Uniform Diagonalization Theorem for Complexity Classes of Promise Problems including Randomized and Quantum Classes

Friederike Anna Dziemba Institut f¨urTheoretische Physik, Leibniz Universit¨atHannover, Germany (Dated: December 19, 2017) Diagonalization in the spirit of Cantor’s diagonal arguments is a widely used tool in theoretical computer sciences to obtain structural results about computational problems and complexity classes by indirect proofs. The Uniform Diagonalization Theorem allows the construction of problems out- side complexity classes while still being reducible to a specific decision problem. This paper provides a generalization of the Uniform Diagonalization Theorem by extending it to promise problems and the complexity classes they form, e.g. randomized and quantum complexity classes. The theorem requires from the underlying computing model not only the decidability of its acceptance and rejec- tion behaviour but also of its promise-contradicting indifferent behaviour – a property that we will introduce as “total decidability” of promise problems. Implications of the Uniform Diagonalization Theorem are mainly of two kinds: 1. Existence of intermediate problems (e.g. between BQP and QMA) – also known as Ladner’s Theorem – and 2. Undecidability if a problem of a complexity class is contained in a subclass (e.g. membership of a QMA-problem in BQP). Like the original Uniform Diagonalization Theorem the extension applies besides BQP and QMA to a large variety of complexity class pairs, including combinations from deterministic, randomized and quantum classes.

I. INTRODUCTION upcoming of quantum computing the concept of promise problems clearly deserves more respect. In the seventies two contrary results gave new insight Promise problems are natural for semantic complex- into the structure of polynomial time reducibility and the ity classes such as randomized and quantum classes due hierarchy between the complexity class P of efficiently to their probabilistic nature. The main purpose of a solvable problems and the broader class NP of efficiently promise is to guarantee that an algorithm of reasonable verifiable problems. While Schaefer’s dichotomy theo- running time can differentiate yes- and no-instances well rem [1] states that every naturally restricted constraint- enough, i.e. adheres the probabilistic error allowed by satisfaction problem outside P belongs immediately to the definition of the complexity class. Without promises the hardest problems of NP (so-called NP-complete), many randomized and quantum classes would not con- Ladner’s theorem [2] shows for every problem outside P tain any canonical problems. For example, neither the the existence of an intermediate problem that is strictly randomized classes BPP and MA of efficiently solvable simpler but still outside P. and verifiable problems nor the quantum analogues BQP In quantum complexity theory Local Hamiltonian and QMA are known to contain any complete decision problems assume the role that constraint-satisfaction problems. But they all contain complete promise prob- problems play in classical complexity theory. The cat- lems, inlcuding problems of high physical relevance like egorization of Local Hamiltonian problems by [3] can be the QMA-complete Local Hamiltonian problem. regarded as a quantum analogue of Schaefer’s dichotomy The adaption of the Uniform Diagonalization Theo- result. In contrast, no quantum version of Ladner’s the- rem to randomized and quantum complexity classes re- orem has been formulated until today. quires two steps: After the preliminary section we will The strongest generalization of Ladner’s Theorem is first extend some necessary terminology originally de- arXiv:1712.07276v3 [cs.CC] 21 Feb 2019 nowadays known as “Uniform Diagonalization Theorem” fined in the context of decision problems to the context [4, 5] and is applicable to a large variety of complexity of promise problems like (total) decidability and recur- classes. By generalizing the reduction notion the authors sive (re-)presentation and show that these properties are of [6–8] even realize a useful application of the theorem to obeyed by standard randomized and quantum complexity complexity classes below P. Unfortunately, the formula- classes. In the next section we can then adapt the proof of tion of the Uniform Diagonalization Theorem only covers the Uniform Diagonalization Theorem to promise prob- classes of decision problems, whereas quantum complex- lems and their complexity classes. ity classes are formed by the broader concept of promise Informally the Uniform Diagonalization Theorem problems. While an algorithm for a decision problem states that for two complexity classes C and C0 and two has to work correctly on every input, an algorithm for a problems A/∈ C and A0 ∈/ C0, there exists another prob- promise problem only has to work correctly on promised lem B which inherits the property not to belong to any of inputs. Because of this “untotal” property, promise prob- the two complexity classes while still being reducible to lems are often avoided and therefore neglected in theo- the marked union of A and A0 (which basically implies re- retical computer sciences. But with the introduction of ducibility to the more difficult problem of the two). The randomized complexity classes and at the latest with the complexity bound on B provided by this reduction is cru- 2 cial, since just finding a problem outside two complexity that maps the current state and the symbol at the head classes is clearly trivial if it can be chosen arbitrarily more position to another state, overwrites the read symbol and difficult. moves the head at most by one cell (left / right / neutral). In the implications section we will show that fixing one If a final state is reached, the machine halts and outputs of the complexity classes and one of the problems leads the string that is written between the head position and to a simplified version of the theorem in the spirit of Lad- the next blank symbol. ner: Given a standard complexity class and a problem A outside this class, one always finds another problem that lies outside the class but is strictly simpler than A. Our versions of the Uniform Diagonalization Theorem and Ladner’s Theorem apply to all previously mentioned randomized and quantum classes and also to combina- tions from both kinds. This means for example that there exists an infinite hierarchy of intermediate prob- FIG. 1. Initial configuration of a Turing machine on input x = lems between QMA and BQP as well as between BQP 110 ... 1 and after one step according to transition function and BPP (understood as set of promise problems) under δ(z0, 1) = (z1, 0,R). the assumption that these classes are unequal. A second branch of implications from the Uniform Di- Given input x we denote the output of a DTM M by agonalization Theorem will cover undecidability results M(x). Note, that M(x) doesn’t have to be defined for for subclass membership, for example: Given a problem all inputs, since a DTM can also run into an infinite loop via a QMA-description it is undecidable if it is contained and never halt. The notion of an input can be extended in BQP. ∗ to a multipartite input x1, x2,... ∈ Σ . These inputs are then written onto the tape successivly separated by a blank symbol and we denote the output accordingly by II. PRELIMINARIES M(x1, x2,... ). There is a correspondance between all strings over the A. Deterministic Turing machines and Properties binary alphabet and Turing machines via the concept of of Functions G¨odelnumbers. A G¨odelnumber is a binary encoding of the transition function together with a specification of Let Σ∗ denote the set of all strings over the binary al- the final states. Note, that the transition function can phabet Σ = {0, 1}. Theoretical computer sciences mainly be encoded by a binary number of finite length since it deals with two kinds of computational tasks: is determined by finitely many transitions. If a binary number does not have the form of a valid G¨odelnumber ∗ ∗ 1. Either computing a function N0 → N0 or Σ → Σ it is interpreted as the encoding of a trivial machine that or always outputs 0. If we state in this paper that “a Turing machine is given” we mean that the G¨odelnumber of the 2. deciding a bipartite question in form of a promise machine is supplied. In the same manner a computable problem A = (Ayes,Ano) with function is always given via the G¨odelnumber of the machine that computes the function. Ayes ∩ Ano = ∅ ∗ Ayes ∪ Ano ⊆ Σ Definition 1. A promise problem A = (Ayes,Ano) is decidable iff there exists a DTM M such that and Ayes ∪ Ano called the promise. ∀x ∈ Ayes M(x) = 1 (“M accepts input x”) Problems for that the last line holds an equality are called ∀x ∈ A M(x) = 0 (“M rejects input x”). decision probems and are a special case of promise prob- no lems. The fundamental computational model for both The concept of G¨odelnumbers allows the construc- computational tasks is the deterministic Turing machine tion of the most famous undecidable decision problem: (DTM): The yes-instances of the Halting problem are exactly (the A deterministic Turing machine is an endless tape ma- G¨odelnumbers of) those DTMs that don’t halt given chine with a reading head starting on the first symbol of their own G¨odelnumber as input. the input x ∈ Σ∗, which is padded at both sides with in- finitely many blank symbols . Moreover, a DTM comes Computability of a function f :Σ∗ → Σ∗ means that along with a finite set of states S including an initial and there exists a DTM that for every input x ∈ Σ∗ outputs possibly several final states. A computational step of the f(x). For functions f : N0 → N0 it means that there machine is described by the transition function exists a DTM that vor every bin(x), x ∈ N0, outputs bin f(x)
, with bin(·) denoting the binary representa- δ : S × {0, 1, } → S × {0, 1, } × {L, R, N} tion of natural numbers without leading zeros. 3

