In this paper, we study the relationship between advice in the form of a training set and classical advice. We do this by analyzing the class $\mathsf{BPP/samp}$ and certain variants of it. Specifically, our main result demonstrates that $\mathsf{BPP/samp}$ is a proper subset of the class $\mathsf{P/poly}$, which implies that advice in the form of a training set is strictly weaker than classical advice. This result remains valid when considering quantum advice and a quantum generalization of the training set. Finally, leveraging the insights from our proofs, we identify both sufficient and necessary complexity-theoretic assumptions for the existence of concept classes that exhibit a quantum learning speed-up. We consider both the worst-case setting, where accurate results are required for all inputs, and the average-case setting.
In this paper, we study the relationship between advice in the form of a training set and classical advice. We do this by analyzing the class BPP/samp and certain variants of it. Specifically, our main result demonstrates that BPP/samp is a proper subset of the class P/poly, which implies that advice in the form of a training set is strictly weaker than classical advice. This result remains valid when considering quantum advice and a quantum generalization of the training set. Finally, leveraging the insights from our proofs, we identify both sufficient and necessary complexity-theoretic assumptions for the existence of concept classes that exhibit a quantum learning speed-up. We consider both the worst-case setting, where accurate results are required for all inputs, and the average-case setting.
We prove the following surprising result: given any quantum state ρ on n qubits, there exists a local Hamiltonian H on poly(n) qubits (e.g., a sum of two-qubit interactions), such that any ground state of H can be used to simulate ρ on all quantum circuits of fixed polynomial size. In terms of complexity classes, this implies that BQP/qpoly⊆QMA/poly, which supersedes the previous result of Aaronson that BQP/qpoly⊆PP/poly. Indeed, we can exactly characterize quantum advice as equivalent in power to untrusted quantum advice combined with trusted classical advice. Proving our main result requires combining a large number of previous tools---including a result of Alon et al. on learning of real-valued concept classes, a result of Aaronson on the learnability of quantum states, and a result of Aharonov and Regev on “QMA+ super-verifiers''---and also creating some new ones. The main new tool is a so-called majority-certificates lemma, which is closely related to boosting in machine learning, and which seems likely to find independent applications. In its simplest version, this lemma says the following. Given any set S of Boolean functions on n variables, any function f∈S can be expressed as the pointwise majority of m=O(n) functions f1,…,fm∈S, such that each fi is the unique function in S compatible with O(log∣S∣) input/output constraints.
In learning theory, the VC dimension of a concept class C is the most common way to measure its “richness.” A fundamental result says that the number of examples needed to learn an unknown target concept c∈C under an unknown distribution D, is tightly determined by the VC dimension d of the concept class C. Specifically, in the PAC model
Θ(dϵ+log(1/δ)ϵ)
examples are necessary and sufficient for a learner to output, with probability 1−δ, a hypothesis h that is ϵ-close to the target concept c (measured under D). In the related agnostic model, where the samples need not come from a c∈C, we know that
Θ(dϵ2+log(1/δ)ϵ2)
examples are necessary and sufficient to output an hypothesis h∈C whose error is at most ϵ worse than the error of the best concept in C. Here we analyze quantum sample complexity, where each example is a coherent quantum state. This model was introduced by Bshouty and Jackson (1999), who showed that quantum examples are more powerful than classical examples in some fixed-distribution settings. However, Atici and Servedio (2005), improved by Zhang (2010), showed that in the PAC setting (where the learner has to succeed for every distribution), quantum examples cannot be much more powerful: the required number of quantum examples is
Ω(d1−ηϵ+d+log(1/δ)ϵ) for arbitrarily small constant η>0.
Our main result is that quantum and classical sample complexity are in fact equal up to constant factors in both the PAC and agnostic models. We give two proof approaches. The first is a fairly simple information-theoretic argument that yields the above two classical bounds and yields the same bounds for quantum sample complexity up to a log(d/ϵ) factor. We then give a second approach that avoids the log-factor loss, based on analyzing the behavior of the “Pretty Good Measurement” on the quantum state-identification problems that correspond to learning. This shows classical and quantum sample complexity are equal up to constant factors for every concept class C.
We prove the following surprising result: given any quantum state rho on n qubits, there exists a local Hamiltonian H on poly(n) qubits (e.g., a sum of two-qubit interactions), such that any ground state of H can be used to simulate rho on all quantum circuits of fixed polynomial size. In terms of complexity classes, this implies that BQP/qpoly is contained in QMA/poly, which supersedes the previous result of Aaronson that BQP/qpoly is contained in PP/poly. Indeed, we can exactly characterize quantum advice, as equivalent in power to untrusted quantum advice combined with trusted classical advice. Proving our main result requires combining a large number of previous tools -- including a result of Alon et al. on learning of real-valued concept classes, a result of Aaronson on the learnability of quantum states, and a result of Aharonov and Regev on "QMA+ super-verifiers" -- and also creating some new ones. The main new tool is a so-called majority-certificates lemma, which is closely related to boosting in machine learning, and which seems likely to find independent applications. In its simplest version, this lemma says the following. Given any set S of Boolean functions on n variables, any function f in S can be expressed as the pointwise majority of m=O(n) functions f1,...,fm in S, such that each fi is the unique function in S compatible with O(log|S|) input/output constraints.
