Quantum Hypothesis Testing with Group Structure

Zane M. Rossi∗ and Isaac L. Chuang Department of Physics, Center for Ultracold Atoms, and Research Laboratory of Electronics Massachusetts Institute of Technology (MIT), Cambridge, Massachusetts 02139, USA (Dated: February 4, 2021) The problem of discriminating between many quantum channels with certainty is analyzed un- der the assumption of prior knowledge of algebraic relations among possible channels. It is shown, by explicit construction of a novel family of quantum algorithms, that when the set of possible channels faithfully represents a finite subgroup of SU(2) (e.g., Cn,D2n,A4,S4,A5) the recently- developed techniques of quantum signal processing can be modified to constitute subroutines for quantum hypothesis testing. These algorithms, for group quantum hypothesis testing (G-QHT), intuitively encode discrete properties of the channel set in SU(2) and improve query complexity at least quadratically in n, the size of the channel set and group, compared to na¨ıve repetition of binary hypothesis testing. Intriguingly, performance is completely defined by explicit group homo- morphisms; these in turn inform simple constraints on polynomials embedded in unitary matrices. These constructions demonstrate a flexible technique for mapping questions in quantum inference to the well-understood subfields of functional approximation and discrete algebra. Extensions to larger groups and noisy settings are discussed, as well as paths by which improved protocols for quantum hypothesis testing against structured channel sets have application in the transmission of reference frames, proofs of security in quantum cryptography, and algorithms for property testing.

I. INTRODUCTION the resulting classical data for as long as she likes, possi- bly using this information to instruct future actions. She Hypothesis testing is a fundamental statistical method is furthermore permitted query access to an oracle whose with wide application in classical and quantum contexts. result is the application of a single-qubit unitary quan- tum channel . This channel is from a publicly known Seminal work [1] has led to a deep information-theoretic Ei understanding of binary hypothesis testing for quan- set S (hereafter the query set) of n distinct unitary chan- nels. Queries consistently apply , and i is unknown. tum states, but only quite recently have analogous lower Ei bounds been proven for error in discrimination among Problem I.1. An S-QHT Problem is any instance quantum channels [2]. This forty-year gap between ma- wherein a party given access to i for unknown i [n] is ture theories for quantum hypothesis testing (QHT), re- tasked with the following: in as fewE queries as possible∈ de- alized as quantum state and channel discrimination re- termine, with certainty, the hidden index i. The minimal spectively, follows from the far richer structure of the expected query complexity the party is able to achieve is latter problem. I.e., general quantum channel discrimi- denoted qs and is taken over an assumption of equal pri- nation protocols may be adaptive, entanglement-assisted, ors on ` ` [n] = S, a set of distinct single-qubit unitary and use auxiliary qubits; moreover, the concomitant opti- quantum{E channels.} ∈ mizations over (possibly adaptive) preparations and mea- surements are computationally expensive. The prefix S in Problem I.1 denotes QHT with respect set It is known that sharpening the problem of quantum to a of quantum channels. This work examines only channel discrimination to narrower settings can drasti- specific subsets of S-QHT games. Moreover, this work cally alter algorithmic efficiency, the requirement of en- considers a specific resource model, described informally tanglement, the requirement of auxiliary qubits, and the at the beginning of this section and depicted in Figure1. ease of both theoretical and computational analysis [3– As described in SubsectionIB, na¨ıve upper and lower 5]. This work considers one such narrower statement of bounds on qs, even for general S, can be computed with- arXiv:2102.02194v1 [quant-ph] 3 Feb 2021 QHT for discriminating quantum channels. out difficulty, although the gap between these bounds is in general large, i.e., exponential in the instance size S [5]. A primary interest is thus to derive a set of properties| | on the set S for which a lower bound for q dependent on A. Problem statement s the structure of S can be both (1) proven and (2) asymp- totically achieved by a quantum algorithm exploiting the We state our problem as a game. Consider a party structure of S to generate a strategy for playing an in- with access to a small (single-qubit) quantum computer; stance of S-QHT (Problem I.1). she is able to apply unitary operations of her choice to This work provides one such sufficient condition on this qubit, measure this qubit in chosen bases, and store S. These constraints not only enable proof of query complexity lower bounds and constructions of algorithms achieving these bounds, but permit the cross-application of diverse methods in abstract algebra and functional ap- ∗ [email protected] proximation theory to quantum information and infer- 2 ence. This work considers the specific constraint that quantum channel discrimination, as well as some condi- S additionally faithfully represents a finite subgroup tions on the set S, e.g., teleportation-covariance (teleco- G < SU(2) (i.e., it is a representation of a finite subgroup variance) and geometric uniform symmetry (GUS) under of the group of single-qubit unitary transformations). which these bounds can be improved upon and, in the former, more restrictive setting of telecovariance, asymp- Definition I.1. A channel set S is said to faithfully totically achieved [2,6]. Such simplifying conditions have represent a finite group G if the elements of S have the also been studied in the multiple unitary channel case for form Ug g G such that, respecting some natural product { } ∈ group covarianct query sets for non-adaptive quantum operation for elements in S, UgUh = Ugh for g, h G, strategies [7]. and moreover that the group homomorphism ∈ is g Ug While it is known that entanglement (and in fact injective, ensuring S = G . 7→ | | | | any resource in a convex resource theory like quantum A variant of S-QHT incorporating the condition dis- mechanics [8,9]) can be useful in quantum hypothe- cussed above is denoted by G-QHT (Problem I.2). While sis testing among non-unitary channels, the performance this work considers groups G < SU(2), this game natu- of entanglement-free or low-entanglement strategies for rally extends to finite representations embedded in any multiple quantum channel discrimination remains largely Hilbert space. unstudied, even in its simplest, unitary form. Namely, while intriguing examples for methods of discrimination Problem I.2. An instance of Problem I.1 with the addi- among large sets of unitary operators where the use of tional constraint that S faithfully represents a finite group entanglement improves query complexity have been given G is an instance of a G-QHT problem or G-QHT game. [5], the necessity of entanglement is not known. More- over, the power afforded to quantum hypothesis testing Before discussing this new game further, it is worth- strategies for quantum channels using entanglement and while to understand previous results in unitary quantum which are also adaptive has been shown to be non-trivial channel discrimination, to which these games have non- in the case of non-unitary channels, where even adaptive- trivial relation. These results support why one should ness alone may assist algorithmic performance [10, 11]. expect that the family of sets S which obeys the prop- Many of the techniques referenced above are agnostic erties of Lemma I.1 is rich enough to furnish non-trivial to the structure of S; however, the notion that the struc- instances of QHT, and why even in a limited resource ture of the query set should inform the structure of op- model algorithms to solve G-QHT efficiently exist. timal procedures to differentiate members of S is an old and clever idea, and indeed can provide optimal hypoth- quantum B. Prior work esis testing protocols for query sets comprising states which are group covariant [12]. It is as a general- ization of this setting to quantum channels that Problem The problem of binary quantum channel discrimina- I.2 (G-QHT) finds its form. Moreover, the study of dis- tion is well-understood under the assumption that the set crete and especially non-abelian algebraic objects in the of possible channels, i.e., the query set, denoted S, com- context of quantum information is not new, and under- prises only unitary channels. Foundational work by Ac´ın 1 lies many open problems, e.g., the dihedral hidden sub- [3] asserts that there is always some finite upper bound group problem [13] and its reductions to various lattice on qs for achieving perfect discrimination for any finite problems [14], as well as the symmetric hidden subgroup S with distinct, known, unitary elements. Moreover it problem and its reductions to graph isomorphism [15]. is known that in the binary case, under the assumption Multiple hypothesis testing for quantum channels that the discriminating party may apply unitary opera- is not merely of independent quantum-information- tions of their choice, neither entanglement nor auxiliary theoretic interest either, but has found use in designing systems nor adaptive protocols are required to achieve protocols for the optimal transmission of reference frames optimal query complexity [4,5]. [16] (i.e., when the query set is a compact group and the For binary discrimination among pairs of general quan- aim is estimation of a fixed unitary transformation). Dis- tum channels, necessary and sufficient conditions are cretized versions of this problem also naturally connect known for the achievability of perfect quantum channel to the study of group frames and SIC-POVMs [17, 18], discrimination in terms of the channel’s respective Choi e.g., as discussed in Lemma VI.1. matrices [2]. Moreover, various general lower bounds are known for the symmetric error of discrimination (given While left as an open extension to this work, applica- a fixed number of channel uses) for binary and multiple tion of methods for quantum hypothesis testing against quantum channels where the n-th channel application de- pends non-trivially on the previous n 1 applications, i.e., memory channels [19] also have application− to proofs 1 This furnishes a loose upper bound for multiple unitary channel of the general impossibility of quantum bit-commitment discrimination as well; one performs perfect discrimination on [20], and are of interest in quantum cryptography. pairs of elements in S, eliminating channels one by one; this is In what follows we more concretely define our algorith- the standard reduction to binary QHT. mic resource model, provide an example of why it might 3 be expected that the question of achievability within function f : 0, 1 m [n] that returns the hidden index the exponential gap between the na¨ıve upper and lower i with certainty,{ } on all→ probable (i.e., non-zero probabil- bounds on query complexity for multiple quantum hy- ity outcomes of) m projective single-qubit measurements pothesis testing is richly structured, and finally give an Λ` ` [m] resulting from the action of i in a serial adap- outline for the methods of proof employed in analyzing tive{ } protocol∈ defined by the quantum algorithmE that uses this structure. qs oracle queries. This definition can be suitably modified replacing S with G, a faithful representation of the group G in a specified Hilbert space. C. Our approach While we will soon be interested in the efficiency of The statement of G-QHT (Problem I.2) together with a single-qubit serial adaptive query model algorithm in the serial adaptive query model depicted in Figure1 deciding a set S which faithfully represents a finite sub- raises the question of whether this model is (1) inter- group G < SU(2), and indeed whether, for these special esting, (2) non-trivial, and (3) tractable to analyze; this sets, query-complexity-optimal, entanglement-free, serial section addresses these questions. adaptive protocols similar to those constructed in [4] are The player challenged in G-QHT to determine the hid- possible to construct, it is worthwhile to look at a sim- ple, concrete instance of our game, and the function f it den index i of the queried channel i is afforded pre- cious few quantum resources. StatingE it another way, induces according to Definition I.2. the player is forced to devise quantum strategies in the We introduce a minimal instance of G-QHT which, in serial adaptive query model. In this model, pictured in addition to demonstrating why the na¨ıve upper bounds Figure1, the player may only intersperse their oracle on query complexity discussed in SubsectionIB are in queries with measurements and unitary operations de- general not tight, also captures some of the intuitive mo- pending on previous measurements. Serially, the querent tivations for the major results of this work for more com- learns progressively more about the hidden index i, adap- plicated query sets. The following example has the added tively modifying her approach. Under the assumption of benefit of (1) requiring no explicit mention of QSP (Sec- a small quantum computer and a reasonable classical one, tionII) in its construction and proof of optimality, and this is the most general approach she may take, assuming (2) providing some intuition for why QSP is natural to all measurements are projective and she wishes to deter- call on to solve the shortcomings that emerge in applying mine i with certainty. Furthermore, in this model, query the strategy of Lemma I.1 to more general query sets. complexity is a reasonable metric by which to judge al- Lemma I.1. For natural numbers n there exists a quan- gorithmic performance. tum algorithm in the serial adaptive query model that per- In addition to the serial adaptive query model, we fectly decides any channel set S that faithfully represents can quickly chart algorithmic schemes for instances of n a cyclic subgroup C2n < SU(2), and which requires 2 1 G-QHT where the querent is afforded a larger quantum oracle queries. − computer. In this case, the possibility for multiple-qubit2 unitaries and collective measurements gives rise to a va- Proof. For C2n , group elements are identifiable with bi- riety of series, parallel, and mixed strategies, which may nary strings of length n of which there are 2n, namely be adaptive or non-adaptive. The relative discriminating labeling according to the angle of rotation in the Bloch 1 n power of these models for specific instances of QHT and sphere in units of 2 − π such that the queried channel ro- specific query sets is not wholly understood. An informal tates about a known fixed axis by this angle. Concretely, depiction of some of these models is give in Figure2. up to overall unitary transformation the query set is As the querent in the course of playing the G-QHT n 1 n game is allowed to store reasonable amounts of classical S = Rx(m π/2 − ) , m [2 ]. (1) information, all that is asked of a successful quantum al- { · } ∈ gorithm for G-QHT in the serial adaptive query model is Any decision protocol using one qubit for readout can that it is able to decide the hidden index i according to provide at most one bit of information as to the n-bit some efficiently computable function on any of its prob- label for the queried group element.3 We read from least able binary qubit measurement outputs. This statement (LSB) to most (MSB) significant bit by the following is made concrete in Definition I.2. algorithm:

n 1 Definition I.2. A quantum algorithm in the serial adap- 1. Prepare 0 . Query the channel 2 − times and tive query model is said to decide on a query set S of measure| ini the standard basis, reading the LSB. distinct unitary quantum channels of size n in qs queries if there exists, for all i [n] a computable deterministic ∈