For the later discussion of complexity classes and the been found until today. But especially when caring about proof of the central theorem we need some terminology complexity it is useful to consider as well other compu- relating to the runtime of a DTM. Most important, we tational models with plausible implementations. These say that a DTM M has a polynomial runtime iff there models aim at deciding decision or promise problems and exists a polynomial p over N0 such that for all inputs x ∈ are usually not used for the computation of functions, Σ∗ M halts in less or equal to p(|x|) computational steps. hence, they just come with a notion of acceptance and Two more properties of functions will become relevant rejection but not with a notion of a function output. in this paper; the first will mainly be used to restrict The first, the probabilistic Turing machine, simply ex- resources (runtime, witness size, etc.) in complexity class tends the model of the deterministic Turing machine by definitions, the second is important for the proof of the allowing a branching at every computational step, which central theorem: is mathematically reflected by a transition function of the form Definition 2. A function f : N0 → N0 is in the set poly iff f is polynomially bounded and there exists a DTM δ : S × {0, 1, } → P(S × {0, 1, } × {L, R, N}) M that for all inputs of length n outputs bin f(n)
in polynomial time. with P denoting the power set. All terminology of DTMs is used accordingly for each Definition 3. A function f : N0 → N0 is called time- constructible, iff there exists a DTM that for each input branch of a PTM. The formalism of G¨odelnumbers can of length n halts exactly in f(n) steps. also be easily extended to PTMs. A PTM as a whole has a probabilistic expression for accepting and rejecting Lemma 1. The following holds for a function f : N0 → an input x determined by the fraction of accepting and N0: rejecting branches: 1. f is time-constructible ⇒ f is computable. # accepting branches P (x) = acc # all branches 2. f is computable ⇒ ∃ time-constructible f 0 ≥ f. # rejecting branches P (x) = . Proof. rej # all branches

1. Let MT (f) be a DTM that time-constructs f. Then The fundamental computing model in quantum infor- the following machine computes f: mation are families of quantum circuits (Cx)x∈Σ∗ that are First, change the input bin(x) on the tape to its made up of a series of unitary quantum gates unary representation 1x. On this new input of U = U ,U ...U U length x simulate MT (f) with the following adap- x x,l x,l−1 x,2 x,1 tion: Interrupt after each original computational acting on a Hilbert space of multiple qubits usually ini- step (by changing into a new “interruption state”), k let the head run to the end of the written tape and tialized in the state |ini = |0 i and closed by a final projective measurement Π := |1i h1| of a designated increment a counter there starting at 0. If the orig- acc out inally subsequent state is a final state, let the head output qubit (e.g. the first one). The probability to mea- sure Πacc is considered the acceptance probability for the remain on the counter and change into the final ∗ state (i.e. the new output is the counter), other- input x ∈ Σ : wise let the head run back to its original position P (x) = hin| U †Π U |ini . and change into the subsequent state. acc x acc x A useful notion of complexity classes demands that the 2. Let Mf be a DTM that computes f. Then the fol- lowing machine time-constructs a function f 0 with allowed gates are of an elementary kind. In this paper we f 0 ≥ f: assume the standard gates H, T and CNOT that act non- trivially only on one or two qubits. In the computational Replace the input bin(x) by its length n = | bin(x)|. basis these gates have the following matrix representa- After simulating Mf , write its output in unary rep- tion: resentation 1f(n) and halt. Since writing onto f(n) cells needs at least time f(n), the new machine 1 1 1 0 0 Hadamard: H = √ time-constructs a function f with f ≥ f. 2 1 1 1 0  T-gate: T = i π 0 e 4 B. Randomized and Quantum Models 1 0 0 0 0 1 0 0 The Church-Turing thesis asserts that every reasonable Controlled NOT: CNOT =   . computational model can be simulated by the model of 0 0 0 1 a deterministic Turing machine and no contradiction has 0 0 1 0 4

The gate set {H, T, CNOT} is a proper choice for univer- C. Complexity Classes and Reductions sal quantum computing since it is known to be universal, i.e. any unitary operation can be approximated arbitrar- A complexity class is any set of promise problems. ily well by these gates. Usually such a set is defined by all problems that can be decided by a certain restricted machine model and Until now we have just described the structure of a can hence be considered of similar “complexity”. We quantum circuit but not the structure of a family of quan- will list here the definitions of the most important com- tum circuits (C ) ∗ whose circuits can basically look x x∈Σ plexity classes, namely those of efficiently decidable and very different for different inputs x. The minimum re- efficiently verifiable problems for deterministic, random- quirement on a circuit family should be the existence of ized and quantum computing. an algorithm that can compute a description of the quan- In accordance with the literature, we define the classes tum circuit C for each input x ∈ Σ∗. Stricly speaking, x P and NP based on deterministic Turing machines as a quantum circuit family is defined via a deterministic sets of decision problems. Historically, the first complex- Turing machine that on input x ∈ Σ∗ outputs an encod- ity classes were defined as sets of decision problems (or, ing of the quantum circuit C in a G¨odelnumber style, x equivalently, of languages, which equal the yes-instances e.g. encode of decision problems), since it is guaranteed that each DTM with a restricted runtime decides a decision prob- lem – in contrast to the later defined probabilistic and quantum machines.

Definition 4. The complexity class P is the set of all decision problems that can be decided by a deterministic Turing machine of polynomial runtime.

Definition 5. The complexity class NP is the set of all as the number decision problems A for that there exists a deterministic Turing machine M of polynomial runtime and an m ∈ 010 11 0 1001 0 110 110 111 0 11010 111. poly such that | {z } | {z } | {z } | {z } H on qubit 2 T on 1 CNOT from 2 to 3 CNOT from 1 to 3 m(|x|) ∀x ∈ Ayes ∃y ∈ Σ : M(x, y) = 1, ∀x ∈ A ∀y ∈ Σm(|x|) : M(x, y) = 0. Always be aware of the difference between the G¨odel no number of (the DTM representing) a quantum circuit The y in the definition of NP is called a witness. In family and the G¨odelnumber of a specific circuit that it contrast to the problem input x we will call the input outputs. An output of a DTM representing a quantum (x, y) the machine actually acts on the computational in- circuit family that is not a valid circuit G¨odelnumber put (here consisting of problem input and witness). is interpreted as an encoding of the trivial circuit that Next, we will define the randomized and quantum ana- never accepts. Due to the Turing machine construction logues of P and NP. With the introduction of random- procedure we will call a quantum circuit family obeying ized classes and at the latest with the upcoming of quan- the restrictions of a complexity class C also a C-machine, tum computing, the definition of complexity classes was e.g. a BQP-machine, though BQP-circuit family might broadened to promise problems. While in quantum in- sound more intuitive. formation it is common sense to understand complexity classes as sets of promise problems, classical computer A polynomial-time generated quantum circuit family scientists are still debating if randomized classes such as (U ) ∗ is a quantum circuit family whose constructing x x∈Σ BPP and MA should be considered as sets of decision DTM has polynomial runtime. Since a circuit’s G¨odel problems or sets of promise problems. To avoid confu- number is always longer than the circuit size (number sion we call the first BPP and MA and the later ones of gates), the size of the constructed circuit is upper PromiseBPP and PromiseMA. bounded by the same polynomial. The bound on the cir- cuit size due to the non-compressive property of the en- Definition 6. The complexity class BPP (PromiseBPP) coding justifies to call polynomial-time generated quan- is the set of all decision (promise) problems A for that tum circuit families “efficient”. Instead of the G¨odel there exists a probabilistic Turing machine M of polyno- scheme presented here one can of course also assume a mial runtime such that different circuit encoding. The specifics of the encod- ing are irrelevant for the definition of usual complexity 2 ∀x ∈ A : P ≥ , classes as long the the encoding is sensible, efficient but yes acc 3 non-compressive and guarantees that “specific informa- 1 ∀x ∈ A : P ≤ . tion about the structure of a circuit [is] computable in no acc 3 polynomial time from an encoding” [9]. 5

