acceptodds
Under review as a conference paper at ICLR 2027

An Information Bottleneck for Learning Quantum Channels via Coherent Information

Abstract

Generalization theory for learning from quantum data has recently been unified under an information theoretic lens with bounds controlling the generalization gap via the mutual information between the hypothesis of a learner and its training data. However, extension of such a framework to dynamic quantum channels plausibly the more practically relevant learning target. Given the ubiquity of parameterized quantum circuits as hypothesis classes for unknown unitary dynamics, it has remained an open problem, and existing bounds for this setting rely on abstract quantum mutual information terms that cannot be measured or actively controlled during training. In this paper, we introduce a framework, Coherently Regularized Quantum Channel learning that replaces this unmeasurable quantity with the coherent information of the hypothesis channel, a defined estimable capacity metric. Using a decoupled reference state construction and quantum optimal transport techniques, and Petz's variational for quantum relative entropy, we prove a high probability generation bound. These bounds confirm that the true risk of a learned hypothesis channel is controlled by its empirical risk and a scaling with its maximum coherent information. Maximum coherent information is measurable, resulting a training time regularizer that acts as an information bottleneck, perturbing hypothesis channels from encoding spurious, query-specific quantum correlations. We validate our approach for learning unknown random Clifford circuits under a restricted query budget, which showcases that coherently regularized training yields a smaller generalization gap than unregularized empirical risk minimization as a function of the gap tracking, and that shallow target circuits are learnable from substantially fewer queries.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Mutual Information, Quantum Channel, empirical risk

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_ywG8OpenReview ↗
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 H(ρR)+IcH(\rho_R)+I_c, 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
  1. 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.

  2. 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
  1. The theoretical justification is somehow incomplete. Section 3.2 defines the single-query “true risk” on ρR⊗N^(ρA)\rho_R\otimes\widehat{\mathcal N}(\rho_A). This removes correlations between a physical input reference and the channel output. It does not, by itself, represent evaluation on fresh data after learning N^\widehat{\mathcal N} from a training set. The decoupled risk is not connected to the held-out risk used experimentally.

  2. A direct counterexample fits the population-average definition in Section 3.2. Take uniformly sampled computational-basis inputs ∣x⟩∣0⟩|x\rangle|0\rangle, with x∈1,…,Dx\in{1,\ldots,D}, an identity target, and a trivial reference. A learner can output a controlled unitary that leaves observed xx unchanged and flips the last qubit for unobserved xx. With the fixed loss L=I⊗∣1⟩⟨1∣L=I\otimes|1\rangle\langle1|, the empirical risk is zero and the population risk is at least 1−m/D1-m/D. Every probe has H(ρR)+Ic=0H(\rho_R)+I_c=0 so Theorem 6 instead predicts a gap of at most 2log⁡(2/δ)/m\sqrt{2\log(2/\delta)/m}. The stated theorem may need substantive additional assumptions.

  3. Theorem 6 also applies concentration without accounting for the hypothesis’s dependence on its training data. Moreover, the practical penalty χ^\hat\chi lacks a corresponding generalization guarantee.

  4. 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
  1. What is the precise map from the random training set to the learned hypothesis, and how do the states σ\sigma and τ\tau represent the training and fresh-data risks used in the experiments?
  2. Which additional assumption or argument addresses the data dependence in Theorem 6 and excludes the counterexample above?
  3. What exactly is the probe set S\mathcal S in Eq. (1)?
  4. 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.
  5. How is the additive H(ρR)H(\rho_R) term absorbed into κmax⁡ψIc\kappa\max_\psi I_c when the maximum is zero?
  6. How do these facts support the stated claim that only a statistical term remains for an anti-degradable target?
  7. Does the regularizer improve held-out accuracy, and what precisely is counted in the claimed query budget?
5 comments
AuthorsOpenReview ↗

Response to Reviewer ywG8: Q1–Q2

Answer to Reviewer ywG8

We thank the reviewer for the careful questions. Our answers follow.

Q1

Let D\mathcal{D} be a distribution over input queries, and let S={∣ψi⟩}i=1m∼D⊗mS = \lbrace|\psi_i\rangle\rbrace_{i=1}^m \sim \mathcal{D}^{\otimes m} be the i.i.d. training sample. The learning algorithm is a regularized map Aλ:S↦W≡θ∗(S)\mathcal{A}_\lambda : S \mapsto W \equiv \theta^*(S). It outputs a hypothesis channel N^W∈HT\hat{\mathcal{N}}_W \in \mathcal{H}_T from a TT-gate parameterized quantum circuit (PQC) class by minimizing R^emp(N^θ;S)+λΩ(N^θ)\hat{R}_{\mathrm{emp}}(\hat{\mathcal{N}}_\theta; S) + \lambda\Omega(\hat{\mathcal{N}}_\theta). Here Ω(N^θ)\Omega(\hat{\mathcal{N}}_\theta) is either max⁡∣ψ⟩∈PIc(N^ϵ,θ,∣ψ⟩)\max_{|\psi\rangle \in \mathcal{P}} I_c(\hat{\mathcal{N}}_{\epsilon,\theta}, |\psi\rangle) (Eq. 1) or the bipartite entangling power χ^(θ)\hat{\chi}(\theta) (Eq. 2).

In the bipartite reference-input formulation (Sections 3.1–3.3), take a probe state ∣ψ⟩RA|\psi\rangle_{RA} purifying the query input. Then σRA′(N^W,ψ)=(idR⊗N^W)(∣ψ⟩⟨ψ∣RA)\sigma_{RA'}(\hat{\mathcal{N}}_W, \psi) = (\mathrm{id}_R \otimes \hat{\mathcal{N}}_W)(|\psi\rangle\langle\psi|_{RA}) is the coupled joint output state, and τRA′(N^W,ψ)=ρR⊗N^W(ρA)\tau_{RA'}(\hat{\mathcal{N}}_W, \psi) = \rho_R \otimes \hat{\mathcal{N}}_W(\rho_A) is the counterfactual state in which query-specific quantum correlations with RR are erased. Theorem 1 bounds their risk deviation: ∣Tr[LτRA′]−Tr[LσRA′]∣≤2(H(ρR)+Ic(N^W,∣ψ⟩))|\mathrm{Tr}[L\tau_{RA'}] - \mathrm{Tr}[L\sigma_{RA'}]| \le \sqrt{2(H(\rho_R) + I_c(\hat{\mathcal{N}}_W, |\psi\rangle))}.

