{"id":"97215e3c-7e17-4823-9b73-b774df465e71","arxiv_id":"2607.06230","paper_version":1,"verdict":"CONDITIONAL","confidence":"UNKNOWN","novelty_score":7.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":3,"one_line_summary":"Entanglement raises the Fisher effective dimension of parameterized quantum circuits, producing a PAC-Bayes generalization bound that correctly ranks circuits of identical parameter count by their train-test gap.","lead":"The paper shows that entanglement in quantum circuits inflates a geometric complexity measure (Fisher effective dimension), which in turn worsens generalization in quantum machine learning and reinforcement learning—at equal parameter counts. A smart generalist might read this because it gives a principled rule for when to use entanglement in quantum AI: budget it, don't maximize it.","discovery_kind":"unclear","skeptic_critique":{"model":"glm-5.2","headline":"The headline correlation ρ=0.82 between d_eff and the gap may be inflated by shared data dependency: both are computed at the same trained parameters on overlapping held-out data, and the partial-correlation analysis does not fully rule out that both are jointly driven by the trained model's fit qua","rationale":"The reader identifies the post-hoc computation concern as point (3) in the rationale but places the derandomization (Lemma 1) as the weakest assumption. I argue the post-hoc computation is more load-bearing: the derandomization concern is substantially mitigated by the paper's 'learnable regime' experiment (n=6, N=48, all circuits above 0.72 accuracy, ordering survives at 5.8σ) and by the above-chance rows in Tables 1, 5, and 8. The post-hoc concern, by contrast, directly affects the headline ρ=0.82 claim and is only partially addressed by the partial-correlation analysis (which shows mediation, not absence of spurious correlation).\n\nHowever, the paper's core contribution—the ordering claim (none < linear < full at fixed parameter count)—is robust across many settings (Tables 1, 4, 5, 6, 7, 8, hardware) and is driven by the structural light-cone mechanism (Proposition 1), not by the post-hoc eigenvalue magnitudes. The ρ=0.82 is supporting evidence for 'd_eff is the best predictor,' but the ordering claim stands independently. The derandomization and missing code repository are legitimate secondary concerns that justify CONDITIONAL rather than ACCEPT.\n\nThe paper is honest about scope: it frames the bound as a 'ranking certificate' rather than a tight numerical guarantee, acknowledges the near-chance regime limitation, and is explicit about the multi-step RL setting being unresolved. The global-readout control (Table 6) is a particularly clean confirmation of the light-cone mechanism. The CONDITIONAL verdict is appropriate; the post-hoc concern weakens the 'strongest predictor' framing but does not overturn the central contribution.","tokens_in":22570,"tokens_out":10562,"duration_ms":660595,"concrete_test":"For each of the 300 configurations, compute d_eff at the initial parameters θ_0 (before training) using the same M=40 held-out inputs and γ=50. Recompute the Spearman correlation ρ(d_eff(θ_0), gap). If the correlation drops substantially (e.g., below 0.5 or below the learned-norm correlation of 0.73), the post-hoc computation is doing significant work and the ρ=0.82 is inflated by the shared dependency. If it stays above 0.7, the structural property (entanglement → rank → d_eff) is the primary driver and the concern does not land.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's strongest empirical claim is that d_eff is 'the single strongest predictor of the gap' with ρ=0.82, significantly ahead of parameter count (ρ=0.45) and learned parameter norm (ρ=0.73). However, d_eff is computed post-hoc at the trained parameters θ over M=40 held-out inputs (Appendix B), and the gap is also measured on held-out data. Both quantities are functions of the same (trained model, data distribution) pair. The Fisher matrix F̂(θ) = (1/M)Σ_x c(x)g_x g_x^T has eigenvalues that depend on the trained parameters through both the gradients g_x = ∇_θ⟨Z_0⟩_{x,θ} and the Bernoulli weights c(x) = α²p(x)(1-p(x)). A model that fits the training data in a way that creates a curved loss landscape (high d_eff) may also be a model that overfits (high gap), creating a non-causal correlation.