Definition 7. The complexity class MA (PromiseMA) QCMA and QMA can both be considered as quantum is the set of all decision (promise) problems A for that analogues of NP. While QCMA is defined with a classical there exists a probabilistic Turing machine M of polyno- witness y, QMA requires a quantum witness |ψi. Note, mial runtime and an m ∈ poly such that that in case of QCMA and QMA the circuit-constructing DTM should not only output the encoding of the gates m(|x|) 2 but also the number of witness qubits m. ∀x ∈ Ayes ∃y ∈ Σ : Pacc ≥ , 3 In extension of the notion “decidability”, we say that m(|x|) 1 the “C-machine M decides the problem A” if M is the ∀x ∈ Ano ∀y ∈ Σ : Pacc ≤ . 3 machine for that the membership of a problem A in one of the above complexity classes C is proven. From calling The probability thresholds c = 2 and s = 1 are called 3 3 a machine either a DTM or a C-machine, it will always completeness and soundness parameter of the complexity be clear if we mean the original notion of decidability or classes. a machine deciding A according to the definition of the Often literature remains unclear, if runtime-bounded complexity class C. probabilistic Turing machines like PromiseBPP- or For abbreviation reasons we denote the set of the in- PromiseMA-machines have to obey the specified run- troduced classes of promise problems by time only on promised inputs or also for non-promised inputs. Since a machine of the first kind can easily be C = {PromiseBPP, PromiseMA, BQP, QCMA, QMA}. transformed into one of the second kind (by counting the computational steps and aborting after the specified These complexity classes together with the correspond- runtime with a default value as output), we assume the ing classical classes of decision problems form the follow- second case. Accordingly we assume that the following ing hierarchy (each class contains the connected classes bounded-runtime generated families of quantum circuits below): obey their runtime specification for all inputs:

Definition 8. The complexity class BQP is the set of all QMA promise problems A for that there exists a polynomial- time generated familiy of quantum circuits (Ux)x∈Σ∗ on k qubits such that QCMA 2 ∀x ∈ A : h0k| U †Π U |0ki ≥ , yes x acc x 3 1 ∀x ∈ A : h0k| U †Π U |0ki ≤ . BQP PromiseMA no x acc x 3 Definition 9. The complexity class QCMA is the set of MA all promise problems A for that there exists a polynomial- PromiseBPP NP time generated familiy of quantum circuits (Ux)x∈Σ∗ on k + m ∈ poly qubits such that BPP m(|x|) ∀x ∈ Ayes ∃y ∈ Σ : 2 P (hy| ⊗ h0k|)U †Π U (|yi ⊗ |0ki) ≥ , x acc x 3 m(|x|) ∀x ∈ Ano ∀y ∈ Σ : Subset relations between classes of decision problems 1 and classes of promise problems are trivially strict when (hy| ⊗ h0k|)U †Π U (|yi ⊗ |0ki) ≤ . x acc x 3 their definitions are taken seriously. For studying hier- archy relations in a reasonable way, even quantum com- Definition 10. The complexity class QMA is the set of plexity classes are sometimes considered as restricted to all promise problems A for that there exists a polynomial- decision problems, e.g. when the question “Is BQP equal time generated familiy of quantum circuits (Ux)x∈Σ∗ on to P?” is asked. k + m ∈ poly qubits such that Reduction notions are a useful tool for a finer and also complexity class independent comparision of problems’ 2m(|x|) ∀x ∈ Ayes ∃ |ψi ∈ C : complexity. A problem A that can be reduced onto a 2 problem B is considered simpler as B since it can be (hψ| ⊗ h0k|)U †Π U (|ψi ⊗ |0ki) ≥ , x acc x 3 decided easily having knowledge about B. 2m(|x|) ∀x ∈ Ano ∀ |ψi ∈ C : Definition 11. A promise problem A is Karp- or m- 1 P (hψ| ⊗ h0k|)U †Π U (|ψi ⊗ |0ki) ≤ . reducible to a promise problem B (notation: A ≤m B) x acc x 3 iff there exists a polynomial-time computable function f : 6

Σ∗ → Σ∗ such that Class Complete problem Ref. P all problems with at least x ∈ A ⇒ f(x) ∈ B , yes yes one yes- and no-instance x ∈ A ⇒ f(x) ∈ B . no no NP k-Satisfiability, k ≥ 3 [10] Definition 12. A promise problem A is Cook- or T - BPP ? P reducible to a promise problem B (notation: A ≤T B) MA ? iff A can be decided by a DTM of polynomial runtime PromiseBPP Acceptance Ratio of DTMs [11] with oracle B (the DTM has access to an oracle state that upon entering replaces every x, x ∈ Byes, at the PromiseMA Stoquastic 6-Satisfiability [12] head position instantenously by 1 and every x, x ∈ Bno, BQP Quadratically Signed [13] by 0). Weight Enumerator Notice, that an oracle is only allowed to be queried QCMA Ground State Connectivity [14] for elements of Byes and Bno. In the case of promise QMA k-Local Hamiltonian, k ≥ 2 [15, 16] problems one has to ensure that the DTM does not query the oracle for any non-promised inputs of B. for BPP and MA. There seems to be a strong belief in the P P Lemma 2. It holds A ≤m B ⇒ A ≤T B. theoretical computer science community that BPP = P and MA = NP which may also be partly due to this fact. Proof. Let f be the polynomial-time computable func- Moreover, we see here a good justifaction why rather tion that reduces A to B. Then A can also be solved by PromiseBPP and PromiseMA should be considered as a polynomial-time DTM first simulating the computa- the proper complexity class definitions. tion of f and then querying the B-oracle on the function Let us close this section by mentioning that some output. promise problems A are considered of such high logical Both introduced reduction notions (and any other rea- or physical relevance that it is worth defining all prob- sonable reduction notion) form a pre-order obeying re- lems reducible to A as new complexity class with an own flexivity and transitivity on the set of problems. For be- name. The complexity class TIM ⊆ QMA is such an ing a partial order the antisymmetric property is missing: example [3], which consists of all problems that can be Problems that can be reduced onto each other can still reduced onto a restricted Local Hamiltonian problem of be different. transverse Ising model form. The notion of completeness A special role is assumed by those problems that are hence allows an alternative, non-machine based approach the most difficult ones in a complexity class: to define complexity classes. Definition 13. A promise problem A is called m-hard (T -hard) for a complexity class C iff all problems in C III. A FRAMEWORK FOR PROMISE can be Karp-reduced (Cook-reduced) to A. If A is a prob- PROBLEMS lem of C itself, A is called m-complete (T -complete) for the complexity class C. A. Extremal Problems and Closure Properties of We denote by C -cm and C -cT, respectively, the set of Complexity Classes all m- and T -complete problems for C. Definition 14. A complexity class C that for every The stricter notion of Karp-reducibility is the standard promise problem (Ayes,Ano) ∈ C also contains every reduction notion for complexity classes above P, since 0 0 P subproblem (Ayes,Ano) with B ∈ C and A ≤m B implies A ∈ C for all complex- ity classes C that can compute a polynomial-time com- 0 A ⊆ Ayes putable function as subroutine (like all previously defined yes 0 classes). The representative nature of Karp-complete Ano ⊆ Ano problems for complexity classes above P is the reason why Karp reductions are also simply called “complete” is called closed under promise restriction. (without the addition “m” or “Karp”) and those that According to definitions 6 - 10 the classes of C are scientists like to find for complexity classes. For most of closed under promise restriction. Since we like to the introduced complexity classes complete problems are uniquely identify only one problem with a machine like in known: the case of P and NP we introduce the notion of extremal The table just contains one examplary complete prob- problems: lem for each complexity class, though for most classes several complete problems are known. For QMA these Definition 15. The promise problem with the smallest are meanwhile several dozens; for NP several hundreds. promise decided by a DTM or a C-machine M, C ∈ C, is Remarkably, no complete (decision) problems are known called the extremal problem of M and denoted by P (M). 7