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|\psi_i\rangle = |\psi_{1,i}\rangle_{A_1} \otimes |\psi_{2,i}\rangle_{A_2} across the balanced cut A=A1⊗A2A = A_1 \otimes A_2. Performance is evaluated on the residual ∣ri⟩=U^(θ)†U∣ψi⟩|r_i\rangle = \hat{U}(\theta)^\dagger U|\psi_i\rangle against ∣ψi⟩|\psi_i\rangle (Appendix A.4). As shown in Definition 4 (lines 193–195), subsystem A2A_2 acts as the internal reference register RR for the reduced erasure channel Eθ:A1→A1′\mathcal{E}_\theta : A_1 \to A_1', obtained by tracing out A2′A_2'. On training queries in R^emp\hat{R}_{\mathrm{emp}}, σA1′A2′(i)=∣ri⟩⟨ri∣\sigma_{A_1' A_2'}^{(i)} = |r_i\rangle\langle r_i| is the coupled output. On fresh product queries from D\mathcal{D}, the ideal generalizable residual is instead a decoupled product state τA1′A2′(i)=ρA1′(i)⊗ρA2′(i)\tau_{A_1' A_2'}^{(i)} = \rho_{A_1'}^{(i)} \otimes \rho_{A_2'}^{(i)}. Penalizing χ^(θ)=1m∑i=1mIc(Eθ,∣ψi⟩)\hat{\chi}(\theta) = \frac{1}{m}\sum_{i=1}^m I_c(\mathcal{E}_\theta, |\psi_i\rangle) suppresses spurious cross-partition entanglement memorized on SS, 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)∣= |\mathbb{E}_S[\hat{R}_{\mathrm{emp}}] - \hat{R}_{\mathrm{emp}}(\hat{\mathcal{N}}_W; S)| in Appendix A.3 (lines 625–630) does require accounting for the data dependence of N^W=Aλ(S)\hat{\mathcal{N}}_W = \mathcal{A}_\lambda(S). In our framework, Aλ\mathcal{A}_\lambda restricts the learner to the coherently regularized PQC sublevel set HT,γ={N^θ∈HT:Ω(N^θ)≤γ}\mathcal{H}_{T,\gamma} = \lbrace\hat{\mathcal{N}}_\theta \in \mathcal{H}_T : \Omega(\hat{\mathcal{N}}_\theta) \le \gamma\rbrace. Data dependence in Term (II) is then controlled in one of two ways. The first is uniform replace-one algorithmic stability β(λ)\beta(\lambda) under McDiarmid's bounded-differences inequality (invoked in the Theorem 6 proof sketch, line 239). The second is uniform convergence over HT,γ\mathcal{H}_{T,\gamma} with effective gate complexity Teff(γ)≤TT_{\mathrm{eff}}(\gamma) \le T (Section 3.4.2, line 204; or 2I(S;W)/m\sqrt{2I(S;W)/m} in expectation).

Combining Term (I) and Term (II) gives the data-dependent bound, which holds with probability at least 1−δ1-\delta:

∣Rtrue(N^W)−R^emp(N^W;S)∣≤2(H(ρR)+Ic(N^W,∣ψ⟩))+O(Teff(γ)log⁡Teff(γ)+log⁡(2/δ)m).\big|R_{\mathrm{true}}(\hat{\mathcal{N}}_W) - \hat{R}_{\mathrm{emp}}(\hat{\mathcal{N}}_W; S)\big| \le \sqrt{2\big(H(\rho_R) + I_c(\hat{\mathcal{N}}_W, |\psi\rangle)\big)} + \mathcal{O}\left(\sqrt{\frac{T_{\mathrm{eff}}(\gamma)\log T_{\mathrm{eff}}(\gamma) + \log(2/\delta)}{m}}\right).

This bound excludes the lookup-table counterexample (U^S\hat{U}_S acting as identity on mm observed basis states ∣x⟩A1∣0⟩A2|x\rangle_{A_1}|0\rangle_{A_2} and flipping the ancilla ∣0⟩A2↦∣1⟩A2|0\rangle_{A_2} \mapsto |1\rangle_{A_2} on all D−mD-m unobserved states) on two independent fronts. First, in Term (I), U^S\hat{U}_S is a multi-controlled gate conditioned on x∉Sx \notin S. Applying it to any superposition product query ∣ψ⟩=∣ψ1⟩A1⊗∣0⟩A2|\psi\rangle = |\psi_1\rangle_{A_1} \otimes |0\rangle_{A_2} maximally entangles the ancilla A2A_2 with A1A_1. This generates large cross-cut entanglement χ^(U^S)=SA1(U^S∣ψ⟩)>0\hat{\chi}(\hat{U}_S) = S_{A_1}(\hat{U}_S|\psi\rangle) > 0, which our regularizer directly penalizes. Second, in Term (II), encoding an arbitrary mm-subset out of DD requires circuit gate count T=Ω(mlog⁡D)T = \Omega(m\log D) and mutual information I(S;W)=log⁡(Dm)=Ω(mlog⁡(D/m))I(S;W) = \log\binom{D}{m} = \Omega(m\log(D/m)). This violates bounded PQC complexity (T≪mT \ll m) and algorithmic stability.

AuthorsOpenReview ↗

Response to Reviewer ywG8: Q3

Q3

To resolve the notational overlap with the mm-query training set SS, we denote the probe set in Eq. (1) by P={∣ψk⟩RA}k=1K⊂HR⊗HA\mathcal{P} = \lbrace|\psi_k\rangle_{RA}\rbrace_{k=1}^K \subset \mathcal{H}_R \otimes \mathcal{H}_A. For an nn-qubit channel, P\mathcal{P} is constructively defined as a fixed, θ\theta-independent ensemble of partially entangled product probes rotated across the three single-qubit Pauli eigenbases:

∣ψk⟩RA=⨂q=1n(p0 ∣0⟩Rq⊗Uk,q∣0⟩Aq+1−p0 ∣1⟩Rq⊗Uk,q∣1⟩Aq),Uk,q∈{I,H,HS†},|\psi_k\rangle_{RA} = \bigotimes_{q=1}^n \left(\sqrt{p_0}\,|0\rangle_{R_q} \otimes U_{k,q}|0\rangle_{A_q} + \sqrt{1-p_0}\,|1\rangle_{R_q} \otimes U_{k,q}|1\rangle_{A_q}\right), \quad U_{k,q} \in \lbrace I, H, HS^\dagger\rbrace,

with a fixed non-uniform Schmidt weight p0=0.7p_0 = 0.7 (matching ρ=diag(0.7,0.3)\rho = \mathrm{diag}(0.7, 0.3) in Section 4.1, Table 3).

Fixing P\mathcal{P} in this manner guarantees three properties. (i) Unlike maximally entangled probes (ρA=I/2n\rho_A = I/2^n), the non-uniform spectrum p0≠0.5p_0 \ne 0.5 combined with the X,Y,ZX, Y, Z eigenbases ensures that no unitary U^(θ)\hat{U}(\theta) commutes with all input marginals or aligns with all noise axes simultaneously. This preserves a non-zero gradient ∇θmax⁡∣ψ⟩∈PIc(N^ϵ,∣ψ⟩)\nabla_\theta \max_{|\psi\rangle \in \mathcal{P}} I_c(\hat{\mathcal{N}}_\epsilon, |\psi\rangle). (ii) Every probe in P\mathcal{P} shares the exact same reference entropy H(ρR)=nh2(0.7)≈0.8813nH(\rho_R) = nh_2(0.7) \approx 0.8813n bits. (iii) Eq. (1) remains directly computable during training without continuous state-space optimization.