3 Note that these don’t need to bits in the label of the queried channel, but rather some set of bits which, at the conclusion of 2 One could of course also imagine access to qudits, or indeed the algorithm, can be taken by the function f to the hidden index stranger Hilbert spaces. i deterministically. 4

U1,1 U1,2 U1,n Ei Ei Ei 1 Ei

U2,1 U2,2 U2,n Ei Ei Ei 2 Ei

Um,1 Um,2 Um,n Ei Ei Ei m Ei

FIG. 1. A general circuit to perform QHT in the serial adaptive query model. The unitary operators Ui,j for i ∈ [nj ], j ∈ [m] may depend on previous single-qubit projective measurements Λk for k < j, for j ∈ [m], communicated by stored classical bit strings of reasonable finite length (represented by arrows). Each row in the figure is a quantum circuit applied to a qubit prepared from classical information depending only on the results of previous measurements. The serial nature of the discrimination protocol to determine the unknown channel is evident; when the protocol terminates a known classical function is computed on the set of measurement results (here, a bit-string of length m), equivalently Λk for k ∈ [m], to infer the hidden channel. Other models one can consider are discussed in Figure2.

(a) serial

(b) parallel (c) mixed

FIG. 2. Simplified illustrations of different models for quantum circuits performing QHT. Depicted are (a) serial adaptive, (b) parallel, and (c) mixed strategies. Given query access to a finite number of applications of the unknown quantum channel (red, outlined), in the figure 6 applications, the querent is conferred the ability to perform unitary operations (blue, non-outlined) of her choice. Blue operations shown are arbitrarily structured and for demonstrative purposes only. For serial adaptive strategies, (a), dashed boxes indicate regions between which only classical information is transmitted (i.e., measurement results). In (b) and (c) entanglement, auxiliary qubits, and collective measurements can, in general, improve the performance of QHT algorithms. Preparations and measurements are not explicitly shown.

n 1 n 2 n 2. Dependent on the measurement in the previous complexity 2 − + 2 − + + 1 = 2 1. Optimality ··· − step the possible query set S0 has description follows from the optimality of phase estimation.  n 2 n 1 R (m π/2 − + π/2 − ) if measured 1 { x · } | i The methods used in the proof of Lemma I.1 illustrate n 2 Rx(m π/2 − ) if measured 0 , an important concept: if the query set S is highly struc- { · } | i tured, binary measurement results can effectively corre- n 1 for m [2 − ]. The latter is a representation of spond to halving the size of the remaining search space ∈ n 1 the cyclic group of order 2 − . The former, if each (or equivalently excluding, with one measurement, half query is preceded by a unitary U = R ( π/2n), is x − of the possible channels). Here, compared to the upper also a representation of this cyclic group. bound given by the standard reduction to binary QHT, n 1 Set U = R ( b π/2 − ), where b was measured we see a square root improvement in the instance size x − · | i in the previous step. C n . Additionally, the function f from the statement of | 2 | n 2 Definition I.2 simply reads the adaptive output measure- 3. Apply U before each of 2 − channel applications ments as a binary string and returns the corresponding to bit-shift the label of the queried group element. integer (the channel’s hidden index). Repeat algorithm for a cyclic group of size 2n 1. − The reason that the simple method of Lemma I.1 works For the cyclic group of order 2, consisting of the identity is because even powers of channel elements are not only n 1 channel and a π-rotation, the decision protocol is obvi- subsets but subgroups of C2n , and specifically 2 − pow- ous. By recursion, the total decision protocol has query ers of group elements are rotations by angles in 0, π , { } 5

a2 a a2 a This work is organized such that algorithms for decid- g g2 ing simpler finite groups can, where applicable, be used 7→ a3 a8 a3 a8 as subroutines for algorithms deciding more complicated groups whose subgroup decomposition is non-trivial. It is this bootstrapped approach that provides novel sufficient a4 a7 a4 a7 conditions under which the open question in Subsection IB can be resolved in the serial adaptive query model. a5 a6 a5 a6 We begin with an overview of the two mathematical techniques that underlie the main results of the paper. Namely, in SectionII we review statements of the main FIG. 3. Subroutine of decision protocol on C8. For the cyclic n 2m theorems of quantum signal processing, their guarantees, group of order 2 , any map g 7→ g for m < n generates a and interpretations. Relatedly, we give a protocol (Algo- cyclic subgroup of order 2n−m. Consequently, as the cyclic group of order 2 has an obvious discrimination strategy, the rithm1) that players of a simplified version of the G-QHT method in Lemma I.1 can recursively determine membership game (Problem I.2) defined in SubsectionIA may use to of the hidden element in cosets of cyclic subgroups of C2n . achieve perfect decision protocols. The theorems of QSP Equivalently, the querent performs binary search, i.e., using (and consequently solutions to the simplified game pro- 2n−1 queries, she can determine membership of the hidden posed in Problem II.1) rely on the existence and efficient element in the red (image) or blue (complement of the image) computability of polynomials over real variables under subset as pictured for the case n = 3, assuming she can solve simple constraints, the properties of which are discussed the n = 2 case. in SectionIII. With both of the mathematical techniques established which give perfectly orthogonal and thus perfectly dis- in SectionsII andIII, the paper proceeds to discuss con- tinguishable states when acting on special known initial crete groups systematically. The statement of Problem pure states. The adaptive protocol permits the querent I.2 as mentioned is simplified to Problem II.1, whose so- to recurse and learn the hidden index by asking individ- lution using the methods of QSP depends solely on the ual questions of coset membership for prime-power order answer to questions in functional approximation. For normal subgroups. each concrete algorithm corresponding to deciding each For cyclic groups of general order, however, this finite subgroup G < SU(2) in SectionIV, we perform method fails. For odd-order cyclic groups, for instance, reductions to decisions on normal subgroups of G where sets of integral powers of group elements do not neces- possible, and restate decision algorithms on G as multi- sarily form non-trivial subgroups by simple consequences ple correlated instances of Algorithm1. Specifically, we of Lagrange’s theorem. The question of bisecting the assert that Algorithm1 and its performance guarantees search space must thus be resolved by other methods; it are integral to the analysis of each algorithm given in is precisely the flexibility of QSP that will permit the re- SectionIV. covery of algorithms of the same flavor as Lemma I.1 for more general groups. That is, to permit the construction Algorithm1 connects decisions on G to problems in of quantum algorithms that act deterministically on not functional approximation which, referring back to the merely subgroups but arbitrarily chosen subsets of the guarantees of SectionIII, determine the query complex- query set. ity of the algorithm deciding on G. This connection is made explicit in Problems III.1 and IV.1. We provide a diagram of the order in which we ad- D. Paper outline and summary of results dress decisions on specific finite subgroups (Figure4) as well as relations between all problems introduced in this The main body of this work describes methods for per- work (Figure5). In turn, the relations between algo- fectly deciding sets of quantum channels (equivalently rithms and problems are summarized in the statement query sets) which faithfully represent finite subgroups of Algorithm1 in conjunction with its accompanying re- G < SU(2) in order of increasing complexity of the fi- marks (Remarks II.1, II.2), toward a coherent framework nite group considered. This culminates in Theorem I.1. for hypothesis testing on discrete query sets. Theorem I.1. [Simplified] There exist quantum algo- For generalizations to larger Hilbert spaces, near- rithms in the serial adaptive query model which perfectly unitary channels, and groups not embeddable in SU(2), decide on all finite subgroups G of SU(2), with the excep- the reader is directed to SectionV. Additionally, Section tion of the simple non-abelian group A5, and which do VI gives a list of open problems in the same vein as the so with asymptotically optimal query complexity. These results presented in this work, suggestions for the shape algorithms each closely track with a single generic al- of their resolution, and instances (e.g., Remark VI.1) in gorithm (Algorithm1), and their individual structure which the methods derived here can be directly applied closely tracks the structure of the considered group. to physical problems. 6

R-QHT P-QHT

? C2n Cpn Cp Cn D2n A4 S4 A5

FIG. 4. The linear flow of this work: deciding on increasingly rich sets of finite subgroups of SU(2). The diagram indicates the order in which instances of G-QHT are solved throughout SectionIV, beginning with cyclic groups and working toward the dihedral and platonic groups; solid arrows indicate increasing complexity of the decision group, while dotted lines indicate where a reduction to an algorithm deciding on the latter group is particularly simple. R-QHT (Problem II.1) and P-QHT (Problems III.1 and IV.1) are developed in parallel to decision protocols on cyclic groups, and are joined for decisions on prime order groups by Theorem IV.1. Applying similar methods to A5 is left to future work.

hidden parameter g of the group action. Naturally, the ideal method for extracting information from the queried channel varies with the structure of G. Taking inspiration S-QHT from algorithms for quantum sensing in the serial query model, we thus might naturally consider the flexible, re- cently developed techniques of quantum signal processing G-QHT (QSP) [21–24]. QSP is a powerful quantum algorithmic primitive to implement matrix polynomials on quantum computers under only mild constraints [24]. Analysis of QSP has enabled intuitive constructions for asymptotically opti- mal algorithms in a range of settings from Hamiltonian P-QHT simulation [22] to the quantum linear system problem [25] in [24, 26, 27]. For our purposes, however, we will need R-QHT only to consider the guarantees of the form of QSP pro- tocols, succinctly stated in the following two theorems. Before this we briefly address an issue of notation. FIG. 5. Inclusion relations among problems formally defined Definition II.1. In this work the convention when re- in this work. Four major problems discussed: S-QHT (Prob- ferring to the Pauli operators is lem I.1), G-QHT (Problem I.2), R-QHT (Problem II.1), and P-QHT (Problems III.1 and IV.1), referring to set, group, ! ! ! 0 1 0 i 1 0 rotation, and polynomial quantum hypothesis testing respec- σx = σy = − σz = , (2) 1 0 i 0 0 1 tively. Each region in the inclusion diagram contains non- − trivial instances. and moreover we will often refer to a linear combination of such operators following the convention

