Morohoshi, Nakayama, Manabe, and Mitarai proposed a physically motivated quantum machine learning problem in which the goal is to predict quantities of the form $\operatorname{Tr}[f(H)ρ]$ from classical descriptions of a Hamiltonian $H$ and a quantum state $ρ$, where $f$ is an unknown function. We call this problem Hamiltonian function learning in this paper. They constructed an efficient quantum learning algorithm under suitable conditions, while leaving a rigorous proof of average-case classical hardness open. In this paper, we rigorously prove the average-case classical hardness for two distribution-specific Hamiltonian function learning problems for $f_{\cos,π}(λ)=\cos(πλ)$ and $f_{\exp,β}(λ)=e^{-βλ}$ discussed in the paper of Morohoshi et al. under the assumption of the average-case hardness of factoring random RSA moduli. More specifically, we show that an efficient classical randomized learner under squared loss whose output hypotheses are evaluable in classical polynomial time for either problem would yield a classical randomized polynomial-time algorithm for factoring random RSA moduli.
Morohoshi, Nakayama, Manabe, and Mitarai proposed a physically motivated quantum machine learning problem in which the goal is to predict quantities of the form Tr[f(H)ρ] from classical descriptions of a Hamiltonian H and a quantum state ρ, where f is an unknown function. We call this problem Hamiltonian function learning in this paper. They constructed an efficient quantum learning algorithm under suitable conditions, while leaving a rigorous proof of average-case classical hardness open. In this paper, we rigorously prove the average-case classical hardness for two distribution-specific Hamiltonian function learning problems for fcos,π(λ)=cos(πλ) and fexp,β(λ)=e−βλ discussed in the paper of Morohoshi et al. under the assumption of the average-case hardness of factoring random RSA moduli. More specifically, we show that an efficient classical randomized learner under squared loss whose output hypotheses are evaluable in classical polynomial time for either problem would yield a classical randomized polynomial-time algorithm for factoring random RSA moduli.
The most general examples of quantum learning advantages involve data labeled by cryptographic or intrinsically quantum functions, where classical learners are limited by the infeasibility of evaluating the labeling functions using polynomial-sized classical circuits. While broad in scope, such results reveal little about advantages arising from the learning process itself. In cryptographic settings, further insight is possible via random-generatability - the ability to classically generate labeled data - enabling hardness proofs for identification tasks, where the goal is to identify the labeling function from a dataset, even when evaluation is classically intractable. These tasks are particularly relevant in quantum contexts, including Hamiltonian learning and identifying physically meaningful order parameters. However, for quantum functions, random-generatability is conjectured not to hold, leaving no known identification advantages in genuinely quantum regimes. In this work, we give the first proofs of quantum identification learning advantages under standard complexity assumptions. We confirm that quantum-hard functions are not random-generatable unless BQP is contained in the second level of the polynomial hierarchy, ruling out cryptographic-style data generation strategies. We then introduce a new approach: we show that verifiable identification - solving the identification task for valid datasets while rejecting invalid ones - is classically hard for quantum labeling functions unless BQP is in the polynomial hierarchy. Finally, we show that, for a broad class of tasks, solving the identification problem implies verifiable identification in the polynomial hierarchy. This yields our main result: a natural class of quantum identification tasks solvable by quantum learners but hard for classical learners unless BQP is in the polynomial hierarchy.
The existence of one-way functions (OWFs) forms the minimal assumption in classical cryptography. However, this is not necessarily the case in quantum cryptography. One-way puzzles (OWPuzzs), introduced by Khurana and Tomer, provide a natural quantum analogue of OWFs. The existence of OWPuzzs implies PP=BQP, while the converse remains open. In classical cryptography, the analogous problem-whether OWFs can be constructed from P=NP-has long been studied from the viewpoint of hardness of learning. Hardness of learning in various frameworks (including PAC learning) has been connected to OWFs or to P=NP. In contrast, no such characterization previously existed for OWPuzzs. In this paper, we establish the first complete characterization of OWPuzzs based on the hardness of a well-studied learning model: distribution learning. Specifically, we prove that OWPuzzs exist if and only if proper quantum distribution learning is hard on average. A natural question that follows is whether the worst-case hardness of proper quantum distribution learning can be derived from PP=BQP. If so, and a worst-case to average-case hardness reduction is achieved, it would imply OWPuzzs solely from PP=BQP. However, we show that this would be extremely difficult: if worst-case hardness is PP-hard (in a black-box reduction), then SampBQP=SampBPP follows from the infiniteness of the polynomial hierarchy. Despite that, we show that PP=BQP is equivalent to another standard notion of hardness of learning: agnostic. We prove that PP=BQP if and only if agnostic quantum distribution learning with respect to KL divergence is hard. As a byproduct, we show that hardness of agnostic quantum distribution learning with respect to statistical distance against PPTΣ3P learners implies SampBQP=SampBPP.