AuthorsOpenReview ↗

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 N\mathcal{N} and unitaries V,WV, W, the Stinespring isometry of W∘N∘V\mathcal{W} \circ \mathcal{N} \circ \mathcal{V} is (W⊗IE)VNV(W \otimes \mathbb{I}_E) V_{\mathcal{N}} V. Its output and environment states at input ρ\rho are unitarily equivalent to those of N\mathcal{N} at input VρV†V\rho V^\dagger, which implies Ic(ρ,W∘N∘V)=Ic(VρV†,N)I_c(\rho, \mathcal{W} \circ \mathcal{N} \circ \mathcal{V}) = I_c(V\rho V^\dagger, \mathcal{N}). Consequently, if noise is placed only once after the full unitary (N^ϵ=Eϵ∘U^(θ)\hat{\mathcal{N}}_\epsilon = \mathcal{E}_\epsilon \circ \hat{U}(\theta), as in the single-qubit illustrative example of Section 4.1) and one maximizes over all possible input density matrices ρ\rho, then max⁡ρIc(ρ,Eϵ∘U^(θ))=max⁡ρIc(ρ,Eϵ)\max_\rho I_c(\rho, \mathcal{E}_\epsilon \circ \hat{U}(\theta)) = \max_\rho I_c(\rho, \mathcal{E}_\epsilon) is independent of θ\theta. In the terminal noise setting, Eq. (1) avoids this invariance by restricting the maximization to the finite, fixed probe set P\mathcal{P} defined in Q3, where U^(θ)\hat{U}(\theta) rotates the fixed probe marginals ρA(k)\rho_A^{(k)} relative to the anisotropic noise eigenbasis of Eϵ\mathcal{E}_\epsilon. More generally, as stated in Proposition 3 (lines 165–166), realistic gate noise in a depth-d′d' PQC acts interleaved per layer rather than only at the terminal output:

N^ϵ(θ)=Eϵ∘U^d′(θd′)∘⋯∘Eϵ∘U^2(θ2)∘Eϵ∘U^1(θ1).\hat{\mathcal{N}}_\epsilon(\theta) = \mathcal{E}_\epsilon \circ \hat{\mathcal{U}}_{d'}(\theta_{d'}) \circ \cdots \circ \mathcal{E}_\epsilon \circ \hat{\mathcal{U}}_2(\theta_2) \circ \mathcal{E}_\epsilon \circ \hat{\mathcal{U}}_1(\theta_1).

Under interleaved noise, unitary invariance removes the dependence on the first layer U^1\hat{\mathcal{U}}_1 under a full-space maximization, but the intermediate unitaries U^2,…,U^d′\hat{\mathcal{U}}_2, \dots, \hat{\mathcal{U}}_{d'} do not commute with the preceding non-unitary noise maps Eϵ\mathcal{E}_\epsilon, leaving genuine θ\theta-dependence even when maximizing over all input states ρ\rho. The table below verifies both regimes numerically under per-qubit dephasing Zp⊗n\mathcal{Z}_p^{\otimes n} at p=0.1p = 0.1 across six Haar-random unitaries per setting:

ChannelnnVaried layermax⁡ρIc\max_\rho I_c (min / max over 6 unitaries)Fixed-probe IcI_c (min / max)
Zp⊗n∘U\mathcal{Z}_p^{\otimes n} \circ U1UU0.53100.5310 / 0.53100.53100.4130.413 / 0.4330.433
Zp⊗n∘U\mathcal{Z}_p^{\otimes n} \circ U2UU1.06201.0620 / 1.06201.06200.8020.802 / 0.8060.806
Zp⊗n∘U2∘Zp⊗n∘U1\mathcal{Z}_p^{\otimes n} \circ U_2 \circ \mathcal{Z}_p^{\otimes n} \circ U_11U1U_10.23220.2322 / 0.23220.23220.1760.176 / 0.1810.181
Zp⊗n∘U2∘Zp⊗n∘U1\mathcal{Z}_p^{\otimes n} \circ U_2 \circ \mathcal{Z}_p^{\otimes n} \circ U_11U2U_20.06900.0690 / 0.29100.29100.0420.042 / 0.2350.235
Zp⊗n∘U2∘Zp⊗n∘U1\mathcal{Z}_p^{\otimes n} \circ U_2 \circ \mathcal{Z}_p^{\otimes n} \circ U_12U1U_10.30550.3055 / 0.30550.30550.1360.136 / 0.1580.158
Zp⊗n∘U2∘Zp⊗n∘U1\mathcal{Z}_p^{\otimes n} \circ U_2 \circ \mathcal{Z}_p^{\otimes n} \circ U_12U2U_20.25700.2570 / 0.33200.33200.0860.086 / 0.2460.246

For post-unitary noise, max⁡ρIc\max_\rho I_c is constant at the dephasing capacity n(1−h2(0.1))=0.5310nn(1 - h_2(0.1)) = 0.5310n while the fixed-probe values vary; under interleaved noise, max⁡ρIc\max_\rho I_c is invariant in U1U_1 but varies substantially with the internal layer U2U_2.

Second, all training experiments in Section 4 use noiseless unitaries trained with χ^(θ)\hat{\chi}(\theta) in Eq. (2). The regularizer χ^(θ)\hat{\chi}(\theta) escapes input-unitary invariance because it is not a supremum over all inputs: it is the average output entanglement SA1(U^(θ)∣ψi⟩)S_{A_1}(\hat{U}(\theta)|\psi_i\rangle) evaluated on fixed product training queries across the fixed bipartition A1∣A2A_1 | A_2. Thus, it is invariant only under post-circuit local unitaries VA1⊗VA2V_{A_1} \otimes V_{A_2}, not under entangling circuit layers in U^(θ)\hat{U}(\theta). Regarding its relation to Corollary 7, by Definition 4 (lines 193–195), SA1(U^(θ)∣ψi⟩)=Ic(Eθ,∣ψi⟩)S_{A_1}(\hat{U}(\theta)|\psi_i\rangle) = I_c(\mathcal{E}_\theta, |\psi_i\rangle) is the coherent information of the reduced erasure channel Eθ:A1→A1′\mathcal{E}_\theta : A_1 \to A_1' with subsystem A2A_2 acting as the reference register RR. Because product queries have zero initial cross-cut entanglement (H(ρRinit)=0H(\rho_R^{\mathrm{init}}) = 0), the Corollary 7 complexity term evaluated on Eθ\mathcal{E}_\theta reduces directly to 2κL2χ^(θ)\sqrt{2\kappa_L^2 \hat{\chi}(\theta)}, as detailed in Q5 below.