II. OVERVIEW OF QUANTUM SIGNAL σξ σx cos ξ + σy sin ξ, (3) PROCESSING ≡ where this construction will often be used in the context of defining a rotation about a fixed axis on the Bloch sphere, We have defined the G-QHT problem (Problem I.2) as namely well as the form that any algorithm in the serial adap- R (θ) exp i(θ/2)σ , (4) tive query model solving this problem must take. We ξ ≡ {− ξ} have not, however, provided a method for analyzing such where this is distinct from the convention of [24]. If the algorithms. For certain groups, e.g., C2n as covered in index is Latin instead of Greek, e.g., Rx(θ), then it is Lemma I.1, we can come up with methods inspired by meant exp i(θ/2)σx : rotation about the xˆ vector. classical algorithms; this intuition breaks down for more {− } Theorem II.1. In [24]. Let k N; there exists Φ complicated groups. In this section we introduce tech- ∈ ∈ Rk+1 such that for all x [ 1, 1] niques toward addressing this breakdown. ∈ − G-QHT might be naturally thought of as a sensing k iφ0σz Y iφj σz
problem: given an unknown g, application of the chan- e W (x) e = nel U (respecting a representation) might be physically j=1 g   explained as the result of probing a system: the action P (x) iQ(x)√1 x2 of the quantum channel contains some information about  −  , (5) 2 the system. Successive queries increase knowledge of the iQ∗(x)√1 x P ∗(x), − 7 iff P,Q C[x] satisfy the following properties: As the rotation operators discussed in the R-QHT ∈ problem satisfy the form expected of the W (x) opera- 1. deg(P ) = k and deg(Q) = k 1. − tor in QSP, the methods of QSP suggest a neat prescrip- 2. P has the same parity as k modulo 2, while Q has tion for a quantum algorithm (Algorithm1) with classi- the opposite parity. cal subroutines such that the output is a solution for the R-QHT problem. We discuss assumptions on the input, 3. For all x [ 1, 1], P and Q satisfy P (x)P ∗(x) + output, and structure of Algorithm1 in Remark II.1, give 2 ∈ − (1 x )Q(x)Q∗(x) = 1. definitions for its classical subroutines in Definition II.2, − and further remark on where the non-trivial aspects of Theorem II.1 asserts that QSP protocols, which involve Algorithm1 lie in Remark II.2. interleaving rotations about orthogonal axes (one of these rotations by a fixed, unknown angle, and the other by an Remark II.1. We present a series of data structures unfixed, known angle) result in unitary operators whose which together define both an instance of the R-QHT elements are polynomials of the unknown rotation an- problem (Problem II.1) and its solution, toward a con- gle. These polynomials are under constraints necessary crete algorithm (Algorithm1). and sufficient to ensure the resulting operator is unitary. While the constraints of Theorem II.1 are non-intuitive • Input: Any instance of R-QHT presupposes access for one wishing to solve the reverse problem (i.e., go from to classical information in the form of a list of dis- polynomial to a unitary operator in which the polynomial tinct angles θ` [0, 2π] , ` [n]. R-QHT also pre- { ∈ } ∈ is embedded), the following theorem addresses precisely supposes access to a quantum oracle which, when this concern. called, applies a quantum channel channel Rξ(θi) for fixed i about some known fixed axis ξ. + Theorem II.2. In [24]. Let k in Z and let P 0,Q0 ∈ R[x]; there exists some P,Q C[x] satisfying the require- • Output: In the serial adaptive query model on ∈ ments of Theorem II.1 such that P 0 = R(P ),Q0 = R(Q) qubits, a projective measurement is an evaluation iff P 0,Q0 satisfy the first two requirements of Theorem of a probabilistic binary function on possible hid- II.1 and additionally P (x)2 + (1 x2)Q (x)2 1. den indices j [n] for the applied channel. An 0 0 ∈ The proof of this statement follows− constructively≤ from R-QHT algorithm’s output is one of these indices, a provably efficient (e.g., polynomial in k) algorithm to where success is dictated by high probability4 of or build the missing complex parts of P,Q. certainty in returning the proper hidden index i. In Theorem II.2 the operator W (x), the signal being • Assumptions: The result of the evaluation of processed, will be analogous to the quantum channel a set of these functions (corresponding to m bi- Ei we wish to discriminate in G-QHT. That said, the utility nary measurements), fj :[n] 0, 1 , j [m] 7→ { } ∈ of these theorems is not immediately clear: the form of on the hidden index i of the queried channel, is W (x) (rotation about a known, fixed axis) is far simpler a composite function g : i 0, 1 m defined as 7→ { } than the members of the query set considered in G-QHT g(i) = f1(i)f2(i) fm(i). ··· for arbitrary finite subgroups of SU(2). If this function is injective for all j [n] then In the interest of making progress, we can thus modify ∈ 5 the algorithm generating the fj solves R-QHT. the statement of Problem I.2 such that QSP stands a fair Equivalently the algorithm computes a series of m chance of providing a solution. Specifically we can write equivalence relations on the set of rotation angles out the generic form of a QSP-based algorithm that per- θ` , ` [n] such that every element is uniquely de- n { } ∈ fectly decides any finite set S = Rx(θ`) ` [n] [ π, π) fined by its membership under these m bisections. { } ∈ ∈ − under the map Rx(θ`) = exp iθ`/2σx . Note that here S need not be a group under{− composition.} This modified Definition II.2. A quantum algorithm solving the R- version of the G-QHT game is discussed in Problem II.1. QHT problem (Problem II.1) is referred to simply as an R-QHT algorithm, where solves indicates that it satisfies Problem II.1. The rotation QHT problem ( R-QHT the input, output, and structural assumptions presented problem) is a simplified version of the G-QHT prob- in Remark II.1. lem (Problem I.2) with the following structure. Given In addition, toward an explicit description of one such query access to a single-qubit quantum channel from R-QHT algorithm (Algorithm1), we define four classical among a finite set S where each channel has again the sub-algorithms whose application together constitutes the form R (θ ) = exp i(θ /2)(cos ξσ + sin ξσ )) for dis- ξ i {− i x y } classical subroutine of Algorithm1). tinct, known θi and known rotation axis ξ, determine the queried channel with certainty in the serial adaptive query model. Note that R-QHT problems are not a subset of G-QHT 4 In the noiseless case, we consider only deterministic algorithms. problems, save in the case that the set of angles θ` are 5 This is a non-trivial condition to satisfy, but in most instances all distinct integral multiples of 2π/n for positive{ integral} can be thought of as assigning a binary tree’s labels to each of n (i.e., S represents a cyclic group). m channels. This is the subject of Remark II.2. 8

• genBisection: Given a group representation G Algorithm 1: A generic algorithm for solving and a (possibly empty) set of evaluations of pre- R-QHT vious binary functions f : S 0, 1 for S j j → { } j ⊆ Sj 1 S1 G, returns a description of Assumptions : Input and output satisfying f − :⊆S · · · ⊆ 0, 1⊆ where S S is a subset assumptions of Remark II.1 j+1 j+1 → { } j+1 ⊆ j of Sj on which fj is constant. Input : A quantum channel oracle Ei for hidden index i; description of n The choice of fj+1 is not arbitrary but instead de- channels {E`}`∈[n]. pends heavily on the embedding of G in a larger Output : The hidden channel index i. continuous group. Examples for methods of choos- for j ← 1 to m do ing these fj can be found in the concrete algorithms of SectionIV. Further discussion of the properties Classical subroutine (see Def. II.2): of these functions is also covered in Remark II.2. fj ← genBisection(G, {f

• genComplexPoly: Given a real polynomial pj satis- fying the constraints of the output of genRealPoly, returns a pair of complex polynomials (Pj,Qj) Remark II.2. The difficulty in Algorithm1 stems from on [ 1, 1], each of definite parity and satisfying selection of the proper functions for sub- −2 2 2 fj : Sj 0, 1 Pj(x) + (1 x )Qj(x) = 1 for x [ 1, 1]. More- → { } − ∈ − sets Sj S of the query set of fixed-axis rotations (equiv- over R(Pj) = Pj0 = pj and R(Qj) = 0. One imple- alently⊆ computing genBisection in Definition II.2). mentation is given in [24]. As each fj takes values on Sj in 0, 1 , they can be thought of as labels dividing or bisecting{ } the query set; • genPhases: Given two polynomials (Pj,Qj) satisfying the constraints on the output of the result of QSP is to make the quantum computation of genComplexPoly, returns a set of phase angles these fj on the hidden index i deterministic. A series of k+1 these f thus form the levels of a binary tree whose bisec- Φj R satisfying Theorem II.1. j ∈ tion condition is the result of a projective measurement This subroutine also returns a classical description onto ψj0 , ψj0⊥ . We discuss the desired properties of of two quantum states, ψj, ψj0 , the former an initial this binary{| i decision| i} tree; these principles foreshadow the state and the latter a state with respect to which a properties discussed in Theorem IV.1. projective measurement is performed to compute fj on the hidden index, i.e., fj(i). These states are ef- ficiently computable and project out p , equivalently j • An efficiently searchable binary tree should be bal- ψ U ψ = p , where U is the QSP unitary j0 Φj j j Φj anced; different channels should have binary labels generatedh | | byi . Φj according to the tree which differ as early as possi- Methods for computing these phase factors are nu- ble, equivalently each fj should divide the remaining merous [24, 26, 28]; all affirm that this computation query set roughly in half. is efficient and stable, using existing techniques in classical optimization. • The discrete fj objects are accessed by interpolat- • We denote by Mψj ( ψ ) the measurement project- ing polynomials in a continuous embedding space, | i ing ψ onto M0,M1 = ψj ψj ,I ψj ψj , and as the minimal degree of such polynomials cor- | i { } {| ih | − | ih |} returning b upon measurement of Mb. respond to algorithmic performance, we desire that the fj subdivide the search space into subsets which 9

have a larger average6 distance between elements in III. CONSTRAINED POLYNOMIAL the natural metric of this space. Equivalently proxi- INTERPOLATION mate elements in the binary tree are also proximate in the embedding space. In the previous section we reduced the solution of Prob- lem II.1, a simplified version of G-QHT, to the existence of interpolating polynomials over real intervals. More- • Each leaf of the binary decision tree must corre- spond to no more than one channel. If each (prob- over we asserted that, despite the restrictive form of the able with respect to measurement) leaf corresponds queried channel W (x) considered in QSP, the guarantees of Theorem II.1 were still strong enough to enable dis- exactly to one channel, then g in II.1 is not only injective but bijective. crimination among channel sets whose structure is richer than rotations about a fixed axis. This section considers one concrete interest of a party playing R-QHT: how can • The fj must have definite parity in the continuous a computationally limited classical party compute Φ for embedding space, here SU(2); this parity constraint, a QSP algorithm such that the resulting matrix polyno- requisite for the use of QSP, follows from properties mials induce measurements obeying the prescriptions of of SU(2). Algorithm1. This is a problem of constrained polynomial interpo- lation. More generally, the field of functional approx- Algorithm1 and its supporting remarks show that, at imation, in which this problem lives, is well-understood least for a special set of channels, our hopes of computing [29–34] given its practical instantiations in classical signal successive equivalence relations on subsets of S to iter- processing and relevance to foundational questions in real atively determine the hidden query element rest on the analysis. We quote the following results in constrained construction of low-degree constrained polynomials over polynomial approximation and present their synthesis as real variables. a new theorem guaranteeing desired properties for the Moreover, as stated in Remark II.2, most of the dif- algorithms that will be constructed in SectionIV for ficulty of this algorithm resides in designing the bi- specific finite groups. Additionally, these results provide nary functions fj. The sequence of equivalence relations quantitative bounds on the query complexity of solutions f1, f2, , fm, which together uniquely define the hidden to the R-QHT problem discussed previously. index i···, need to be properly chosen such that (1) the de- We present a further sharpening of R-QHT (Problem grees of their polynomial interpolations are not too large, II.1); this new problem, P-QHT, is similar to R-QHT but and (2) that the concatenation of their evaluations is in- provides a new quantitative condition on the performance vertible on every i; luckily these conditions are not so of an algorithm solving R-QHT. complicated to achieve in practice. Problem III.1. The polynomial QHT problem, or P- E.g., we can see one such set of fj in observing the QHT problem, answers the following question. Given an ‘QSP-free’ decision algorithm for C2n in Lemma I.1, instance of the R-QHT problem (Problem II.1), which j namely fj(i) = i (mod 2 ) for j [n]. Evidently in this ∈ implicitly defines a set of angles θ` , what is an upper simplest case the family of fj define precisely a binary { } bound on the sum of degrees of the set of polynomials pj search on the hidden channel index (and consequently 7 { } which interpolate binary functions f1, f2, , fm satisfy the equator of the Bloch sphere under the map i i). ··· Remark II.2. This upper bound depends only on θ` . What remains to be shown is the generalization of7→ such E { } a search. Toward analyzing the minimal degree of such interpo- It turns out that Algorithm1 can indeed be extended lating polynomials as desired in Problem III.1, we give a to more interesting channel sets than single-axis rota- series of older results from works in constrained interpo- tions (i.e., that we can lift R-QHT problems to G-QHT lation. problems). However, before investigating the flexibility Theorem III.1. In [30] Let Ξ = x : x < x < < { i 1 2 ··· of Algorithm1 as a subroutine, we first briefly address xn a set from the real interval [a, b] and the set of methods in constrained polynomial interpolation. This polynomials.} For all  > 0 and for each f PC0[a, b], the analysis, in addition to closing the loop on the R-QHT continuous functions on [a, b], there exists∈ p such problem and its query complexity, will demonstrate the that the following conditions are satisfied: ∈ P methods by which the optimal query complexity of R- 1. p is interpolating: p(x ) = f(x ) x . QHT is computed, and provide a foundation for general- i i ∀ i izing to G-QHT.