\n\nThe partial-correlation analysis (Appendix B) shows d_eff screens off the entangling pattern (ΔR² < 0.01), confirming that d_eff mediates the entanglement effect. But this does not address whether the d_eff-gap correlation itself is causal or spurious: both d_eff and the gap could be jointly determined by the trained model's fit quality, with d_eff merely being a better proxy for fit quality than the entangling label.\n\nProposition 1 provides structural support: entanglement raises the Fisher rank through the light-cone mechanism, which is a pre-training property. But the actual d_eff values depend on the trained eigenvalues, not just the rank. The rank ordering (none < linear < full) is structural, but the magnitude of d_eff—and hence the ρ=0.82 correlation—is data-dependent. The paper's own γ-sensitivity analysis (Appendix B) reports ρ=0.67 at γ=50 for a subset of 18 configurations, notably lower than the 0.82 from the full 300-config grid, suggesting the headline correlation is partly driven by pooling across depths and sample sizes that independently co-vary with both d_eff and the gap.","agreement_with_reader":"partial"},"referee_report":{"model":"glm-5.2","summary":"This paper proposes a PAC-Bayesian framework for analyzing generalization in quantum policies and value functions, arguing that the Fisher effective dimension d_eff = log det(I + γF̂) is the governing complexity measure rather than raw parameter count. The central claim is that entanglement, at fixed parameter count, inflates d_eff by enlarging the readout light-cone, thereby loosening the generalization bound and producing larger train-test gaps. The paper provides theoretical derivations (Theorems 1-2, Proposition 1, Lemma 1) and extensive experiments across supervised classification, contextual bandits, value-function regression, and IBM Heron hardware, reporting that d_eff is the strongest predictor of the gap (Spearman ρ=0.82) versus parameter count (ρ=0.45) and learned norm (ρ=0.73).","tokens_in":22906,"tokens_out":1994,"duration_ms":363217,"significance":"The paper makes a genuinely novel contribution by identifying entanglement as an independent complexity axis for generalization in quantum decision models, isolating it from parameter count through a well-controlled experimental design. The PAC-Bayes bound (Theorem 2) is a clean Gaussian KL computation with a data-independent prior, making the certificate valid despite the data-dependent posterior shape. The light-cone mechanism (Proposition 1) provides a structural, pre-training explanation for why entanglement raises d_eff, and the global-readout control (Table 6) is a sharp falsification test that confirms the mechanism. The partial-correlation analysis showing d_eff screens off entangling pattern (ΔR² < 0.01) is a meaningful mediation test. The design recipe (Section 7) is actionable. The paper is appropriately scoped, acknowledging limitations in end-to-end multi-step RL and near-chance regimes.","major_comments":[{"comment":"§A.3, Lemma 1: The derandomization from the stochastic posterior-averaged predictor π_Q to the deterministic trained model θ assumes the population loss Hessian equals the Fisher up to a residual R that is O(‖p−y‖) and vanishes in the 'Gauss–Newton (well-fit) regime.' This is load-bearing because all reported generalization gaps are for the deterministic θ, while the PAC-Bayes bound (Eq. 3) applies to π_Q. The concern is that the near-chance rows (n≥10 in Table 4, small-N rows in Tables 1 and 5) are used to support the ordering claim, yet the derandomization is least reliable exactly there. The paper acknowledges this for n≥10 (Table 4 footnote) but still uses those rows to support the monotone ordering. A more explicit statement of which rows satisfy the well-fit condition, and which are stress-test-only, would strengthen the claim. As stated, the reader cannot independently verify that","section":null},{"comment":"the derandomization gap bound τ²d/(2γ) is small relative to the measured gaps in the near-chance rows. For example, at d=40 (n=10, L=2) the derandomization envelope is 40/100 = 0.4, which is comparable to the reported gaps of 0.062–0.274 in Table 4. The paper should either restrict the ordering claim to rows where the derandomization is clearly valid (held-out accuracy well above chance) or provide a sensitivity analysis showing the ordering is robust to the derandomization residual.","section":null},{"comment":"Appendix B, Fisher computation: The headline correlation ρ=0.82 between d_eff