a life (Berlant, paniment of ‘fast motion’ videos being separated 2014, p. 52). That being said, I believe Berlant’s from its place in Minaj’s discography like in Figure concept of intuition can be taken further, as inter- 2 5. Virtual intuition also involves negotiating new acting with memes demonstrates what I call a form meme norms, which may emerge alternatively to of virtual intuition. Because memes can solely be the real world but with their own standardized logic. accessed via digital technology which limits the Whatever the originally descriptive situation might depth of our sensual engagement, this virtual intu- have been (I do not claim to have ever seen the ori- ition emerges as not only a way of trying to catch gins of this trend in this research), virtual intuition up to the present, but also to sense a world beyond has also made space for similar fast motion videos that which appears before us. to be shared which are then contextualized as lacking or missing the song’s presence. This shows how Memes provide a scene, wherein the promises the ephemeral emergence of a meme form can so- that cluster around a meme can be reproduced lidify into a trend which trains virtual intuition to and re-encountered by seeing something famil- make connections based on the funny memes one iar in new contexts, allowing these elements to has seen towards content one encounters. There- gain new meaning through conversation with other fore, we come to sense the contours of digital con- users and memes simultaneously. Therefore, par- tent for its possible connections through our virtual ticipating in memes requires a sense of virtual in- intuition which is developed by our interaction with tuition – to savor the humor, to search for more memes and then expected as projected forward. comedy, to continue in this collective production of The content then returns to and fits into our world, humor which ideally will make everyone (or those left open for connections yet solidifying into a trend you would group in with ‘us’) laugh. Virtual intu- which may become viral, or even reflective of the ition, then, is like suspending disbelief for the sake year it circulated. The meme, through its perceived of seeing what comes out the other side, to allow accuracy and ability to describe (or what we intuit for the virtual relations which have not mattered to be an inner logic from the chaos), can trans- or have not yet emerged to exist and be shared. form even the most absurd or ambiguous content This is not to say that one will believe anything a into a stable, comedic trope solely through its re- meme shows, but that the meme’s job is to provide peated and frequent repetition and remixing. While enough of a form around disparate content to hint formal meme norms may solidify through replica- at a relation which virtual intuition identifies and tion, these norms cannot prevent the content from measures against the world. As memes are each grasping for or attaching to alternate possibilities, made by someone with a particular subjectivity, reimagined through this virtual intuition.

5See Chapter 3 for this argument.

127 Capstone Issue Vol. 12, 2019 Cullen Ogden

and relations that I do not consider. When a meme We can also see how virtual intuition is impor- ‘works’, we find the expectation that if one thinks in tant for accessing and navigating the claims under- the ‘right way’ (or right in relation to the creator of lying a meme. We can do this by considering the the meme), a meme will ‘work’ which requires con- experience of encountering a political meme from sidering what sorts of accounts one aligns with and an ideology that the viewer does not subscribe the sorts of communities one wants to connect to. to, and seeing where the joke may be lost. This As much as an escape can provide an unexpected demonstrates how the unpredictability of memes source of humor, memes produce a product of in- makes the pleasure of identification with the joke tuition; where one’s ability to virtually sense out a just as contingent as the meme itself. Although possible meme or to create a meme can somehow memes have the potential to recognize the viewer, still make someone feel recognized, and somehow it is up to the viewer to intuit their relation to this confirm their place in this world. On the other hand, recognition, and particularly to sense out how the memes also provide a humble way of indoctrinat- location of this affect and its network effects the ing someone into a particular mode of intuition, in meme. This offensive potential of the meme, if the ways we may not be fully able to recognize in the comedy does not foster a sense of intersubjectiv- moment. ity, is where the subject can experience the feeling of being the butt of the joke made by an excluding ‘them’. Feeling Funny: Comedy and Intuition

To investigate this sense of virtual intuition, I de- Compared to the refinement and training that voted time to exploring Far-right or Alt-right memes intuition usually operates through, comedy is about on Instagram through searching for ‘right memes’, surprise, the unexpected, the ability of a genre to which generated multiple pages. When I tried to collapse the space between objects, to allow them look at these memes, I wrote this: to play out in new and surprising ways, and to ul- “As a liberal/left person, I find that these memes timately develop joy or laughter from what feels annoy me because there is something so humble like a freedom that surpasses the limitations of the or inviting to the meme, and to see a meme from world causing a sort of disturbance (Berlant Ngai, a view one does not support either makes you feel 2017). It is difficult, then, to consider how we can like the butt of the joke or like something about be both expected and pressured to perform com- it is too far from the reality we live in, that the edy in a growing number of zones of life while also story is missing that part that makes justice pos- holding that comedy itself involves the pleasure of sible. Funny stuff doesn’t need to make sense, sensitive and contingent relations being disturbed but unfunny stuff makes what feels like the ‘wrong in different ways. This, I argue, is the importance sense’.” (observations, May 4th, 2019) of virtual intuition; in being attuned to the digital flows, as well as the real world experiences, one While Althusser said that interpellation happens is able to combine both disclosure of experience to us without even being able to resist it, I see vir- with comedic interpretation, projecting a sense of tual intuition as a more active form of intuiting the self that is comedic from their participation in the world, as the digital provides the space for a buffer network. As comedy upends assumptions to chal- of composure and the ability for a user to control lenge form, virtual intuition is the process whereby their own experience, actively affecting their fu- more flexible relations to the world further protect ture experiences through judging in the now. In the subject from its closeness, while feeling out the case of my fieldwork, I found a collection of these assumptions and identifying where comedy memes that seemed similar to those I had seen may disturb (hopefully) without rupture. Comedy before, but the content had been shifted while ap- is the consolation prize for adapting intuition to the propriating a form for purposes that, to me, were changing world falling apart, because if one can not accurate. As a result, I do not follow or like stage the scene in just the right light, there will these sorts of posts, but it made me think about be pleasure derived from it. In social media, espe- this process of seeing memes that somehow re- cially, these applications arrange the posts in any sist my desire implicates a further network of users variety of ways, but virtual intuition looks for the

128 Capstone Issue Vol. 12, 2019 Cullen Ogden

patterns in the roll of dice, unintentionally creating This is where the sociality of the internet is incredi- new connections and breaking old ones. bly important and virtual intuition necessitated, as the networks and relations one develops play a role It is the perfect multiplicity of the internet - in in how we expect or understand certain content to its accessibility, its (assumed) representation, and be funnier than other content. access - that we find a sudden appreciation for the flaws and errors that at once hurt and re-enact I’m not saying that memes change the way failure as a performance. We can fight, love, con- we see the world, but I do believe that they have fess, and proclaim to the public in unprecedented changed the way that we relate to and affectively ways and the worst that can happen is, ultimately, encounter the world around us. Memes paint the having one’s view challenged for deviating from world not in events and processes, but by recog- the norms of others. Instead of developing new nizing how the crisis ordinariness of life today is stories and new characters to seek representation already funny enough. Thus, connecting to oth- through, memes take the world as a means of ex- ers rather than to pre-produced genres of the past pressing itself. They draw on our own lives which does not make promises it cannot keep. As long frame our experience of the world for inspiration as one is able to feel out this scene, the worst that as our own precarity in the contemporary present can happen is that on the other side of the meme, is funny enough. In hopes of connecting through a things are still the same but with a new way of pos- shared sense of pleasure from self-connection, the sibly and potentially seeing it. The prize, however, expectation developed around memes is that these is the affirmation of one’s mode of intuition itself digital objects will be funny, as much so as we can through widespread normalization. ‘get it’ or ‘like it’, if getting the joke feels worth it.

Conclusion: Between the World virtual intuition useful for explicating the meme, and Me(me) there is much more that can be said about memes, even about the ones featured here. Instead, I encourage those who read this work to consider how As much as theory can produce a deeper unit may speak to their own experience. I hope that derstanding and appreciation for the world, it can it may help some to return to their disciplinary ap- also warp and distort reality in that same pursuit. proaches and incorporate an affective approach to As comedy emerges like an unexpected pleasure memes in their study, as considerations of affect of failures and faults, theory can work to retrospec- make tangible that which underlines our experi- tively speak to comedy. Even then, to address and ence of a range of memes. Ultimately, the meme is not totalize the subjectivity at hand is difficult. By a tool for considering the overlooked, attuning the speaking to the affective structure rather than sub- self more flexibly to not only the world, but to the jective experience, this research has demonstrated sorts of perceptions that are being circulated at the that we can recognize the facets and dynamics at time. If we think of the meme as an affective medi- hand in memes without forcing memes to behave ator between the world and me, we can finally see in a specific manner. Instead, embracing the imma- their importance within the digital, the everyday, nence of affect, this theory I have extended from and in this contemporary present as grassroots Berlant provides the starting point for further con- interpretation and critique through comedic judg- sideration of the affective contours of memes and ment. What is a meme without a me, after all? the digital in shaping our present as we continue to shape them in turn.

To conclude, I would like to allow - in the canon of cultural analysis and style of its founder, Mieke Bal - for the memes to, “speak back” to my own exhibiting of their form, by providing a series of memes drawn from my ﬁeldwork which have contributed to the development of this theory (Bal, 1999). As much as I consider self-connection and

129 Capstone Issue Vol. 12, 2019 Cullen Ogden

11th, 2019)

Figure 5 – “How’s Life?” (observation, March 11th, 2019)

Figure 8 – head cold (observation, March 11th, Figure 6 – “Me at my next job interview” 2019) (observation, March 11th, 2019)

Figure 9 – smallest number (observation, March 11th, 2019)

Figure 7 – “I’m so broke” (observation, March

130 Capstone Issue Vol. 12, 2019 Cullen Ogden

Banet-Weiser, S., Miltner, K. M. (2015). Masculinitysofragile: Culture, structure, and networked misogyny. Fem- inist Media Studies, 16(1), 171-174. doi:10.1080/14680777.2016.1120490 Barker, C. (2004). Genre. In The Sage Dictionary of Cultural Studies (pp. 74-75). London: Sage. Figure 10 – American dream (observation, March Barker, C. (2004). Subjectivity. In The Sage Dic- 11th, 2019) tionary of Cultural Studies (pp. 194-195). London: Sage. Bauman, Z. (2018). Liquid modernity. Cambridge: Polity Press. Benedict, R., Goldenweiser, A. (1937). Patterns of culture. American Sociological Review, 2(5), 802. doi:10.2307/2083847 Berlant, L. (2011). Cruel Optimism. Duke University Press. Blackmore, S. J. (2000). The Meme Machine. Ox- ford: Oxford University Press. Figure 11 – Be a lot cooler if you did (observation, Blackmore. (2001). Evolution and memes: The hu- March 11th, 2019) man brain as a selective imitation device. Cybernetics and Systems, 32(1-2), 225- 255. doi:10.1080/019697201300001867 Boellstorff, T. (2012). Ethnography and virtual worlds: A handbook of method. Princeton: Princeton University Press. Borenstein. (2004). Survival of The Catchiest: Memes and Postmodern Russia. The Slavic and East European Journal, 48(3), 462-483. doi:www.jstor.org/stable/3220071 Brinkema, E. (2014). The forms of the affects. Durham: Duke Univ. Press. Figure 12 – Why do you want this job? Clough, P. T. (2008). The affective turn. (observation, March 11th, 2019) Theory, Culture Society, 25(1), 1-22. doi:10.1177/0263276407085156 Works Cited Coker, C. (2008). War, memes and memeplexes. International Affairs, 84(5), 903-914. Adams, T. E., Ellis, C., Jones, S. H. doi:10.1111/j.1468-2346.2008.00745.x (2017). Autoethnography. The In- Comaroff, J., Comaroff, J. (2003). Ethnography on ternational Encyclopedia of Commu- an awkward scale. Ethnography, 4(2), 147- nication Research Methods, 1-11. 179. doi:10.1177/14661381030042001 doi:10.1002/9781118901731.iecrm0011 Conte, R. (2001). Memes through (social) Ahmed, S. (2015). The Cultural Politics of Emo- minds. Darwinizing CultureThe Sta- tion. New York, New York: Routledge. tus of Memetics as a Science, 83-119. doi:10.4324/9780203700372 doi:10.1093/acprof:oso/9780192632449.003.0005 Bal, M. (1999). The practice of cultural Analy- Cvetkovich, A. (2012). Depression: A public feel- sis: Exposing Interdisciplinary INTERPRE- ing. Durham, North Carolina: Duke Univer- TATION. Edited by Mieke Bal. Retrieved sity Press. June 2, 2019, from http://www.sup.org/bo Dagta¸s.ˇ (2016). ’Down With Some Things!’ The oks/title/?id=851 Politics of Humour and Humour as Politics