Understanding the computational complexity of learning efficient classical programs in various learning models has been a fundamental and important question in classical computational learning theory. In this work, we study the computational complexity of quantum state learning, which can be seen as a quantum generalization of distributional learning introduced by Kearns et.al [STOC94]. Previous works by Chung and Lin [TQC21], and Bădescu and O′Donnell [STOC21] study the sample complexity of the quantum state learning and show that polynomial copies are sufficient if unknown quantum states are promised efficiently generatable. However, their algorithms are inefficient, and the computational complexity of this learning problem remains unresolved. In this work, we study the computational complexity of quantum state learning when the states are promised to be efficiently generatable. We show that if unknown quantum states are promised to be pure states and efficiently generateable, then there exists a quantum polynomial time algorithm A and a language L∈PP such that AL can learn its classical description. We also observe the connection between the hardness of learning quantum states and quantum cryptography. We show that the existence of one-way state generators with pure state outputs is equivalent to the average-case hardness of learning pure states. Additionally, we show that the existence of EFI implies the average-case hardness of learning mixed states.
Despite significant effort, the quantum machine learning community has only demonstrated quantum learning advantages for artificial cryptography-inspired datasets when dealing with classical data. In this paper we address the challenge of finding learning problems where quantum learning algorithms can achieve a provable exponential speedup over classical learning algorithms. We reflect on computational learning theory concepts related to this question and discuss how subtle differences in definitions can result in significantly different requirements and tasks for the learner to meet and solve. We examine existing learning problems with provable quantum speedups and find that they largely rely on the classical hardness of evaluating the function that generates the data, rather than identifying it. To address this, we present two new learning separations where the classical difficulty primarily lies in identifying the function generating the data. Furthermore, we explore computational hardness assumptions that can be leveraged to prove quantum speedups in scenarios where data is quantum-generated, which implies likely quantum advantages in a plethora of more natural settings (e.g., in condensed matter and high energy physics). We also discuss the limitations of the classical shadow paradigm in the context of learning separations, and how physically-motivated settings such as characterizing phases of matter and Hamiltonian learning fit in the computational learning framework.
Given that quantum computers are naturally suited to simulate the behavior of quantum many-body systems, an immediate question arises: can one formulate physically motivated quantum machine learning (QML) tasks that exhibit learning separations? We address this problem by studying the learnability of quantum many-body dynamics from the perspective of probably approximately correct (PAC)-learning. Concretely, we devise a supervised learning problem where the training set consists of specifications of randomized stabilizer probe states, evolution times sampled uniformly from a polynomially large time interval [0,T], coupled with expectation values of certain observables evaluated on the resulting time-evolved state under an unknown Hamiltonian. For this learning task, we provide an efficient quantum procedure whose training phase learns the underlying Hamiltonian from short-time training samples, and whose deployment phase combines Hamiltonian simulation with the classical shadows protocol to perform inference on a newly given data point. By contrast, the existence of O(poly(n))-time instances ensures classical hardness: by embedding a BQP-complete computation into the polynomially long time-dynamics of a low-intersection variant of the Feynman-Kitaev clock Hamiltonian construction, we show that, for a certain family of input distributions, no randomized classical polynomial-time algorithm can fulfill our learning condition, unless BQP⊆P/poly. Furthermore, we show that the classically hard instance maintains quantum learnability. We also give an interpretation of our results in learning-assisted certified quantum simulation. Taken together, our results demonstrate a rigorous learning separation for a natural ML task based on Hamiltonian evolution, while building connections between quantum learning theory, quantum simulation, and QML.
Our main result is a reduction from worst-case lattice problems such as GapSVP and SIVP to a certain learning problem. This learning problem is a natural extension of the `learning from parity with error' problem to higher moduli. It can also be viewed as the problem of decoding from a random linear code. This, we believe, gives a strong indication that these problems are hard. Our reduction, however, is quantum. Hence, an efficient solution to the learning problem implies a quantum algorithm for GapSVP and SIVP. A main open question is whether this reduction can be made classical (i.e., non-quantum). We also present a (classical) public-key cryptosystem whose security is based on the hardness of the learning problem. By the main result, its security is also based on the worst-case quantum hardness of GapSVP and SIVP. The new cryptosystem is much more efficient than previous lattice-based cryptosystems: the public key is of size O~(n2) and encrypting a message increases its size by a factor of O~(n) (in previous cryptosystems these values are O~(n4) and O~(n2), respectively). In fact, under the assumption that all parties share a random bit string of length O~(n2), the size of the public key can be reduced to O~(n).