Accordingly, we call the decision problem decided by a Obviously, for decision problems it holds P- or NP-machine M extremal and denote it by P (M). ANB = A\B = B\A = A4B, Definition 16. We denote by C∗, C ∈ C the restriction of the according complexity class to its extremal problems. while for general promise problems A\B 6= B\A and the symmetric notion ANB is only a subset of the symmetric Of course, the notion of extremal problems can easily notion A4B: be adapted to other “machine-based” complexity classes. One might wonder why C∗, C ∈ C, is not used as the proper definition for the according randomized or quan- A4B tum complexity class. Indeed, logically there is no rea- son to artificially demand a larger promise from a prob- lem than necessary, but practically one usually starts by A\B B\A defining a logically or physically interesting problem and then aims at proving membership for this problem in a certain complexity class. These proofs often involve A B many implication arguments and approximations to fi- N nally show that an algorithm accepts with a sufficiently high or low probability. But this does usually not rule Definition 18. A complexity class C (of decision prob- out that the algorithm also accepts some other, non- lems) is closed under finite variations (c.f.v.) iff A ∈ C promised instances with similar high or low probability. implies B ∈ C for every (decision) problem B that equals Even in the case of the k-Local Hamiltonian problem, in A almost everywhere. which the fundamental algorithm accepts with a proba- bility that trivially relates to the promise [15], the final amplification to achieve the standard completeness and B. Total Decidiability soundness parameters involves Chernoff’s bound [15, 17]. Hence, even the k-LH problem is probably not an ex- tremal problem according to its usual definition. But the Many standard literature uses the notion of decidabil- advantage of its usual definition is that the promise on ity only for decision problems (e.g. [18]). Those who the ground energy is simply physically describable. use the notion in the context of promise problems (e.g. [19, 20]), agree on the one we gave in definition 1. The We stress these subleties since structural results like disadvantage of this definition is the arbitrary behaviour the Uniform Diagonalization Theorem make actually of the machine on non-promised inputs. It is therefore statements about the structure of machine sets and hence reasonable to introduce a stricter version of decidability about extremal classes. To formulate structural results that we will call total decidability: correctly we use the above notations.

We will close this subsection by defining another clo- Definition 19. A promise problem A = (Ayes,Ano) is sure property that even applies to all the complexity totally decidable iff there exists a DTM M such that classes in this paper: the closure under finite variations. For two languages (or equivalently, decision problem) ∀x ∈ Ayes M(x) = 1 there exists a common definition of their symmetric dif- ∀x ∈ Ano M(x) = 0 ference [5]. For promise problems we see two reasonable ∗ possibilities to extend this definition. We define closure ∀x ∈ Σ \(Ayes ∪ Ano) M(x) = 10. under finite variations via the wider notion of the total Obviously total decidability implies decidability and symmetric difference of two promise problems, since this the two notions are identical in the case of decision prob- is the notion that will become relevant in the proof of the lems. There are reasonable examples for both promise Uniform Diagonalization Theorem. problems that are totally decidable and promise prob- Definition 17. For promise problems A and B we define lems that are decidable but not totally decidable:

Example 1. The promise problem A = (Ayes,Ano) with ANB := {Ayes ∩ Bno} ∪ {Ano ∩ Byes} (sym. diff.) A\B := {Ayes\Byes} ∪ {Ano\Bno} (difference) Ayes = {DTM with even runtime on its G¨odelnumber}

A4B := (A\B) ∪ (B\A) (total sym. diff.) Ano = {DTM with odd runtime on its G¨odelnumber}

We say “A equals B almost everywhere (a.e.)” iff A4B is decidable but not totally decidable. is finite. Proof. Clearly, A is decidable by simply counting the run- Note, that the right side of the above difference expres- ning time of the given machine. sions is always a set, despite the fact that the promise The set of non-promised inputs consists of exactly problems on the left side correspond to tuples of sets. those DTMs that don’t halt on their own G¨odelnum- 8 ber as input. If A was totally decidable, one could hence in the case of QCMA and decide the Halting problem. k † k Pacc := highest eigenvalue of Q := h0 | UxΠaccUx |0 i The purpose of the promise of the above problem is to avoid undecidable instances, which prevents the prob- in the case of QMA. lem from being totally decidable. The original purpose We recall that the generated quantum circuit just con- of promise problems was however to exclude instances for sists of H-, T- and CNOT-gates. In case of BQP the that checking membership up to a certainty / accuracy acceptance probability Pacc equals hence a sum of prod- required by the randomized or quantum complexity class ucts of elements from the field under consideration would exceed the runtime restriction 1 of that class. Take for example the k-local Hamiltonian Q(√ , i) problem: In the standard protocol [15] the promised gap 2 on the ground energy of the Hamiltonian directly relates (notice that the phase eiπ/4 of the T-gate can be written to the acceptance probability or – if this is improved to as √1 (1 + i)). obey the standard completeness and soundness parame- 2 ter – to the runtime overhead caused by amplification. This finite field extension can be handled as 4- Luckily we don’t care about runtime when asking for dimensional vector space V over the rational numbers total decidability, hence, the extremal problems of most with the abstract basis vectors standard complexity classes are totally decidable: 1 i 1, √ , i, √ . Lemma 3. The extremal problems of PromiseBPP- and 2 2 PromiseMA-machines are totally decidable. A DTM can compute operations on the coefficients ex- Proof. The extremal problem of a PromiseBPP-machine actly by storing two integers and a sign for each rational can be totally decided by a DTM that simulates all number and it can carry out the finitely many differ- branches of the PTM (which always halt per definition) ent products of basis vectors abstractly. There is con- and checks its fraction of accepting branches. In case of sequently a DTM that can compute Pacc as a rational linear combination of the abstract vectors 1 and √1 (the a PromiseMA-machine this algorithm simply has to be 2 repeated for each of the possible 2m witnesses. imaginary vector obviously vanishes for the acceptance probability). If the coefficient of the √1 -vector vanishes and the co- Note, that for the correctness of the above lemma the 2 concrete form of the runtime bound r and the concrete efficient of the 1-vector equals c or s, the DTM can accept values of the completeness and soundness parameters c or reject directly. Otherwise, it is clear that Pacc unequals and s are irrelevant. Since the acceptance probability c and s. Since there exists methods (e.g. the babylonian of a halting PTM is always an exactly computable ra- / Heron method) to compute monotonous sequences of tional number, the statement holds accordingly for any converging upper and lower bounds of square roots, the complexity class that replaces r and c > s by other com- DTM can compute improving lower and upper bounds putable functions and in addition if one or both of the cri- on Pacc until their exceeding of or falling below c or s teria Pacc ≥ c and Pacc ≤ s are changed to strict inequal- discloses if Pacc < s, Pacc > c or Pacc ∈ ]s, c[. ities. The statement also remains valid for computable To totally decide the extremal problem of a QCMA- functions c = s with maximally one of the criteria con- circuit family the above algorithm simply has to be run taining a strict inequality. These generalizations hold as for all possible witnesses y ∈ {0, 1}m. well for the quantum analogue of the lemma, which we In case of a QMA-problem the following equivalences will prove next: hold: Lemma 4. The extremal problems of BQP-, QCMA- s I −Q positive semi-definite ⇐⇒ Pacc ≤ s and QMA-machines are totally decidable. c I −Q not positive definite ⇐⇒ Pacc ≥ c Proof. By simulating the circuit-constructing DTM we obtain the G¨odelnumber of a quantum circuit with gate and consequently Pacc ∈ ]s, c[ if none of the two con- series U on k plus possibly m witness qubits. We have ditions holds. Consequently, we have to argue that the x positive semi-definiteness of s −Q and the positive def- to differentiate if the acceptance probability Pacc is ≤ s, I ∈ ]s, c[ or ≥ c with s = 1 , c = 2 and initeness of c I −Q are decidable. Sylvester’s criterion 3 3 states the positive definiteness of a matrix is equivalent P := h0k| U †Π U |0ki to the positivity of all its principal minors and positive acc x acc x semi-definiteness to the non-negativity of all its leading in the case of BQP, principal minors. It is simple to decide the positivity and non-negativity of minors (determinants of submatrices) k † k by computing improving bounds in the vector space W Pacc := max (h0 | ⊗ hy|)UxΠaccUx(|0 i ⊗ |yi) y∈{0,1}m as described above. 9