131 Capstone Issue Vol. 12, 2019 Cullen Ogden

in Turkey’s Gezi Protests. Etnofoor, 28(1), and memes. Retrieved June 2, 2019, 11-34. from http://henryjenkins.org/blog/2009/02/ Dennett C. (1990). Memes and the exploita- if_it_doesnt_spread_its_dead_p.html tion of imagination. The Journal of Aes- Jorgensen, D. L. (2015). Participant obser- thetics and Art Criticism, 48(2), 127-135. vation. Emerging Trends in the So- doi:10.2307/430902 cial and Behavioral Sciences, 1-15. DeWalt, K. M., DeWalt, B. R. (2011). Participant doi:10.1002/9781118900772.etrds0247 observation: A guide for ﬁeldworkers. Lan- Koláˇrová. (2017). The inarticulate Post-socialist ham, Maryland: Altamira Pres, Md. Crip on the cruel optimism of NEOLIB- Falzon, M. (2016). Multi-Sited Ethnography: The- ERAL transformations in the Czech Re- ory, Praxis and Locality in Contemporary public. Culture - Theory - Disability. Research. Routledge. doi:10.14361/9783839425336-013 Gal, N., Shifman, L., Kampf, Z. (2016). Lefebvre, H. (1992). Critique of everyday life (Vol. “It gets better”: Internet memes and 1) (J. Moore, Trans.). London: Verso. the construction of collective identity. Leys. (2011). The turn to affect: A cri- New Media Society, 18(8), 1698-1714. tique. Critical Inquiry, 37(3), 434-472. doi:10.1177/1461444814568784 doi:10.1086/659353 Geiger, R. S., Ribes, D. (2011). Trace Ethnogra- Malinowski, B. (2015). Argonauts of the western phy: Following Coordination through Docu- paciﬁc. Routledge. mentary Practices. 2011 44th Hawaii Inter- Marwick, A. (2013). Memes. Contexts, 12(4), 12- national Conference on System Sciences. 13. doi:10.1177/1536504213511210 doi:10.1109/hicss.2011.455 Murthy. (2008). Digital ethnography: An Examina- Godwin, M. (1994, January 10). Meme, Counter- tion of the Use of NEw Technologies for So- Meme. Wired. Retrieved June 2, 2019, from cial Research. Sociology, 42(5), 837-855. www.wired.com/1994/10/godwin-if-2/ doi:10.1177/0038038508094565 Highmore, B. (2008). Introduction: Questioning ev- Narayan, K. (1993). How native is a eryday life. In The everyday life reader (pp. "Native" Anthropologist? Ameri- 1-34). London: Routledge. can Anthropologist, 95(3), 671-686. Hjorth, L., Cumiskey, K. M. (2018). Mobiles doi:10.1525/aa.1993.95.3.02a00070 facing death: Affective witnessing and Ngai, S. (2007). Ugly feelings. Cambridge, Mas- the intimate companionship of devices. sachusettes: Harvard University Press. Cultural Studies Review, 24(2), 166-180. Ngai, S., Berlant, L. (2017). Comedy has is- doi:10.5130/csr.v24i2.6079 sues. Critical Inquiry, 43(2), 233-249. Hjorth, L., Pink, S. (2013). New visualities and the doi:10.1086/689666 digital wayfarer: Reconceptualizing cam- Nissenbaum, A., Shifman, L. (2015). Inter- era phone photography and locative me- net memes as contested CULTURAL cap- dia. Mobile Media Communication, 2(1), ital: The case of 4chan’s /b/ board. 40-57. doi:10.1177/2050157913505257 New Media Society, 19(4), 483-501. Hjorth, L., Horst, H. A., Galloway, A., Bell, G. doi:10.1177/1461444815609313 (2017). The Routledge Companion to Digi- Postill, J., Pink, S. (2012). Social metal Ethnography. New York: Routledge. dia ETHNOGRAPHY: The DIGITAL re- Holdcroft, D., Lewis, H. (2000). Memes, minds searcher in a messy web. Media In- and evolution. Philosophy, 75(2), 161-182. ternational Australia, 145(1), 123-134. doi:10.1017/s0031819100000231 doi:10.1177/1329878x1214500114 Jameson, F. (2018). Postmodernism and consumer Puar, J. (2012). Precarity Talk: A Virtual Roundtable society. In V. B. Leitch (Author), Norton An- with Lauren Berlant, Judith Butler, Bojana thology of Theory and Criticism (pp. 1758- Cvejic,´ Isabell Lorey, Jasbir Puar, and Ana 1771). New York: Norton Company Lim- Vujanovic.´ TDR/The Drama Review, 56(4), ited, W. W. 163-177. doi:10.1162/drama00221 Jenkins, H. (2009, February 11). If it doesn’t Rossolatos, G. (2014). The ice-bucket chal- spread, it’s Dead (part ONE): MEDIA viruses lenge: The legitimacy of the MEMETIC

132 Capstone Issue Vol. 12, 2019 Cullen Ogden

mode of CULTURAL reproduction is the MIT Press. message. SSRN Electronic Journal. Shifman, L. (2014). The cultural logic of doi:10.2139/ssrn.2505616 PHOTO-BASED MEME GENRES. Jour- Saito, A. P. (2017). Moe and internet nal of Visual Culture, 13(3), 340-358. MEMES: The resistance and accommo- doi:10.1177/1470412914546577 dation of Japanese popular culture in Sterelny, K. (2006). Memes revisited. The British China. Cultural Studies Review, 23(1). Journal for the Philosophy of Science, 57(1), doi:10.5130/csr.v23i1.5499 145-165. doi:10.1093/bjps/axi157 Sampson, T. D., Maddison, S., Ellis, D., Seig- worth, G. J. (2018). Affect and social me- Taussig, M. T. (2016). The Nervous System. Lon- dia emotion, mediation, anxiety and conta- don: Routledge. gion. London: Rowman Littleﬁeld Interna- Zenner, E., Geeraerts, D. (2018). One tional. does not simply process memes: Im- Schroeder, R. (2018). Social theory after the inter- age macros as multimodal construc- net: Media, technology and globalization. tions. Cultures and Traditions of Word- London: UCL Press. play and Wordplay Research, 167-194. Shifman, L. (2014). Memes in digital culture. The doi:10.1515/9783110586374-008

133 Humanities

Pieces of Resistance: Protest Signs as Objects of Dissent

Nina Klaff

Supervisor Dr. Erinç Salor (AUC) Reader Dr. Marco de Waard (AUC) Capstone Issue Vol. 12, 2019 Nina Klaff

Abstract

While much has been written about protest and its place in the new media age, little research makes use of visual methods. I argue the importance of the use of visual analytic processes to understand protest culture by investigating protest signs as politico-cultural objects. I argue that the protest signs made and shown by protesters themselves are artifacts of individual expression through which people perform both their political and personal identities. I ﬁrst present a ﬁrm theoretical grounding of protest marches. The research methodology includes attendance of protest events, as well as cultural and formal analyses of their artifacts. Two case studies are analysed. First, the People’s Vote March, demanding a second vote on the U.K.’s Brexit Referendum, provides a basis for investigating how protest culture has evolved with social media by analysing the rhetoric and aesthetics employed in the signs to draw attention to their cause. Second, the Global Women’s Marches (2017-present), an international movement of feminist marches, demonstrates how movements and their signs travel transnationally through social media. I then examine political statements made through visual media in online identities to argue that the space of appearance itself is digitally networked and that our conception of protest signs can be applied to other forms of aesthetic expression, which include those created in the digital sphere.

Keywords and phrases: materiality, embodiment, performativity, aesthetics, assembly

A note on form Introduction

Parts of this paper are punctuated with short On the 20th of October 2018, 700,000 peo- texts in black boxes. These are words, slogans, ple marched the streets of London demanding a sentences I have seen on protest signs. These second vote on the Brexit Referendum that en- have been placed to interrupt the reading process, tailed the U.K.’s leaving of the European Union (BBC their messages becoming part of the spectacle of News). I was one of them. While the gathering of this paper, to emulate the prominence of signs in so many bodies in the public space was, as it was the constellation of a demonstration. Inspired by meant to be, a beguiling carnivalesque spectacle, Michael Taussig’s “I’m So Angry I Made A Sign,” in the most striking visual components of this march which he includes some texts from signs held by were the signs made and carried by the protesters protesters at the Occupy Wall Street occupation of themselves. The prevalence of protest signs is not Zuccotti Park, New York (2011) in these same black particular to this event. It is traditional in protest boxes. This strips them of their aesthetic, embod- culture to carry signs that demonstrate one’s par- ied, and material dimension, and lets the words ticipation within a protest event, which express the speak for themselves. Consider what has been lost, views of the individuals who carry them as well and what has been gained, by doing so. It is hard as the character of the movement itself. Protest to give credit to each of their authors, so instead I signs signal not only participation within a particu- attribute credit to the movement within which they lar movement, but are also signals of dissent them- were shown. These signs are here to be read along- selves. side the words of other scholars, researchers, and people of note. As Taussig said, “I don’t think you In Charlie Chaplin’s Modern Times (1936), Chap- will confuse them, but it’s better that you do” (56). lin is seen walking down the street among a mon- tage of modernity: factory machinery, cars, and crowded streets. A truck drives past him, a flag perched precariously on the back. As the truck drives off, the flag falls. Chaplin runs into the road and picks it up. As he waves it in the air to alert the driver, a crowd appears behind him. Chaplin is oblivious. The crowd is holding signs. As he continues to chase the car, waving the flag, viewers Figure 1: Klaff, Nina. Illustration of a sign at are made to understand that he has inadvertently Occupy Wall Street as reported in Taussig’s “I’m found himself at the head of a protest march as So Angry I Made A Sign,” 2019. the police exclaim, “so, you’re the leader!” (Mod-