Reviewer_8zN5OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes coherent information as a complexity measure and regularizer for learning quantum channels. It derives a single-query bound using a decoupled reference state and attempts to extend it to a finite-sample generalization guarantee. The manuscript correctly identifies that coherent information is parameter-independent for unitary hypotheses, then considers noisy hypotheses and a separate output-entanglement regularizer. Experiments learn random Clifford targets with parameterized unitary circuits. They report mixed generalization-gap changes at 3–5 qubits and smaller gaps at 10–12 qubits using the entanglement regularizer.

Critical strengths

The manuscript identifies important degeneracies and reports negative results. Proposition 2 correctly establishes that coherent information is constant across unitary hypotheses for a fixed probe. Lemma 5 also correctly identifies cancellation of the proposed sample-size improvement under the stated tensor-product argument. Reporting the mixed small-system results and depolarizing-noise control is useful and helps expose the limitations of the approach.

Critical weaknesses
  1. The claimed generalization guarantee does not adequately account for learning from the evaluated data. Section 3.2 defines “true risk” through a decoupled reference–output state, but does not establish that this corresponds to population prediction risk on fresh queries. Theorem 6 and Appendix A.3 then apply concentration to training losses as if they were independent bounded variables, despite the hypothesis being fitted to those same queries. Fixing the trained hypothesis does not restore independence of its training data. This is a central dependence that information-theoretic generalization frameworks explicitly account for. See Caro et al. (2024) (https://arxiv.org/abs/2311.05529). A different proof or additional assumptions are necessary.

  2. Neither the proposed repair nor the experiments establish the advertised coherent-information regularization benefit. For the fixed post-unitary noise model E∘Uθ\mathcal E\circ\mathcal U_\theta, maximizing coherent information over all inputs remains invariant under Uθ\mathcal U_\theta, since the unitary simply reparameterizes the input set. Thus, parameter dependence for one fixed dephasing probe does not establish parameter dependence of the maximum. The experiments instead regularize output entanglement, explicitly without a corresponding bound. Tables 1–2 emphasize generalization gaps without corresponding aggregate held-out risk improvements; a smaller gap could result from worse fitting rather than better prediction.

Questions for authors
  1. What operational learning model makes the decoupled-state risk equal to population risk, and how does Theorem 6 handle the dependence of the learned hypothesis on its training queries?
  2. For the stated model E∘Uθ\mathcal E\circ\mathcal U_\theta, what restriction on the probe set or change in the noise placement prevents the maximized coherent information from being invariant under unitary preprocessing?
  3. Can the authors report training and held-out risks separately, across additional independent seeds, to show whether the entanglement penalty improves prediction rather than merely reducing the generalization gap?
3 comments
AuthorsOpenReview ↗

Response to Reviewer 8zN5: Q1

Answer to Reviewer 8zN5

We thank the reviewer for the insightful feedback and constructive questions. Our point-by-point responses follow.

Q1

The decoupled state τ\tau matches fresh-query population evaluation in two ways, one for each setting we study: the noisy bipartite formulation (Sections 3.1–3.3) and the other is the unitary product-state compilation experiments (Sections 3.4.2, 3.7, and 4).

In the bipartite query-index view (paralleling Xu & Raginsky, 2017 and Caro et al., 2024), the reference register RR purifies the training sample distribution and tracks the query realization to which the learned hypothesis N^W\hat{\mathcal{N}}_W is fitted. The coupled state σRA′=(idR⊗N^W)(∣ψ⟩⟨ψ∣RA)\sigma_{RA'} = (\mathrm{id}_R \otimes \hat{\mathcal{N}}_W)(\vert{}\psi\rangle\langle\psi\vert{}_{RA}) represents evaluating N^W\hat{\mathcal{N}}_W on the entangled training realization. The decoupled state τRA′=ρR⊗N^W(ρA)\tau_{RA'} = \rho_R \otimes \hat{\mathcal{N}}_W(\rho_A) erases sample-specific correlations with RR and evaluates N^W\hat{\mathcal{N}}_W on an independently drawn fresh query marginal.

In the product-state compilation protocol of Section 4, queries are drawn from the distribution D\mathcal{D} of pure product states ∣ψi⟩=∣ψ1,i⟩A1⊗∣ψ2,i⟩A2\vert{}\psi_i\rangle = \vert{}\psi_{1,i}\rangle_{A_1} \otimes \vert{}\psi_{2,i}\rangle_{A_2} across a balanced bipartition A=A1⊗A2A = A_1 \otimes A_2. Loss is then measured on the residual ∣ri⟩=U^(θ)†U∣ψi⟩\vert{}r_i\rangle = \hat{U}(\theta)^\dagger U\vert{}\psi_i\rangle against ∣ψi⟩\vert{}\psi_i\rangle (Appendix A.4). By Definition 4 (lines 193–195), subsystem A2A_2 serves as the internal reference register RR for the reduced erasure channel Eθ:A1→A1′\mathcal{E}_\theta : A_1 \to A_1' obtained by tracing out A2′A_2'.

Every fresh query that is drawn from D\mathcal{D} is a product state across A1∣A2A_1 \vert{} A_2, and the true target residual U†U∣ψ⟩=∣ψ⟩U^\dagger U\vert{}\psi\rangle = \vert{}\psi\rangle is strictly unentangled across the cut. So any cross-partition entanglement that U^(θ)\hat{U}(\theta) generates in the coupled state σA1′A2′(i)=∣ri⟩⟨ri∣\sigma_{A_1' A_2'}^{(i)} = \vert{}r_i\rangle\langle r_i\vert{} is simply a memorization of the training set SS. Projecting onto the decoupled product state τA1′A2′(i)=ρA1′(i)⊗ρA2′(i)\tau_{A_1' A_2'}^{(i)} = \rho_{A_1'}^{(i)} \otimes \rho_{A_2'}^{(i)} isolates the unentangled residual component which generalizes to fresh product queries from D\mathcal{D}. Product queries have zero initial entanglement (H(ρRinit)=0H(\rho_R^{\mathrm{init}}) = 0). So applying Theorem 1 to the erasure channel Eθ\mathcal{E}_\theta directly bounds this cross-cut risk deviation by 2χ^(θ)\sqrt{2\hat{\chi}(\theta)}.