As we discussed before, the above two lemmata hold is not even necessary for the total decidability of the ex- accordingly for randomized and quantum complexity tremal problems; the only necessary condition is the com- classes with different runtime bounds and completeness putability of the Stinespring dilation (a representation of and soundness parameters c and s as long as they are the channels by ancilla qubits and a unitary operation) computable functions. In case of quantum complexity in the abstract representation of algebraic numbers: classes one might also ask if the last lemma still holds under the assumption of a different gate set providing Lemma 5. The extremal problem of a noisy QMAE - the basic elements of a quantum circuit. From the proof machine is totally decidable if E = (Em)m∈N is a quantum we can see that the answer is obviously yes as long as channel family whose Stinespring dilation over the field the matrix entries of these gates are from a field exten- of algebraic numbers is computable. sion of with countable degree such that the product Q Proof. Given a QMA -machine outputting a quantum of any two basis vectors is computable abstractly and a E circuit with gate series U on k plus m witness qubits, its converging lower and upper bound for each basis vector x extremal problem can be totally decided by DTM sim- is computable. ilar to the one defined in the proof of lemma 4. After An example for such an infinite field extension is the the computation of the Stinespring dilation with l an- field of all algebraic numbers which is made up by all A cilla qubits and unitary Vm of the channel Em defined by roots of polynomials over Q. Instead of writing elements the witness size m the DTM has to check if the highest of A as linear combinations of abstract basis vectors, one eigenvalue of can also identify each element directly by a unique ratio- k l † † k l nal polynomial and a sufficiently small isolating rectangle (h0 | ⊗ h0 |)VmUxΠaccUxVm(|0 i ⊗ |0 i) in the complex plane. The author of [21] shows that sum- mation, substraction, multiplication, division and power- 2 is ≥ 3 and if the highest eigenvalue of ing in this representation only needs rational computa- k † k tions (including the computation of resultants of polyno- h0 | UxΠaccUx |0 i mials which is a new polynomial whose coefficients are an expression in the coefficients of the original ones) and 1 is ≤ 3 . can hence be done exactly. The final isolating rectangle Both these questions can be answered deterministi- can be narrowed enough to allow the approximation of cally by computing in the field of algebraic numbers as the final number to arbitrary precision, while the exact described before (the initial channel unitary is nothing root candidates c and s can be ruled out as before by else then an additional gate). If the first question is an- simple insertion. swered with “yes”, then the input is a yes-instance. If One case in which a complexity class crucially changes the first question is answered with “no” but the second with the form of the gate set is a one-sided complexity with “yes”, the input is a no-instance. If both questions class such as QMA1 which equals QMA with c = 1. To are answered with “no” the input is a non-promised in- achieve this perfect completeness a QMA1-circuit family stance. is usually allowed to contain any matrix element from the field the problem is formulated in. The authors of [22] advocate the algebraic numbers as largest reasonable C. Recursive (Re)presentation field to define QMA1. If one naively thought to allow any field that can be “described by words”, this would The Uniform Diagonalization Theorem will apply to instead lead to an easy construction of circuits whose those complexity classes for that the extremal problems extremal problem is not decidable at all. Consider e.g. are totally decidable and the corresponding machines a field containing a Chaitin’s number whose i’th digit enummerable. These two properties together define a equals 1 if i is a yes-instance of the Halting problem and class as recursively (re)presentable, which we introduce 0 otherwise. below as an extension of the well-known notion for deci- sion problems [4, 5]: That total decidability still holds for extremal prob- lems of QMA-like machines with gates over algebraic Definition 20. A complexity class C of promise prob- numbers, also implies the total decidability for QMA- lems is recursively presentable, iff there exists a com- variants where the witness in the completeness case (in- putable series M0,M1,M2 ... of halting DTMs such put is a yes-instance) passes a natural quantum chan- that C contains exactly those promise problems A = (Ayes,Ano) for that there exists an i ∈ 0 such that nel. These noisy QMA-classes, denoted by QMAE with N E = (Em)m∈N the quantum channel family, have the ad- vantange that they describe a large variety of classes be- ∀x ∈ Ayes : Mi(x) = 1, low QMA including QCMA (choosing Em as fully dephas- ∀x ∈ Ano : Mi(x) = 0, ing channel) and BQP (choosing Em as fully depolarizing ∀x∈ / Ayes ∪ Ano : Mi(x) = 10. channel). In [23] these classes were introduced for inde- pendent and identical qubit noise. But this restriction We call a class C0 recursively representable iff it equals 10 the closure of a recursively presentable class C under is that we don’t know how to decide if the extremal prob- promise restriction. lem of a polynomial-time PTM is a decision problem, i.e. 2 1 if the machine accepts with probability ≥ 3 or ≤ 3 on Computability of the series M0,M1,M2,... means of all inputs. This relates to the missing knowledge of com- course the computability of the function i → Mi. Ex- plete problems for BPP and MA, since a complete prob- pressing it as a series just reflects better the enummer- lem would provide another possibility to prove recursive ability property. presentation, namely by enummeration of all reduction The property of recursive representability can obvi- functions: ously only be held by complexity classes of promise prob- lems, since classes of decision problems are not closed un- Lemma 7. For a totally decidable problem A the set der promise restriction by definition. Reversely, recursive P ≥m P presentability can not only apply to classes of decision A := {promise problem B | B ≤m A} problems but also to classes of promise problems, espe- cially to those that are restricted to extremal problems. is recursively representable.

∗ If A is a decision problem, then the series of DTMs Lemma 6. The complexity classes P, NP and any C ≥P that recursively represent A m recursively presents all de- with C ∈ C are recursively presentable. cision problems that can be m-reduced to A. Proof. All deterministic Turing machines can be simu- Proof. Let M the DTM that totally decides A and lated given their G¨odelnumber and all polynomials over A (f ) the computable series of all polynomial-time form a computable series (p ) since polynomials i i∈N0 N0 i i∈N0 computable functions Σ∗ → Σ∗. Then (M ) with whose coefficients add up to the same sum form a finite i i∈N0 set and these are obviously enummerable. Hence, one can
Mi(x) = MA fi(x) define a computable series (Mi)i∈N0 for all P-machines with Mi, i interpreted as pair (j, k), the P-machine sim- is a recursive representation (presentation) of all (deci- ulating the DTM with G¨odelnumber j up to runtime pk sion) problems B that are reducible to (the decision prob- (with returning a default value, if the DTM doesn’t halt lem) A. on 0 or 1 in time pk). Note that by dropping or adapting the restriction on This approach allows to prove recursive the output values we can as well obtain a computable (re)presentation for complexity classes that are de- series for all polynomial-time computable functions Σ∗ → fined via a complete problem like TIM, even if they are Σ∗. If we consider the index i even as tuple (j, k, l) and missing a machine-based definition. simulate the DTM with G¨odelnumber j up to time pk The enumerability of reduction functions (and and output the minimum of the obtained output and pl, polynomial-time oracle machines) can also be used to we even construct a computable series (fi)i∈N0 for all prove a result the other way around: all problems of functions in the set poly. usual complexity classes that are more difficult than a

A computable series (Mi)i∈N0 of all DTMs that decide problem A are recursively presentable as well. The proof NP-problems is realized by defining Mi, i interpreted as basically resembles the proof for Cook-complete decision tuple (j, k, l), as the DTM that checks the acceptance problems from [4]: probability of the DTM with G¨odelnumber j limited to Lemma 8. Let C be a recursively presentable complexity time pk for each witness of length fl. class c.f.v. and A ∈ C be totally decidable. Then