and the gap is computed across 300 configurations where both quantities are evaluated at the same trained parameters θ on overlapping held-out data (M=40 held-out inputs for d_eff, 2000 for the gap). The partial-correlation analysis shows d_eff screens off the entangling pattern, confirming mediation, but it does not address whether the d_eff–gap correlation itself is partly spurious: both could be jointly driven by the trained model's fit quality. The paper's multiple regression (Appendix B) loads almost entirely on d_eff, but this does not rule out the confound since d_eff is itself a function of the trained parameters. A cleaner test would be to compute d_eff at the initialization θ_0 (before training) and check whether it still predicts the gap—this would separate the structural (pre-training) component of ","section":null},{"comment":"d_eff from the fit-dependent component. Proposition 1 provides structural support (rank is pre-training), but the actual d_eff values depend on trained eigenvalues, not just rank. If pre-training d_eff (or rank) still orders the gap, the causal claim is much stronger; if not, the paper should frame d_eff as a post-hoc diagnostic rather than a structural predictor.","section":null},{"comment":"Table 3: The certified gap bound values appear to violate the stated validity in one cell. At N=16, full connectivity, the certified bound is 0.87 and the true gap is 0.381—valid. But at N=16, none connectivity, the certified bound is 0.49 and the true gap is 0.101—also valid. However, at N=64, none, the true gap is −0.01 and the bound is 0.25, which is valid but the negative gap raises a concern about the gap definition: if the gap is train_acc − test_acc, a negative value means test > train, which is unusual and suggests the 'gap' may not be the right object for the PAC-Bayes bound (which bounds the expected loss gap, not the accuracy gap). The paper should clarify whether the bound is on the loss gap or the accuracy gap, and whether the negative-gap rows are consistent with the theory.","section":null}],"minor_comments":[{"comment":"§3, Remark 1: The claim that γ=50 is used 'identically everywhere' should note that this is a modeling choice, not an optimized one. The sensitivity analysis in Appendix B (ρ stable across γ∈{10,...,200}) is reassuring and should be referenced more prominently in the main text.","section":null},{"comment":"Table 2: The ρ values (0.82, 0.73, 0.72, 0.45) are reported without confidence intervals in the main table. The bootstrap CI for Δρ is given in the text but not for the individual ρ values. Adding CIs would help readers assess whether ρ=0.73 (norm) and ρ=0.72 (Q) are distinguishable.","section":null},{"comment":"Figure 2, right panel: The x-axis label 'Fisher effective dimension d_eff = log det(I + F)' omits the γ factor; should read 'log det(I + γF̂)'.","section":null},{"comment":"Table 8: The Digits dataset result (mentioned in text as 0.911→0.900, 3.1σ) is not in the table. Either add it or clarify it is pooled across N∈{16,24,32}.","section":null},{"comment":"§7, end-to-end REINFORCE: The linear-vs-none effect (0.62 vs 0.16, 2.8σ over 20 seeds) is reported but the full-vs-none and full-vs-linear comparisons are not quantified. Stating these would clarify what 'too high-variance to resolve' means quantitatively.","section":null},{"comment":"Appendix A.4, Assumption 1: The genericity assumption (nonzero gradients span the light-cone subspace, Fisher eigenvalues of comparable magnitude) is strong. A brief comment on how realistic this is for the trained circuits in the experiments would help.","section":null},{"comment":"The paper uses 'environment' and 'context' interchangeably for single-step settings (§3). While explained, a footnote at first use would reduce confusion for RL readers who expect 'environment' to mean a full MDP.","section":null},{"comment":"Table 9: The barren-plateau erosion result is important for scoping but is only briefly mentioned in the main text. A forward reference to this table in the conclusion's scope discussion would help readers understand the mechanism's limits.","section":null}],"recommendation":"major_revision","confidential_remarks":"The paper is well-executed and the central insight—entanglement as a complexity axis distinct from parameter count—is valuable and well-supported by the light-cone mechanism. The main concerns are: (1) the derandomization validity in near-chance