7 Note that for our purposes it is often not important to distin- guish between {`} the set of indices and {θ`} the set of angles. While the degree of the interpolating polynomial depends on 6 This is purposefully left ambiguous at this moment; we wish to these angles, this dependence can be simplified by promises on lower the required derivative of the interpolating polynomial. separations between neighboring θ`. 10

S0 f (s) = 0 ψ ψ⊥ f (s) = 1 0 | 0i | 0 i 0

0 1 S1 S1 0 0 1 1 ψ ψ ⊥ ψ ψ ⊥ | 0i | 1 i | 0i | 1 i

00 01 10 11 S2 S2 S2 S2 00 00 01 01 10 10 11 11 ψ ψ ⊥ ψ ψ ⊥ ψ ψ ⊥ ψ ψ ⊥ | 0 i | 1 i | 0 i | 1 i | 0 i | 1 i | 0 i | 1 i

000 001 010 011 100 101 110 111 S3 S3 S3 S3 S3 S3 S3 S3

FIG. 6. Binary search as enacted by Algorithm1. Algorithm1 takes binary functions fj on subsets of S0, specifically f

+ 2. The polynomial p uniformly approximates f on Theorem III.3. In [31] Let n Z and let xj = cos θj [a, b], where θ < θ < < θ [0, 2π∈] and the minimum sep- 1 2 ··· n ∈ aration between adjacent θj (on the unit circle) is given max p(x) f(x) . by δ > 0. Given any real function f C([ 1, 1]) there x [a,b] | − | ≤ ∈ − ∈ exists a polynomial p such that the following conditions hold: 3. The polynomial p obeys the additional constraint 1. p is interpolating: p(x ) = f(x ) x . j j ∀ j max p(x) = max f(x) . x [a,b] | | x [a,b] | | 2. The polynomial p is of degree 2m c/δ where c > 0 ∈ ∈ is some absolute constant. ≤ Theorem III.2. In [35] Let ν index an increasing se- 3. The following inequality holds where the infimum is quence of finite dimensional approximation subspaces Nν taken over the space of all polynomials q of degree in C(T ), for T a topological space, whose union N is at most 2m and k is a constant independent of f dense in C(T ). If T is compact Hausdorff then the degree and n: of approximation with Lagrange (function value) interpo-   latory side conditions Eν (f, A) is related to the degree of max f(x) p(x) k inf max f(x) q(x) x [ 1,1]| − | ≤ q x [ 1,1]| − | the unrestricted approximation Eν (f) by the inequality ∈ − ∈P ∈ −

E (f, A) Theorem III.4. Let Ξ = xj j [n] where xj = cos θj ν { } ∈ lim sup 2 f C(T ) N, and where θ1 < θ2 < < θn [0, 2π] such that the ν Eν (f) ≤ ∀ ∈ \ ··· ∈ →∞ minimum separation between adjacent θj (on the unit where the constant 2 cannot be decreased in general, and circle) is given by δ > 0. Then given any real function f C([ 1, 1]) there exists a polynomial p such that the is the best possible in the uniform approximation of (1) ∈ − entire periodic functions by trigonometric polynomials following conditions hold: and (2) entire functions on any closed finite interval by 1. p is interpolating: i.e., p(xj) = f(xj) xj. algebraic polynomials. ∀ 2. The polynomial p is of degree m = (1/δ). Corollary III.2.1. In the context of constrained polyno- O mial interpolation the statement of Theorem III.2 can be 3. The polynomial p satisfies the following inequality made less general as follows: Given a real interval [a, b] max p(x) = max f(x) . and a real polynomial f of degree d which interpolates a x [ 1,1]| | x [ 1,1]| | ∈ − ∈ − function g on [a, b] at d distinct points in [a, b], the min- imal degree of a polynomial which interpolates g at these Proof. The existence of this polynomial is assured by same points and has norm strictly less than g on [a, b] Theorem III.1, the scaling of degree of the unconstrained is bounded above by 2d as d goes to infinity andk k moreover (uniformly approximating) polynomial is given by Theo- this bound cannot be decreased in general. rem III.4, and that the constrained polynomial’s degree 11 does not grow too large with respect to the unconstrained sets of rotations about known, fixed axes for which the polynomial’s is given by Theorem III.2.  polynomials that interpolate each fj can be explicitly de- scribed, computed, and characterized in terms of degree. Finally, we present a lemma which permits us to apply Once this degree is known, the expected query complex- all of the above results in the context, mandated by QSP, ity of these algorithms follows by the results of Section that the constrained interpolating polynomials used have III. Before this, however, we extend the statement of P- definite parity. QHT (Problem III.1), which as stated applied only to Lemma III.1. If there exists a polynomial of degree n sets of rotations about a fixed axis, to sets which obey interpolating a set of points which has (the point set) def- more general structure. inite parity, and the polynomial is of fixed norm, then there exists a polynomial of degree m n which still in- Problem IV.1. The P-QHT problem (Problem III.1) terpolates the points and which has the≤ same parity as the can be extended given the following prescription on a so- points. Proof follows by re-expressing the polynomial as lution form. We begin with the standard statement of G- a sum of terms with definite parity and observing that the QHT: given query access to one quantum channel among component of parity matching those of the interpolation a faithful representation of a finite group G < SU(2) de- points still satisfies the desired properties. termine the optimal query complexity of an adaptive se- rial query model algorithm that determines the hidden The results of this series of theorems, and particularly index of the queried channel with certainty. the assurances of Theorem III.4, permit us to justify the Importantly, however, for P-QHT to provide a solu- idealized claims of the classical program discussed in Al- tion, one must be able to transform the query set in a gorithm1, at least for cyclic groups. I.e., given that special way; this reduction follows from the conditions the quantum channels considered can be (at least for the given below: case that G is cyclic) distinguished by their eigenvalues, the methods of QSP and the assurances of Theorem III.4 1. There must exist a series of protocols, given query together imply that their exist computationally cheap, access to a channel set S, for generating compound flexible quantum algorithms whose measurement results queries8 (see Definition IV.1) whose structure is are themselves deterministic functions on the discrete set (1) precisely a set of rotations by known angles of possible channels. around a fixed axis (i.e., inputs to the R-QHT prob- With respect to a resolution of Problem III.1, this sec- lem satisfying Remark II.1), or (2) a subset of a tion has provided a key observation: the minimal degree finite group G0 for which a decision algorithm is of the interpolating polynomial on a set of angles θ` , as already known. in the R-QHT, problem is linear in both the number{ } of interpolation points and max 1/ θ θ , the minimal `,k | ` − k| 2. In the case of (1) as given above there must exist separation between (distinct) queried angles. a solution for P-QHT (Problem III.1) for the new Once the interpolating polynomials pj are computed, query set. There must also exist some additional the path to generating QSP angles Φj is well under- assumption, specific to the structure of S, that per- stood and computationally efficient (i.e., polynomial in mits the compound query map used to be invertible. the degree of the interpolating polynomial). There are This is accomplished in different ways for different many ways to perform such a computation, both an- groups, e.g., under the assumption that the repre- alytically [24] and by numerically stable computations sented group is a semi-direct product, as in Theo- [26]. Moreover, the interpolating polynomials can be rem IV.2. computed in any number of ways, usually relating to a modified Remez-type algorithm [36, 37]. Definition IV.1. A compound query with respect to a quantum channel : A B is a quantum circuit : A B which usesE a finite→ number of copies of Cas IV. DECISION PROTOCOLS ON FINITE well→ as a finite number of additional unitary operatorsE SUBGROUPS OF SU(2) independent of . Compound queriesE are often used by quantum algo- We now close the loop on our simplification of G-QHT rithms (e.g., Algorithm1) in place of bare queries, i.e., in Problem I.2 to R-QHT in Problem II.1 and finally, simply i. Usually, useful compound query circuits do through Algorithm1 to a problem in polynomial inter- not actE injectively on the query set. polation where the degree of these polynomials relates directly (by the results of Problem III.1) to the query complexity of the solution to R-QHT. In this section we finally address the more general problem of G-QHT for 8 In simple terms one may think of these as small quantum circuits small groups G. which employ a small number of queries to the original oracle, For each finite subgroup G < SU(2), we provide con- and may be used as subroutines replacing oracle calls for a pro- structive proof that there exists a series of binary func- tocol expecting queries of a different form. Multiple physical tions f , f , , f and a series of protocols to access queries can form one compound query. 1 2 ··· m 12

Remark IV.1. The extended statement of the P-QHT Proof. Any positive integer n has a unique decomposition problem (Problem IV.1) exists to answer the following into a product of unique primes as given in (6), where ri question: how far can Algorithm1 be taken beyond its is the multiplicity of the i-th smallest prime dividing n, role as a solution to R-QHT? pi, and s is the largest index for which pi divides n at Consequently each of the algorithms discussed in this least once. section is, in truth, simply (1) a procedure for reduction Assuming the existence of a deterministic algorithm to R-QHT, followed by (2) application of Algorithm1. C that can perfectly decide Cp , elements of the group A pi i When reduction is made to deciding a simpler group, the Cn are decided according to the following protocol: application of Algorithm1 is hidden behind algebraic ab- straction. 1. If the multiplicity ri of pi in n is one, in the place of the query usually made by the protocol C , A pi We go through the finite list of distinct families of fi- query the oracle n/pi times. This compound query nite subgroups of SU(2) in order of increasing complexity, may be conjugated by a known unitary to achieve recovering instances of Problem IV.1 as stated above. As the representation that p expects. A i a road-map we provide the following lemma, which com- pletely characterizes the finite subgroups of SU(2). A 2. If the multiplicity of pi in n (ri) is greater than one, the same method presented in the Lemma I.1 diagram of the path of these reductions was given in Fig- ri ure4. is applied to compound queries of order n/pi to read off successive bits (this time in base pi) of ri,