By interpreting G¨odelnumbers as encodings of prob- P ≤m P abilistic or quantum circuit generating Turing machines AC : = {B ∈ C | A ≤m B}, we can construct in a similar way computable series of all P ≤T P C-machines, C ∈ C. Since the extremal problems of these AC : = {B ∈ C | A ≤T B} machines are totally decidable according to lemmata 3 are recursively presentable. and 4 the machines in these series can be replaced by the DTMs that totally decide their extremal problems to ob- Proof. Let (Mi)i∈N0 be a recursive presentation of C, tain a recursive presentation of the complexity class. (fi)i∈N0 the computable series of all polynomial-time computable functions Σ∗ → Σ∗ and (O ) the com- Corollary 1. The complexity classes of C are recursively i i∈N0 putable series of all polynomial-time oracle Turing ma- representable. chines. Let MA be the DTM that totally decides A. ≤P Corollary 2. For quantum channel families E whose m For the recursive presentation of the set AC we define Stinespring dilation over the field of algebraic numbers the DTM Ni, i = (j, k), that on input x checks forall ∗ is computable, the complexity class QMAE is recursively |y| ≤ |x| if presentable and QMAE hence recursively representable. y ∈ A ⇒ f (y) ∈ P (M ) Note, that we do not know how to recursively present yes j k yes BPP or MA. The problem of the straight forward method y ∈ Ano ⇒ fj(y) ∈ P (Mk)no 11

n is fulfilled. If yes, it outputs Mk(x), otherwise MA(x). step is efficient in |x|. Since r (0) ≥ n, the number of it- ≤P For the recursive presentation of the set A T we define eration steps is limited by |x|+1 and the above algorithm C is an efficient decision algorithm for (G[r], G[r]). the DTM Ni, i = (j, k), that on input x checks forall |y| ≤ |x| if Now, we can finally prove the Uniform Diagonalization Theorem and construct a problem B from two problems 0 0 0 y ∈ Ayes ⇒ Oj with oracle P (Mk) on input y A/∈ C and A ∈/ C that inherits the property B/∈ C ∪C 0 only queries the oracle for promised while being Karp-reducible to the marked union A ⊕ A . inputs and accepts Definition 22. The marked union A⊕A0 of two promise 0 y ∈ Ano ⇒ Oj with oracle P (Mk) on input y problems A and A is defined as the promise problem D with only queries the oracle for promised 0 inputs and rejects Dyes = {0x|x ∈ Ayes} ∪ {1x|x ∈ Ayes} 0 Dno = {0x|x ∈ Ano} ∪ {1x|x ∈ A }. is fulfilled. If yes, it outputs Mk(x), otherwise MA(x). no P P ≤m ≤T (Ni)i∈N0 is a recursive presentation of AC (AC ) Theorem 1 (Uniform Diagonalization Theorem). Let C, 0 since P (Ni) = P (Mk) if P (Mk) for i = (j, k) is a problem C be complexity classes closed unter finite variations of of C on that A can be m-(T -)reduced while otherwise which each is recursively presentable or recursively repre- 0 0 P (Ni) = A almost everywhere. sentable. Let A/∈ C, A ∈/ C be totally decidable promise problems. Then there exists a totally decidable promise Corollary 3. Let C be a recursively presentable com- problem B such that plexity class c.f.v. with at least one totally decidable m- 0 P 0 (T -)complete problem. Then C -cm (C -cT) is recursively B/∈ C ∪ C and B ≤m A ⊕ A . presentable. If A and A0 are extremal for one of the complexity classes from C or decision problems, then so is B. IV. PROOF OF THE UNIFORM 0 0 0 Proof. Let M0,M1,M2,... and M0,M1,M2,... be re- DIAGONALIZATION THEOREM cursive representations (presentations) for the complex- ity classes C and C0, respectively. Due to A/∈ C, every With the extended definitions and new notations in- Mi does not (totally) decide correctly some input with troduced in the last section the proof of the extended regard to the problem A (for a recursively representable Uniform Diagonalization Theorem resembles the original class C such a “contradicting” element can only be from one limited to decision problems [4, 5]. Ayes ∪Ano; for a recursive presentable class also instances Before we actually prove the Uniform Diagonalization from Ayes ∪ Ano can be contradicting, namely iff Mi ac- Theorem we first restate the definition and an efficiency cepts or rejects). The same holds for C0 and A0. The condition for the so-called gap language G[r], which al- construction idea for the new problem B is to “mix” A lows us later to mix the two problems stated in the Uni- and A0 such that B inherits a “contradicting” element 0 form Diagonalization Theorem by restricting them to cer- for each Mi and Mi . tain alternating intervals: To define a valid promise problem we mix A and A0 by restricting them to alternating intervals via the pre- Definition 21. Let r ∈ N0 → N0 be a computable func- viously defined gap language, i.e. tion with r(m) > m for all m ∈ N0. The gap language 0 generated by r is defined as the set Byes = (G[r] ∩ Ayes) ∪ (G[r] ∩ Ayes) 0 G[r] := {x ∈ Σ∗ | rn(0) ≤ |x| < rn+1(0) for n even} Bno = (G[r] ∩ Ano) ∪ (G[r] ∩ Ano) with rn denoting the n-fold concatenation of r. with a properly chosen function r : N0 → N0. Clearly, in this form B = (Byes,Bno) is a valid promise problem. The function r has to be chosen such that for each M the Lemma 9. If r : N0 → N0 with r(m) > m is time- i union of all even intervals, i.e. G[r], contains an element constructible, then (G[r], G[r]) ∈ P. 0 contradicting A and that for all each Mi the union of all Proof. Compute iteratively r(0), r2(0), r3(0) ... like in odd intervals, i.e. G[r], contains an element contradicting the proof of lemma 1. Abort the iteration if the counter A0 (see figure 2). k during the computation of r (0) reaches |x|. Accept if We achieve this by defining the function q : N0 → N0, k − 1 = n is even, otherwise reject. q(n) := max{|zi,n|} + 1 Revisiting the proof of lemma 1.1 shows that the com- i≤n putation of r(m) is efficient in r(m) and so is an aborted ∗ simulation in the final counter. Hence every iteration with zi,n ∈ Σ the smallest word according to the usual 12

FIG. 2. Mixing of the two problems A and A0 including all necessary elements to push it outside C and C0.

∗ ∗ binary order such that |zi,n| > n and (G[r], G[r]) ∈ P and hence f :Σ → Σ with ( ( A\P (Mi) if C recursively representable 0x if x ∈ G[r] zi,n ∈ f(x) := A4P (Mi) if C recursively presentable. 1x if x ∈ G[r]