135 Capstone Issue Vol. 12, 2019 Nina Klaff

ern Times 1936). On his own, Chaplin was a mere As Sartwell argues, “not all art is political, but all passer-by. The flag, left on the back of the truck, politics is aesthetic” (qtd. in Werbner 1). Protests or lying on the street, held little meaning. How- are political manifestations, and are thus, in this ever, as soon as Chaplin picked it up, its symbol- understanding, also aesthetic. I understand aes- ism was catalysed, even without it carrying any thetic under Rieff’s definition: “that activity in textual or other symbolic cues. The flag made a which, by means of external signs, human be- protester of Chaplin, otherwise an uninvolved by- ings communicate feelings to each other” (479). stander; and Chaplin, by inadvertently joining an I thus pay due attention to protest signs as ex- assembly, made a protest sign out of an otherwise ternal devices used by ‘the people’ to communi- insignificant flag. cate their politics. The signs act as expressions of the thoughts and beliefs of the carriers, who use them to signify what they want to say. Protesters make them by hand, by writing, painting, drawing, or sometimes printing a statement onto paper or cardboard. They decorate them with images and other symbols for the purpose of holding them up and carrying them at protest events. Sign-holding is an embodied, performative, material, and aesthetic action of political expression. I argue that these signs, imbued with political meaning, are not only important objects for visual and cultural analy- Figure 2: Still from Modern Times, 1936, retrieved ses, but are also emblematic of protest culture and via YouTube. “Charlie Chaplin. Modern Times. that their production and display in protest activi- Protest Sign.” YouTube, 00:54, uploaded by user ties is a revolutionary practice itself. I first establish GeslominaAM, 10 Mar. 2014. Accessed a firm theoretical framework, drawing from protest 30/05/2019. discourse in the fields of visual culture, anthropol- The embodied fusion of Chaplin and this object ogy, sociology and materiality, to explore how con- gave political significance to them both. The mere temporary protest movements express their aims term protest sign demonstrates that these objects (e.g. Butler Notes on Assembly; Reiss textitThe function as universal signals of protest; these are Street as Stage). Drawing from Tufekçi’s analy- important objects of political and human history in sis of 21st century social movements, I argue that popular culture. protest signs allow a movement to set their own In Twitter and Tear Gas, Tufekçi demonstrates narrative, and thus epitomise their “narrative ca- that “collective actions, social movements, and pacity” (192). I build on Mitchell’s seminal ques- revolutions are woven into the fabric of human his- tion, “what do pictures want?” to posit the consider- tory. They have been studied at great length and ation of what protest signs want (71). I further pre- for good reason: they change history” (xix). How- vious protest scholarship to establish what is meant ever, as Philipps argues, their scholarship “rarely by the space of appearance and how movements employs visual analytic procedures” (3). I argue use protest marches to express their aims (Arendt that this is a drastic oversight. The visual objects The Human Condition). deployed by protest movements give great insight into their nature and culture, and are themselves I then explore the innovation in contemporary objects of political significance. As Khatib states, protest signs by analysing their role within two “the image is at the heart of political struggle,” movements.1 Firstly, the People’s Vote marches which, she argues, is “an inherently visually pro- in London in 2018 and 2019, which demanded a ductive process” itself, because it “is a struggle second vote on the Brexit referendum, were lim- over presence, over visibility” (qtd. in Werbner 2). ited to nation-specific politics. I analyse the move- Thus, visual devices, such as protest signs, are im- ment’s signs in terms of materiality and politi- portant political objects. cised humour, which entails a consideration of

1The two case studies in question (the People’s Vote March and the Women’s March) are non-violent, peaceful, and democratic dissenting events. Of course, other forms of protests feature signs, but for the sake of the clarity and brevity of this paper, I focus mainly on these protest marches.

136 Capstone Issue Vol. 12, 2019 Nina Klaff

social media aesthetics. To facilitate an analy- fines whether the movement can “disrupt the sta- sis of how social media might be utilised by con- tus quo;” and “electoral or institutional capacity,” temporary movements, I consider the signs at refers to a movement’s ability to incite electoral or the Global Women’s Marches (2017-present): a institutional change (Tufekçi 192). Under this def- transnational movement that exemplifies how so- inition, “disruptive capacity” outlines whether the cial media has changed the organisation, execu- movement can “disrupt the status quo;” “electoral tion, and the very nature of contemporary protest or institutional capacity,” refers to a movement’s movements, demonstrating that our conception of ability to incite electoral or institutional change; protest culture must now include the digital sphere. and “narrative capacity refers to the ability of the Such an analysis serves to highlight the ubiquity of movement to frame its story on its own terms, protest signs in contemporary protest culture, and to spread its worldview” (Tufekçi 192). I argue encourage an understanding of their importance as that while the movements in point might not have objects of cultural interest and of socio-political im- overtly disrupted the status quo or incited tangible port. electoral or institutional change, they have been effective in setting their own narrative. The protest sign was one of the key devices deployed to do so. I. Placing Signs in Protest I argue that these objects contribute invalu- ably to the “narrative capacity” of social move- In Twitter and Tear Gas, Tufekçi defines a so- ments. They are objects of what Jenkins might cial movement as “a claim made to a public that call “participatory culture,” whereby people actu- a wrong should be righted or a change should be ally contribute and participate in the production of made” (8). I thus propose a consideration of the communication (qtd. in Gerbaudo 22). They always social movements use protest signs to de- low each sign-maker to set their own personal nar- mand such change. Tufekçi notes how important rative, which in turn shapes the narrative of the it is to recognise the risks and dangers associated movement itself. Tufekçi argues that speaking up with protest in certain contexts (Tufekçi 87). In “in a challenging way in public” tends to be the re- democratic societies such as Britain, the Nether- serve of the privileged (100). 2 Protest signs en- lands, and the United States, demonstrations in the able people who might not otherwise feel comfort- public sphere remain legal, and entail less risk of able to speak up, to carry their own voices, loud danger. However, both the People’s Vote Marches above their heads, and parade them through the and the Women’s Marches in part arose out of the public sphere, demonstrating their own convictions fear of such liberty being curtailed under new gov- and potentially persuading others of their beliefs. ernments and legislations. Thus, even in peaceful As Garlough argues, “rhetoric is the art of persua- and non-violent protests, the stakes remain high. sive speaking or writing, [but] it can be applied to Moreover, as Tufekçi argues, “protests have always other forms of aesthetic performances that aim to had a strong expressive side, appealing to peo- persuade” (272). Protest signs, as aesthetic objects ple’s sense of agency” (8). Furthering this, I argue that express political views, can be considered for that protest signs are emblematic of the expres- their performative rhetoric of political persuasion. sive qualities of a movement, as they facilitate the creative expression of individual agency. Tufekçi Texts and their rituals demonstrates that markers of a social movement’s success and strength “are more complex than indi- It is first important to consider signs in the con- cators like headcounts or number of protests” and text in which they are displayed. Exploring the outlines “three crucial capabilities of social move- materiality of text-based rituals in Judaism, Stolow ments from the point of view of power: narrative argues that “texts are governed by various proto- capacity, disruptive capacity, and electoral and/or cols that make them perceptible; engagement with institutional capacity” (191-2). Under this defini- them is not only intellectual but also embodied” tion, “narrative capacity refers to the ability of the (316). He outlines the rituals surrounding the cre- movement to frame its story on its own terms, ation and production of the Torah scroll, such as to spread its worldview;” “disruptive capacity” de- purification rituals that scribes must observe (319).

2As she also notes, in her own experience, it also tends to be the province of men.

137 Capstone Issue Vol. 12, 2019 Nina Klaff

These contribute to the religious significance of the 2). The street is an ideal stage for protest events holy text in its material form. I use this illustration in its symbolic meaning as a fundamentally pub- to demonstrate that there are rituals surrounding lic space and its material affordance as a facilitator textual objects. Protest signs become perceptible for the movement of bodies from one place to an- when they are displayed at protest events. Thus, other. As Reiss argues, “the street is a stage on any analysis of signs requires a consideration of which feelings and convictions can be expressed, the context in which they are displayed. As Tufekçi issues addressed, demands presented, and sup- argues, protest ceremonies have their own rituals, port solicited” (2). A people is ruled by those who which she terms a movement’s “repertoires of con- rule the streets upon which they walk. By gather- tention” (89). The protest sign is a text that has its ing on the streets, they reclaim that public sphere own rituals of production and use. It is customary as their own, thus symbolically reclaiming agency to make signs, and carry them at marches, holding over their governance.3 It is a demand for their them up high, marching them through the streets. presence in society to be acknowledged, for their They have become as much a part of the march demands as political agents to be heard, and for as the people who gather together in participation, their shared desires as individuals within a larger and integral textual objects in the ritual of protest group to be recognised.4 marches. Thus, if the street is to be thought of as a stage, we can begin to understand protest activities as Signs in Context: Protest Marches performances on that stage and ascertain how the devices used as props to that performance might In Dissensus, Rancie`re argues that “politics be- function. gins when those who were destined to remain in the domestic and invisible territory of work and The Embodied Assembly reproduction, and prevented from doing ‘anything else,’ take the time that they ‘have not’ in order to What marks the difference between a protest affirm that they belong to a common world” (139). placard and a banner or poster that is simply hung This is what happens at protest marches: a group out of a window, displayed on the street, or plas- of people takes the time to come together, united tered on a wall, is that signs are held by protesters on the common space of the streets, to affirm their in the context of demonstrations. Understanding unity of thought on a certain issue - which is most their significance thus requires careful considera- often political. tion of embodiment in protest rituals and an under- Protest events offer, as Reiss argues, “the op- standing that the gathering of bodies in the pub- portunity for social or political movements to show- lic sphere is itself a political act. It is the perfor- case themselves to the public in various ways and mance of “the right to appear, a bodily demand for solicit support” (3). The idea of a showcase is a more livable set of lives” (Butler 25). Indeed, one of particular interest as the language of protest merely has to cast an eye over the history of po- discourse is permeated with the lexicon of per- litical activism to conclude the importance of the formance. For the sake of this research, I fo- human body in socio-political contention (Kraidy cus on two protest marches that were peaceful, 2018). By participating in a march, as Reiss posits, non-violent performances of dissent in the pub- “the marchers embody not only the issue they lic sphere. Marches are a “relatively new phe- represent, but also the movement behind it” (3). nomenon:” until the 20th century, other forms This embodied co-presence of people in the public of public representation were favoured (Reiss 2). sphere is a bodily show of popular dissent in the Protest marches are “understood as organised and lexicon of politics. Butler reminds us that an as- choreographed processions of groups in the public sembly speaks volumes: “the coming together of a sphere with the aim of making a statement” (Reiss crowd has, as John Inazy contends, ‘an expressive

3Mitchell, in his analysis of the Occupy Wall Street movement, turns our attention to the empty, negative space against which assemblies appear and argues that the demand of “occupatio,” of embodied presence in the public sphere, is made in the full knowledge that public space is, in fact, pre-occupied by the state and the police” (10). 4I here refer only to democratic societies, such as the contemporary UK and United States, where protest marches remain legal. Indeed, the state and the police is often given fair warning of such events, in order for the proper preparations to be put in place. There is an interesting paradox in a government actually allowing and protecting citizens even as they voice dissent against them.

138 Capstone Issue Vol. 12, 2019 Nina Klaff

function’ prior to any particular claim or utterance essarily and not forever” (199). Its embodied na- it may make” (161). There is a powerful statement ture entails a materiality that is of great import. being made each time people gather, as they do, The space of appearance is thus not just public to express themselves, which is heightened when space, but is also created and sustained by the as- it is placed in the public sphere. As Butler argues, sembly of bodies. The “space of appearance,” she “the assembly is already speaking before it utters contends, “unlike the spaces which are the work any words” (156). An assembly itself is a demon- of our hands,” “disappears with the dispersal” of stration of solidarity and a way of communicating those that created it (Arendt 199). The collective a people’s message. It is a show of collective unity communion of bodies is thus a signiﬁcant site of between individual people, united in a singular aim political dissent. who are willing to put their bodies on the line to Durkheim coined the term “collective efferves- make a political statement. cence” in 1912 to explain the ebullient energy that I argue that holding a protest sign not only em- emerges in a gathering of people which is the feel- phasises the claim to the right to appear, but also ing of being “swept through one’s actions by a substantiates that claim as well as furthering what larger power when one is ‘within a crowd moved by is being expressed in an individual’s participation in a common passion’” (Durkheim quoted in Tufekçi an embodied protest. Arendt marks an important 89). Durkheim was speaking then of religious life; difference between individual strength and collec- however, as Tufekçi notes, he himself “showed tive power: “while strength is the natural quality that rituals also extend to secular activities” (89). of an individual seen in isolation, power springs up In Durkheim’s own words, “the very fact of as- between men when they act together and vanishes sembling is an exceptionally powerful stimulant” the moment they disperse” (200). Power, accord- (Durkheim qtd. in Gerbaudo 39). “Collective ef- ing to Arendt, necessitates a multiplicity of people: fervescence” can be felt whenever there is a crowd the crowd invokes the power of the people in the of people gathered together, united in one singular strength of numbers and through their union in one aim, belief, or ritual: as Tufekçi posits, “that tran- aim, and it is only when “men” - or indeed just peo- scendent feeling of being part of something larger ple - “are together in the manner of speech and than oneself applies also to protests, however secu- action,” that “the space of appearance comes into lar their aims may be” (89). The protest march is an being” (199). embodied tradition, a ritual of dissent, in which an assembly of individuals gathers together to demonstrate their unity as a group, united in a common aim, and “this relationship between belonging and individual expression is a key component of protest participation” (Tufekci 89). It is within this coming together, by gathering in the public sphere, in the creation of the space of appearance, that individual freedom, and indeed, dissent, is expressed, performed, and embodied. I argue that protest signs Figure 3: Klaff, Nina. Illustration of Patti Smith sign heighten “collective effervescence” as they signi- seen at the People’s Vote March, March 2019. ﬁes participation, which entails a feeling of belong- Indeed, the more people are present, the louder ing, and that these are, most importantly, objects their claims are made. Thus, while one individ- of individual expression. ual, or sign, might have strength, it is through their display in communion in numbers that the power Protest Marches as Carnivalesque Spec- of the voices they carry is truly enacted. Disre- tacles garding for a moment any other devices might be used to express political claims and demands, the Marches are performative spectacles; and as mere gathering of bodies in the public sphere cre- Werbner states, “voting lacks the spectacular as- ates and sustains the space of appearance, which pects of packed bodies in the square” (10). These Arendt argues is “wherever people gather together, “revolutionary moments,” Vaneigem argues, “are it is potentially there, but only potentially, not nec- carnivals in which the individual life celebrates its