Lemma IV.1. The finite subgroups of SU(2) are in bi- using the assumed subroutine for deciding Cp . A i jection with the finite subgroups of SO(3) under the stan- Compound queries allow access to prime-power-order dard double covering SU(2) SO(3). These finite sub- cyclic subgroups of C whose decision algorithms are groups are thus completely described→ by five families: (1) n + strictly simpler and reducible to decisions on Cp for p The cyclic groups of order n, Cn, n Z . (2) The di- ∈+ prime.  hedral groups of order 2n, D2n, n Z . (3) The alter- ∈ nating group on four elements, A4. (4) The symmetric We proceed by considering a result concerning the group on four elements, S4. (5) The alternating group smallest non-trivial cyclic group, C3, with which to play on five elements A5. G-QHT. This can be thought of as a base case for our eventual reduction from decision protocols on large cyclic groups to smaller ones. A. Cyclic groups Discriminating between quantum channels represent- ing C3 has some precedent in prior work: such chan- Peres- Before lifting the methods of Lemma I.1 from C2n to nels are precisely those which can generate the general cyclic groups we provide a few lemmas. Wootters states [38, 39] or equivalently Mercedes-Benz frames [40, 41] (for their threefold symmetry). Lemma IV.2. The cyclic group of order n is isomorphic Lemma IV.4. There exists an algorithm that per- to the direct product of cyclic groups AC3 fectly decides C3 (or rotations about a fixed axis on the Bloch sphere by one angle among the three angle set Cn = Cpr1 Cpr2 Cprs , ∼ 1 × 2 × · · · × s 0, 2π/3, 4π/3 ) using at most 6 oracle queries. This al- { } iff the unique prime decomposition of n is gorithm is said to solve the three angle problem.

s Proof. Without loss of generality the group C3 Y ri is represented by the set of quantum channels n = pi , (6) i=1 R0(0),R0(2π/3),R0(4π/3) . Consider the QSP se- quence{ defined by QSP phase} list Φ = 0, α, α, 0 using { − } for distinct primes pi. I.e., Cn is isomorphic to a di- the convention of Theorem II.1, i.e., the product rect product of cyclic groups of prime-power order for all maximal prime powers dividing . This is one statement UΦ = Rx(θ)Rz(α)Rx(θ)Rz( α)Rx(θ), (7) n − of the Chinese remainder theorem. for any angle θ. It is not hard to explicitly compute the

We proceed to analyze decisions on Cn by a series of top left component of this unitary operator, and specif- reductions to decisions on the more restricted (albeit in- ically for the special angle α = arccos(cos θ/[1 cos θ]), which is real whenever π/3 θ 5π/3, the top− left com- finite) family of cyclic groups of prime order. ≤ ≤ ponent of this unitary 0 UΦ 0 is 0. Consequently with Lemma IV.3. If there exists a family of algorithms = three queries to the oracle,h | and| i α = arccos( 1/3), the F − Cp that each perfectly decide Cp for all primes p then transition probability 0 0 is 1 if θ = 0 and 0 if θ {Athere} exists an algorithm that perfectly decides C 2π/3, 4π/3 . Consequently| i 7→ | threei additional queries are∈ ACn n { } for n N, and which is asymptotically optimal in query enough, possibly replacing Rx(θ) with Rx(θ)Rx( 2π/3) complexity∈ if the algorithms in are also optimal. in (7), to completely and perfectly determine the− hidden F 13

angle. Over equal priors the expected query complexity + | i of this technique is 5. Alternatively in the language of Theorem II.1, we in- tend that the top left element of UΦ, under the map a cos θ/2 x, has the form 7→ b 4 f (x) = x (x 1/2)(x + 1/2), 1 3 − which is a polynomial9 that takes modulus 1 at x = 1 − and x = 1, has definite parity, and takes value 0 at x = c 1/2. This, along with f1(x) under the map θ θ 2π/3 produces± a pair of binary measurements for7→ which− the map10 S M is injective where M is the set of binary |−i measurements.7→  (a) Geometric interpretation. 1, 0 θ = 0  { } + UΦ + , + U 0 + = 0, 1 θ = 2π/3 {h | | i h | Φ| i} { } a  0, 0 θ = 4π/3, { }

where UΦ0 is the aforementioned pre-rotation replacing Rx(θ) with Rx(θ)Rx( 2π/3) or equivalently θ θ a 2π/3. A visual depiction− of this algorithm is given7→ in− Figure8, and a table relating this Lemma’s construction a2 a3 directly to Algorithm1 is given in TableI.  a The functional intuition of protocols deciding on repre- 2 3 sentations of cyclic groups is depicted in Figure8. As dis- a a cussed previously, QSP protocols take equiangular rota- tions about different axes in equator of the Bloch sphere (see Figure7), interleave them with rotations about or- a2 a3 thogonal axes on the Bloch sphere, and give efficient methods for forcing the corresponding matrix elements of (b) Algebraic interpretation. the final, composite rotation to be desired trigonometic polynomials in the unknown rotation angle. Figure8 FIG. 7. Geometric (a) and algebraic (b) depictions of the demonstrates that polynomials which have modulus 0 or proof of Lemma IV.4. Unitary representations of C3 in SU(2) 1 at specific angles result in deterministic protocols for are, without loss of generality, equivalent to a set of rotations dividing the search space. The work remaining is to sys- which cycles states (a, b, c) as shown on the Bloch sphere in (a). Moving away from the Bloch sphere, any sequence of tematize sub-protocols of this form to generate efficient quantum channel discrimination protocols whose binary PVM decision protocols on the entire query set. output differs on subsets of quantum channels representing Finally we can provide a proof for perfect decision pro- C3 (e.g., partitions C3 elements into red and blue subsets as tocols on all prime order cyclic groups, and in fact this pictured, and as proven in Lemma IV.4), also determines the shows an even stronger result as the same method goes queried quantum channel perfectly. The partitions indicated through for cyclic groups of any odd order. However, in (b) are generated by polynomials given in Figure8. given the results of Lemma IV.3, QSP is only a neces- sary tool in the prime-order case, when compound queries provide no helpful simplifications. Proof. The proof follows from the existence of a family of polynomials f , f , , f whose moduli take values in Theorem IV.1. There exists a family of determinis- 1 2 ··· m 0, 1 on a finite set of subsets Sj for j [m] of the set tic algorithms = C for all primes p, where C { } { } ∈ F {A p } A p of p possible phases S0 induced by queries to the oracle, perfectly decides Cp, with asymptotically optimal query namely complexity.  πn  S = cos , n [p] , 0 p ∈ 9 Note that (4/3)(x − 1/2)(x + 1/2) also satisfies constraints re- such that that the successive subsets S0 S1 Sm quired by QSP, and indeed this lemma can be shown using only 11 ⊇ ⊇ · · · ⊇ 4 maximum (10/3 expected) oracle queries, though the resulting have the following properties: protocol is less geometrically obvious. 10 S is overloaded here: both channel elements and the continuous real parameter θ characterizing these elements. Note also that this map can be written as a binary tree as in Figure6. 11 Also described in Remark II.2 and Figure6. 14

0 Index Query map pj |ψj i |ψj i 3 j = 1 Ei 7→ Ei (4x − x)/3 |+i |±i 7→ {0, 1} 3 j = 2 Ei 7→ EiRx(−2π/3) (4x − x)/3 |+i |±i 7→ {0, 1}

 {0, 1} 7→ R (0)  x {f1(Ei), f2(Ei)} = {1, 0} 7→ Rx(2π/3) (Inverse map)   {1, 1} 7→ Rx(4π/3)

TABLE I. The use of Algorithm1 as a subroutine for solving the three angle problem (Lemma IV.4) in tabular form. As pj for j ∈ [m] completely define both fj and the corresponding QSP-derived objects given in Definition II.2, they provide a minimal 0 explicit demonstration of the use of Algorithm1. Included are quantum states for preparation, |ψj i, and measurement, |ψj i, as well as the compound query map (Definition IV.1), where Algorithm1 is fed compound queries. Finally, an inverse map is given to recover the hidden channel.

• Bisecting: The order of Sj+1 should satisfy that space, and is subsequently dropped as this iterative di- Sj+1 (1/c) Sj for some fixed constant c = vision continues. The functions fj are defined as poly- | (1).| ≤ | | nomials which interpolate any binary function on the set O Sj 1 which alternates maximally with definite parity on • Density reducing: The minimum separation be- [ −1, 1] (f will share this parity). These functions have − j tween elements of Sj+1 on which the modulus of the explicit description, e.g., when given some subset Sj of interpolating polynomial fj+1 takes distinct values size 2n + 1, indexing by ` for increasing s` in [ 1, 1]. should increase exponentially in j. − ( 1 ` 2 [1 + ( 1) ] 1 ` n • Totally deciding: Constructing a family of inter- fj(x`) = − ≤ ≤ 1 ` 1 polating polynomials pj whose moduli take values 2 [1 + ( 1) − ] n + 1 ` 2n + 1. in the set 0, 1 on S is equivalent to computing − ≤ ≤ { } j a family of binary functions fj on Cp; the evalua- This evidently preserves parity and confers the right tion of these binary functions on the hidden channel properties on successive subsets. In plain terms this is corresponding to g Cn should uniquely identify a binary search whose constituent sub-searches grow ex- g. I.e., this map g ∈ 0, 1 m should be injective ponentially cheaper in query complexity, and whose base 7→ { } (see Figure6). case is handled by Lemma IV.4. 

• Parity preserving: The elements of Sj should be Finally, by the previous results we can make a state- of definite parity for all j; this parity is shared by ment for all cyclic groups, and proceed to richer sub- all pj. groups of SU(2). If all of these conditions are satisfied by some judicious Corollary IV.1.1. For all n N, there exists a deter- ministic algorithm which∈ perfectly decides C , with sequence of Sj the result follows if the number of such ACn n non-trivial subsets of S, given by m, is asymptotically asymptotically optimal query complexity. This follows di- j rectly from Lemma IV.3 and Theorem IV.1. log p and the degree of the polynomial pj goes as (p/c ) in which case the entire protocol has query complexityO linear in p. The existence of these interpolating polynomials is B. Dihedral groups guaranteed by the results of SectionIII, while their asymptotic query complexity follows directly from ex- We consider the dihedral groups of order 2n; it is not ponentially increasing promised gaps between elements too difficult to see that each bit-string label for an ele- of S . For a given group C these subsets S have the ment g D2n requires exactly one more bit to uniquely j p j ∈ explicit, measurement dependent, form describe the element, corresponding to membership of g in one of two cosets of the normal cyclic subgroup Cn/Dn. S0 = S0 We show that this bit can be recovered in one additional measurement, and that our protocol is thus optimal as- 0 Sj = sk Sj 1, fj 1(sk) = 0 suming the optimality of the protocol which decides Cn. { ∈ − − } 1 Sj = sk Sj 1, fj 1(sk) = 1 Theorem IV.2. Assuming existence of an algorithm { ∈ − − } Cn that perfectly decides Cn there exists an algorithm where the new S ’s upper index indicates the measure- A that calls as a sub-routine and perfectly de- j AD2n ACn ment result of the QSP sequence dividing the search cides D2n, the dihedral group of order 2n, with one addi- 15

Amplitude + UΦ + for the Three Angle Problem (I) The lemma follows if there exists a simple protocol to, h | | i 1 given query access to an unknown element g D2n, de- termine membership of the queried element g ∈among the two cosets of Cn < D2n. Assume U(g) is the unitary operation corresponding to the group element g embedded in SU(2) as stated. Then 0 HU( σ )H 0 = 0 (10) h i | i | i HU(τ)U( σ )H 0 = 1 , (11) h i | i | i Transition amplitude where H is the Hadamard gate and U( σ ) represents some unitary operation within the subgrouph i σ gener- 1 ated by σ. Intuitively, H rotates 0 to anotherh i state − π 2π 4π 5π | i 0 3 3 π 3 3 2π insensitive to the action of the cyclic index 2 subgroup of Rotation angle, θ D2n. This follows from the lack of irreducible represen- tations of Cn in SU(2). (a) Phase I If 0 is measured then Cn can be applied as normal to future| i queries, respectingA the embedding of the subgroup Amplitude + U + for the Three Angle Problem (II) h | Φ| i σ . Otherwise any query made to the oracle U(g) is h i 1 prefaced by Rx(π), reducing to a decision on Cn. Only one additional query is needed by to decide D (a AD2n 2n group with twice the size). 