is a valid Karp-reduction from B to A⊕A0. This complex- Notice that zi,n always exists. Otherwise, ity bound on B is the essence of of the Uniform Diago- nalization Theorem, since just finding a problem outside A ⊆ P (Mi) a.e. if C recursively representable two complexity classes is trivial when it can be chosen A = P (Mi) a.e. if C recursively presentable arbitrarily more difficult. G[r] ∈ P also implies that B is extremal if A and A0 and since C is closed under finite variations (and promise are extremal for a complexity class C ∈ C, since every restriction in the first case) this implies A ∈ C, which class C ∈ C is capable of performing the polynomial-time contradicts the hypothesis of the theorem. decision algorithm for G[r] as initial subroutine before simulating the algorithm for A or A0. The total decidability of A and the existence of z i,n If one had chosen the interval jumps not at r(n) but imply the computability of q. Analogously, the function exactly at q(n) and q0(n), i.e. exactly at the highest 0 0 necessary contradicting element, then the determination q (n) := max{|zi,n|} + 1 i≤n of the interval containing an input x can be far from efficient. For this one actually would have to compute 0 ∗ 0 is computable with zi,n ∈ Σ defined accordingly for A the contradicting elements which means a simulation of and M 0. i all machines Mi with i ≤ n for which we cannot fix a general polynomial runtime bound. The trick is that We now choose our desired function r as the time- r(n) is defined larger than the maximum of these contra- constructible function dicting elements and the runtime that it needs to com- r(n) ≥ max{q(n), q0(n)} pute them (recall the proof of Lemma 1). So usually r(n)  q(n) and the contradicting elements in figure 2 that exists according to lemma 1. The definitions of q(n) should be drawn cumulated at the lower interval limits. and q0(n) imply r(n) > n. Hence, the gap language G[r] This is also the reason why the proof technique is some- is well-defined. And since it is decidable, B is totally times referred to as “delayed diagonalization” [24]. The decidable. check if the next interval limit lies above the input x can now not only be answered positively by the output of the Notice that the most important step of the proof is iterative computation of contradicting elements but also indeed to choose the time-constructible function r in- when this computation exceeds a runtime of |x|, which is stead of q and q0 for defining the interval jumps of the obviously efficient. gap language. For this function lemma 9 tells us that After we proved the reduction statement of the theo- 13 rem, it only remains to show rigorously that the defined of the QMA-machine that decides the Local Hamiltonian problem B lies indeed outside the two complexity classes problem) and the respective complexity classes C and C0. Assume B ∈ C. This means there exists an ∗ 0 ∗ ∗ i ∈ N0 such that C = QMA -cm C = QMA \ QMA -cm, ∗ 0 ∗ ∗ C = QMA -cT C = QMA \ QMA -cT . B\P (Mi) = ∅ if C recursively representable B4P (Mi) = ∅ if C recursively presentable. Notice that these classes are closed under finite variations since the assumption QMA is strictly more powerful than Let m be an even integer such that n := rm(0) ≥ i. As we the closure of P under promise restriction implies that m- have seen above there exists zi,n ∈ [n, r(n)[ with zi,n ∈ and T -complete problems have infinitely many yes- and A\P (Mi)(zi,n ∈ A4P (Mi)). But since zi,n ∈ G[r], no-intances. this implies zi,n ∈ B\P (Mi)(zi,n ∈ B4P (Mi)) which is impossible. Hence, our initial assumption is wrong The above results on their own might not seem very and B/∈ C. Analogously, it can be proven that B/∈ C0. intriguing, but they reveal their whole power in the fol- This completes the proof of the Uniform Diagonaliza- lowing implications: tion Theorem. Corollary 6. Given a QMA-machine it is undecidable if its extremal problem is in BQP assuming BQP ( QMA. V. IMPLICATIONS Proof. Assume BQP ( QMA and it is decidable whether the extremal problem of a QMA-machine is in BQP. Let ∗ This section briefly lists the most important impli- M1, M2, ... be a recursive presentation of QMA . By cations of the Uniform Diagonalization Theorem for- substituting every Mi whose extremal problem is in BQP mulated for the classes QMA and BQP. This list is by the QMA-machine deciding the k-LH∗ problem, we ∗ ∗ far from being comprehensive. While QMA and BQP obtain a recursive presentation of QMA \ BQP which is can be substituted by many other pairs of recursively a contradiction to corollary 4. Hence, it is undecidable if (re)presentable classes, we like to stress again that these the extremal problem of a QMA-machine is in BQP. implication are not known to hold for BPP and MA Corollary 7. Given a QMA-machine it is undecidable due to our lacking knowledge about their recursive pre- if its extremal problem is m-complete (T -complete) under sentability and complete problems. the assumption that QMA does not equal the closure of In contrast to proving recursive representability like in P under promise restriction. lemma 6, we can use the Uniform Diagonalization Theo- rem firstly as a tool to prove that certain classes are not Proof. Analogously. recursively presentable: Originally [4] proved the above corollaries for combina- ∗ ∗ Corollary 4. The class QMA \ BQP is not recursively tions of the complexity classes NP P, PSPACE and PH presentable. as well as complement classes such as co-NP. We omit a ∗ ∗ corresponding version here, since complement classes are Proof. If BQP = QMA, then QMA \ BQP is empty and not very common to consider for sets of promise prob- therefore not recursively presentable. Let’s hence con- lems. One reason might be that some structural con- sider the case BQP ( QMA. Then there exists a problem 0 ∗ ∗ ∗ ∗ sequences that hold for complement classes of decision A ∈ QMA \ BQP . Let’s assume that QMA \ BQP problems do not hold for promise problems. E.g., if a is recursively presentable. Clearly, A = (∅, Σ∗), A0, ∗ ∗ 0 ∗ problem in co-NP turned out to be NP-complete under C = QMA \ BQP and C = BQP fulfill the hypothesis Cook reductions, then this would imply NP = co-NP of the Uniform Diagonalization Theorem. The problem [20, 25]. But there is no analogous implication known for B constructed in the Uniform Diagonalization Theorem 0 classes of promise problems like for QMA and its com- is Karp-reducible to A⊕A and hence in QMA. Indeed, it plement. is even in QMA∗ since A and A0 are extremal for QMA. On the other hand, the Uniform Diagonalization The- Undecidability results form a first branch of implica- 0 ∗ orem tells us that B/∈ C ∪ C = QMA which is a tions of the Uniform Diagonalization Theorem. The sec- ∗ ∗ contradiction. Hence, QMA \ BQP is not recursively ond branch of implications – proving the existence of in- presentable. termediate problems – is established by Ladner’s simpli- ∗ ∗ fication of the theorem: Corollary 5. The classes QMA \ QMA -cm and ∗ ∗ QMA \ QMA -cT are not recursively presentable under Theorem 2 (Ladner’s theorem). Let A be a promise the assumption that QMA does not equal the closure of problem in QMA∗ \ BQP∗. Then there exists a problem ∗ ∗ P P P under promise restriction. B ∈ QMA \ BQP with B ≤m A and A T B. Proof. This follows analogously to corollary 4 by substi- Proof. C := BQP∗ is recursively presentable according 0 ∗ 0 ∗ P tuting the problem A = k-LH (the extremal problem to lemma 6 and so is C := {D ∈ QMA | A ≤T D) 14 according to lemma 8. These complexity classes and A promise problems and showed that standard random- and A0 := (∅, Σ∗) hence fulfill the hypothesis of the Uni- ized and quantum complexity classes fulfill the recursive form Diagonalization Theorem. Moreover, A and A0 are (re)presentability property required by the Theorem. In extremal for QMA. order not to overload the paper we explicitly showed this Consequently, there exists a problem B ∈ QMA∗ such for the classes PromiseBPP, PromiseMA, BQP, QCMA ∗ P P 0 that B/∈ BQP , A T B and B ≤m A⊕A . The last con- and QMA, but the argumentation is easily adaptable P dition simplifies to B ≤m A since the reduction function to most natural randomized and quantum complexity can be concatenated by the polynomial-time computable classes including classes without probability gap such as function that maps every string with an initial 0x to x PP and classes with different time- or space-restrictions and every 1 to a default no-instance of A (which has to and an interactive or multi-witness extension. We also exist due to A/∈ BQP∗). argued that recursive (re)presentability is still fulfilled if the gate set of the quantum computing model is extended Corollary 8. If BQP ( QMA, then there exists an in- to all matrices over algebraic numbers, which is relevant finite hierarchy of intermediate problems between QMA for partly-deterministic classes such as QMA1. and BQP (regarding both Karp- and Cook-reductions). Besides the formulation of further corollaries for spe- Since Ladner’s Theorem constructs the intermediate cific interesting pairs of classes like [4, 5] list for deci- problem as a mixture of the hard problem A and the sion problems, it remains the non-trivial task to adapt constant-no problem A0, the series of intermediate prob- the here presented form of the Uniform Diagonalization lems between A = k-LH∗ and BQP are variants of the Theorem to more stricter reduction notions than Karp to Local Hamiltonian problem with more and more yes- make the theorem applicable to randomized and quan- instances turned into no-instances. This is why Ladner’s tum classes with a runtime time below polynomial. For original proof is also called the method of “blowing holes decision problems and log-space and log-time-reductions into the complete problem” [24]. this has been achieved by [7, 8]. Does this descriptive property of the intermediate While it is clear that different reduction notions are problems tell us something about the difficulty of specific needed for complexity classes below P, complexity classes Local Hamiltonian instances? Unfortunately, the criteria above are usually considered within the framework of of kicking certain Local Hamiltonian instances out of the Karp reductions. That this might be to strict shows the yes-instances is far from having any physical meaning. the consistency problem for local density matrices [26] Instead it results from the fact that a binary string is in- which is not known to be QMA-complete under Karp but terpreted as two unrelated encodings. If a Local Hamil- under randomized Cook reductions. For classes above P tonian yes-instance remains an intermediate problem’s it should be legitimate to consider weaker reduction no- yes-instance depends on the behaviour of certain Tur- tions (e.g. randomized or quantum) as long as they are ing machines on certain inputs, both determined by the “simulable” within this class. Since these reduction no- encoding of the Local Hamiltonian instance and the spe- tions are implied by Karp, we can simply substitute them cific G¨odelnumbering chosen for Turing machines. Due in the formulation of the Uniform Diagonalization Theo- to the large degree of freedom in both encoding schemes, rem. Ladner’s Theorem holds for these reductions as well the holes blown into the Local Hamiltonian problem to as long as the complexity class problems harder than the make it easier are rather artificial than physically mean- supplied problem remain recursively presentable. ingful. Another field of investigation arising from this paper is the structure of non-extremal problems. Several ques- Since recursive presentability of extremal problems tions arise that do not seem trivial to answer. To list also holds for all other complexity classes of promise only some of them: problems introduced in this paper, it follows: • The advantage of our extremality definition is that Corollary 9. The above corollaries 4, 6, 8 and Lad- it allows the recursive presentation of the extremal ner’s Theorem hold accordingly for every pair of the com- problems for standard complexity classes. The dis- plexity classes P, NP, PromiseBPP, PromiseMA, BQP, advantage is that the notion only applies to classes QCMA, QMA and the noisy QMA-variant QMAE with with a machine-based definition (and hence e.g. not to TIM). Alternatively, a problem could be defined E = (Em)m∈N a quantum channel family whose Stine- spring dilation over the field of algebraic numbers is com- as extremal for a complexity class, if the class does putable. not contain any supproblem. Are the such defined Corollaries 5 and 7 also hold for any of the beforemen- extremal problems recursively presentable? tioned classes with a complete problem instead of QMA. • Does there exists a machine-based definition for TIM and hence a reasonable notion of extremality? VI. DISCUSSION • Do there exist non-complete problems that are only decidable by machines whose extremal prob- We adapted the Uniform Diagonalization Theorem and lem is complete? Notice, that if yes, it is in- its most important implications to complexity classes of deed important to state in Ladner’s Theorem that 15