On data dependence in Theorem 6, we do agree that applying a single-hypothesis Hoeffding bound directly to Term (II) =∣Rpop(N^W)−R^emp(N^W;S)∣= \vert{}R_{\mathrm{pop}}(\hat{\mathcal{N}}_W) - \hat{R}_{\mathrm{emp}}(\hat{\mathcal{N}}_W; S)\vert{} in Appendix A.3 (lines 625–630) requires accounting for the dependence of N^W=Aλ(S)\hat{\mathcal{N}}_W = \mathcal{A}_\lambda(S) on SS. In our framework, minimizing the regularized objective actually restricts the learner to the coherently constrained PQC sublevel set HT,γ={N^θ∈HT:Ω(N^θ)≤γ}\mathcal{H}_{T,\gamma} = \lbrace\hat{\mathcal{N}}_\theta \in \mathcal{H}_T : \Omega(\hat{\mathcal{N}}_\theta) \le \gamma\rbrace. Data dependence in Term (II) is then controlled in one of three ways:(1) the uniform replace-one algorithmic stability β(λ)\beta(\lambda) under McDiarmid's bounded-differences inequality (as stated in the Theorem 6 proof sketch, line 239); (2) the uniform convergence over the restricted sublevel set HT,γ\mathcal{H}_{T,\gamma} with effective entangling gate count Teff(γ)≤TT_{\mathrm{eff}}(\gamma) \le T (Section 3.4.2, line 204); or (3) the algorithmic mutual information 2I(S;W)/m\sqrt{2I(S;W)/m} of Caro et al. (2024) in expectation (Section 3.5, line 264).

After combining Term (I) and Term (II) using the triangle inequality gives with at least probability of 1−δ1-\delta, the following:

∣Rtrue(N^W)−R^emp(N^W;S)∣≤2(H(ρR)+Ic(N^W,∣ψ⟩))+O(Teff(γ)log⁡Teff(γ)+log⁡(2/δ)m).\big\vert{}R_{\mathrm{true}}(\hat{\mathcal{N}}_W) - \hat{R}_{\mathrm{emp}}(\hat{\mathcal{N}}_W; S)\big\vert{} \le \sqrt{2\big(H(\rho_R) + I_c(\hat{\mathcal{N}}_W, \vert{}\psi\rangle)\big)} + \mathcal{O}\left(\sqrt{\frac{T_{\mathrm{eff}}(\gamma)\log T_{\mathrm{eff}}(\gamma) + \log(2/\delta)}{m}}\right).

AuthorsOpenReview ↗

Response to Reviewer 8zN5: Q2

Q2

Suppose that the noise acts only once after the full unitary (N^ϵ=Eϵ∘U^(θ)\hat{\mathcal{N}}_\epsilon = \mathcal{E}_\epsilon \circ \hat{U}(\theta)) and coherent information is maximized over the entire continuous Hilbert space HR⊗HA\mathcal{H}_R \otimes \mathcal{H}_A. Then U^(θ)\hat{U}(\theta) would simply reparameterize the set of all input states and leave the global supremum sup⁡∣ψ⟩RAIc(Eϵ∘U^(θ),∣ψ⟩)\sup_{|\psi\rangle_{RA}} I_c(\mathcal{E}_\epsilon \circ \hat{U}(\theta), |\psi\rangle) independent of θ\theta. Our framework avoids this unitary preprocessing invariance in two ways.

First is that, for the terminal noise model N^ϵ=Eϵ∘U^(θ)\hat{\mathcal{N}}_\epsilon = \mathcal{E}_\epsilon \circ \hat{U}(\theta) the Eq. (1) does not maximize over the full unitary-invariant state space. It instead restricts the maximum to a fixed, finite, θ\theta-independent probe ensemble P={∣ψk⟩RA}k=1K⊂HR⊗HA\mathcal{P} = \lbrace|\psi_k\rangle_{RA}\rbrace_{k=1}^K \subset \mathcal{H}_R \otimes \mathcal{H}_A of partially entangled product states rotated across the three single-qubit Pauli eigenbases:

∣ψk⟩RA=⨂q=1n(p0 ∣0⟩Rq⊗Uk,q∣0⟩Aq+1−p0 ∣1⟩Rq⊗Uk,q∣1⟩Aq),Uk,q∈{I,H,HS†},|\psi_k\rangle_{RA} = \bigotimes_{q=1}^n \left(\sqrt{p_0}\,|0\rangle_{R_q} \otimes U_{k,q}|0\rangle_{A_q} + \sqrt{1-p_0}\,|1\rangle_{R_q} \otimes U_{k,q}|1\rangle_{A_q}\right), \quad U_{k,q} \in \lbrace I, H, HS^\dagger\rbrace,

with a fixed non-uniform Schmidt weight p0=0.7p_0 = 0.7 (matching ρ=diag(0.7,0.3)\rho = \mathrm{diag}(0.7, 0.3) in Section 4.1, Table 3). Because P\mathcal{P} is a finite discrete subset of non-maximally entangled states instead of a full unitary orbit, so applying U^(θ)\hat{U}(\theta) rotates the input marginals ρA(k)\rho_A^{(k)} out of P\mathcal{P} relative to the anisotropic noise axes of Eϵ\mathcal{E}_\epsilon. Consequently, max⁡∣ψ⟩∈PIc(Eϵ∘U^(θ),∣ψ⟩)\max_{|\psi\rangle \in \mathcal{P}} I_c(\mathcal{E}_\epsilon \circ \hat{U}(\theta), |\psi\rangle) is not invariant under U^(θ)\hat{U}(\theta) and hence retains a non-vanishing parameter gradient, as verified numerically in Table 3.

Secondly, on noise placement, Proposition 3 (lines 165–166) specifies that physical gate noise in a depth-d′d' PQC acts per gate or per layer and not only at the terminal output:

N^ϵ(θ)=Eϵ∘U^d′(θd′)∘⋯∘Eϵ∘U^2(θ2)∘Eϵ∘U^1(θ1).\hat{\mathcal{N}}_\epsilon(\theta) = \mathcal{E}_\epsilon \circ \hat{\mathcal{U}}_{d'}(\theta_{d'}) \circ \cdots \circ \mathcal{E}_\epsilon \circ \hat{\mathcal{U}}_2(\theta_2) \circ \mathcal{E}_\epsilon \circ \hat{\mathcal{U}}_1(\theta_1).

The intermediate parameterized unitary layers U^2(θ2),…,U^d′(θd′)\hat{\mathcal{U}}_2(\theta_2), \dots, \hat{\mathcal{U}}_{d'}(\theta_{d'}) do not commute with the preceding non-unitary noise maps Eϵ\mathcal{E}_\epsilon. The interleaved channel therefore cannot be factored into the form E~∘U^total(θ)\tilde{\mathcal{E}} \circ \hat{U}_{\mathrm{total}}(\theta). Changing the internal gate parameters (θ2,…,θd′)(\theta_2, \dots, \theta_{d'}) alters the relative alignment between consecutive noisy layers which then changes the singular-value spectrum and quantum capacity of the composite channel itself. So under interleaved per-layer noise, even the unconstrained maximum sup⁡∣ψ⟩RAIc(N^ϵ(θ),∣ψ⟩)\sup_{|\psi\rangle_{RA}} I_c(\hat{\mathcal{N}}_\epsilon(\theta), |\psi\rangle) over all bipartite states depends non-trivially on θ\theta.