ba 0 a ba2 ba7 a2 a7 a Transition amplitude a2 a7

a3 a6 a3 a6 3 6 1 ba 4 5 ba − π 2π 4π 5π a a 0 3 3 π 3 3 2π Rotation angle, θ a4 a5 (b) Phase II ba4 ba5

FIG. 8. Quantum response function employed in the proof of Lemma IV.4 (a), and its shifted version (b). On the left is FIG. 9. Two presentations of Cayley graphs for the cyclic the polynomial, in cos (θ/2), which is generated as the top left group of order 7 and the dihedral group of order 14. The component of the single-qubit unitary UΦ corresponding to observation that the cyclic group admits no irreducible repre- the angles Φ indicated in the first QSP subroutine of Lemma sentations in SU(2) allows the perfect determination, in one IV.4. On the right is the same protocol using a pre-rotation additional query, of coset membership for the maximal cyclic by 2π/3, permitting a unique binary labeling of each channel subgroup of D2n, partitioning the red and blue sets. after two measurements. tional oracle query. A depiction of the overarching idea of this algorithm is given in Figure9. C. Platonic groups

Proof. Without loss of generality D2n has oracle access Finally we address protocols for the finite subgroups of A to a channel in a representation of D2n whose cyclic sub- SU(2) that do not fit into countably infinite families, and group Cn in SU(2) has representation: exhibit a richer non-abelian structure than the dihedral R (m 2π/n) , m [n]. (8) group. These are often referred to as the platonic groups { z · } ∈ due to their appearance in the study of symmetry groups The SU(2) embedding of D2n that contains our embed- of platonic solids. Before discussing protocols for decid- ding of Cn as a subgroup is generated by a generator of ing A4, S4 and A5 we define two basic group theoretic this Cn, σ, and another group element, τ, which without concepts that will aid in their construction. loss of generality has representation R (π). The standard x Definition IV.2. (Cycle decomposition). Let be a presentation of D is S 2n finite set, e.g., the integers 1, 2, , n , and σ a permu- D σ, τ σn = τ 2 = τστσ = e , (9) tation S S. The cycle decomposition{ ··· } of σ expresses 2n ≡ { | } → 16

0 Index Query map pj |ψj i |ψj i

j = 1 Ei 7→ Rξ(−π/2)EiRξ(π/2) x |+i |±i 7→ {0, 1} ......

( ? {0, · · · } 7→ e ∗ Cn {f1(Ei), f2(Ei)} = ? (Inverse map) {1, · · · } 7→ τ ∗ Cn

TABLE II. The use of Algorithm1 as a subroutine for deciding on dihedral groups D2n (Theorem IV.2) in tabular form. The table proceeds until reduction to Cn is achieved (i.e., after the first query); this query rotates to the basis in which σ (the generator of Cn /D2n) acts trivially on {|±i}. Once coset membership in the maximal cyclic subgroup of the queried element is known, it can be inverted and applied to form compound queries that reduce the query set to Cn, given in Corollary IV.1.1. ? Here Cn is an unknown power of σ.

σ as a product of disjoint cycles. For instance, if S has Given that A4 D4 oC3, all elements g can be written size 4 and the action of σ swaps pairs 1, 2 and 3, 4, then in the form kh where≡ h is drawn from a chosen normal the cycle decomposition of σ is denoted (1, 2)(3, 4), where D4 subgroup of the representation of A4 and k is from the order of tuples is not uniquely defined. a realized C3 subgroup. By pre-applying powers of a generator of one of these C3 subgroups, the D4 algorithm Definition IV.3. (Cycle type). Let S be a finite set, for on cubes of queries will consistently compute the binary instance the integers 1, 2, , n , and σ a permutation function of membership of the queried element g in a S S. The cycle type{ 12 ···of σ is} a tuple indicating the → particular coset of the normal D4 /A4. Assuming equal number of cycles of each given length in the cycle decom- priors, such an algorithm is expected to13 terminate in of . E.g., for the example given in Definition position σ 14.5 queries.  IV.2, the cycle decomposition (1, 2)(3, 4) has cycle type (0, 2, 0, 0), indicating two length-two cycles. Definition IV.4. We give a name to a particular sub- Note that the set of all possible cycle types is in routine presented in Theorem IV.3, whose use is indi- bijection with unordered partitions of the integers in cated in Table III. 1, 2, , n : i.e., for cycle type tuple c, the sum of cj j The function correctCoset takes as input the evalua- for{ j ···[n] is} simply n. · tion of the three binary measurements given in Theorem ∈ IV.3 to determine which element of the D4 normal sub- Theorem IV.3. There exists a deterministic algorithm group of A4 enters into the chosen semi-direct product

A4 that perfectly decides A4 with asymptotically optimal D C and returns the representation of the inverse of A 4 o 3 query complexity. This algorithm is additionally given in this element. Table III. Theorem IV.4. There exists a deterministic algorithm Proof. The elements of A can be classified according to 4 which perfectly decides S , with asymptotically opti- their cycle type as permutations on four elements. For S4 4 malA query complexity. A4 these types are (1, 0, 1, 0), (0, 2, 0, 0) and (4, 0, 0, 0) (the last being the identity permutation). Proof. Squares of elements in S4 necessarily fall in the Cubes of any element in A4 have cycle type (0, 2, 0, 0) alternating group A4, though this mapping is not al- or (4, 0, 0, 0) only, meaning that if the queried element ways invertible. It is invertible, however, when the g is already in one of three representations for the D4 queried element g in S4 has the cycle type (1, 0, 1, 0). normal subgroup of A4 then running the D4 algorithm For any element in S4 there exists an element h of cy- on cubes of physical query elements gives the correct an- cle type (2, 1, 0, 0) for which the product gh is of cy- swer, and otherwise acts as if the queried element were cle type (1, 0, 1, 0). Consequently there exists an algo- the identity. This element can be determined in at most rithm that, for every element h of cycle type (2, 1, 0, 0), three compound queries deterministically, measuring in of which there are six, pre-applies h to queries (and re- three mutually unbiased bases on the Bloch sphere, cor- peats this process to generate squares of this query el- 2 responding to the eigenstates of each of the generators of ement, ghgh = (gh) ) and runs the A4 algorithm on the chosen D4 subgroup. this compound query, which recovers perfectly in finitely

12 The cycle type is sometimes defined as a tuple of the lengths of 13 The explicit calculation is (1/4)·6+(1/4)·12+(1/4)·18+(1/4)· each cycle in the cycle decomposition, rather than the number 22 = 29/2 for the 3 + 3 + 3 non-trivial elements of the cosets of of cycles of each given length. the normal D4 followed by 3 trivial elements. 17

0 Index Query map pj |ψj i |ψj i 3 j = 1 Ei 7→ (Ei) x |+i |±i 7→ {0, 1} 3 j = 2 Ei 7→ Rx(π/2)(Ei) Rx(−π/2) x |+i |±i 7→ {0, 1} 3 j = 3 Ei 7→ Rx0 (π/2)(Ei) Rx0 (−π/2) x |+i |±i 7→ {0, 1} 3 j = 4 Ei 7→ EicorrectCoset(f<4(Ei)) (4x − x)/3 |+i |±i 7→ {0, 1} 3 j = 5 Ei 7→ Rx00 (2π/3)EicorrectCoset(f<4(Ei)) (4x − x)/3 |+i |±i 7→ {0, 1}

 e ?  {0, 0, 0, · · · } 7→ D4 ∗ C3  a ?  {1, 1, 0, · · · } 7→ D4 ∗ C3 {f1(Ei), f2(Ei), f3(Ei), ···} = b ? (Inverse map)  {1, 0, 1, · · · } 7→ D4 ∗ C3  ab ?  {0, 1, 1, · · · } 7→ D4 ∗ C3

TABLE III. The use of Algorithm1 as a subroutine for deciding on A4 as in Theorem IV.3 in tabular form. As pj for j ∈ [m] completely define both fj and the corresponding QSP-derived objects given in Definition II.2, they provide a minimal explicit demonstration of the use of Algorithm1. The pj given here have also had their QSP angles explicitly given in Lemma IV.4. g ? Here D4 ∗ C3 indicates a group element in the semi-direct product defining A4 which is the product of g, an element of the 0 chosen D4 normal subgroup in terms of generators {a, b} and an unknown element of the chosen C3 subgroup. Axes x, x are 00 chosen such that these rotations generate the chosen D4 subgroup, and x the axis of rotation for the chosen C3.

many queries the hidden element g when the image (gh)2 groups follows, in each instance described in this section, has cycle type (1, 0, 1, 0). Namely there exists a sub- from the following sketched protocol. routine which determines coset membership for cosets of For each finite group presented in SectionIV, we must the normal A4 /S4, and proceeds by reduction to deci- provide (1) a small quantum circuit to produce compound sion on A4. This protocol is expected to terminate in 34 queries (Definition IV.1) satisfying the input assump- 14 queries.  tions of Algorithm1, (2) apply Algorithm1 and keep track of its minimal required query complexity, and fi- We note that the two protocols given above make no nally (3) verify the satisfaction of conditions under which reference to the mechanisms of QSP, but are instead com- the compound query map is invertible (e.g., as in Remark pletely algebraic in form, exploiting the simple canonical II.2), these conditions remaining unchanged despite the subgroup towers of A4 and S4 to reduce decisions on rep- introduction of compound queries. resentations of these groups to those on their normal sub- Whether this protocol is possible to perform for gen- groups. It is the small size of these non-abelian groups eral groups is an open question, and indeed the methods in particular which, unfortunately, bring the following of this section relied on the fact that the finite groups in- remark. Resolving this problem is left open as stated in vestigated were non-simple and often semi-direct products SectionVI. of abelian groups. Remark IV.2. The alternating group on five elements has, unfortunately, no simple reduction to an algorithm of the previous, smaller groups, in part because A5 is the V. EXTENDING QHT PROTOCOLS TO smallest simple non-abelian group, and thus permits no LARGER GROUPS AND NOISY SETTINGS non-trivial decompositions in terms of a canonical tower of subgroups. It is natural to consider generalizations to the setting in Before concluding this section we give an overview (Re- which the results of SectionIV were derived. This section mark IV.3) of the major technique which has permitted concerns itself with two: (1) the inclusion of noise and the extension of algorithms solving R-QHT (i.e., Algo- (2) extension to larger finite groups. rithm1) to those solving G-QHT.