B ∈ QMA∗ \ BQP∗. The weaker statement B ∈ extremal? QMA \ BQP would be meaningless, since B could simply be a nonextremal problem of QMA whose extremal supproblem is as least as difficult as A.

P ACKNOWLEDGMENTS ≤m • We proved recursive presentability of AC and P ≤T AC for recursively presentable complexity classes I thank Tobias J. Osborne for our regular discussions C. Is the analogous set for recursively representable and Alex Grilo for several useful remarks. This work was classes also recursively presentable? supported by the ERC grants QFTCMPS, and SIQS, and • Is every complete problem extremal? through the DFG by the cluster of excellence EXC 201 Quantum FQ Engineering and Space-Time Research, and • Are all totally decidable problems of a class C ∈ C the Research Training Group 1991.

[1] T. Sch¨afer,“The complexity of satisfiability problems,” the local hamiltonian problem,” SIAM Journal on Com- 10th ACM Symposium on Theory of computing, pp. 216– puting, vol. 35, no. 5, pp. 1070–1097, 2006. arXiv:quant- 226, ACM, 1978. ph/0406180. [2] R. E. Ladner, “On the structure of polynomial time re- [17] A. Y. Kitaev, A. H. Shen, and M. N. Vyalyi, Classical ducibility,” JACM, vol. 22, pp. 155–171, Jan. 1975. and Quantum Computation (Graduate Studies in Math- [3] T. Cubitt and A. Montanaro, “Complexity classification ematics), vol. 47. AMS, June 2002. of local hamiltonian problems,” SIAM Journal on Com- [18] C. H. Papadimitriou, Computational Complexity. Addi- puting, vol. 45, no. 2, pp. 268–316, 2016. arXiv:1311.3161. son Wesley Longman, 1994. [4] U. Sch¨oning,“A uniform approach to obtain diagonal [19] S. Arora and B. Barak, Computational Complexity – A sets in complexity classes,” Theoretical Computer Sci- Modern Approach. Cambridge University Press, 2009. ence, vol. 18, pp. 95 – 103, 1982. [20] O. Goldreich, Computational Complexity – A Conecep- [5] J. L. Balczar, J. Diaz, and J. Gabarr´o, Structural Com- tual Perspective. Cambridge University Press, 2008. plexity I. Springer, 1995. [21] A. W. Strzebo´nski,“Computing in the field of complex [6] D. Schmidt, “The recursion-theoretic structure of com- algebraic numbers,” Journal of Symbolic Computation, plexity classes,” Theoretical Computer Science, vol. 38, vol. 24, no. 6, pp. 647 – 656, 1997. pp. 143 – 156, 1985. [22] S. Bravyi, “Efficient Algorithm for a Quantum Analogue [7] H. Vollmer, “The Gap-Language-Technique Revisited,” of 2-SAT.” arXiv:quant-ph/0602108, February 2006. 4th Workshop on Computer Science Logic, pp. 389–399, [23] F. A. Dziemba, “Robustness of QMA against witness Springer, 1991. noise,” Quantum Info. Comput., vol. 17, no. 13&14, [8] K. W. Regan and H. Vollmer, “Gap-languages and log- pp. 1167–1190. arXiv:1611.07332. time complexity classes,” Theoretical Computer Science, [24] R. Downey and L. Fortnow, “Uniformly hard languages,” vol. 188, no. 1, pp. 101 – 116, 1997. Theor. Comput. Sci., vol. 298, pp. 303–315, Apr. 2003. [9] J. Watrous, Quantum Computational Complex- [25] S. Even, A. L. Selman, and Y. Yacobi, “The complex- ity, pp. 7174–7201. Springer New York, 2009. ity of promise problems with applications to public-key arXiv:0804.3401. cryptography,” Information and Control, vol. 61, no. 2, [10] S. A. Cook, “The complexity of theorem-proving proce- pp. 159 – 173, 1984. dures,” 3rd STOC ’71, (New York, USA), pp. 151–158, [26] Y.-K. Liu, Consistency of Local Density Matrices is ACM, 1971. QMA-Complete, pp. 438–449. Springer Berlin Heidel- [11] O. Goldreich, Theoretical Computer Science, ch. On berg, 2006. Promise Problems: A Survey, pp. 254–290. Springer, 2006. [12] S. Bravyi, D. P. DiVincenzo, R. I. Oliveira, and B. M. Terhal, “The Complexity of Stoquastic Local Hamilto- nian Problems,” Quantum Info. Comput., vol. 8, no. 5, pp. 0361–0385, 2008. arXiv:quant-ph/0606140. [13] E. Knill and R. Laflamme, “Quantum computing and quadratically signed weight enumerators,” Information Processing Letters, vol. 79, no. 4, pp. 173 – 179, 2001. arXiv:quant-ph/9909094. [14] S. Gharibian and J. Sikora, “Ground state connectivity of local hamiltonians,” Automata, Languages, and Pro- gramming: 42nd International Colloquium, ICALP 2015, pp. 617–628, Springer, 2015. arXiv:1409.3182. [15] D. Aharonov and T. Naveh, “Quantum NP - A Survey.” arXiv:quant-ph/0210077, October 2002. [16] J. Kempe, A. Kitaev, and O. Regev, “The complexity of