A basic question in the PAC model of learning is whether proper learning is harder than improper learning. In the classical case, there are examples of concept classes with VC dimension d that have sample complexity Ω(εdlogε1) for proper learning with error ε, while the complexity for improper learning is O(εd). One such example arises from the Coupon Collector problem. Motivated by the efficiency of proper versus improper learning with quantum samples, Arunachalam, Belovs, Childs, Kothari, Rosmanis, and de Wolf (TQC 2020) studied an analogue, the Quantum Coupon Collector problem. Curiously, they discovered that for learning size k subsets of [n] the problem has sample complexity Θ(klogmin{k,n−k+1}), in contrast with the complexity of Θ(klogk) for Coupon Collector. This effectively negates the possibility of a separation between the two modes of learning via the quantum problem, and Arunachalam et al.\ posed the possibility of such a separation as an open question. In this work, we first present an algorithm for the Quantum Coupon Collector problem with sample complexity that matches the sharper lower bound of (1−ok(1))klnmin{k,n−k+1} shown recently by Bab Hadiashar, Nayak, and Sinha (IEEE TIT 2024), for the entire range of the parameter k. Next, we devise a variant of the problem, the Quantum Padded Coupon Collector. We prove that its sample complexity matches that of the classical Coupon Collector problem for both modes of learning, thereby exhibiting the same asymptotic separation between proper and improper quantum learning as mentioned above. The techniques we develop in the process can be directly applied to any form of padded quantum data. We hope that padding can more generally lift other forms of classical learning behaviour to the quantum setting.
Despite years of effort, the quantum machine learning community has only been able to show quantum learning advantages for certain contrived cryptography-inspired datasets in the case of classical data. In this note, we discuss the challenges of finding learning problems that quantum learning algorithms can learn much faster than any classical learning algorithm, and we study how to identify such learning problems. Specifically, we reflect on the main concepts in computational learning theory pertaining to this question, and we discuss how subtle changes in definitions can mean conceptually significantly different tasks, which can either lead to a separation or no separation at all. Moreover, we study existing learning problems with a provable quantum speedup to distill sets of more general and sufficient conditions (i.e., ``checklists'') for a learning problem to exhibit a separation between classical and quantum learners. These checklists are intended to streamline one's approach to proving quantum speedups for learning problems, or to elucidate bottlenecks. Finally, to illustrate its application, we analyze examples of potential separations (i.e., when the learning problem is build from computational separations, or when the data comes from a quantum experiment) through the lens of our approach.
$ \newcommand{\eps}{\varepsilon} $In learning theory, the VC dimension of a concept class C is the most common way to measure its "richness." In the PAC model $$ Θ\Big(\frac{d}{\eps} + \frac{\log(1/δ)}{\eps}\Big) $$ examples are necessary and sufficient for a learner to output, with probability 1−δ, a hypothesis h that is $\eps$-close to the target concept c. In the related agnostic model, where the samples need not come from a c∈C, we know that $$ Θ\Big(\frac{d}{\eps^2} + \frac{\log(1/δ)}{\eps^2}\Big) $$ examples are necessary and sufficient to output an hypothesis h∈C whose error is at most $\eps$ worse than the best concept in C. Here we analyze quantum sample complexity, where each example is a coherent quantum state. This model was introduced by Bshouty and Jackson, who showed that quantum examples are more powerful than classical examples in some fixed-distribution settings. However, Atici and Servedio, improved by Zhang, showed that in the PAC setting, quantum examples cannot be much more powerful: the required number of quantum examples is $$ Ω\Big(\frac{d^{1-η}}{\eps} + d + \frac{\log(1/δ)}{\eps}\Big)\mbox{ for all }η> 0. $$ Our main result is that quantum and classical sample complexity are in fact equal up to constant factors in both the PAC and agnostic models. We give two approaches. The first is a fairly simple information-theoretic argument that yields the above two classical bounds and yields the same bounds for quantum sample complexity up to a $\log(d/\eps)$ factor. We then give a second approach that avoids the log-factor loss, based on analyzing the behavior of the "Pretty Good Measurement" on the quantum state identification problems that correspond to learning. This shows classical and quantum sample complexity are equal up to constant factors.