Finally, in the noiseless unitary regime of Eq. (2) and Section 4, the regularizer χ^(θ)=1m∑i=1mSA1(U^(θ)∣ψi⟩)\hat{\chi}(\theta) = \frac{1}{m}\sum_{i=1}^m S_{A_1}(\hat{U}(\theta)|\psi_i\rangle) is backed by a bound. By the erasure-channel identity of Definition 4 (lines 193–195), SA1(U^(θ)∣ψi⟩)=Ic(Eθ,∣ψi⟩)S_{A_1}(\hat{U}(\theta)|\psi_i\rangle) = I_c(\mathcal{E}_\theta, |\psi_i\rangle) is the exact coherent information of the reduced channel Eθ:A1→A1′\mathcal{E}_\theta : A_1 \to A_1', with A2A_2 serving as the reference register RR. Since H(ρRinit)=0H(\rho_R^{\mathrm{init}}) = 0 on product queries. Theorem 1 applied to Eθ\mathcal{E}_\theta therefore directly yields the 2χ^(θ)\sqrt{2\hat{\chi}(\theta)} bound on cross-partition overfitting.

AuthorsOpenReview ↗

Response to Reviewer 8zN5: Q3

Q3

We designed and ran a new experiment to answer this question. On n=3n=3 qubits we train the d′=4d'=4 ansatz (T=45T=45) on m∈{1,2,4,8}m \in \lbrace 1, 2, 4, 8\rbrace product-Haar queries. Here the problem is under-determined and an inductive bias can therefore matter. There are three targets: Haar random product unitaries (zero entangling power); depth-1 random Clifford circuits (mean output entanglement 0.520.52 bits on held-out inputs); and depth-2 random Clifford circuits (0.720.72 bits). There are two initializations, near identity scale 0.10.1 and an uninformative scale 1.51.5. Each condition has 20 independent seeds and each seed redraws the target, training set, held-out set and initialization. Within a seed, plain ERM, weight decay (0.01∥θ∥22/T0.01\lVert\theta\rVert_2^2/T) and the penalty λχ^\lambda\hat{\chi} for λ∈{0.02,0.1,0.3}\lambda \in \lbrace 0.02, 0.1, 0.3\rbrace all see identical data and identical initial parameters. We report training risk and exact population risk R(θ^)=1−Tr[(W⊗W†)Πn]R(\hat{\theta}) = 1 - \mathrm{Tr}[(W \otimes W^\dagger)\Pi_n] (where W=U†Uθ^W = U^\dagger U_{\hat{\theta}} and Πn=6−n∑A⊆[n]FA\Pi_n = 6^{-n}\sum_{A \subseteq [n]} \mathbb{F}_A, with FA\mathbb{F}_A swapping the two copies on the qubits in AA), pooled over mm (80 paired runs per cell). We then report every λ\lambda, so no hyperparameter is selected on the held-out data.

InitTargetRegularizerΔ\Delta train riskΔ\Delta population risk vs. ERM (±\pm s.e.)better / worse (of 80)Wilcoxon pp
0.1Clifford-1λ=0.1\lambda=0.1+0.007+0.007−0.023±0.006-0.023 \pm 0.00658 / 223.1×10−53.1 \times 10^{-5}
0.1Clifford-1λ=0.3\lambda=0.3+0.044+0.044−0.063±0.008-0.063 \pm 0.00867 / 131.3×10−91.3 \times 10^{-9}
1.5Clifford-1λ=0.1\lambda=0.1+0.006+0.006−0.022±0.006-0.022 \pm 0.00662 / 186.5×10−66.5 \times 10^{-6}
1.5Clifford-1λ=0.3\lambda=0.3+0.043+0.043−0.053±0.007-0.053 \pm 0.00766 / 142.6×10−92.6 \times 10^{-9}
0.1Clifford-2λ=0.3\lambda=0.3+0.051+0.051+0.012±0.006+0.012 \pm 0.00625 / 550.0110.011
1.5Clifford-2λ=0.3\lambda=0.3+0.051+0.051+0.011±0.006+0.011 \pm 0.00630 / 500.0170.017
0.1 / 1.5Clifford-1/2weight decay<0.001<0.001−0.011-0.011 to +0.002+0.002–\text{--}≥0.07\ge 0.07
0.1 / 1.5Productany≈0\approx 0no consistent effect (∣Δ∣≤0.05\lvert\Delta\rvert \le 0.05); WD worseWD worse in 54 and 64WD p≤0.007p \le 0.007

Per-mm results for the most informative case show that the effect grows with mm. For the Clifford-1 target at initialization 1.51.5, ERM's population risk is 0.846,0.779,0.724,0.6490.846, 0.779, 0.724, 0.649 at m=1,2,4,8m=1, 2, 4, 8, versus 0.837,0.743,0.672,0.5350.837, 0.743, 0.672, 0.535 with λ=0.3\lambda=0.3 (s.e. ≤0.013\le 0.013; p=1.9×10−6p=1.9 \times 10^{-6} at m=8m=8). Training risk at m=8m=8 increases from 0.2640.264 to 0.3090.309. So the gain in population risk comes with a deliberate loss of fit. For the Clifford-2 target at m=8m=8 the penalty is harmful as the population risk is 0.606→0.6270.606 \to 0.627 (p=0.002p=0.002). In that same case the gap shrinks by 0.0400.040 which is a direct instance of the reviewer's warning. Product targets are learned exactly for m≥4m \ge 4 by every method. So there is nothing to improve. Weight decay has no reliable effect on the Clifford targets and hurts on product targets. Our answer to Q3 is therefore that the penalty improves prediction when its bias matches the target and harms it otherwise. Note that the Clifford-2 case shows a smaller gap together with a larger population risk.

Reviewer_oa1sOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies Coherently Regularized Quantum Channel Learning as a quantum analogous of the information theoretic generalization bound analysis in learning theory. The authors propose measuring a property of the trained model called “coherent information” and using it both to predict how much the model will overfit and as a penalty during training to discourage overfitting.

Critical strengths

The problem of an operational information theoretic quantity that controls generalization in channel learning is indeed open and interesting.