139 Capstone Issue Vol. 12, 2019 Nina Klaff

unification with a regenerated society” (qtd. in Tan- emerging lens of materiality offers interesting per- cons 297). The analogy of the carnival, the popu- spectives on cultural objects such as protest signs: lar festival, is also often employed in protest dis- as Trentmann proposes, “after the turn to discourse course: these “festivals of resistance” are empow- and signs in the late twentieth century, there is ering occasions that, through their carnivalesque a new fascination with the material stuff of life” qualities, might fortify one’s personal identity while (283). Under this light, we may observe the ma- simultaneously consolidating one’s feelings of be- teriality of protest signs themselves: these are of- longing (Graeber 23). Werbner et al. remind us ten stuck on wooden or plastic sticks and bars to that “as Bakhtin showed for the medieval carnival, hold above the body. Activists might gather before carnivals are contained moments of ‘rebellion:’” a march to create their signs together, in solidar- against the rigidity of the quotidien, a popular cele- ity with each other, and the signs move with the bration of us against them (12). Tancons, however, protesters through the crowd, as much a part of the warns of the danger of the carnivalesque analogy constellation of the event as the bodies gathered undermining protest activities, and maintains that together of which it is contingent. Taussig, observ- the analogy of the carnivalesque is not meant to ing the Occupy Wall Street protesters in Zuccotti undermine the gravitas of a movement’s claim: Park, sees “the sign as an extension of the human “carnival is serious business” (Tancons 291). Car- Figure,” and the protester/sign fusion as “centaur- nival is a revolutionary performance practice that like” (75). An understanding of the symbolism they turns the status quo on its head and the perfor- carry thus requires a consideration of their nature mance of protest activities are not to be thought as material objects in their formal analyses, such of as in any way frivolous. In Tate’s words, peo- as in Chapters One and Two. ple “don’t do demos or entertain police assault for abstract carnivalesque goals” (quoted in Tancons 302). Simply by gathering in the public sphere, Mediation individuals make their potentially overlooked pres- In contradiction to Arendt, I propose that the ence in society overtly visible. Protest marches spectacle does not end when the crowd disperses. are feasts for the senses, and placards add to the For De Kloet, walking through the streets of Hong visual spectacle. Through them, protesters can Kong during the Umbrella Revolution, “the move- demonstrate their otherwise private individual po- ment felt like a theater show, a spectacle of resis- litical thoughts and emotions. Rancière defines pol- tance, built not only to be experienced, but above itics as when people “make the invisible visible, all to be mediated” (159). The aim of protest events and make what was deemed to be the noise of suf- is to interrupt the flow of daily life by creating a fering bodies heard as a discourse concerning the spectacle that will encourage the communication of ‘common’ of the community” (139). I argue that the reasons for the performance of dissent, in the protest marches serve to make the potentially in- hope that this will subvert the status quo. Werbner visible aims and desires of its participants visible, et al. remind us of the work of Juris, who “proposes and that what can differentiate between this noise that these ‘image events’ communicate to wider and the discourse concerning the ‘common’ is an audiences by ‘hijacking’ the global media, while at object as simple as a protest sign. the same time creating affective solidarity through performance” (11). This consideration of the atten- Materiality tion drawn by these protest events used as image events and the solidarity created through perfor- The protest sign is that which links the symbolic mative devices leads to a consideration of the aes- assertions, claims and demands made by the atten- thetic inventions that people use as props to their dance of a march with the claim of rights and free- performance of protest. As Butler recognises, “ev- dom of expression made simply through the cre- ery activist needs to negotiate how much exposure, ation of the space of appearance by the gathering and in what way, is necessary to achieve his or her of bodies. Within these signs, the strength of each political goals” (55). The ethical concerns of me- individual protester can be read, and when gath- dia framing - who gets to be seen and in what way ered together they represent the power and the - is a major concern, but as Serafini argues, social character of the movement itself. Moreover, the media “have enabled activists to record and then

140 Capstone Issue Vol. 12, 2019 Nina Klaff

share creative spectacles as never before” (321). point into the understanding of the importance of Moreover, activists can control how an event itself protest signs as cultural objects. I thus propose a appears by peppering the spectacle with signs that consideration of the People’s Vote March and the clearly say what they want to say, and in their own Women’s March to demonstrate how the aims and voices. Indeed, Taussig argues that in a protest character of a movement can be read in the protest event, “the sign holder is posing for photographers, signs they display. or, rather, the sign is being made to pose for the camera” (Taussig 75). The protest sign is no longer, nor perhaps has it ever been, conﬁned to the em- II. The People’s Vote March bodied spectacle, but is an object that demands to be remembered, mediated, and shared.

What Do Pictures Want?

In his seminal 1996 essay, Mitchell turns visual culture traditions on their head by displacing the Figure 4: Illustration of a sign seen at the People’s notions of desire that are usually read into images Vote March, 23rd March 2019. and returning agency to the images themselves. He asks: “What Do Pictures Really Want?” and con- Context cludes, through tracing visual culture “away from meaning and power to the question of desire”, that In June 2016, the United Kingdom held an ad- “what pictures want in the last instance, then, is visory referendum on its membership of the Eu- simply to be asked what they want, with the under- ropean Union. The British public voted to leave standing that the answer may well be, nothing at the EU, “forcing the resignation of Prime Minis- all” (Mitchell 81-82). Protest signs as political ob- ter David Cameron and dealing the biggest blow jects make political demands; they don’t ever want since World War Two to the European project of nothing. They demand that the voices of the people forging greater unity” (Faulconbridge). The Leave be listened to, and this depends on the individual campaign, popularised by then UKIP leader - now and the movement within which they are shown. leader of the Brexit Party - Nigel Farage, won with Stolow argues that texts, as they are “materially 51.9% of the votes, to Remain’s 48.1% (BBC News). embodied,” in books or digital media, or even in The marginal outcome of the referendum has been protest signs, “possess a sort of agency, demand- the subject of long-running contention, with the ing from their human users the acquisition of cer- Leave campaign then under legal scrutiny by the tain competencies and skills, and enabling various British High Court following allegations of “signif- modes of performance and action” (318). I thus icant overspending, data breaches and possibly use Mitchell’s question in specific relation to plac- Russian involvement in the referendum” (Bowcott). ards, and ask what do protest signs really want? I These “corrupt and illegal practices” have been argue that the desires of protest signs, as images ruled by the Court to “undermine the validity of the and texts at once, can only be read in the context decision to leave the EU” in a referendum that re- of the movement within which they are used. We mains advisory rather than regulatory (Bowcott). must first ascertain what the movement is asking Nonetheless, Theresa May’s government contin- for. ued to persist in its attempts to negotiate a deal. Before we can adequately answer what protest The People’s Vote organisation, led by James Mc- signs want, we first have to address what the gath- Grory, has been campaigning for a second refer- ered assembly itself wants, which common passion endum that they believe would yield a different unites them. The gathering of people in the pub- outcome. The movement gathered an estimated lic sphere makes demands in itself, but the explicit 700,000 people on the 20th of October 2018, and aims of that assembly are particular to each move- over a million again on the 23rd of March 2019 who ment. I posit a consideration of the elements in- marched the streets of London in the demand of a trinsic to protest marches as rituals of expressing People’s Vote. An analysis of this movement serves political dissent, such as protest signs, as an entry to give insight into some of the aesthetics, rhetoric,

141 Capstone Issue Vol. 12, 2019 Nina Klaff

and power at play in contemporary protest signs.5

Materiality: The Streets, The Signs

A city embodies the history and culture of its citizens and as Kapferer notes, “as a built forma- Figure 5: Ben Rabinovich. People’s Vote March tion, the city materializes, houses, and gives sub- route, Google Maps, 19 Oct. 2018. Accessed stance to its social and ideological structure” (69). 14/03/2019. I here consider protest activities through the lens of To access Park Lane by tube most conveniently, materiality to demonstrate how the concrete struc- one must alight at Marble Arch station, a stone’s tures of cities and their histories and geographies throw from the famous Speaker’s Corner, which are used by political movements as platforms upon Kapferer describes as “traditionally a venue for the which to express dissent through other material de- expression of popular debate and the airing of con- vices. As Tufekçi argues, cities “alter how we in- temporary contention” (77). Protesters then moved teract by gathering people in large numbers and to Picadilly, a landmark London road dotted with creating places for interaction outside of private high-end shops and hotels such as the famous The spaces” (5). The People’s Vote Marches took place Ritz, and then went down St James’s Street, past in London: the importance of its location lies not St James’ Palace, one of the oldest royal build- only in this being the capital of the United King- ings in the city, to Trafalgar Square. Kapferer re- dom: by considering the geography of the city and minds us that Mace calls the latter “The Heart of its landmarks, we may regard the route chosen as the Empire,” and herself describes its aura as “half- a symbolic reclaiming of the social and ideological tourism, half-ideology” (69; 75). Of the square, structure of the capital. Kapferer conducts an in- Kapferer says “first and foremost, it has long been depth analysis of the socio-political history of Lon- the great gathering ground of the British people in don’s monuments, which she posits “as crucibles times of triumph, jubilation, tribulation, and rage” for the expression, symbolization, formation, and (69). Its monuments, namely Nelson’s Column, “all re-formation of the social orders of the city and the unequivocally bespeak the power of an imperial state” (Kapferer 67). An examination of the geo- past” (Kapferer 75). Its central location, proximity graphical setting of the People’s Vote march and to parliament, and dotting with monuments to his- the landmarks they encountered and occupied will torical Figures are significant in that each of these serve to illustrate that the message conveyed by factors contribute to it being chosen as one of the assemblies is heightened against the backdrop of most common backdrops for voicing dissent. Thus, concrete symbols. it is in the concreteness of monuments, landmarks, Both the march on the 20th October 2018 and and areas of a city, and their durability against the that on the 23rd March 2019 started in the same fragile mortality of the human body, that protests place - and at the same time of day. Protesters first enact their symbolic value. This lens of materiality gathered on Park Lane, a road synonymous with af- also facilitates a consideration of the materiality of fluence, that borders Hyde Park - London’s iconic protest signs. Royal Park - and its vast spread of greenery adjoins London’s wealthiest and most influential areas.

5This analysis has been conducted in full recognition of the fact that every demonstration has many, many signs - especially one as large as the People’s Vote March - and that no one person could ever witness, let alone analyse, every one of them. I have thus selected those I observed personally and deemed of particular interest for the purpose of this paper. However, so long as these objects are in production, so long will their analysis require revisiting and revising. I thus encourage further research into other signs carried there and at other marches.

