Quantum PAC Learning Overview

Updated 7 February 2026

Quantum PAC Learning is the extension of classical PAC learning to quantum objects, defining learning problems for states, channels, and measurements.
It formalizes sample complexity and computational efficiency using quantum-specific metrics like fat-shattering dimensions and trace distance bounds.
The framework integrates statistical learning, quantum information, and cryptographic principles to tackle challenges in security, privacy, and computational hardness.

Quantum PAC Learning is the extension of the Probably Approximately Correct (PAC) learning framework to quantum information-processing tasks, encompassing the learning of quantum states, processes (channels), and measurement devices. Quantum PAC learning formalizes sample complexity and computational efficiency bounds for identifying unknown quantum operations based on observed input-output behavior, generalizing classical notions of concept classes and hypothesis testing to noncommutative settings involving density matrices, channels, and positive-operator-valued measures (POVMs). The field synthesizes statistical learning theory, quantum information, computational complexity theory, and aspects of quantum tomography.

1. Quantum PAC Learning: Model Definitions

Quantum PAC learning generalizes the classical PAC model by replacing classical concepts with quantum objects such as channels, states, or measurements, and by defining the sample space in terms of quantum preparations and measurements.

Formal Model:

Quantum Channel Learning: Let $C$ be a finite class of quantum channels $c:\mathcal D(\mathbb C^{d_1})\to\mathcal D(\mathbb C^{d_2})$ (completely positive, trace-preserving maps). A learner receives $T$ i.i.d. input-output pairs $(x_i, c^*(x_i))$ , where $x_i$ are pure input states sampled from an unknown distribution $D$ , and $c^*(x_i)$ are the corresponding output density matrices from the unknown target channel $c^*$ . The goal is to output a hypothesis $h\in C$ such that the expected trace distance $\Delta_D(h,c^*)\le\epsilon$ with probability at least $1-\delta$ (Chung et al., 2018).
Quantum PAC State Learning: For an unknown $n$ -qubit quantum state $\rho$ , drawn measurements $E_i$ from an unknown distribution $D$ yield observations $y_i = \mathrm{Tr}(E_i \rho)$ . The learner must output $\sigma$ so that for a new test $E$ , $|\mathrm{Tr}(E \sigma) - \mathrm{Tr}(E\rho)|\le\epsilon$ with high probability (Rocchetto, 2017).
Quantum PAC Learning for POVMs: Given a class $\mathcal H$ of two-outcome POVMs and training pairs (quantum state, label), the task is to select $h\in\mathcal H$ minimizing classification risk on fresh data (Magner et al., 2023).
Distributional Quantum PAC Learning: Given sample access to a quantum process or distribution, the learner outputs a quantum/classical generator (or evaluator) closely approximating the unknown target distribution under an appropriate metric, e.g., total variation distance (Hinsche et al., 2021, Sweke et al., 2020).

Key Quantitative Measures:

Sample Complexity: The minimal number of quantum examples (copies or uses of the process) required to achieve $(\epsilon,\delta)$ -PAC accuracy for all possible targets and input distributions.
Computational Efficiency: Whether a learning algorithm runs in poly( $n,1/\epsilon,1/\delta$ ) time (where $n$ is the relevant system size).

2. Sample Complexity and Information-Theoretic Bounds

Quantum PAC learning satisfies sample complexity bounds analogous to classical PAC, but quantum-specific phenomena arise.

Scenario	Sample Complexity	Notes
Pure-state quantum channels	$O((\ln\|C\|+\ln(1/\delta))/\epsilon^2)$	Distance amplification + maximum-likelihood
Mixed-state quantum channels	$O(\ln^3\|C\|(\ln\|C\|+\ln(1/\delta))/\epsilon^2)$	BSD/PGM; partition-and-prune
General quantum concepts (VC dim $d$ )	$\Theta((d+\log(1/\delta))/\epsilon)$	Match classical up to constants

The optimal PAC sample complexity is dictated by the VC dimension (for classical functions) or, in quantum settings, fat-shattering dimension and covering numbers of operator-valued function classes (Magner et al., 2023, Caro et al., 2023).

Information-Theoretic Techniques:

Lower bounds employ reductions to quantum state identification using the pretty-good measurement (PGM) and mutual/Holevo information arguments (Arunachalam et al., 2016, Hadiashar et al., 2023).
These show that quantum examples cannot drive down the distribution-independent sample complexity below the classical $\Theta(d/\epsilon)$ or $\Theta(d/\epsilon^2)$ thresholds, except in special query or fixed-distribution models.

3. Computational Complexity and Hardness Results

While sample efficiency is achievable in quantum PAC learning, computational tractability can diverge sharply for different concept classes.

Efficient (Proper) Learning:

Stabilizer States: The class of $n$ -qubit stabilizer states is both sample- and polynomial-time efficiently PAC-learnable; Pauli measurement data suffices, with linear algebraic reconstruction (Rocchetto, 2017).
Clifford Circuits: Proper PAC-learning of CNOT/Clifford circuits is feasible in sample complexity but is NP-hard for classical randomized algorithms, and quantumly only possible if NP $\subseteq$ RQP (Liang, 2022).
Parametric Quantum Circuits: Circuits with polynomially many gates can be PAC-learned (improperly, via variational ansatz), with sample complexity scaling $\widetilde{O}(n^{c+1}/\epsilon^2)$ (Cai et al., 2021).