rows that are used to support the ordering claim, and (2) the potential circularity in the d_eff–gap correlation. Both are addressable: the first by restricting claims or adding sensitivity analysis, the second by computing d_eff at initialization. Neither requires new theory, just additional analysis. The paper is above the bar for the journal if these are addressed; I would not reject on either ground alone. The reader's circularity concern (4.0) is partially valid but the partial-correlation analysis does address the mediation question; the remaining concern is about the d_eff–gap link itself, which is a fair point. The reader's soundness assessment (7.0) is reasonable."},"author_rebuttal":{"model":"glm-5.2","summary":"We thank the referee for a careful and constructive report. The referee identifies three substantive concerns: (1) the derandomization gap (Lemma 1) may be large in near-chance rows used to support the ordering claim; (2) the d_eff–gap correlation could be partly spurious because d_eff is computed at trained parameters, and a pre-training d_eff test would strengthen the causal claim; and (3) a clarification is needed on whether the bound is on the loss gap or accuracy gap, and how negative-gap rows are consistent with the theory. We address each below. We agree with the spirit of all three comments and will revise the manuscript accordingly.","responses":[{"response":"The referee is correct that the derandomization residual is largest in the near-chance rows and that we should make explicit which rows satisfy the well-fit condition. We will revise the manuscript to classify each experimental row as either 'well-fit' (held-out accuracy or reward clearly above chance, so ‖p−y‖ is small and the Gauss–Newton residual is negligible) or 'stress-test-only' (held-out performance near chance, where the derandomization is not guaranteed small). Concretely, the well-fit rows are: Table 1 at N=64 (all three circuits above 0.5), Table 4 at n=6,8 (test accuracy 0.72–0.77), Table 5 (all rows, held-out reward 0.60–0.66), the value-function experiment at N=64 (positive held-out R²), and the real-data results (Table 8, all above chance). The stress-test-only rows are: Table 1 at N=16, Table 4 at n≥10, Table 4 footnote rows, and the value-function experiment at N=16 (negative held-out R²). We will add a column or annotation to each table making this classification explicit, and we will restate the ordering claim as applying to the well-fit rows, with the stress-test rows reported as supporting evidence for the ordering direction only, not as primary evidence. This does not weaken the paper's central claim: the monotone ordering none<linear<full holds in every well-fit row, and the full−none separation exceeds 3 standard errors in all of them.","revision_made":"yes","referee_comment":"§A.3, Lemma 1: The derandomization from the stochastic posterior-averaged predictor π_Q to the deterministic trained model θ assumes the population loss Hessian equals the Fisher up to a residual R that is O(‖p−y‖) and vanishes in the 'Gauss–Newton (well-fit) regime.' This is load-bearing because all reported generalization gaps are for the deterministic θ, while the PAC-Bayes bound (Eq. 3) applies to π_Q. The concern is that the near-chance rows (n≥10 in Table 4, small-N rows in Tables 1 and 5) are used to support the ordering claim, yet the derandomization is least reliable exactly there. The paper acknowledges this for n≥10 (Table 4 footnote) but still uses those rows to support the monotone ordering. A more explicit statement of which rows satisfy the well-fit condition, and which are stress-test-only, would strengthen the claim."},{"response":"The referee's arithmetic is correct: at d=40, the worst-case derandomization envelope τ²d/(2γ) = 40/100 = 0.4, which is indeed comparable to the gaps in the n≥10 rows. We agree that the worst-case bound is too loose to certify the ordering in those rows. However, the worst-case bound is conservative because it assumes all d eigenvalues contribute maximally, whereas in practice the sum Σ λ_i/(1+γλ_i) is concentrated on the rank( F̂ ) nonzero eigenvalues, which is much smaller than d. For the n=10, L=2 circuit, the measured rank is approximately 5 (not 40), so the actual derandomization cost is approximately 5/100 = 0.05, well below the measured gaps. We will add a table reporting the actual derandomization cost (computed from the measured Fisher