- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
The paper proposes CRQCL, aiming to control generalization through the coherent information of a learned channel. It constructs a reference-output state and the product of its marginals, relates their relative entropy to , and states single-query and high-probability finite-sample bounds. It identifies the invariance of coherent information under unitary hypotheses and discusses two responses: noisy hypotheses and a separate internal-cut output-entanglement penalty. The actual channel-learning experiments use the latter penalty to fit random Clifford targets with unitary PQCs, using a local fidelity surrogate for larger systems. Evaluation covers n=3,4,5 and n=10,11,12, together with gradient-variance diagnostics and a single-qubit noise example. Reported gap reductions are mixed and statistically insignificant in the smaller-system sweep, but consistent across the larger-system matched runs.
Critical strengths
-
This manuscript has a useful conceptual observation. In Proposition 2, the author explains why coherent information is constant for a fixed probe under unitary hypotheses. This provides a useful warning for designing regularizers for quantum circuits.
-
The research pattern is clear. The larger-system experiments report consistent reductions in the generalization gap, while the paper also acknowledges the mixed results at smaller scales. This provides an initial empirical motivation for the entanglement penalty.
Critical weaknesses
-
The theoretical justification is somehow incomplete. Section 3.2 defines the single-query “true risk” on . This removes correlations between a physical input reference and the channel output. It does not, by itself, represent evaluation on fresh data after learning from a training set. The decoupled risk is not connected to the held-out risk used experimentally.
-
A direct counterexample fits the population-average definition in Section 3.2. Take uniformly sampled computational-basis inputs , with , an identity target, and a trivial reference. A learner can output a controlled unitary that leaves observed unchanged and flips the last qubit for unobserved . With the fixed loss , the empirical risk is zero and the population risk is at least . Every probe has so Theorem 6 instead predicts a gap of at most . The stated theorem may need substantive additional assumptions.
-
Theorem 6 also applies concentration without accounting for the hypothesis’s dependence on its training data. Moreover, the practical penalty lacks a corresponding generalization guarantee.
-
Tables 1–2 report generalization gaps without separate training and test infidelities, so better test accuracy is not established. The limited number of seeds and unclear accounting of target-channel queries also weaken the claims of robustness and query efficiency.
Questions for authors
- What is the precise map from the random training set to the learned hypothesis, and how do the states and represent the training and fresh-data risks used in the experiments?
- Which additional assumption or argument addresses the data dependence in Theorem 6 and excludes the counterexample above?
- What exactly is the probe set in Eq. (1)?
- Is the intended noise model a fixed channel after the full unitary or noise interleaved between parameterized gates? Please explain how the resulting regularizer avoids input-unitary invariance and how it relates to Corollary 7.
- How is the additive term absorbed into when the maximum is zero?
- How do these facts support the stated claim that only a statistical term remains for an anti-degradable target?
- Does the regularizer improve held-out accuracy, and what precisely is counted in the claimed query budget?
5 comments
Response to Reviewer ywG8: Q3
Q3
To resolve the notational overlap with the -query training set , we denote the probe set in Eq. (1) by . For an -qubit channel, is constructively defined as a fixed, -independent ensemble of partially entangled product probes rotated across the three single-qubit Pauli eigenbases:
with a fixed non-uniform Schmidt weight (matching in Section 4.1, Table 3).
Fixing in this manner guarantees three properties. (i) Unlike maximally entangled probes (), the non-uniform spectrum combined with the eigenbases ensures that no unitary commutes with all input marginals or aligns with all noise axes simultaneously. This preserves a non-zero gradient . (ii) Every probe in shares the exact same reference entropy bits. (iii) Eq. (1) remains directly computable during training without continuous state-space optimization.
Response to Reviewer ywG8: Q4
Q4
We address both the physical noise placement in Proposition 3 and how our noisy and noiseless regularizers avoid input-unitary invariance and connect to Corollary 7.
First, regarding noise placement and input-unitary invariance in Eq. (1), for any channel and unitaries , the Stinespring isometry of is . Its output and environment states at input are unitarily equivalent to those of at input , which implies . Consequently, if noise is placed only once after the full unitary (, as in the single-qubit illustrative example of Section 4.1) and one maximizes over all possible input density matrices , then is independent of . In the terminal noise setting, Eq. (1) avoids this invariance by restricting the maximization to the finite, fixed probe set defined in Q3, where rotates the fixed probe marginals relative to the anisotropic noise eigenbasis of . More generally, as stated in Proposition 3 (lines 165–166), realistic gate noise in a depth- PQC acts interleaved per layer rather than only at the terminal output:
Under interleaved noise, unitary invariance removes the dependence on the first layer under a full-space maximization, but the intermediate unitaries do not commute with the preceding non-unitary noise maps , leaving genuine -dependence even when maximizing over all input states . The table below verifies both regimes numerically under per-qubit dephasing at across six Haar-random unitaries per setting:
| Channel | Varied layer | (min / max over 6 unitaries) | Fixed-probe (min / max) | |
|---|---|---|---|---|
| 1 | / | / | ||
| 2 | / | / | ||
| 1 | / | / | ||
| 1 | / | / | ||
| 2 | / | / | ||
| 2 | / | / |
For post-unitary noise, is constant at the dephasing capacity while the fixed-probe values vary; under interleaved noise, is invariant in but varies substantially with the internal layer .