142 Capstone Issue Vol. 12, 2019 Nina Klaff

Aesthetic Regime

Protest signs are, first and foremost, material things. Fashioned out of found materials, these objects of artivism are made by the people, and this is reflected in their characteristic DIY aesthetic. Ran- cie`re argues that under the aesthetic regime of art, “art is art to the extent that it is something else than art” (118). Under this definition, I thus con- Figure 9: Klaff, Nina. “The £350m on the bus was sider protest signs as forms of art in that that they lies, lies, lies” sign, unpublished photograph, also serve a political function. Indeed, Rancie`re People’s Vote March, 20 Oct. 2018. asks us to reconsider our understanding of the “po- In Fig.9, the sign is made of cardboard, attached litical import of artistic practices” (135). I argue to a stick of bamboo on which an anthropomor- that the creation of a protest sign is both an artistic phised bus looks worried. The words “the £350m practice and a political act of protest in itself. It is on the bus was lies, lies, lies!” are handwritten in through the production of the objects that activists felt-tip in different sizes, and musical notes such are able to enact their political autonomy and ex- as quavers and beams decorate the background. press their own individual political thoughts. Car- This refers to the fact that in the months leading up rying them within the context of a protest march, to the 2016 Referendum, the Vote Leave campaign they participate in the production of the spectacle deployed a bus that traveled around the country and contribute to the character of the movement plastered with the promise that £350 million paid as a whole. by the United Kingdom to remain members of the European Union would be reallocated to saving the Some protest signs are straightforward, focus- British National Health Service on the event of U.K. ing simply on expressing the desires of the move- leaving the European Union (Getty Images). ment in a clear and focused manner, such as those in Figures 6-8.

Figure 6;7;8: Illustrations of clear signs seen at People’s Vote March, March 2019. Figure 10: Getty Images. Boris Johnson in front of However, it has also become customary to put the £350 million promise bus. The Independent, 8 effort into producing high-quality signs that empha- Sept. 2018. Accessed 29/05/2019. sise the aims of a movement and express an in- Garlough argues that “playfulness and creativ- dividual’s relationship with that movement in ever ity, drawing upon news events, historical materi- more innovative and engaging ways. There is a als, folklore and popular culture” is a practice in growing culture of producing effective protest signs many aesthetic productions of protest movements that will capture the attention of the media and (270). On the placard in Fig.20, the musical notes, onlookers. Multiple media platforms published se- as visual symbols, imply a melody, and the bus, as lections of “The Best Signs” shown at the People’s well as the text written on it, is evocative of British Vote March; from The Guardian’s “‘Fromage not nursery rhyme “The Wheels on the Bus,” in which Farage:’ the best signs and sights on the People’s the wheels go “round and round,” a line that is re- Vote march,” to The Slate’s “People’s Vote March: peated to a sing-song melody. Thus, the bus sub- The Best Signs From the Huge Anti-Brexit March verts a song that many will remember from their in London,” the signs themselves can make head- childhood, combines it with the anger at a news lines. event, and the subsequent effect is also humorous

143 Capstone Issue Vol. 12, 2019 Nina Klaff

to an extent. that aim to express opinion and solicit support for Fig.11 is a rather innovative sign. It is not made a movement (Tancons 296). Putting effort into the out of paper or cardboard but out of cloth, attached creation of a sign can indeed be a way of demon- to two sticks, its message patchworked out of felt strating one’s commitment to a cause through cre- fabric and marker pen. ative aesthetic practice. However, there is also a culture of resistance against this very phenomenon. Protesters have taken to making self-deprecating jokes about the limits of their signs (see Fig.13-14).

Figure 11: Nina Klaff. It’s A Stitch-Up Give us a FAIR say! Unpublished photograph, London, People’s Vote March, March 2019. The sign has two sticks, one on either side, which means it requires being held up by two people to be visible. It is held way above the heads Figure 13: Klaff, Nina. “Sh*t placard for a sh*t of the protesters. In this image, the sign is visible show.” Unpublished photograph, People’s Vote against the backdrop of the street, with a branch of March, London, March 2019. the prominent British bank HSBC discernible in the background. Taussig proposes that “it is the hand- madeness of the signs, their artisanal crudity, art before the age of mechanical and digital reproduction” that make them so impactful (75). Here, the “artisanal crudity” lies in its handmade quality. The sign is hand-stitched and handwritten; it declares the Brexit referendum to have been a “stitch-up,” meaning that it was deceptive. Philipps argues that “visual methods can add insight to the analysis of text-based protest messages” (8). The sign in Fig. Figure 14: Klaff, Nina. “My placard is a mess but 8 deploys a self-referential visual and verbal pun it’s not as bad as Brex-shit.” Unpublished uniting the medium and the message. However, photograph, People’s Vote March, London, March this visual and material pun is lost in Fig.12, when 2019. the text is alienated from its original manifestation.

Figure 12: Klaff, Nina. Illustration as transcript of Figure 15: Klaff, Nina. Illustration transcript of “It’s A Stitch Up” sign, People’s Vote March, 2019. “Placard is a Mess” sign. The “Stitch Up” sign is of arguably high quality, However, even as the protesters use profanities evidently well thought-out and executed. Protest to state they consider their placards artistically lim- signs are just one of the “countless examples of ited, they still objectively act as effective conveyors personal ingenuity” displayed in protest activities of dissent and signs of political contestation. What

144 Capstone Issue Vol. 12, 2019 Nina Klaff

Fig.13-14 might lack in artistic quality, they make mour as a revolutionary tool” and argues that hu- up for in an important aesthetic quality: humour. mour is thought “to exist as an entirely liberatory force in the aid of ‘the people’” (Holm 41; 38). Politicised Humour

Figure 17: Klaff, Nina. Illustration: Mark Twain, Figure 16: Klaff, Nina. Illustration of “Every Joke is “Against the Assault of Laughter Nothing Can a Tiny Revolution,” George Orwell (qtd. in Bruce Stand” (qtd. in Holm 48). 1). Just as the production and carrying of protest Contemporary protest signs often employ hu- signs is an expression of independent thought in mour: through the use of quick quips, written and the constellation of group political dissent, the visual puns, they can be fun and playful objects use of humour within those signs is also a rev- of dissent. Holm argues that humour is “an un- olutionary act. Holm reminds us that for Min- avoidable aspect of how we approach and under- dess, “humour operates as a key means by which stand the world as a site of meaning, politics, and a liberal subject might recognise and realise her life itself” (8). The humour in signs in no way triv- own absolute autonomy” (43). In peaceful non- ialises protest activities, however. As Holm pro- violent demonstrations, in which political animos- poses that we think of humour itself as politics ity is expressed through the gathered assembly (43), the humour in protest signs functions polit- and marching through the streets in a performative ically. Holm establishes the notion of a “political spectacle of protest, humour is a means of express- aesthetics” of humour, which he argues demon- ing individual autonomy. Holm argues that when strates “the idea that the aesthetic aspect of a text humour is used in this way it “is humour as a force - its form, style, palette, rhythm, narrative, struc- of freedom: not as an addendum, but as the very ture and form - can do political work, by which [he heart of a liberally focussed culture and politics of means] it can intercede in the negotiation, contes- well-mannered and reasonable dissent” (42). I ar- tation and distribution of power” (12). However, gue that protest signs contain humour in order to what we ﬁnd funny is, of course, limited to partic- do positive political work, to reclaim the narrative ular contexts. Thus, Holm differentiates between and to subvert the status quo through subversion humour, “an aesthetic quality operative at cultural of authority in a reasonably well-mannered fashion. level,” and “funniness,” which is “a particular sub- I again stress that humour is serious political jective reaction to those texts” (19). Thus, I shall work. The aptly named Montaignian essay “Comme not address whether the protest signs are funny, nous pleurons et rions d’une même chose” (“As but maintain that they often do contain humour, we cry and laugh at one same thing,” translation which can operate as a form of protest itself. Holm my own) demonstrates how outrage and despair deﬁnes “humour that directly addresses the con- can inspire comedic activities. In “The Anatomy tent of the political sphere,” including but not lim- of Schadenfreude,” Carroll Simon works closely ited to governments, is “politicised humour” (61). with Montaigne’s theories of comedy. He works to Thus, the humour in protest signs is politicised as demonstrate how laughter and crying, rather than they are objects that belong to the public sphere. being at opposite ends of the human emotive spec- In a New York Times article about the use of trum, actually are akin to each other. He argues jokes in the 2011 Egyptian revolution, Slackman that “for Joubert, the guffaw encompasses the sob, states that “the organizers used humor as part of but the reverse is not the case”; Joubert dubs this their communications strategy, to motivate people “a battle of two feelings” (Carroll Simon 268). He and bring out the crowds.” Holm interprets Slack- draws from Joubert to argue that laughter is “a gift man’s article as a demonstration of “the role of hu- of rest from the body to the mind” (Carroll Simon

145 Capstone Issue Vol. 12, 2019 Nina Klaff

268). These protest signs are not always laugh-out- loud funny. The jokes in protest signs are gifts of rest to the minds of people preoccupied with situations of political unrest, such as those disillusioned with the U.K. government’s handling of the Brexit referendum. They jest to emphasise the serious- ness of the situation. It is an act that subverts authority, as is visible in the double entendre in Fig.18.

Figure 20: Unknown Author. “The Bullingdon club of ‘87.” Ben Fenton, The Telegraph, 12 Feb. 2007. Accessed 19/05/2019. Eton is considered the pinnacle of the British elite. Eton mess is a symbol of British indulgence that has become part of its gastronomic hegemony. The sign declares the referendum and its consequences to be an Eton mess, a farrago instigated and exacerbated by the leaders who were educated there. Understanding this sign requires an integral understanding of British culture. It also requires a deep understanding of the widespread criticism of the current British government as being led by the elite - and poorly at that. Just as funniness is dependent on shared cultural norms and values, the rhetoric of protest signs often reflects the context in which they were created. The People’s Vote March, Figure 18: Klaff, Nina. Eton Mess. Unpublished as an Anglo-specific movement, has a very British photograph, London, March 2019. character in both its humour and popular culture references, but it has also been influenced by the more global culture of social media aesthetics.

Social Media Rhetoric and Aesthetics

Figure 19: Klaff, Nina. Illustration of Eton Mess sign at People’s Vote March, March 2019. Eton Mess refers to a traditional English dessert made of strawberries, meringues, and whipped cream, which is said to have originated at Eton College, the notoriously expensive private school for boys. The school has produced an extraordi- nary number of British leaders with David Cameron (photographed below, circled in the top row), member of its exclusive Bullingdon Club and the instiga- tor of the Referendum, was the school’s 19th alum- nus to act as Prime Minister, and Boris Johnson (also Figure 21: Klaff, Nina. Tweet by Liam Gallagher as photographed below, circled in the bottom row), protest sign, unpublished photograph, London, who became Eton’s 20th Prime Minister (Moss). 2019.

146 Capstone Issue Vol. 12, 2019 Nina Klaff

In the image in Fig. 21, the sign is made longer than usual signs. It is a full sentence, while of a large scale print out of a tweet by British most signs contain short quips and slogans. Finding celebrity and former Oasis band member Liam Gal- the maker of the sign, and quantifying how many lagher glued onto a whiteboard, which reads sim- times it had been shared on social media, proved ply “Foooooooooooking Hell” (@Liam Gallagher). It difficult: as an image, locating it through social me- was posted on the 20th of March 2019, three days dia searches is only possible if it is accompanied by before the protest and the date that Prime Minis- keywords, but there are a number of related tweets ter Theresa May requested a delay in Brexit proof interest. Fig. 24 depicts a Tweet from 2016, mere ceedings from the European Union. Not only are his days after the referendum, comparing Halliwell’s words included, but so is a screengrab of the origi- belief in her solo career to the Brexit referendum. nal tweet, along with the date, handle, and number of retweets and likes the post received. It remains recognisable as a tweet, exemplifying how, protest signs can be said to have evolved to include social media aesthetics. Perhaps, a more pertinent example of this is Figure 24: Screenshot of a tweet by user Fig.22, which is a sign that captured my own atten- @etunstead (Eoin). “Remember when. . . ” Twitter, tion at the People’s Vote March in October 2018. 24 Jun. 2016, 2.47 p.m. Accessed 29/05/2019. It received only two likes in response: while this is perhaps an insignificant first iteration of this nature, it was by no means the last. User @James- Melville tweeted, on the 6th of August 2018 (see Fig. 25):

Figure 22: Klaff, Nina. Halliwell sign, unpublished photograph, London, Oct. 2018.