Remark IV.3. Extending the recursive bisection de- A. Noisy channels and noisy quantum gates picted in Figure6, which in turn demonstrates the meth- ods of Algorithm1, to representations of non-cyclic The algorithms of SectionIV relied on the fact that compound queries to the oracle could, under the assump- tion of access to unitary channels, make perfect use of the 14 Again this number is arrived at by explicitly writing a table of algebraic relations which were a priori known among the elements of S4 and running them through the protocol as given query set. These relations led to effective query access until it terminates. to simpler query sets for whom the optimal hypothesis 18 testing algorithm was known. Naturally, however, real- be re¨expressedand bounded above according to istic quantum computers and quantum channels exhibit noise, and one might be concerned about two different U U (ρ ) U 0 U 0 (ρ ) k 2 ◦ Em ◦ 1 ◦ Em j − 2 ◦ Em ◦ 1 ◦ Em j k sources of error as summarized below. U (ρ ) 0 U 0 (ρ ) ≤ kEm ◦ 1 ◦ Em j − Em ◦ 1 ◦ Em j k U (ρ ) U 0 (ρ ) ≤ kEm ◦ 1 ◦ Em j − Em ◦ 1 ◦ Em j k 1. The queried elements Ug may not perfectly satisfy + 0 U1 m(ρj) 0 U1 0 (ρj) the conditions imposed on a faithful representation kEm ◦ ◦ E − Em ◦ ◦ Em k (ρ ) 0 (ρ ) of G, but may instead only approximately satisfy ≤ kEm j − Em j k them, i.e., + m[U1 0(ρj)] m0 [U1 m0 (ρj)] kE ◦ E − E ◦ E k 2 m m0 , ≤ kE − E k UgUh  Ugh g, h G, ≈ ∀ ∈ where the inequalities, in order from top to bottom, fol- where the approximate equality is with respect to low from (1) the monotonicity of the trace distance, (2) some reasonable norm, here the diamond norm. the triangle inequality, (3) monotonicity with respect to Alternatively one can consider that the channels the CPTP map m0 U1, and (4) that the diamond dis- tance dominatesE the◦ trace distance on any particular ini- themselves are only near-unitary, i.e., that Ug0  ≈ tial state. This result can be iterated for arbitrarily many Ug for all g G where the norm is again reasonable. Such a channel∈ might be given by its operator-sum channel applications where the coefficient on the diamond representation distance goes as nj where nj is the discrimination algo- rithm’s j-th subpart’s query complexity.  Z Ug fg(h)Uh dµ(h), For the second instance, where the querent’s own uni- ≡ h tary operations are only close to the ideal operations, an analogous argument to that used in [42] permits us where f (h) is some probability density function de- g to bound error to a multiple of the per-gate error  fined suitably on elements of SU(2) which is peaked (usually computed in terms of a trace distance between about g to induce an operator whose diamond norm the intended and applied channel) where this multiple with U is suitably small. Here µ is some suitable g is proportional to the QSP sequence’s length. Conse- measure over SU(2). quently under reasonable assumptions of noise in both the queried channel and the locally applied unitary op- 2. The unitary operators applied by the querent may, erators, the methods presented in the previous section in general, also not be perfect. This is the state- do no worse than expected, accruing error linearly in se- ment that the rotations normally applied in a QSP quence length for reasonable norms. sequence as per the statement of Algorithm1 may again only satisfy U 0 U for all indices j in the j ≈ j QSP sequence. We denote by Uj0 the applied uni- B. Extensions to larger groups tary and by Uj the intended unitary. The methods of SectionIV use compound queries (e.g., We consider the first instance, namely the physically positive integer powers of queries), defined in Problem realistic scenario that the ideal query set S is not the IV.1, to access representations of subgroups of G. It is sampled query set, but instead that physical queries may thus of interest to determine when one is to expect that be slightly perturbed from ideal queries. I.e., the phys- (1) subsets of m-th powers of group elements generate ical queries i0 are such that the diamond distance proper subgroups, and (2) what information can be ex- {E } i i0  for some small  > 0. In this case, which tracted under the assumption of the ability to decide on kE − E k ≤ encompasses all small perturbations, methods analogous said subgroups. We state a series of related lemmas re- to the ‘peeling lemma’ in [2], permit us to bound our new garding these questions, assuming a basic understanding error in discrimination. of group theory. The following two lemmas in particular discuss suffi- Lemma V.1. Fixing an initial state ρj the trace distance cient conditions under which a known normal subgroup ρ ρ0 between the serial quantum channel discrimina- of G permits query access to compound queries that re- k − k tion protocol defined by the interspersed unitaries Ui,j side in said normal subgroup. These lemmas capture the { } acting on ρj where the queried channel set is i versus underlying mechanism of the protocols given previously {E } i0 is bounded above by nj m m0 nj. for deciding D2n and A4. {E } kE − E k ≤ Lemma V.2. If a finite group G admits a normal sub- Proof. In the case that the QSP sequences used are length group N of index m then the subset of m-th powers of G, 2, we show the result, and show that the method gener- equivalently Sm = gm, gm, , gm for all n elements { 1 2 ··· n } alizes to length n sequences. The distance ρ ρ0 can of G generates a proper subgroup G0 N < G. j k − k ≤ 19

Proof. Proof follows from recognizing that elements of on groups which are semi-direct products grow easier. m the form gi are in the kernel of the group homomorphism Thus, while extension to larger Hilbert spaces may not re- G G/N and thus Sm is a (possibly non-proper) solve the discussion of efficient decision algorithms on all subgroup→ of the normalh subgroupi N of G, equivalently larger groups in the serial adaptive query model, it may m S N < G.  reasonably result in interesting quantitative statements h i ≤ on the entanglement or auxiliary system size necessary Lemma V.3. If a finite group G admits a normal sub- to achieve efficient (query-complexity-wise) discrimina- group N of index m then the subset of m-th powers of tion dependent on the nature of the represented group. G, i.e., the group generated by Sm as in Lemma V.2, is contained within the intersection of all index m normal subgroups of G. Proof again follows by the isomorphism VI. DISCUSSION AND CONCLUSIONS theorems.

Furthermore we give a lemma that describes the un- In this work we have provided a constructive approach derlying behavior of the protocol given previously for de- for achieving efficient quantum multiple hypothesis test- ciding on S4 (Theorem IV.3). It is the statement of this ing for query sets whose algebra faithfully represents a lemma, as well as the two preceding it, that precludes a finite subgroup of SU(2). The nature of this construc- solution for deciding on A5, which admits no non-trivial tion centers on the use of Algorithm1, a quantum algo- normal subgroups. rithm for solving the simpler R-QHT problem (Problem II.1), as a subroutine along with methods for exploit- Lemma V.4. If the m-power map g gm applied to 7→ ing known algebraic structure of the query set to enable elements of G generates a proper subgroup G0 < G, and reductions to R-QHT. This reduction is summarized in there exists a group element h G such that for some Remark IV.3. subset of the map ∈ m is invertible for all S G s (sh) Concretely, when the represented group G is either , and there exists a quantum7→ protocol for deciding s S abelian or both non-abelian and non-simple the proto- ∈, then there exists a quantum protocol for deciding the G0 cols we construct achieve optimal query complexity with- query set G0 S. ∪ out the use of auxiliary systems or entanglement; this Proof. Constructing compound queries of m-th powers statement is equivalent to a statement about the minimal of physical queries allows access (at m times the query degree of constrained interpolating polynomials, and re- solves an open question in [5], as well as generalizes an old complexity) to a representation of G0. The statement of the lemma with respect to the set S says merely that result in [12] to quantum channels. Moreover, the bridge pre-application of h G before each query s is, under that Algorithm1 and its derivate algorithms demonstrate the map given, invertible,∈ and thus unique identification between quantum information and functional approxima- tion theory indicates a rich variety of novel instantiations of s is also possible with knowledge of h.  of the basic ideas of QSP [22, 24]. The statements given in the lemmas above do not de- In addition to achieving efficient quantum channel dis- pend on particularly complicated notions in group the- crimination for a family of channel sets in a serial adap- ory; instead, we have simply asked which simple opera- tive query model, we show that our protocols can be tions can be performed in our limited resource model to aborted early while still accomplishing useful tasks; this faithfully simplify the query set. In most instances, these follows simply from the nature of the binary search dis- simplifications correspond to the existence of normal sub- cussed in Remark II.2. For instance, the reductions pro- groups (equivalently kernels of group homomorphisms). vided throughout SectionIV are directly realizable as For statements beyond those given here, especially those coset membership testing procedures in general, or pe- concerning the conditions under which the assumptions riod finding for the case of cyclic groups. of Lemma V.4 hold beyond S4, we define a selection of In the following remarks and problem definitions, we open problems in SectionVI. provide one more direct application of the methods dis- The procedure outlined in Lemma V.4 is also not the cussed in this work to a problem in quantum communi- most general one; indeed, compound queries can be built cation. from general products of known unitary operations (some of which may coincide with the query set) and possibly Remark VI.1. As mentioned in [16, 20], efficient pro- multiple copies the queried channel. Conditions under tocols for the estimation of unitary processes have use in which such a map is invertible relate intimately to the the transmission of reference frames as well as various study of characters in representation theory, and provide proofs of insecurity for device-independent protocols for exciting avenues for improved protocols for larger finite quantum bit-commitment. groups, e.g., G < SU(n). Moreover, when considering We give one example for how this work can be ap- larger Hilbert spaces, in analogy to the algorithms decid- plied to a discretized version (e.g., group frames [17, 18], ing on the dihedral groups D2n, the family of finite groups which share close relation with SIC-POVMs) of reference which permit no irreducible representation in said larger frame-sharing (Problem VI.1 and Lemma VI.1). Hilbert space grows richer, and correspondingly decisions 20

Problem VI.1. Consider two separated parties, Alice sequence defined by Φ0 acts identically given access to and Bob; each is able to (1) perform single-qubit unitaries parties of relative angular displacement θ as the sequence and (2) transmit qubits noiselessly to the other party. Al- defined by Φ acts given query access to an equiangular ice and Bob agree on a shared z-axis but are rotated with rotation exp ( i[θ/2]σ ) in the setting of standard QSP. − z respect to each other by some unknown angle θ about this Consequently a protocol solving Problem VI.1 follows axis. Moreover, the possible θ lie within a discrete set Θ directly from a protocol solving Problem II.1 under the of size , known to both parties. n prescription (following Algorithm1) defined by Φ j,k0 = Alice and Bob can come to agreement on the unknown Φj,k + π for k 1, , nj and Φj,0 0 = Φj,0.  angle θ with certainty in a finite length interactive pro- ∈ { ··· } tocol; this protocol is denoted dual QSP due to its simi- larities with standard QSP [21–24], and is said to solve Remark VI.2. We can analyze the performance of the the dual QSP problem. protocol given in Lemma VI.1 in two ways: (1) in com- parison na¨ıverepetition of binary hypothesis testing and Lemma VI.1. There exists a finite length interactive in- (2) in comparison to phase estimation, the continuous teractive protocol by which two parties playing the game analogue of the problem statement. defined in Problem VI.1 can win with certainty and with asymptotically optimal round number (under the restric- • The results of [5] assert that the query complex- tion of sending single qubits). ity for distinguishing two distinct unitary opera- tors U, V scales as (1/Θ[U †V ]) where Θ[W ] is the Proof. Proof proceeds by direct construction. Beginning length of the smallestO arc containing all the eigen- with some initial state ψ , Alice applies to it a rotation | 0i values of W on the unit circle in the complex plane. about her local x axis, namely exp (iφ0σx), and sends this qubit to Bob. Bob applies a rotation about his local When phrased as a decision on a representation16 x-axis by another specified angle φ1, or equivalently ac- of Cn, eliminating one possible quantum channel cording to Alice (if she knew the angle θ) Bob appears to at a time gives a query complexity that scales as 1 2 apply exp (iφ [cos θσ + sin θσ ]) = U exp (iφ σ )U − (n ) (as (n) such discrimination procedures are 1 x y B 1 x B O O where UB = exp ( i[θ/2]σz). required, each costing (n) queries). As shown in In other words,− Alice and Bob can, according to some the constructions leadingO to Corollary IV.1.1, how- previously agreed upon prescription of real angles Φ = ever, decisions on Cn and consequently also dis- φ0, φ1, , φm , collaboratively compute the unitary op- crete reference-frame sharing, have query complex- erator{ 15··· } ity scaling as (n) (up to logarithmic factors) cour- O iφ σ i[θ/2]σ iφ σ i[θ/2]σ iφ σ tesy of the implicit binary search in Algorithm1. U = e m x e− z e 3 x e z e 2 x Φ ··· i[θ/2]σz iφ1σx i[θ/2]σz iφ0σx e− e e e . (12) • A feature of Lemma VI.1 is that it yields a de- terministic quantum algorithm. If one only wishes Moreover, following the final application of exp (iφmσx) to determine the relative rotation with high confi- and measurement against ψ1 , Alice can sample from the Bernoulli distribution defined| i by the transition probabil- dence, one can use phase estimation and achieve 2 the same (n) query complexity scaling [43] using ity p = ψ1 UΦ ψ0 . |h | | i| (log n +O log (1/)) qubits for confidence . This The construction above is almost a vanilla QSP se- O quence. It is not so difficult to see that if Alice also matches the performance of the estimation pro- cedure in [16]. Thus while estimative methods per- and Bob additionally apply the rotation exp i[π/2]σx {± } form similarly in the cyclic group case to G-QHT- respectively, locally, after their φj rotation for j 1, 2, , m , the collaborative sequence instead be-∈ derived methods, the methodology of Lemma VI.1 is comes{ ··· } tailored to the statement of discrete reference-frame sharing, can be done serially, and can be extended iφ σ i[θ/2]σ iφ σ i[θ/2]σ iφ σ U 0 = e m x e z e 3 x e z e 2 x Φ ··· to richer finite groups. ei[θ/2]σz eiφ1σx ei[θ/2]σz eiφ0σx , (13) The methods of Lemma VI.1 suggest a useful tech- which is of the form of a standard QSP sequence. Con- nique; namely, whenever a suitable sensing problem can sequently we see concrete connection between dual QSP be (1) discretized and (2) made coherent, the ability to, and standard QSP: i.e., a redefinition of QSP phase an- by a simple quantum process, induce a phase on, e.g., a gles. single qubit, allows all of the mechanisms built in earlier Given a standard QSP strategy, defined by an angular sections to be directly applied with concomitant state- sequence Φ, there exists an angular sequence Φ0 follow- ing the prescription given above such that the dual QSP