Critical weaknesses

Typically in learning theory we have a target label yy and a prediction that aims to predict the true label as accurately as possible so that the loss is EX,Y[l(g(X),Y)]E_{X,Y}[l(g(X), Y)] where ll is the loss function.
But this paper studies a conceptually different model: The loss is defined as the difference between two expectation values: ∣Rtrue(N^)−R^(N^)∣|R_{true}(\hat{\mathcal N}) - \hat{R}(\hat{\mathcal N})| where R^(N^)=Tr(L(I⊗N^)(∣ψ⟩⟨ψ∣)\hat{R}(\hat{\mathcal N})=Tr(L (I\otimes \hat{\mathcal N})(|\psi\rangle \langle\psi|) for some state ∣ψ⟩RA|\psi\rangle_{RA}, and Rtrue(N^)=Tr(L(ρR⊗N^(ρA)))R_{true}(\hat{\mathcal N}) = Tr(L (\rho_R \otimes \hat{\mathcal N}(\rho_A))), where ρR\rho_R and ρA\rho_A are the marginals ∣ψ⟩RA|\psi\rangle_{RA}. In short I am not convinced if this is a natural model to think about. If N^=N\hat{\mathcal N}=\mathcal N then are the two expectations equal?

For a unitary hypothesis with a maximally entangled probe the bound in Theorem 1 reads 2n2\sqrt n bits, far above the range of a bounded loss. For the product-state queries actually used in all experiments there is no reference register, σ=τ\sigma =\tau and Theorem 1 “gap” is 0 by construction. The gap reported in Tables 1–2 is a different quantity to which no theorem in the paper applies.

There are other comments:

  • Proposition 2 is trivial!
  • In numerical results only the gap is shown, never the true risk under CRQCL.
Questions for authors

see my comments above.

2 comments
AuthorsOpenReview ↗

response to reviewer oa1s: (Part I)

We thank the reviewer for the direct and constructive feedback. We address all five concerns below.

Conceptual Risk Model and Behavior When N^=N\hat{\mathcal{N}} = \mathcal{N}

Whether R^(N^)\hat{R}(\hat{\mathcal{N}}) and Rtrue(N^)R_{\mathrm{true}}(\hat{\mathcal{N}}) coincide when N^=N\hat{\mathcal{N}} = \mathcal{N} depends on the setting we evaluate: the raw bipartite probe of Section 3.2, or the residual compilation channel of Sections 3.7 and 4.

On an entangled bipartite probe ∣ψ⟩RA|\psi\rangle_{RA}, suppose N\mathcal{N} preserves entanglement with RR and N^=N\hat{\mathcal{N}} = \mathcal{N}. Then σRA′=(idR⊗N^)(∣ψ⟩⟨ψ∣RA)\sigma_{RA'} = (\mathrm{id}_R \otimes \hat{\mathcal{N}})(|\psi\rangle\langle\psi|_{RA}) remains entangled, while τRA′=ρR⊗N^(ρA)\tau_{RA'} = \rho_R \otimes \hat{\mathcal{N}}(\rho_A) is decoupled. So Tr[LσRA′]\mathrm{Tr}[L\sigma_{RA'}] and Tr[LτRA′]\mathrm{Tr}[L\tau_{RA'}] are not equal in general. Here Tr[LτRA′]\mathrm{Tr}[L\tau_{RA'}] is a counterfactual decoupling risk (analogous to contrasting P(X,Y)P(X,Y) with P(X)P(Y)P(X)P(Y) in Xu & Raginsky, 2017 and Caro et al., 2024). It isolates how strongly the channel relies on correlations with RR. We will update Section 3.2 to distinguish this counterfactual decoupling risk from the fresh-query population risk Rpop(N^)=Eψ∼D[ℓ(N^,ψ)]R_{\mathrm{pop}}(\hat{\mathcal{N}}) = \mathbb{E}_{\psi \sim \mathcal{D}}[\ell(\hat{\mathcal{N}}, \psi)].

In the residual formulation used in Sections 3.7 and 4, the two expectations do coincide when N^=N\hat{\mathcal{N}} = \mathcal{N}. For product queries ∣ψi⟩=∣ψ1,i⟩A1⊗∣ψ2,i⟩A2|\psi_i\rangle = |\psi_{1,i}\rangle_{A_1} \otimes |\psi_{2,i}\rangle_{A_2} across a balanced cut A=A1⊗A2A = A_1 \otimes A_2, loss is measured on the residual ∣ri⟩=U^(θ)†U∣ψi⟩|r_i\rangle = \hat{U}(\theta)^\dagger U|\psi_i\rangle against ∣ψi⟩|\psi_i\rangle (Appendix A.4). Whenever U^(θ)=U\hat{U}(\theta) = U (up to global phase), ∣ri⟩=∣ψi⟩|r_i\rangle = |\psi_i\rangle is strictly unentangled across A1∣A2A_1 | A_2. So σA1′A2′(i)=∣ri⟩⟨ri∣\sigma_{A_1' A_2'}^{(i)} = |r_i\rangle\langle r_i| equals τA1′A2′(i)=ρA1′(i)⊗ρA2′(i)\tau_{A_1' A_2'}^{(i)} = \rho_{A_1'}^{(i)} \otimes \rho_{A_2'}^{(i)}, and both expectations vanish identically.

The 2n2\sqrt{n} Bound for Unitary Hypotheses on Maximally Entangled Probes

We agree. For a unitary hypothesis on a maximally entangled nn-qubit probe (H(ρR)=nH(\rho_R) = n), Proposition 2 gives Ic=nI_c = n, so Theorem 1 evaluates to 2n>12\sqrt{n} > 1 for any [0,1][0,1]-valued loss.

This vacuousness is the degeneracy Section 3.4 is designed to expose: a naive coherent-information bound on noiseless unitary PQCs is both vacuous and parameter-independent (∇θIc≡0\nabla_\theta I_c \equiv 0). The bound becomes hypothesis-discriminating once that degeneracy is broken, which happens in two ways. The first is interleaved physical noise (Proposition 3 and Table 3), where IcI_c drops below H(ρR)H(\rho_R) and varies with θ\theta, tightening below 11 as noise drives IcI_c toward −H(ρR)-H(\rho_R). The second is the product-state residual regime of Definition 4 (H(ρRinit)=0H(\rho_R^{\mathrm{init}}) = 0), which reduces the correlation term to 2χ^(θ)\sqrt{2\hat{\chi}(\theta)}.

We will add an explicit remark after Theorem 1 stating that the 2n2\sqrt{n} value acts as a certificate of this unitary degeneracy.

Product-State Queries (σ=τ\sigma = \tau) and the Theoretical Link to Tables 1 and 2

We agree that product-state queries without an external register RR yield σ=τ\sigma = \tau trivially on an external cut. In Section 4, however, Definition 4 (lines 193–195) partitions the nn-qubit register itself into A=A1⊗A2A = A_1 \otimes A_2, and subsystem A2A_2 serves as the internal reference register RR for the reduced erasure channel Eθ:A1→A1′\mathcal{E}_\theta : A_1 \to A_1'.