Figure 25: Screenshot of a tweet by user @JamesMelville (James Melville). “Brexit is a bit like. . . ” Twitter, 6 Aug. 2018, 7.49 a.m., Accessed 29/05/2019. It was retweeted 165 times and received 628 Figure 23: Klaff, Nina. Illustration of Halliwell sign likes. It did not go viral, but it received sufficient at People’s Vote March, 2019. shares for it to be at least possible that the author I was drawn to this sign because, quite simply, I of the sign was overtly or subconsciously inspired personally found it funny. The Spice Girls were idols by this tweet - or perhaps others. Indeed, the day for those of my generation, children of the 90s, and before the first People’s Vote March, on the 19th their split was a British tragedy that resounded on of October, Twitter erupted with the debates of this the global stage. The U.K. is here likened to pop comparison. Halliwell fans also took to Twitter to singer Geri Halliwell and the E.U. likened to the defend their idol. Some used it as an argument in Spice Girls band, implying that the U.K. has overes- favour of leaving the EU. @CravenPartners stated timated its strength as a lone country once it is no that “When Geri left the Spice Girls she flourished” longer a member of the European Union. Consider and that through this move “she took back control” the sign’s length: at 102 characters, it is notably (Twitter, see Fig.26).

147 Capstone Issue Vol. 12, 2019 Nina Klaff

while the march was still happening, in real time.

Figure 26: Screenshot of a tweet by user @CravenPartners (Paul Craven). “When Geri left...” Twitter, 19 Oct. 2018, 8.43 p.m. Accessed 20/05/2019. Others took the jibe more literally, such as Figure 28: Screenshot of a tweet by user @husseybyname, who swore that if but one sign @bobbyalbon (Bob Albon RevokeArticle50). were to “wrongly suggest that Geri’s solo career “Brexit is like when Geri Halliwell. . . ” Twitter, wasn’t successful,” they promised to storm the 3.19pm, 20/05/2019. 6 march and lecture on the value of her song “Ride The sign went on to create a social media storm. It” (Twitter, see Fig.27). Albon’s tweet was retweeted over 792 times and received 3,042 likes. Stylist Magazine published an article entitled “Forget Brexit, this sign about Geri Halliwell is what’s really dividing voters,” which was shared all over Facebook (Murray 2018). Humorous engagement of that kind adds to the carnivalesque aspects of protest marches, which has been made possible through social media. The networked public sphere presents a new way of expressing political views through digital subversion and humour that prolongs the life of the protest well after the assembly itself has dispersed.

Figure 27: Screenshot of a tweet by user III. The Global Women’s March @husseybyname (A). “I swear to fuck...” Twitter, 19 Oct. 2018, 2.58pm. Accessed 20/05/2019. It appears whoever made the sign did not take heed of @husseybyname’s warning in Fig.27. I was not aware of this context when I took the photo, nor was I the only one drawn to the placard itself. Twit- ter user @bobbyalbon posted the tweet in Fig.28

6As I observed Fig.29 more closely, attempting to decipher any points of note regarding the image itself, I couldn’t help but notice that my mother and I are in the background. She is wearing sunglasses and a blue hat (bottom right), and the stars of my own matching blue E.U. beret are just barely visible behind the person in the yellow shirt with grey hair.

148 Capstone Issue Vol. 12, 2019 Nina Klaff

Figure 29: Klaff, Nina. Illustration: “I can’t keep public sphere” (or “networked public sphere”) to quiet” sign seen at Women’s March Amsterdam, explain this new conFiguration of the public sphere 11 Mar. 2017. (6). She argues that this “does not mean ‘online- only’ or even ‘online-primarily’”: It encompasses how the online and face-to-face worlds interact in Context social movements (Tufekçi 6). Thus, Tufekçi pro- The first Women’s March was held in Washing- poses that “the twenty-first-century public sphere ton D.C. on the 21st of January 2017, the day af- is digitally networked and includes mass media and ter Donald Trump was inaugurated as the President public spaces, such as the squares and parks where of the United States. The movement began on so- many protests are held, as well as new digital me- cial media and its marches and sister marches are dia” (6, emphasis my own). That is certainly the still organised online. The movement describes its case in the Women’s March, a movement that was aims in “Unity Principles”: ending violence, fight- catalysed and sustained through digital interaction. ing for reproductive rights, LGBTQIA rights, work- In late 2016, Theresa Shook, a lawyer based in ers’ rights, civil rights, disability rights, immigrant Hawaii, took to social media to voice her outrage at rights, and environmental justice. They want to the outcome of the 2016 U.S. presidential election. “create a global community in which women — One significant affordance of digital technologies including Black women, indigenous women, poor for social movements is that they give “ordinary women, immigrant women, disabled women, les- people [...] the potential to reach millions of people bian, queer, and trans women, and women of every at once” (Tufekçi 6). According to fellow Women’s religious, non-religious, and atheist background — March organiser Mrinalini Chakraborty, Shook “cre- are free” (Women’s March Global). An analysis of ated a private Facebook event page for the march the movement will aim to demonstrate how what and invited a few dozen online friends to join before was “heralded as the largest single-day synchro- going to sleep” (Tembe 224). As Tufekçi notes, “dig- nized global mass mobilization ever” has continued ital technologies” now “are especially important to mobilise transnationally, and on a global scale, during the initial formation of social movements” through social media (Tambe 223). I base some (6; 10). It is understood that Shook first posted the of this analysis on my attendance of the 2017 and event in a private Facebook group called “Pantsuit 2019 “chapters” in Amsterdam, The Netherlands. Nation,” set up for supporters of Hillary Clinton, who ran against Donald Trump in the 2016 election, to communicate with each other (Kairney; Tembe). Towards a Networked Public Sphere As Sandoval-Almazan and Gil-Garcia argue, “online Echoing the Arendtian supposition that the social media expands the channels for and veloc- space of appearance requires bodily co-presence ity of the message,” and its benefits are that it in- addressed in Chapter One, Gerbaudo notes that curs “no information costs” and presents “a com- “social movements have traditionally relied on the mon ground of technology” (46). Word then spread existence of local face-to-face networks” of embod- overnight through Pantsuit Nation and other similar ied co-presence (30). He argues that these “have online groups and Shook reported that the move- been considered almost unanimously by scholars ment went viral overnight: “when I woke, up it had as the most important channel for mobilisation” gone ballistic,” with hundreds of thousands of users (Gerbaudo 30). However, while both the People’s declaring their intent to attend the event (Kearney Vote March and Women’s Marches culminated in 2016). Then, seeing the viral response, “veteran embodied marches, the digital engagement be- activists and organizers began planning a large- tween the participants was, and continues to be, scale event scheduled for January 21, 2017, the just as important as their embodied manifestation. day after Inauguration Day” (“HISTORY” Women’s I argue that social media contributes to, and is in- March). extricable from, these movements’ embodied man- Moreover, not only can digital technologies in- ifestations. As Tufekçi argues, “the whole public crease the speed of communication, but they also sphere, as well as the whole way movements oper- “alter our sense of space” (Tufekçi 7). Our logic ate, has been reconFigured by digital technologies” of place has changed (Gerbaudo). What began as (6). Tufekçi proposes the term “digitally networked a politically charged Facebook post, urging one in-

149 Capstone Issue Vol. 12, 2019 Nina Klaff

dividual’s online friends to mobilise in what was with different views, keep taking to the albeit ge- then called the Million Women March in one Amer- ographically different streets, to enact the same ican city, is now a global movement that operates claims of justice. The movement is serialised in transnationally, composed of 52 “chapters,” or edi- that it continues to create new chapters, hold as- tions, of the movement, with a total of 673 “sis- semblies in new places, and garner support from ter marches” that gathered 4,814,000 marchers new participants all over the world. I argue that in cities around the world (“Women’s March” Web- the Women’s Marches’ aesthetic formation can be site). It is now known as the Global Women’s March. thought of in terms of the ebb and flow of a wave. I argue that the global nature of this movement was As Peeren et al. eloquently put it, “while a single only possible because of social media. wave may fizzle out on the shore, the sea never However, Tufekçi warns that while social me- stops moving” (12). Like a wave, the Women’s dia can help movements spread their message in March movement continues to swell in size, only unprecedented ways, with these affordances also taking respite until its next iteration, in a different “comes weakness, some of it unexpected” (70). place, becomes active. She argues that “the tedious work performed during the pre-internet era served other purposes as Global Aesthetic Formations well” (xiii). It “acclimatized people to the processes of collective decision making and helped One of the main aims of the Women’s March create the resilience all movements need to sur- movement is to create a global community of vive and thrive in the long term” (Tufekçi xiii). Ac- women. In Anderson’s Imagined Communities, he cording to Chakraboty, “the spontaneous nature of demonstrates how print media changed our con- the march’s origin meant that not a lot of thought ception of nationhood. He argues that communities had gone into either the logistical planning of such such as nations are “imagined” as many members a mass mobilization or other crucial details, such will likely never meet, “yet in the minds of each as the name” (225). As Tufekçi notes, “having lives the image of their communion” (49). Build- arisen so suddenly and grown so quickly,” many ing on Anderson, Tufekçi demonstrates that “peo- “networked movements have few means of dealing ple who would never expect to meet in person or with the inevitable internal conflicts of politics” (xiii, to know each other’s name come to think of them- 270). Indeed, the women’s movement came under selves as part of a group through the shared con- fire in 2019 as co-president Tamika Mallory report- sumption of mass media” (5-6). She argues that edly posted a photograph of herself with Louis Far- in the 21st century, it is “digital connectivity” that rakhan, the leader of the Nation of Islam, whose has reconFigured our conception of communities extremist views lead to his organisation being de- and reshaped how movements connect, organize, clared a “hate group.” Vesoulis reports that “in and evolve during their lifespan” (Tufekçi xix). In a the past, Farrakhan compared Jewish people to ter- movement as large as the Women’s March, which is mites and described them as ‘satanic’”. Mallory’s dispersed across different continents, most partici- post was captioned with “GOAT” - a contemporary pants will indeed never know many of their fellow- slang appraisal, acronym for Greatest of All Time. members. Indeed, Meyer argues that Anderson’s Mallory’s association with a known anti-semite re- fleeting remark that “communities are to be distin- flected badly on the Women’s March organisation. guished "by the style in which they are imagined" It was vilified in the media and even drew criti- (1991, 6, emphasis my own), hints at the impor- cism from Theresa Shook herself, who again took tance of scrutinizing how the binding of people into to Facebook to argue that this was “in opposition to imagined communities actually occurs and is real- our Unity Principles” and called for Mallory and oth- ized in a material sense” (6). However, Meyer asks ers to step down, stating that they had “steered the that we move away from the term “community,” Movement away from its true course” (Vesoulis). as she believes it is limiting: it suggests a “fixed, Despite a fall in participation numbers since the bounded social group” that is too rigid (6). She controversy, the Women’s March continued under posits instead the term “formation,” which, she ar- the same names, and with the same principles of gues, “refers both to a social entity (as in social equality. Continuing to build their community over formation) - thus designating a community - and social media, people around the world, at times to processes of forming” (Meyer 6-7, emphasis in

150 Capstone Issue Vol. 12, 2019 Nina Klaff

original). Rather than thinking of these formations anger even. The expression here represents a uni- as imagined, Meyer argues, we must acknowledge versal disillusionment with the status quo. The how communities and their imaginations “materi- piece of found cardboard, crumpled at the bottom alise through media” (7). This necessitates a more and with a line down the middle where it has been explicit focus on things and their thingness itself, bent in two, is held up defiantly by an activist, act- but allows us to “[retain] Anderson’s view that the ing as an extension of their Figure. The efforts to making of bonds in modern times depends on me- make the words look like printed typing demon- dia and mediation” (Meyer 7). Thus, a considera- strate the crossover of social media and the inter- tion of the Women’s March, a movement that was net aesthetic in the grassroots ethos of the move- organised, dissipated, and sustained through so- ment. Thus, the Women’s Marches’ protest signs, cial media on a global scale, requires a consid- just as those of the People’s Vote march, have also eration of the influence of social media and dig- become influenced by contemporary social media ital practices in binding people within the move- aesthetics. ment. Meyer argues, “more attention needs to be Another key way in which this is visible is in the paid to the role played by things, media, and the similarity of the signs visible in the chapters in en- body in actual processes of community making” tirely different locations. Consider the sign below, (6). She posits the term “aesthetic formations” to held by an activist at the Amsterdam chapter in argue that “community thus evolves around shared March 2019 (Fig.31) images and other mediated cultural forms” (9). I argue that the aesthetic formation of the Women’s Marches developed through social media, and that many of the shared images that constructed their aesthetic formations were protest signs that were shared online.