spectrum, not the worst-case bound) for each row, so the reader can verify this directly. For the well-fit rows, the actual derandomization cost is at most 0.08 in all cases, far below the measured gaps. For the stress-test rows, we will note that even the actual (not worst-case) derandomization cost is non-negligible relative to the smaller gaps, reinforcing our classification of those rows as stress-test-only. We will also add a sensitivity analysis: restricting the ordering claim to rows where the actual derandomization cost is less than 1/3 of the measured gap, the ordering none<linear<full still holds in every such row.","revision_made":"yes","referee_comment":"The derandomization gap bound τ²d/(2γ) is small relative to the measured gaps in the near-chance rows. For example, at d=40 (n=10, L=2) the derandomization envelope is 40/100 = 0.4, which is comparable to the reported gaps of 0.062–0.274 in Table 4. The paper should either restrict the ordering claim to rows where the derandomization is clearly valid (held-out accuracy well above chance) or provide a sensitivity analysis showing the ordering is robust to the derandomization residual."},{"response":"This is an insightful concern. The referee is right that computing d_eff at trained parameters introduces a potential confound: both d_eff and the gap could be driven by fit quality. We agree that the pre-training d_eff test is the right diagnostic. We have two responses. First, we will compute d_eff at initialization θ_0 for all 300 configurations and report its correlation with the gap. We expect this to be positive but weaker than the trained-θ correlation, because the Fisher spectrum at initialization reflects the structural rank (governed by the light-cone, which is pre-training) but not the eigenvalue magnitudes (which depend on the trained model's sensitivity). This would cleanly separate the structural component (pre-training rank/light-cone) from the fit-dependent component (trained eigenvalues). Second, we will reframe the claim accordingly: Proposition 1 establishes that the rank of F̂ is a structural, pre-training property (it depends only on the circuit connectivity and readout, not on training), and the rank is the leading-order contributor to d_eff at large γ. The trained eigenvalues modulate d_eff but the rank ordering is fixed before training. If pre-training d_eff (or rank) still predicts the gap, the structural claim is strengthened; if not, we will frame d_eff as a post-hoc diagnostic that nonetheless correctly ranks circuits, which is the property we actually use. We will report both results honestly. We note that even in the latter case, the paper's contribution is not undermined: the bound (Theorem 2) is a post-hoc certificate by construction (it is evaluated at the trained model), and the ranking property is what we claim, not a causal prediction from pre-training structure alone.","revision_made":"yes","referee_comment":"Appendix B, Fisher computation: The headline correlation ρ=0.82 between d_eff and the gap is computed across 300 configurations where both quantities are evaluated at the same trained parameters θ on overlapping held-out data (M=40 held-out inputs for d_eff, 2000 for the gap). The partial-correlation analysis shows d_eff screens off the entangling pattern, confirming mediation, but it does not address whether the d_eff–gap correlation itself is partly spurious: both could be jointly driven by the trained model's fit quality. The paper's multiple regression (Appendix B) loads almost entirely on d_eff, but this does not rule out the confound since d_eff is itself a function of the trained parameters. A cleaner test would be to compute d_eff at the initialization θ_0 (before training) and check whether it still predicts the gap—this would separate the structural (pre-training) component of"},{"response":"We agree with the framing the referee proposes. To be precise about what we can and cannot claim: Proposition 1 establishes that rank(F̂) is a structural property determined by the readout light-cone, which is fixed by the circuit architecture before training. The rank ordering none<linear<full is therefore a pre-training prediction. However, the actual d_eff values depend on the trained eigenvalues, which are fit-dependent. We will run the pre-training d_eff computation and report the result. If pre-training d_eff (or rank alone) predicts the gap, we will strengthen the