Technical Summary: Quantum Channel Learning with Limited Parallel Access

Problem Statement

This paper addresses the fundamental limits of learning unknown quantum channels under constraints on parallel access. The central task is defined as Channel Learning (Problem II.1): estimating the absolute values of entries in a generalized transfer matrix (or function) that characterizes a quantum channel $E$. This transfer matrix describes the channel's action on an orthogonal operator basis (Heisenberg-Weyl generators for qudits and bosonic modes).

The study focuses on $c$-copy protocols, where a learner has simultaneous access to $c$ copies of the unknown channel $E$ (or the paired resource $E \otimes E^*$, where $E^*$ is the complex-conjugate channel) in each measurement round. The learner may use arbitrary ancillary systems, adaptively prepare input states, and perform adaptive measurements (POVMs) based on past outcomes. The goal is to determine the sample complexity (number of channel uses) required to estimate these transfer matrix entries to an additive accuracy $\epsilon$ with success probability $1-\delta$.

The paper specifically investigates how the availability of the complex-conjugate channel $E^*$ and the number of parallel copies $c$ affect the difficulty of the learning task across three system types:

1. Qubit systems ($d=2$).
2. Qudit systems ($d$-level systems, where $d$ is prime or square-free).
3. Bosonic systems (continuous-variable multimode systems).

Methodology

1. Formalism: Transfer Matrices and Choi States

The authors utilize the Choi-Jamiołkowski isomorphism to reformulate channel learning as the learning of the channel's Choi state.

• Qudits: The channel is characterized by a Displacement Transfer Matrix $C_E$, defined via the expectation values of Heisenberg-Weyl displacement operators $\hat{D}_{d,m}(q, p)$ on the Choi state.
• Bosons: Due to the unbounded nature of bosonic operators, the authors define a TMSV Transfer Function $C^{TMSV,r}_E$ using a two-mode squeezed vacuum (TMSV) state as the probe. This ensures the learning task remains physical and well-defined while still providing a complete description of the channel.

2. The Master Lemma (Lower Bound)

The core theoretical contribution is Lemma III.1, a "Master Lemma" that establishes a general lower bound for the depth $T$ of any $c$-copy learning protocol.

• Hypothesis Testing Framework: The lemma reduces the channel learning problem to a "many-one channel discrimination" task (Problem III.1). The learner must distinguish between a null hypothesis (a replacement channel $E_0$) and a set of alternative hypotheses ($E_{u,v}$) parameterized by random variables.
• Mechanism: The bound relies on the operator norms of specific tensor products of displacement operators. Specifically, it analyzes the quantity $\Delta$, which captures the distinguishability of hypotheses. The bound states that $T = \Omega(1/\Delta)$.
• Generality: This lemma applies to arbitrary input/output Hilbert spaces (finite or infinite dimensional) and allows for arbitrary ancilla assistance and adaptive strategies. It extends previous state-learning lower bounds (e.g., [11–13]) to the more general channel setting.

3. Construction of Hard Channels

To prove the lower bounds, the authors construct specific families of "hard-to-learn" channels (entanglement-breaking channels) that have sparse representations in the transfer matrix basis.

• Qudits: Channels are constructed using sums of displacement operators $\hat{D}_{d,m}(q, p)$ with random parameters.
• Bosons: Channels are constructed using Gaussian mixtures of displacement operators.
• Self-Conjugate Channels: For scenarios without access to $E^*$, the authors construct channels where $E \equiv E^*$ to test if self-conjugacy alone aids learning.

Key Contributions and Results

The paper derives tight sample-complexity bounds for estimating transfer matrix entries, revealing a strict hierarchy of resources required for efficient learning.

1. Access to Complex-Conjugate Channel ($E \otimes E^*$)

• Result: If the learner has access to the complex-conjugate channel $E^*$ (effectively access to $E \otimes E^*$), efficient learning is possible.
• Scaling: The sample complexity scales as $\Theta(\epsilon^{-4})$ (Theorem IV.1, IV.6, IV.7).
• Significance: This holds for qubits, qudits, and bosonic systems. The presence of $E^*$ allows the learner to measure commuting observables efficiently, bypassing the exponential complexity seen in other settings. The $\epsilon^{-4}$ scaling is shown to be tight (matching the upper bound).

2. Limited Parallel Access without $E^*$ (Qudits)

• Result: Without access to $E^*$, the number of copies $c$ is critical.
• Case $c < d$: For $d$-level systems (where $d$ is prime), if the learner has fewer than $d$ copies ($c < d$), the sample complexity is exponential in the number of qudits $m$ (Theorem IV.2). Specifically, $T = \Omega(d^m c^{-2} \epsilon^{-2})$.
• Case $c = d$: Once the learner has access to $d$ copies, the complexity drops to a polynomial scaling of $\Theta(\epsilon^{-2d})$.
• Square-free $d$: This hierarchy extends to square-free integers $d$ (e.g., $d=6$), where $d-1$ copies are insufficient, but $d$ copies enable efficient learning.
• Self-Conjugate Channels: Even if the channel satisfies $E \equiv E^*$, having only 1-copy access ($c=1$) is insufficient for efficient learning; the complexity remains exponential (Theorem IV.4, IV.5). Efficient learning for self-conjugate channels requires 2-copy access.

3. Bosonic Systems

• Result: For bosonic channels, the difficulty persists even with multiple copies if the number of copies is limited relative to the accuracy.
• Scaling: For any $c = O(1/\epsilon)$, the sample complexity remains exponential in the number of modes (Theorem IV.3, IV.5).
• Implication: Unlike the qudit case where increasing $c$ eventually leads to efficiency, bosonic channels exhibit persistent hardness for limited parallel access regimes, requiring exponential resources unless $E^*$ is available.

4. Hierarchy of Learning Resources

The paper establishes a clear hierarchy (visualized in Fig. 4):

• 1-copy hard: General multi-qubit channels and self-conjugate channels require at least 2 copies for efficiency.
• $(d-1)$-copy hard: General multi-qudit channels (local dimension $d$) require $d$ copies for efficiency.
• Always hard (for $c=O(1/\epsilon)$): Multi-mode bosonic channels.
• Efficient with $E \otimes E^*$: All channel types become efficiently learnable with access to the conjugate channel.

Significance and Claims

The authors claim that this work establishes a hierarchy of channel-learning resources, demonstrating that:

1. Channel learning is strictly harder than state learning: While state learning is a special case of channel learning (where the channel is a replacement channel), the ability to probe the channel with arbitrary inputs makes the task more encompassing. The lower bounds derived here are stronger than standard state-learning bounds because the learner controls the input states.
2. The power of the conjugate channel: Access to $E^*$ is a powerful resource that collapses the complexity hierarchy, enabling efficient learning for all system types with a tight $\epsilon^{-4}$ scaling.
3. Dimension-dependent thresholds: For qudits, there is a sharp transition in complexity at $c=d$, driven by the commutativity of observables when $d$ copies are available. This generalizes previous state-learning results to the channel domain.
4. Fundamental limits of parallel access: The results highlight that without specific resources (like $E^*$ or sufficient copies $c \ge d$), learning quantum channels in high-dimensional or continuous-variable systems is fundamentally inefficient, requiring exponential resources.

The paper concludes that while the task is technically "learning a state" (the Choi state), the constraints on channel access create a unique landscape of complexity that differs significantly from standard state tomography, particularly regarding the necessity of specific copy numbers or conjugate access to achieve efficiency.