Hardness:

Free-Fermionic States: PAC-learning (proper, i.e., Gaussian hypothesis) is NP-hard, due to the nonlinear Pfaffian constraints governing consistency of empirical correlation data with a Gaussian state (Bittel et al., 2024).
Shallow Classical Circuits: Quantum PAC learners for classes such as $\mathrm{AC}^0$ , $\mathrm{TC}^0$ , or $\mathrm{TC}_2^0$ are impossible under standard post-quantum cryptographic assumptions (LWE, RLWE), mirroring classical hardness barriers (Arunachalam et al., 2019).
Clifford Circuits: No poly-time proper PAC learning unless RP=NP (classical) or NP $\subseteq$ RQP (quantum), even though the fat-shattering dimension is $O(n^2)$ and sample complexity is small (Liang, 2022).

4. Learning Quantum Measurement and Process Classes

Quantum PAC learning can encompass classes of measurement devices (POVMs) and more general quantum processes.

Fat-Shattering and Joint Measurability: For quantum measurements, PAC learnability is characterized by finite fat-shattering dimension (probabilistic separation on quantum states) and approximate finite partitionability into sets of jointly measurable POVMs. ERM fails due to measurement disturbance, but a denoised ERM procedure (DERM) achieves optimal sample complexity (Magner et al., 2023).
Quantum Empirical Risk Minimization (ERM): With compatible POVM classes, empirical risk minimization achieves sample complexity $O((\log|\mathcal C|+\log(1/\delta))/\epsilon^2)$ , mirroring classical results, but sample reuse among incompatible hypotheses is limited by non-commutativity (Heidari et al., 2021).
Information-Theoretic Generalization: Recent frameworks give explicit generalization error bounds for quantum learners in terms of both classical and quantum mutual information between training data and learned hypothesis, underpinning generalization in state discrimination, parameter estimation, and process learning scenarios (Caro et al., 2023).

5. Quantum PAC Learning of Distributions and Generative Tasks

PAC Distribution Learning: For classes such as Born distributions of local quantum circuits, strong separations are established between what can be PAC-learned with sample access versus statistical queries, and between classical and shallow quantum generators (Hinsche et al., 2021, Pirnay et al., 2024).
- Clifford Circuits: PAC-learnable by classical algorithms with $O(n)$ samples in the sample-model, but not in the statistical-query model for super-logarithmic depth.
- Constant-Depth Quantum vs Classical Circuits: Unconditional PAC distribution-learning advantage exists: for certain target distributions, depth- $O(1)$ quantum circuits (QNC $^0$ ) can achieve small total variation distance, but any comparable-depth classical circuit (NC $^0$ ) is bounded away from this, with explicit separation gap (Pirnay et al., 2024).
Cryptographic Hardness and Quantum Advantage: Under standard cryptographic assumptions (e.g., DDH for PRFs), there are families of distributions which are not efficiently PAC-learnable classically, but can be reconstructed by efficient quantum algorithms with only a single sample when provided access to quantum subroutines (e.g., discrete logarithm) (Sweke et al., 2020).

6. Security, Privacy, and Advanced Variants

Quantum Secure PAC Learning: Security notions are formalized via upper and lower sample bounds. The no-broadcasting theorem enables protocols where any eavesdropping or sample duplication is detected through induced disturbance. This can be operationalized as a sample budget interval: legitimate learners succeed, while adversaries who must induce more disturbance require more samples than are available (Song et al., 2019, Bang, 4 Nov 2025).
- The security condition is physically rooted in quantum information principles (e.g., Holevo bounds) and cannot be obtained classically.
- Stopping and certification rules naturally integrate runtime channel statistics (such as quantum bit error rate) and authentication via basis sifting.
Quantum Statistical Query (QSQ) Model: The QSQ model restricts access to expectation estimation of quantum examples, making it a quantum analog of the classical SQ model. It admits quantum advantages for learnability relative to tolerance and SQ-dimension, provides privacy guarantees, and yields tight sample complexity bounds for classes like parities and DNF (Arunachalam et al., 2020).

7. Open Problems and Outlook

Separation Regimes: Despite information-theoretic equivalence in sample complexity (up to constants) for many classes, efficient and proper quantum PAC learning (in polynomial time) is provably hard for a wide range of concept classes unless classical complexity collapses.
Improvements under Structure or Query Access: Provable quantum advantages persist for specific tasks or access models (e.g., equivalence queries with oracular access enable quadratic speedup in $\epsilon$ for generic VC classes (Salmon et al., 2023)).
Generalization and Unification: The unifying information-theoretic perspective connects risk, mutual information, and privacy for both classical and quantum regimes, with research converging on robust sample and generalization guarantees under quantum data constraints (Caro et al., 2023).
Practical Implications: The distinction between sample, computational, and security efficiency, and between proper and improper PAC learning, is a driving theme in current work, influencing the design of quantum learning algorithms for state tomography, process discrimination, measurement learning, and generative modeling.

In summary, Quantum PAC Learning establishes a rigorous theoretical framework for learning in quantum information settings, characterizing the sample and computational complexity for a spectrum of learning problems involving quantum states, operations, and measurements. While quantum examples do not generically reduce distribution-independent sample complexity below classical rates, quantum protocols can achieve provable advantages under structural, cryptographic, or resource-theoretic conditions, with novel security and privacy properties inherently unavailable in classical PAC learning. Recent advances continue to refine the understanding of fat-shattering, generalization, and hardness, setting a unified stage for quantum learning theory across domains (Chung et al., 2018, Arunachalam et al., 2016, Magner et al., 2023, Caro et al., 2023, Liang, 2022, Heidari et al., 2021).