16 This merely connects n in a reasonable, i.e., reciprocal, functional 15 Here assuming that m is even, i.e., that the protocol ends with map to a factor defining the difficulty of discrimination, in which Alice receiving the qubit. the stated quadratic improvement is always possible. 21

ments about query complexity or round complexity17 op- tial discrimination protocols, e.g., beyond those timality. provided for deciding coset membership, which de- Beyond direct applicability to discrete versions of prob- cide other interesting properties of the group rep- lems defined in prior work (e.g., reference frame sharing), resented by the query set while not totally deciding several fundamental open problems remain whose solu- on the group? tion might lie in methods related to those discussed in this work; we outline a few of them below. • Estimating compact group elements: Can the performance of quantum channel estimation pro- • Decisions on the subgroup tower: In anal- tocols for compact groups G, e.g., as in [16], be ogy to the protocol given for deciding the dihe- suitably recovered by employing a method similar dral group in SubsectionIVB, are there families to those of this work to systematically divide the search space up to within a specified error? Un- of larger groups G0 whose lack of irreducible repre- sentation in the natural Hilbert space of multiple der what assumptions about the compact group is 2 n d this decision-to-estimation conversion in the serial qubits ([C ]⊗ ) or qudits (C ) permits groups G adaptive query model still efficient? whose canonical subgroup tower includes G0 to be decided by reduction to decisions on G ? What are 0 To summarize, major avenues for extending this work sufficient conditions under which protocols deciding lie in (1) natural generalizations to higher dimensional G can, even inefficiently, be reduced to protocols Hilbert spaces and (2) characterizations of richer finite for deciding normal subgroups of G? Small exam- groups which find natural representations in higher di- ples of this phenomenon are given in the lemmas of mensional Hilbert spaces. Improvements in methods to SubsectionVB. address these questions have implications in quantum al- • Optimal G-QHT with bounded entangle- gorithms for problems in discrete algebra, and this sub- ment: Given the procedure in the above part, field in turn has potential application, following transla- does there exist a quantifiable trade-off between tion of G-QHT-like problems to novel contexts (e.g., as the serial and parallel query model query complex- in Lemma VI.1), to useful quantum algorithms for cryp- ities required for deciding groups G given access tography, communication, and sensing. to Hilbert spaces in which no representation of G is irreducible? If entanglement is required for opti- mal QHT algorithms on large or highly non-abelian VII. ACKNOWLEDGMENTS query sets, are there methods to quantify the re- quired minimum entanglement? This work was supported in part by the NSF Center for Ultracold Atoms (CUA), the NSF EPiQC Expedition • Quantum property testing: Do there exist par- in Computing, and NTT Research.

[1] C. W. Helstrom, Quantum Detection and Estimation [6] Q. Zhuang and S. Pirandola, Ultimate limits for multiple Theory, Mathematics in Science and Engineering: A Se- quantum channel discrimination, Phys. Rev. Lett. 125, ries of Monographs and Textbooks (Academic Press, Inc., 080505 (2020). 1976). [7] T. Hashimoto, A. Hayashi, M. Hayashi, and M. Horibe, [2] S. Pirandola, R. Laurenza, C. Lupo, and J. L. Pereira, Unitary-process discrimination with error margin, Phys. Fundamental limits to quantum channel discrimination, Rev. A 81, 062327 (2010). npj Quantum Inf. 5, 50 (2019). [8] R. Takagi, B. Regula, K. Bu, Z. Liu, and G. Adesso, Op- [3] A. Ac´ın, Statistical distinguishability between unitary erational advantage of quantum resources in subchannel operations, Phys. Rev. Lett. 87, 177901 (2001). discrimination, Phys. Rev. Lett. 122, 140402 (2019). [4] R. Duan, Y. Feng, and M. Ying, Perfect distinguishabil- [9] R. Takagi and B. Regula, General resource theories in ity of quantum operations, Phys. Rev. Lett. 103, 210501 quantum mechanics and beyond: Operational character- (2009). ization via discrimination tasks, Phys. Rev. X 9, 031053 [5] R. Duan, Y. Feng, and M. Ying, Entanglement is not (2019). necessary for perfect discrimination between unitary op- [10] A. W. Harrow, A. Hassidim, D. W. Leung, and J. Wa- erations, Phys. Rev. Lett. 98, 100503 (2007). trous, Adaptive versus nonadaptive strategies for quan- tum channel discrimination, Phys. Rev. A 81, 032339 (2010). [11] M. F. Sacchi, Optimal discrimination of quantum opera- tions, Phys. Rev. A 71, 062340 (2005). 17 In the methods given, query complexity and round complexity [12] E. Davies, Information and quantum measurement, IEEE are precisely the same (under the map from dual QSP to standard Trans. Inf. Theory 24, 596 (1978). QSP): transmission of the shared qubit is necessary to enact a [13] G. Kuperberg, A subexponential-time quantum algo- unitary operation dependent on the relative rotation. rithm for the dihedral hidden subgroup problem, SIAM 22

J. Comput. 35, 170 (2005). [29] P. Erdos, On some convergence properties of the inter- [14] O. Regev, Quantum computation and lattice problems, polation polynomials, Ann. Math 44, 330 (1943). SIAM J. Comput. 33, 738– (2004). [30] W. Wolibner, Sur un polynome d’interpolation, Collo- [15] A. M. Childs and W. van Dam, Quantum algorithms for quium Mathematicae 2, 136 (1951). algebraic problems, Rev. Mod. Phys. 82, 1 (2010). [31] H. N. Mhaskar, F. J. Narcowich, N. Sivakumar, and J. D. [16] G. Chiribella, G. M. D’Ariano, and M. F. Sacchi, Optimal Ward, Approximation with interpolatory constraints, estimation of group transformations using entanglement, Proc. Am. Math. Soc. 130, 1355 (2001). Phys. Rev. A 72, 042338 (2005). [32] H. W. McLaughlin and P. M. Zaretzki, Simultaneous ap- [17] S. Waldron, Group frames, in Finite Frames, Theory and proximation and interpolation with norm preservation, Applications, edited by P. G. Casazza and G. Kutyniok J. Approx. Theory 4, 54 (1971). (Birkh¨auser,New York, 2013) Chap. 5, pp. 171–192. [33] F. Deutsch and P. D. Morris, On simultaneous approx- [18] J. Kovaˇcevi´cand A. Chebira, An introduction to frames, imation and interpolation which preserves the norm,J. Found. Trends Signal Process. 2, 1 (2008). Approx. Theory 2, 355 (1969). [19] G. Chiribella, G. M. D’Ariano, and P. Perinotti, Memory [34] H. Yamabe, On an extension of the helly’s theorem, Os- effects in quantum channel discrimination, Phys. Rev. aka J. Math 2, 15 (1950). Lett. 101, 180501 (2008). [35] R. K. Beatson, Degree of Approximation Theorems for [20] G. M. D’Ariano, D. Kretschmann, D. Schlingemann, and Approximation with Side Conditions, Ph.D. thesis, Uni- R. F. Werner, Reexamination of quantum bit commit- versity of Canterbury (1977). ment: The possible and the impossible, Phys. Rev. A [36] E. Remez, Sur le calcul effectif des polynomes dapprox- 76, 032328 (2007). imation de tchebichef, C. R. Acad. Sci. Paris 199, 337 [21] G. H. Low, T. J. Yoder, and I. L. Chuang, Methodology (1934). of resonant equiangular composite quantum gates, Phys. [37] F. Grenez, Design of linear or minimum-phase fir filters Rev. X 6, 041067 (2016). by constrained chebyshev approximation, Signal Process. [22] G. H. Low and I. L. Chuang, Optimal hamiltonian sim- 5, 325 (1983). ulation by quantum signal processing, Phys. Rev. Lett. [38] P. W. Shor, The adaptive classical capacity of a quan- 118, 010501 (2017). tum channel, or information capacities of three symmet- [23] G. H. Low and I. L. Chuang, Hamiltonian simulation by ric pure states in three dimensions, IBM Journal of Re- qubitization, Quantum 3, 163 (2019). search and Development 48, 115 (2004). [24] A. Gily´en,Y. Su, G. H. Low, and N. Wiebe, Quantum [39] A. Peres and W. K. Wootters, Optimal detection of quan- singular value transformation and beyond: exponential tum information, Phys. Rev. Lett. 66, 1119 (1991). improvements for quantum matrix arithmetics, in Pro- [40] B. Parvathalu and P. S. Johnson, Construction of mer- n ceedings of the 51st Annual ACM SIGACT Symposium cedes–benz frame in R , Int. J. Appl. Comput. Math 3, on Theory of Computing, STOC 2019 (Association for 511 (2017). Computing Machinery, 2019) p. 193–204. [41] S. A. Mohammad-Abadi and M. Najafi, Type of equian- n [25] A. W. Harrow, A. Hassidim, and S. Lloyd, Quantum al- gular tight frames with n + 1 vectors in R , Int. J. Appl. gorithm for linear systems of equations, Phys. Rev. Lett. Math. Res. 1, 391 (2012). 103, 150502 (2009). [42] A. Y. Kitaev, A. H. Shen, and M. N. Vyalyi, Classical [26] Y. Dong, X. Meng, B. Whaley, and L. Lin, Efficient phase and Quantum Computation, Graduate Studies in Mathe- factor evaluation in quantum signal processing (2020), matics, Vol. 47 (American Mathematical Society, 2002). arXiv Preprint. https://arxiv.org/abs/2002.11649. [43] M. A. Nielsen and I. L. Chuang, Quantum Computation [27] L. Lin and Y. Tong, Solving quantum linear system prob- and Quantum Information: 10th Anniversary Edition, lem with near-optimal complexity (2019), arXiv Preprint. 10th ed. (Cambridge University Press, USA, 2011). https://arxiv.org/abs/1910.14596. [44] A. Childs, J. Preskill, and J. Renes, Quantum informa- [28] J. Haah, Product decomposition of periodic functions in tion and precision measurement (1999), arXiv Preprint. quantum signal processing, Quantum 3, 190 (2019). https://arxiv.org/abs/quant-ph/9904021v2.