language to 'structural predictor.' If it does not, we will reframe d_eff as a post-hoc diagnostic that correctly ranks circuits of equal parameter count—which is the property the bound provides and the property we use in the design recipe. Either way, we will be explicit about which framing the evidence supports. We note that the paper already hedges appropriately in some places (e.g., 'the bound is primarily a ranking certificate') but uses 'structural' language in others (e.g., 'entanglement as a structural complexity axis'); we will make the usage consistent with the evidence.","revision_made":"yes","referee_comment":"d_eff from the fit-dependent component. Proposition 1 provides structural support (rank is pre-training), but the actual d_eff values depend on trained eigenvalues, not just rank. If pre-training d_eff (or rank) still orders the gap, the causal claim is much stronger; if not, the paper should frame d_eff as a post-hoc diagnostic rather than a structural predictor."},{"response":"The referee raises an important clarification. The PAC-Bayes bound (Theorem 2) is on the expected loss gap: it bounds E[L_test] − E[L_train] for the loss L (cross-entropy in classification, negative reward in RL). The 'gap' we report in Tables 1, 3, 4, 5 is the accuracy gap (train_acc − test_acc) or reward gap, which is a monotone transform of the loss gap but not identical to it. The negative gap at N=64, none (−0.01) means test accuracy slightly exceeds train accuracy, which is perfectly consistent with the theory: the bound is on the expected loss, and a small negative accuracy gap corresponds to a small positive loss gap (since cross-entropy is convex and the model is not at the global optimum on either set). The bound is valid because it upper-bounds the loss gap, which is non-negative in expectation. We will clarify this in the manuscript: (i) state explicitly that the bound is on the loss gap, (ii) note that the reported accuracy/reward gaps are a monotone proxy for the loss gap, (iii) explain that negative accuracy gaps are consistent with the theory because they correspond to small positive loss gaps, and (iv) add the loss-gap values for the Table 3 rows so the reader can verify the bound directly on the object it certifies. We will also note that the negative gap at N=64, none is within one standard error of zero (as the manuscript already states), so it is not a meaningful deviation.","revision_made":"yes","referee_comment":"Table 3: The certified gap bound values appear to violate the stated validity in one cell. At N=16, full connectivity, the certified bound is 0.87 and the true gap is 0.381—valid. But at N=16, none connectivity, the certified bound is 0.49 and the true gap is 0.101—also valid. However, at N=64, none, the true gap is −0.01 and the bound is 0.25, which is valid but the negative gap raises a concern about the gap definition: if the gap is train_acc − test_acc, a negative value means test > train, which is unusual and suggests the 'gap' may not be the right object for the PAC-Bayes bound (which bounds the expected loss gap, not the accuracy gap). The paper should clarify whether the bound is on the loss gap or the accuracy gap, and whether the negative-gap rows are consistent with the theory."}],"tokens_in":22817,"tokens_out":2614,"duration_ms":304618,"standing_objections":[]},"desk_editor":{"model":"glm-5.2","letter":"The main thing to know: this paper identifies entanglement as an independent complexity axis for generalization in quantum policies, and the mechanism is clean. At fixed parameter count, adding entangling gates enlarges the backward light-cone of the readout observable, which raises the rank of the Fisher information matrix, which inflates the effective dimension d_eff = log det(I + γF̂), which loosens the PAC-Bayes bound. Proposition 1 and Lemma 2 give the structural argument; the experiments confirm the ordering none < linear < full across classification, contextual bandits, value-function regression, and IBM Heron hardware. This is a real contribution — Abbas et al. used effective dimension as a capacity measure but never connected it to entanglement, and Rodriguez-Grasa et al. gave PAC-Bayes bounds for quantum classifiers without isolating entanglement. The paper is also unusually honest about scope: it explicitly says the bound is a ranking certificate, not a tight numerical guarantee, and it flags the near-chance rows as ordering