Social Media, Global Culture

Much of the aesthetic visible in the signs carried by protesters was evidently informed and inﬂu- enced by social media rhetoric, literacy, and aesthetics. In the sign below, an activist holds up a piece of cardboard inscribed with the words “JUST, Figure 31: Klaff, Nina. “NASTY WOMEN.” UGH.” in capital letters (Fig.30). Unpublished photograph, Women’s March, Amsterdam, 11 Mar. 2017. The phrase refers to an instance in which Trump called his opponent Hillary Clinton a “nasty woman” during one of the ﬁnal presidential debates. The Guardian reports that “NastyWoman was soon trending as people reacted against Trump” (Woolf). It became a slogan of the movement. It was turned into a poem by Nina Mariah Donovan and performed emphatically by Ashley Judd on the 21st January 2017, the video receiving over a million views (YouTube). The term also appeared on a Figure 30: Klaff, Nina. “JUST, UGH.” Unpublished protest sign in Amsterdam, and emerged in differ- photograph, Women’s March, Amsterdam, 9 Mar. ent in marches in other cities - in other countries. 2019. This is a turn of phrase commonly used on the internet to express exasperation. The ono- matopoeic “UGH” is a sound that hits the back of the throat, an expression of disgust, frustration,

151 Capstone Issue Vol. 12, 2019 Nina Klaff

terms of bodily co-presence: this excludes those not able to access these spaces. The Arendtian conception of the space of appearance as contingent on embodied co-presence public sphere does not account for those who are not able to be present for whatever reasons. Indeed, in Sick Woman Theory, Hedva takes issue with Arendt’s presupposition that political activism requires bodily presence in the public sphere: she says, “Arendt said, just get your body into the street, and bam: political” (Hedva). However, as Hedva demonstrates, The inevitability of violence at a demonstration – especially a demonstration that emerged to insist upon the importance of bodies who’ve been vio- Figure 32: Artisan. Scan of Page 22 of Why We lently un-cared for – ensures that a certain amount March: Signs of Protest and Hope - Voices from the of people won’t, because they can’t, show up. Cou- Women’s March. Artisan, New York, 2017. ple this with physical and mental illnesses and dis- In Why We March: Signs of Protest and Hope - abilities that keep people in bed and at home, and Voices from the Women’s March, an entire book de- we must contend with the fact that many whom voted to the protest signs of the Women’s Marches these protests are for, are not able to participate all over the world, the signs that refer to the “Nasty in them – which means they are not able to be visi- Woman” debacle are given a whole page spread of ble as political activists. (“Sick Woman Theory”) their own. They are here documented as appearing As Hedva pertinently asks, “how do you throw in London, Ottawa, and L.A. This demonstrates that a brick through the window of a bank if you can’t the rhetoric of protest signs is inﬂuenced by trend- get out of bed?” (Hedva). If political activism re- ing and viral events in mainstream and social me- quires embodied participation, how can one partic- dia and the humour is subversive. The images and ipate if one’s body is not safe in the public sphere, slogans its participants produce have been shared or if the space of appearance itself is not accessi- and spread widely and across distances, countries ble? Indeed, how can one hold up a protest sign, if and continents. Thus, just as the movement claims one can’t get out of bed? The answer may well lie globality, so too might its aesthetic formation. in digital media once again. In the Amsterdam chapter of the Women’s March, the marching assembly convened on Mu- Epilogue: Online Protest Signs seumplein, where protesters gathered to hear speeches by organisers and activists. The second speaker was Annika Mell, a disability activist who pertinently explained: “I didn’t march with you today” and stressed that this was “not because I don’t care” (Mell). She told the crowd how her physical disability prevented her from participating in the march, and that her experience is not an isolated case: Figure 33: Klaff, Nina. Illustration of “accessibility “We” didn’t march with you today because I am sets us free” sign, as seen at Women’s March not alone in this. We didn’t march with you today Amsterdam, 9 Mar. 2019. because we are too sick, too tired, too fragile, in too As their name implies, marches entail walking: much pain, too anxious, too depressed, too afraid through public space, in protest, together. The abil- of big cities, too afraid of big crowds to participate ity to do so is often taken for granted. If marches in a massive event like this. We didn’t march with continue to be the epitome of politics, the space you today because by its very nature marches such of appearance continues to be thought of only in as this are not accessible to everyone. [...] We

152 Capstone Issue Vol. 12, 2019 Nina Klaff

didn’t march with you today because we are sick and tired of being excluded from feminist spaces. (Mell 2019, transcript and emphases my own)

Mell’s vocal condemnation of the persistent in- accessibility of not only the Women’s March but the feminist movement in general demonstrates an inconsistency in the claims of equality and inter- sectionality. The Women’s March movement have made efforts to address this through holding online Disability Marches. Figure 34: Screenshot of “Emily’s” participant contributions to the Disability March. Disability The Washington Disability March began on De- March, 21 Jan. 2019. cember 21st 2018. According to their website, by While she was not able to join the march in per- 7pm on 21st of January, they had “1,654 entries son, her voice is not to be discounted: here, on published, and thousands more” that were not pro- the Disability March, she can vocalise her desire cessed (Disability March). That number represents “to Fight for so many other women and their rights” those who could not, for whatever reason, attend (Fig.34). As Tufekçi argues, the “availability” of me- the embodied march, but wished to remain visible dia “not only affects the kinds of discussions that in the space of appearance. On this page, partic- can be had, [but also] the kinds of people who can ipants posted representations of themselves that have them” (7). Media can thus be positive and acted as signs of their participation. They uploaded inclusionary forces within movements. The stan- images of themselves. Recall here Sontag’s claim dalone website devoted to those not able to march that “a photograph is both a ‘pseudo-presence’ and gave platform for their expression of their partici- a ‘token of absence’” (qtd. in Villi 2). Building pation in a movement, and I thus argue that these on this, Villy argues, “photographs offer presence- posts are to be thought of in equal standing as in-absence” (Villi 3). Moreover, these “tokens of those displayed at the embodied assemblies. More- absence” on the DM are not only visual, but also over, online assembled collectivities remain long textual. I argue that these would be the Disability after the gathering of an assembly has dispersed March’s equivalent to the protest sign. While they as their digital footprints remain online, and they did not participate in the embodied marches, their hold the potential to be re-engaged with time and presence within the movement is enacted through time again. their online postings. While the protest placard is traditionally limited to short, memorable slogans, With this in mind, I argue that there are other the posts on both the Washington and Amsterdam objects that might be considered as protest signs sites go into great detail. They describe their rea- that were not covered in the limited scope of this sons for absence, which are often extremely per- paper. The People’s Vote March, for example, cre- sonal. They explain what has kept them from ated a petition to “Revoke Article 50 and Remain marching (e.g. physical or mental illness, absence in the E.U.,” which, despite receiving over 6 mil- from country), and why they want to participate in lion digital signatures, was unsuccessful in its aim the movement. “Emily” posted, “I deal with sciatica (Geordianou).7 The government’s response was and uterine ﬁbroids, and on top of that don’t have again to assert its promise “to honour the result the funds to join the march in person. I want to of the 2016 referendum” (Geordianou). However, Fight to keep healthcare affordable and accessible each of those signatures acts as an autonomous to everyone” (Fig. 34). sign of dissent. While these digital signatures do not encompass the embodied, expressive, and aesthetic aspects of protest signs, they are of equal standing as their markers of participation within a movement; they are also persuasive acts. Individ- uals may use visual media in their online identity

7The government made this statement in response: “This Government will not revoke Article 50. We will honour the result of the 2016 referendum and work with Parliament to deliver a deal that ensures we leave the European Union” (Geordianou).

153 Capstone Issue Vol. 12, 2019 Nina Klaff

to perform protest digitally, by using filters on their “in spite of EVERYTHING” she was “taught by the Facebook profile pictures for example. media to hate” (@iWeigh). Through her already vast online platform, this image instigated a movement of more than 742,000 followers who participate by sharing similar photos of themselves. I argue that these are also to be considered aesthetic and embodied expressions of dissent that contribute to the creation of aesthetic formations of movements. While this project has focused on material and embodied protest signs, the digitally networked nature of the 21st century public sphere makes such limited consideration insufficient. I therefore encourage further research into the digital mutations of embodied protest signs that goes Figure 35: User Kate Mitchell’s profile picture with beyond sharing images of signs on social media. climate movement Extinction Rebellion’s filter, Facebook, 12 Feb. 2019. Accessed 30/05/2019. Included with permission from user. Moreover, there are also many, many ways of Conclusion: What do Protest Signs using social media to build collectivities that are Want? not aimed to be conducive to culminating in embodied assemblies, but are created only to operate online. One such example is the selfie-feminism @iWeigh movement, founded by celebrity Jameela Jamil, who posted the image below on her Insta- gram story out of outrage at the societal obsession with women’s weight.

Figure 37: Klaff, Nina. Illustration paraphrase of George Orwell, “Every Joke is a Tiny Revolution.” 2019. Social movements arise when people believe that “a wrong should be righted or a change should be made” (Tufekçi 25). In the two case studies I have addressed, ﬁrstly, the British people gathered in London to demand that the Brexit wrongs be righted, and secondly, at Women’s Marches all over the world, people have been demanding equal rights for women everywhere. I reiterate that while these were peaceful demonstrations, both movements arose out of a perceived threat to the liberty and safety of a people, and the stakes remain high Figure 36: Screenshot from @iWeigh instagram for both. I have not aimed to ascertain whether or page: Jamil, Jameela. “I Weigh. . . Fucking KG.” not these movements have been successful. This is Instagram, 2018. Accessed 29/05/2019. because I do not believe that the success of protest Jamil posted an image of herself with all the movements can merely be assessed on “instru- other things she believes to be more important mental aims” (Tufekçi 89). While the British govern- than the numbers on a scale, such as her “lovely ment remains resolute in proceeding with Brexit, relationship,” “great friends,” and her liking herself Donald Trump is still in ofﬁce, and the patriarchal