Second, all training experiments in Section 4 use noiseless unitaries trained with in Eq. (2). The regularizer escapes input-unitary invariance because it is not a supremum over all inputs: it is the average output entanglement evaluated on fixed product training queries across the fixed bipartition . Thus, it is invariant only under post-circuit local unitaries , not under entangling circuit layers in . Regarding its relation to Corollary 7, by Definition 4 (lines 193–195), is the coherent information of the reduced erasure channel with subsystem acting as the reference register . Because product queries have zero initial cross-cut entanglement (), the Corollary 7 complexity term evaluated on reduces directly to , as detailed in Q5 below.
Response to Reviewer ywG8: Q1–Q2
Answer to Reviewer ywG8
We thank the reviewer for the careful questions. Our answers follow.
Q1
Let D be a distribution over input queries, and let S={∣ψi⟩}i=1m∼D⊗m be the i.i.d. training sample. The learning algorithm is a regularized map Aλ:S↦W≡θ∗(S). It outputs a hypothesis channel N^W∈HT from a T-gate parameterized quantum circuit (PQC) class by minimizing R^emp(N^θ;S)+λΩ(N^θ). Here Ω(N^θ) is either max∣ψ⟩∈PIc(N^ϵ,θ,∣ψ⟩) (Eq. 1) or the bipartite entangling power χ^(θ) (Eq. 2).
In the bipartite reference-input formulation (Sections 3.1–3.3), take a probe state ∣ψ⟩RA purifying the query input. Then σRA′(N^W,ψ)=(idR⊗N^W)(∣ψ⟩⟨ψ∣RA) is the coupled joint output state, and τRA′(N^W,ψ)=ρR⊗N^W(ρA) is the counterfactual state in which query-specific quantum correlations with R are erased. Theorem 1 bounds their risk deviation: ∣Tr[LτRA′]−Tr[LσRA′]∣≤2(H(ρR)+Ic(N^W,∣ψ⟩)).
In the product-state experiments of Section 4 (Sections 3.4.2 and 3.7), queries are product states ∣ψi⟩=∣ψ1,i⟩A1⊗∣ψ2,i⟩A2 across the balanced cut A=A1⊗A2. Performance is evaluated on the residual ∣ri⟩=U^(θ)†U∣ψi⟩ against ∣ψi⟩ (Appendix A.4). As shown in Definition 4 (lines 193–195), subsystem A2 acts as the internal reference register R for the reduced erasure channel Eθ:A1→A1′, obtained by tracing out A2′. On training queries in R^emp, σA1′A2′(i)=∣ri⟩⟨ri∣ is the coupled output. On fresh product queries from D, the ideal generalizable residual is instead a decoupled product state τA1′A2′(i)=ρA1′(i)⊗ρA2′(i). Penalizing χ^(θ)=m1∑i=1mIc(Eθ,∣ψi⟩) suppresses spurious cross-partition entanglement memorized on S, so the learned unitary generalizes to fresh product inputs.
Q2
Applying a single-hypothesis Hoeffding bound directly to Term (II) =∣ES[R^emp]−R^emp(N^W;S)∣ in Appendix A.3 (lines 625–630) does require accounting for the data dependence of N^W=Aλ(S). In our framework, Aλ restricts the learner to the coherently regularized PQC sublevel set HT,γ={N^θ∈HT:Ω(N^θ)≤γ}. Data dependence in Term (II) is then controlled in one of two ways. The first is uniform replace-one algorithmic stability β(λ) under McDiarmid's bounded-differences inequality (invoked in the Theorem 6 proof sketch, line 239). The second is uniform convergence over HT,γ with effective gate complexity Teff(γ)≤T (Section 3.4.2, line 204; or 2I(S;W)/m in expectation).
Combining Term (I) and Term (II) gives the data-dependent bound, which holds with probability at least 1−δ:
Rtrue(N^W)−R^emp(N^W;S)≤2(H(ρR)+Ic(N^W,∣ψ⟩))+O(mTeff(γ)logTeff(γ)+log(2/δ)).
This bound excludes the lookup-table counterexample (U^S acting as identity on m observed basis states ∣x⟩A1∣0⟩A2 and flipping the ancilla ∣0⟩A2↦∣1⟩A2 on all D−m unobserved states) on two independent fronts. First, in Term (I), U^S is a multi-controlled gate conditioned on x∈/S. Applying it to any superposition product query ∣ψ⟩=∣ψ1⟩A1⊗∣0⟩A2 maximally entangles the ancilla A2 with A1. This generates large cross-cut entanglement χ^(U^S)=SA1(U^S∣ψ⟩)>0, which our regularizer directly penalizes. Second, in Term (II), encoding an arbitrary m-subset out of D requires circuit gate count T=Ω(mlogD) and mutual information I(S;W)=log(mD)=Ω(mlog(D/m)). This violates bounded PQC complexity (T≪m) and algorithmic stability.