tests only, not gap-magnitude evidence. The global-readout control (Table 6) is a nice adversarial test — it confirms the light-cone mechanism by showing that activating parameters directly through the readout absorbs the entanglement contrast at small n. Two soft spots. First, the derandomization (Lemma 1) assumes the population loss Hessian equals the Fisher up to a residual that vanishes only in the well-fit regime. This is load-bearing because the deterministic trained model's gap is what gets reported, while the bound applies to the stochastic posterior-averaged predictor. The paper acknowledges this, and the deliberately-learnable regime (n=6, N=48, 5.8σ separation) shows the ordering survives where the model genuinely generalizes. But the near-chance rows at n≥10 and small-N bandit settings are exactly where the derandomization is least reliable, and some ordering evidence sits there. Second, the headline ρ=0.82 between d_eff and the gap has a shared-dependency concern: both are computed at the same trained parameters on held-out data, and the partial-correlation analysis confirms d_eff mediates the entangling pattern but does not rule out that both are jointly driven by fit quality. The γ-sensitivity analysis (ρ drops to 0.67 for 18 configs vs 0.82 for 300) suggests pooling across depths and sample sizes inflates the correlation. That said, the paper's actual claim rests on the ordering, not the correlation magnitude — and the ordering is supported both structurally and empirically. No code repository despite a reproducibility statement is a minor but real gap. This is for quantum ML researchers and anyone designing quantum policy ansätze. The structural result and the design recipe (budget entanglement against d_eff cost) are the lasting contributions. It deserves a serious referee — the core argument holds up, and the soft spots are the kind that a good referee can push the authors to address (tighter derandomization analysis, code release, disentangling the correlation) without undermining the central claim.","headline":"Solid paper connecting entanglement to generalization via Fisher effective dimension in quantum policies. The structural argument is genuinely new; the empirical work is extensive and honest about its limits. Two real soft spots, neither fatal. Deserves a serious referee.","tokens_in":23498,"tokens_out":1287,"would_cite":true,"duration_ms":128051,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["03.67.-a","03.67.Lx"],"model":"glm-5.2","headline":"Entanglement, not parameter count, drives quantum model overfitting","keywords":["parameterized quantum circuits","PAC-Bayes","generalization bound","Fisher information","entanglement","quantum reinforcement learning","effective dimension","readout light-cone"],"falsifier":"If two circuits with identical parameter count but different entanglement levels were found to have the same d_eff (e.g., because the extra entangling gates do not enlarge the readout light-cone), and yet still showed different generalization gaps, the central claim that d_eff is the governing complexity measure would be challenged.","tokens_in":22666,"feed_emoji":"🔗","tokens_out":1363,"duration_ms":160495,"temperature":0.7,"pith_summary":"This paper argues that entanglement is an independent axis of generalization complexity for parameterized quantum circuits used as policies and value functions. The core mechanism is that entangling gates enlarge the backward light-cone of the readout observable, which raises the rank of the Fisher information matrix and hence the Fisher effective dimension d_eff = log det(I + γF̂). Because d_eff enters the PAC-Bayes generalization bound as the complexity term, more entangled circuits provably incur larger train-to-test gaps—even when the raw parameter count is held identical. The authors validate this across supervised classification (4–16 qubits), a quantum contextual bandit, multi-step value-function regression, real benchmark datasets (Iris, Breast Cancer, Wine, Digits), and an IBM Heron quantum processor, finding that d_eff is the single strongest predictor of the generalization gap (Spearman ρ = 0.82), well ahead of parameter count (ρ = 0.45) and learned parameter norm (ρ = 0.73). The bound functions primarily as a ranking certificate: it correctly orders circuits of identical parameter count by their generalization gap, which no parameter-counting bound can do. The practical