154 Capstone Issue Vol. 12, 2019 Nina Klaff

system that the Women’s Marches so fervently aim at both the People’s Vote and Women’s Marches to undermine is still very much in operation, as marches were featured in articles in news multiple Tufekçi argues, “protests can be ends as well as news outlets, from The Guardian to Vogue, and Ar- means. People wish to belong, especially to com- tisan even published a complete anthology of the munities that make them feel good” (Tufekçi 90). I Women’s March signs, called Why We March: Signs argue that protest marches, as performative rituals of Protest and Hope - Voices from the Women’s - carnivalesque spectacles of popular dissent - can March, which demonstrates their significance as be events in which people come together, united emblems of dissent (2018). in a singular aim, and that this in itself can be a The messages they carry continue to resonate positive force for good. These “festivals of resis- as they are published and shared on the global tance” are indeed carnivals “in which the individ- media stages, even as their function as embodied ual life celebrates its unification with a regenerated protest placards might vanish with the dispersal of society” in contemporary protest culture (Graeber the assembly, and become akin to the banner or 23; Vaneigem qtd. in Tancons 297). I argue that the poster. creation and display of protest signs are an important way of making a people’s unification and their desire for a regenerated society explicit. Indeed, as Rieff argues, “art is a means of union among [people], joining them together in the same feelings,” which is “indispensable for their community, however that community is conceived” (479). The artistic production and emphatic display of protest signs unites people in the protest event. They are thus to be understood as forms of activist art, or artivism, truly indispensable in the creation of “aesthetic formations” of social movements (Meyer 9). They are invaluable props to the dramaturgy of protest events through which individuals express their political thoughts and emotions. They are integral objects to the “reper- Figure 38: Klaff, Nina. Protest signs stuck to the toires of contention” of many social movements gates of the Houses of Parliament after the that allow for participants to display their auton- People’s Vote March, unpublished photograph, 23 omy through creative originality, even as they be- Mar. 2019. come one with a crowd and swept up in “collec- The Arendtian supposition of the “space of ap- tive effervescence” (Tufekçi 89; Durkheim). They pearance” being ephemeral and contingent on the are also important objects of cultural production embodied co-presence in one place no longer holds that reflect the norms and values of the context true (The Human Condition). Indeed, as Tufekçi ar- in which they are displayed: the politicised, and gues, the 21st century public sphere is “digitally notably British, humour in the signs at the Peo- networked,” as mainstream and social media en- ple’s Vote March, and the prevalence of social me- able its extension in both time and space (6). dia aesthetics and rhetorical influences in the signs Moreover, while the two movements in ques- both in London and at Women’s Marches around the tions might not have achieved their explicit aims, world, display their nature as artifacts of “participa- there is another marker of their success. I agree tory culture,” that reflects the context of the move- with Tufekçi that a key indicator of the strength of ment and the shared norms and values of its partic- a movement lies in its “narrative capacity,” how ipants (Jenkins). Moreover, because they are stan- it might “frame its story on its own terms” and dard features in contemporary protests, the signs “spread its worldview” (Tufekçi 192). I argue that are easily recognisable objects of dissent that of- protest signs are invaluable devices in construct- ten become the subject of attention themselves: ing this narrative capacity, and that protest signs they might go viral on social media, or be used as should be the subject of more research as material headline photos for news reports. The “best signs” and cultural objects of socio-political import within

155 Capstone Issue Vol. 12, 2019 Nina Klaff

protest discourse. I propose that further research Bowcott, Owen. “‘Corrupt’ Vote Leave campaign might include the signs produced and circulated undermines Brexit vote, court told.” The online-only. Guardian Online, 7 December 2018. Ac- People march because they care. They use cessed 5/05/2019. protest signs to demonstrate why and in what way Butler, Judith. Notes Toward a Performative The- they care. Aesthetic devices, under Rieff’s defini- ory of Assembly. Harvard University Press, tion, communicate feelings, in the hope that those Cambridge, MA, 2015, 154-92. to whom the messages are transmitted “also expe- Carroll Simon, David. "The Anatomy of Schaden- rience them” (479). When Mitchell asks “what do freude; or, Montaigne’s Laughter." Critical pictures really want?” he accepts that the “answer Inquiry 43, no. 2, Winter 2017, pp. 250- may well be, nothing at all” (81). However, protest 280. signs are objects of artistic expression, as well as “Charlie Chaplin - Protest Scene - Modern Times expressions of aesthetics, rhetoric, and persuasion; 1936.” YouTube, uploaded by Charlie Chap- what they want is never nothing at all. They want lin Official, 4th February 2019, Retrieved for the people to be asked what they want and for 1/3/2019. this to be heard. When people march, they are Couldry, Nick, et al. “Collectivities.” The Mediated making demands. They express their demands in Construction of Reality, 2017. the signs they make and hold themselves, and they Disability March, Disability March, 21 Jan. 2019. Ac- must be given attention. Whether they are humor- cessed 19/05/2019. ous or just plain angry, to paraphrase Orwell: every Downey, Anthony. Art and Politics Now. Thames sign is a tiny revolution. Hudson, London, 2014. Durkheim, Emile. The Elementary Forms of Works Cited Religious Life. translated by Karen E. “About.” PussyHat Project: Design Innovations for Fields, The Free Press, New York, 1995. Social Change. mbaker/cshs503/durkheimreligiouslife.pdf.style Adams, Bruce. Tiny Revolutions in Russia: Emily. “Why I Am Joining the March.” Dis- Twentieth-century Soviet and Russian his- ability March, Posted by FEMTEACH on tory in anecdotes, Routledge Curzon, 2005. 29/01/2019. Accessed 5/06/2019/. Anderson, Benedict. Imagined Communities: Re- @etunstead (Eoin). “Remember when Geri Halliwell flections on the Origin and Spread of Na- thought she could go solo? Yeah. Brexit.” tionalism, London: Verso, 1983. Twitter, 24 Jun. 2016, 2.47 p.m. Accessed Arendt, Hannah. The Human Condition. 2nd ed., 29/05/2019. University of Chicago Press, Chicago, 1958. Faulconbridge, Guy, and Kate Holton. “’Explo- Artisan. Why We March: Signs of Protest and Hope sive shock’ as Britain votes to leave EU, - Voices from the Women’s March. Artisan, Cameron quits.” Reuters, 23rd June 2016. New York, 2017. Accessed 8/05/2019. BBC News. “People’s Vote march: Hundreds of “‘Fromage not Farage’: the best signs and sights on thousands attend London protest.” BBC the People’s Vote march’.” The Guardian, News Online, 20/10/2018. 23 March 2019. Accessed 7/05/2019. BBC News. “EU Referendum: Results in Full.” BBC Garlough, Christine. “Vernacular Culture and News Online, 2019. Accessed 7/05/2019. Grassroots Activism: Non-Violent Protest Berlant, Lauren and Sianne Ngai. "Com- and Progressive Ethos at the 2011 Wiscon- edy Has Issues." Critical Inquiry 43, sin Labour Rallies.” The Political Aesthetics no. 2, Winter 2017, pp. 233-249. of Global Protest: the Arab Spring and Be- https://doi.org/10.1086/689666 yond, edited by Pnina Werbner et al., Ed- @bobbyalbon (Bob Albon RevokeArticle50). “Brexit inburgh: The Edinburgh University Press, is like when Geri Halliwell overestimated 2014, her viability as a solo artist and left the Gerbaudo, Paolo. Tweets and the Streets: Social spice girls. Best placard on the Peo- Media and Contemporary Activism, Pluto plesVoteMarch today.” Twitter, 3.19pm, Press, London, 2012, pp. 18–47. JSTOR. 20/05/2019. Accessed 21/2/2019.

156 Capstone Issue Vol. 12, 2019 Nina Klaff

Geordianou, Margaret Anne. “Revoke Article 50 Stylist, 2018. and Remain in the E.U.” Petitions, UK Gov- Peeren, Esther et al. Global Cultures of Contes- ernment and Petitions, 20 Aug. 2019. Ac- tation. Springer International Publishing, cessed 31/05/2019. 2018. Holm, Nicholas. Humour as Politics: the Political Philipps, Axel. “Visual Protest Material as Em- Aesthetics of Contemporary Comedy. Pal- pirical Data.” Visual Communication, vol. grave Studies in Comedy, 2017. 11, no. 1, 2012, pp. 3–21. https://doi- @JamesMelville (James Melville). “Brexit is a bit org.proxy.uba.uva.nl:2443/10.1177/1470357211424675 like when Geri Halliwell overestimated her Rabinovich, Ben. “Brexit: Everything you need to potential as a solo artist and subsequently know about the People’s Vote UK march,” decided to leave the Spice Girls.” Twit- Mail Online, 19th October 2018. Accessed ter, 6 Aug. 2018, 7.49 a.m. Accessed 14/03/2019. 29/05/2019. Rancie`re, Jacques. Dissensus: On Politics and Aes- Jamil, Jameela. @iWeigh. Instagram Account thetics, ed. and trans. Steven Corcoran, Jenkins, Henry, et al. Confronting the Challenges of Continuum, 2010. Participatory Culture: Media education for Reﬂect. “Ashley Judd’s EPIC "Nasty Woman" Speech the 21st century. MIT Press, 2009. At The Women’s March On Washington.” Kapferer, Judith. “Constructing a Public Sphere: YouTube, 21 Jan. 2017. Accessed on Materiality and Ideology.” Social Analysis: 21/05/2019. The International Journal of Social and Cul- Reiss, Matthias. ed. The Street as Stage: Protest tural Practice, vol. 51, no. 1, 2007, pp. Marches and Public Rallies since the Nine- 68–85. JSTOR, teenth Century. Oxford University Press, Kearney, Laila. “Hawaii grandma’s plea launches 2007. women’s march in Washington.” Reuters, Rieff, Philip. “Aesthetic Functions in Modern Poli- Kentish, Benjamin. “Brexit: Campaign to prosecute tics.” World Politics, vol. 5, no. 4, 1953, Boris Johnson over £350m NHS bus claim pp. 478–502. JSTOR, raises £24,000 in two days.” The Indepen- Seraﬁni, Paula. “Subversion through Performance: dent, 8 Sept. 2018. Accessed 29/05/2019. Performance Activism in London.” The Po- McQuiston, Liz. Visual Impact: Creative Dissent in litical Aesthetics of Global Protest: the Arab the 21st Century. Phaidon, London, 2015. Spring and Beyond, edited by Pnina Werb- Mell, Annika. Speech at Women’s March Amster- ner et al., Edinburgh: The Edinburgh Uni- dam, 10th March 2019. Footage available versity Press, 2014. on YouTube. Women’s March on Amster- SharpEdges. “Women’s March Amsterdam 2019. dam 2019 | Annika Mell speech | March 9 De speech van Naomie Pieter (Naomi 2019, Uploaded by Women’s March, The Balentien).” YouTube, 9 Mar. 2019. Re- Netherlands, published 14th April 2019. trieved 5/5/2019. Accessed 6/05/2019. “Sister Marches.” Women’s March On Washington. Meyer, Birgit. Aesthetic Formations. Palgrave Archived from the original on January 23, Macmillan, 2009. 2017. Accessed 7/5/2019. Mitchell, W. J. T. “Image, Space, Revolution: The Slackman, Michael. “When a Punch Line is Arts of Occupation.” Critical Inquiry 39.1, No Longer a Lifeline for Egyptians.” New 2012, 8-32. York Times, 5 April 2011. Accessed Mitchell, W. J. T. “What Do Pictures ‘Really’ Want?” 28/05/2019. October, vol. 77, 1996, pp. 71–82. JSTOR, Tambe, Ashwini. “The Women’s March on Washing- Moss, Paul. “Why has Eton produced so many ton: Words from an Organizer: An Interview prime ministers?” BBC News, 12/05/2010. with Mrinalini Chakraborty.” Feminist Stud- Montaigne. “Comme nous pleurons et rions d’une ies, vol. 43, no. 1, 2017, pp. 223–229. même chose.” The Complete Essays of JSTOR, Montaigne. Tancons, Claire. “Occupy Wall Street: Carni- Murray, Megan. “Forget Brexit, this sign about Geri val Against Capital? Carnivalesque as Halliwell is what’s really dividing voters.” Protest Sensibility.” The Political Aesthetics

157 Capstone Issue Vol. 12, 2019 Nina Klaff

of Global Protest: the Arab Spring and Be- Power and Fragility of Networked Protest. yond, edited by Pnina Werbner et al., Ed- Yale University Press, New Haven London, inburgh: The Edinburgh University Press, 2017. 2014. Vesoulis, Abby. “Women First Marched to Challenge Taussig, Michael. “I’m so Angry I Made a Sign.” Trump. Now They Are Challenging Each Critical Inquiry, vol. 39, no. 1, 2012, pp. Other.” Time Magazine, 19/01/2019. Ac- 56–88. JSTOR, cessed 11/05/2019. Trentmann, Frank. “Materiality in the Future of His- tory: Things, Practices, and Politics.” Jour- Werbner, Pnina, et al. The Political Aesthetics of nal of British Studies, vol. 48, no. 2, 2009, Global Protest: the Arab Spring and Be- pp. 283–307. JSTOR, yond. Edinburgh: The Edinburgh University Tufekçi, Zeynep. Twitter and Tear Gas: The Press, 2014.

158