Whenever an overparameterized U^(θ)\hat{U}(\theta) generates spurious cross-cut entanglement on training queries, σA1′A2′(i)≠τA1′A2′(i)\sigma_{A_1' A_2'}^{(i)} \neq \tau_{A_1' A_2'}^{(i)}. Because H(ρRinit)=0H(\rho_R^{\mathrm{init}}) = 0 on product inputs, Theorem 1 applied to Eθ\mathcal{E}_\theta directly bounds this deviation by 2χ^(θ)\sqrt{2\hat{\chi}(\theta)}, where χ^(θ)=1m∑i=1mSA1(U^(θ)∣ψi⟩)=1m∑i=1mIc(Eθ,∣ψi⟩)\hat{\chi}(\theta) = \frac{1}{m}\sum_{i=1}^m S_{A_1}(\hat{U}(\theta)|\psi_i\rangle) = \frac{1}{m}\sum_{i=1}^m I_c(\mathcal{E}_\theta, |\psi_i\rangle).

Penalizing χ^(θ)\hat{\chi}(\theta) also restricts the learner to the regularized PQC sublevel set HT,γ={N^θ∈HT:χ^(θ)≤γ}\mathcal{H}_{T,\gamma} = \lbrace\hat{\mathcal{N}}_\theta \in \mathcal{H}_T : \hat{\chi}(\theta) \le \gamma\rbrace, with effective gate complexity Teff(γ)≤TT_{\mathrm{eff}}(\gamma) \le T (Section 3.4.2, line 204). Combining the physical correlation term with uniform convergence over HT,γ\mathcal{H}_{T,\gamma} (or replace-one stability via McDiarmid's inequality, line 239) gives the data-dependent bound governing Tables 1 and 2:

∣Rtrue(N^W)−R^emp(N^W;S)∣≤2(H(ρR)+Ic(N^W,∣ψ⟩))+O(Teff(γ)log⁡Teff(γ)+log⁡(2/δ)m).\big|R_{\mathrm{true}}(\hat{\mathcal{N}}_W) - \hat{R}_{\mathrm{emp}}(\hat{\mathcal{N}}_W; S)\big| \le \sqrt{2\big(H(\rho_R) + I_c(\hat{\mathcal{N}}_W, |\psi\rangle)\big)} + \mathcal{O}\left(\sqrt{\frac{T_{\mathrm{eff}}(\gamma)\log T_{\mathrm{eff}}(\gamma) + \log(2/\delta)}{m}}\right).

AuthorsOpenReview ↗

Response to Reviewer oa1s: (Part II)

Role of Proposition 2

We agree that Proposition 2 has an elementary four-line proof. We included it as an operational diagnostic, not for proof depth. It warns that turning coherent information directly into a regularizer for unitary PQCs yields an identically zero gradient, which motivates both Proposition 3 and Definition 4. We will demote Proposition 2 to a brief Remark or Lemma in the revision.

True Population Risk Under CRQCL

We agree that reporting only the gap ∣Rtrue−R^emp∣|R_{\mathrm{true}} - \hat{R}_{\mathrm{emp}}| is insufficient, because underfitting can shrink the gap without improving test accuracy. To test true prediction directly, we ran a 20-seed study on n=3n=3 qubits (d′=4d'=4, T=45T=45, m∈{1,2,4,8}m \in \lbrace 1, 2, 4, 8\rbrace product-Haar queries). It covers three target families: Haar-random product unitaries, depth-1 random Cliffords (0.520.52 bits mean held-out entanglement), and depth-2 random Cliffords (0.720.72 bits). Each family was run at initialization scales 0.10.1 and 1.51.5. We compared ERM (λ=0\lambda=0), weight decay (0.01∥θ∥22/T0.01\lVert\theta\rVert_2^2/T), and λχ^\lambda\hat{\chi} (λ∈{0.02,0.1,0.3}\lambda \in \lbrace 0.02, 0.1, 0.3\rbrace) on identical data and seeds.

We evaluate exact population risk via the product-Haar twirl R(θ^)=1−Tr[(W⊗W†)Πn]R(\hat{\theta}) = 1 - \mathrm{Tr}[(W \otimes W^\dagger)\Pi_n] (W=U†Uθ^W = U^\dagger U_{\hat{\theta}}, Πn=6−n∑A⊆[n]FA\Pi_n = 6^{-n}\sum_{A \subseteq [n]} \mathbb{F}_A). Results are pooled over mm (80 paired runs per cell), with no held-out hyperparameter selection:

InitTargetRegularizerΔ\Delta train riskΔ\Delta population risk vs. ERM (±\pm s.e.)better / worse (of 80)Wilcoxon pp
0.1Clifford-1λ=0.1\lambda=0.1+0.007+0.007−0.023±0.006-0.023 \pm 0.00658 / 223.1×10−53.1 \times 10^{-5}
0.1Clifford-1λ=0.3\lambda=0.3+0.044+0.044−0.063±0.008-0.063 \pm 0.00867 / 131.3×10−91.3 \times 10^{-9}
1.5Clifford-1λ=0.1\lambda=0.1+0.006+0.006−0.022±0.006-0.022 \pm 0.00662 / 186.5×10−66.5 \times 10^{-6}
1.5Clifford-1λ=0.3\lambda=0.3+0.043+0.043−0.053±0.007-0.053 \pm 0.00766 / 142.6×10−92.6 \times 10^{-9}
0.1Clifford-2λ=0.3\lambda=0.3+0.051+0.051+0.012±0.006+0.012 \pm 0.00625 / 550.0110.011
1.5Clifford-2λ=0.3\lambda=0.3+0.051+0.051+0.011±0.006+0.011 \pm 0.00630 / 500.0170.017
0.1 / 1.5Clifford-1/2weight decay<0.001<0.001−0.011-0.011 to +0.002+0.002–\text{--}≥0.07\ge 0.07
0.1 / 1.5Productany≈0\approx 0no consistent effect (∣Δ∣≤0.05\lvert\Delta\rvert \le 0.05); WD worseWD worse in 54 and 64WD p≤0.007p \le 0.007

For Clifford-1 (init 1.51.5), CRQCL (λ=0.3\lambda=0.3) lowers exact population risk at every query budget. At m=1,2,4,8m=1, 2, 4, 8 it goes from 0.846,0.779,0.724,0.6490.846, 0.779, 0.724, 0.649 (ERM) to 0.837,0.743,0.672,0.5350.837, 0.743, 0.672, 0.535 (p=1.9×10−6p = 1.9 \times 10^{-6} at m=8m=8), while training risk rises modestly from 0.2640.264 to 0.3090.309.

On the more entangling Clifford-2 target at m=8m=8, however, λ=0.3\lambda=0.3 raises population risk (0.606→0.6270.606 \to 0.627, p=0.002p=0.002) while shrinking the gap by 0.0400.040. This directly confirms the reviewer's caution. Product targets are learned exactly for m≥4m \ge 4 by all methods.

We will report empirical training risk and held-out true risk side by side in all tables of the revised manuscript.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.