prescription is to budget entanglement against its effective-dimension cost rather than maximizing expressivity.","feed_headline":"Entanglement, not parameter count, drives quantum model overfitting","feed_subtitle":"A PAC-Bayes bound shows entangling gates inflate Fisher effective dimension, provably worsening generalization at fixed circuit size across ","key_machinery":"The Fisher effective dimension d_eff = log det(I + γF̂), where F̂ is the loss Fisher information matrix of the quantum policy's Bernoulli readout head, evaluated at the trained parameters. The light-cone structure of the readout observable governs the rank of F̂, and entanglement enlarges that light-cone.","core_discovery":"The paper identifies a precise causal chain: entangling gates enlarge the readout light-cone (Lemma 2), which raises the rank of the Fisher information matrix (Proposition 1), which inflates the Fisher effective dimension d_eff = log det(I + γF̂), which tightens (loosens) the PAC-Bayes generalization bound (Theorem 2, Equation 3). This chain operates entirely at fixed parameter count, making entanglement an independent complexity axis that parameter-counting bounds are structurally blind to. The derandomization step (Lemma 1) connects the stochastic posterior-averaged predictor that the bound applies to, with the deterministic trained model whose gap is actually measured, showing the two gap","pith_inferences":["If d_eff is the governing complexity measure, then circuit architectures that achieve high expressivity with low Fisher rank—e.g., structured ansätze that keep most parameters outside the readout light-cone—should generalize better than hardware-efficient ansätze of equal parameter count, a testable prediction for ansatz design.","The light-cone mechanism suggests that readout choice is a complexity knob co-equal with entanglement: a designer could trade entangling gates for a modestly larger readout observable to control d_eff, which the paper's global-readout experiment supports but does not fully systematize.","The erosion of d_eff under barren plateaus implies a non-monotonic relationship between circuit width and generalization: adding qubits could initially worsen generalization (more entanglement, higher d_eff) but eventually improve it (barren plateau suppresses d_eff), suggesting an optimal width for generalization at fixed depth.","The derandomization gap bound τ²d/(2γ) depends on the posterior hyperparameters; if γ is chosen data-adaptively (via the union-bound grid the paper mentions), the derandomization cost shrinks, potentially tightening the connection between the stochastic bound and the deterministic gap in regimes where the model fits poorly."],"forward_implications":["Quantum circuit ansatz design should optimize a Pareto frontier of validation performance versus d_eff, not versus parameter count or gate count, since d_eff is what the generalization bound actually charges for.","For local single-qubit readouts (standard in quantum RL), entanglement is the dominant structural driver of d_eff; switching to a global readout can partially substitute for entanglement by directly activating parameters, confirming the light-cone mechanism rather than refuting it.","Barren plateaus erode the mechanism's substrate: at fixed depth, d_eff decreases with qubit number because Fisher eigenvalues are suppressed faster than rank grows, so the entanglement–generalization trade-off is sharpest in the shallow, trainable regime and weakens at scale.","The ranking certificate property means d_eff can serve as a model-selection criterion that distinguishes circuits of identical size, which parameter-count-based generalization bounds cannot do.","On real quantum hardware (IBM Heron), the entanglement–gap ordering is preserved under genuine gate and readout noise, suggesting the trade-off is not a simulation artifact."],"fun_headline_variants":["Entanglement inflates Fisher dimension, worsening quantum model generalization","Entangling gates raise Fisher effective dimension, hurting quantum generalization","Entanglement, not parameter count, governs quantum model generalization","Fisher effective dimension links entanglement to worse quantum generalization","Quantum generalization bound ranks circuits by entanglement, not parameter count"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The derandomization step (Lemma 1) assumes the population loss Hessian equals the Fisher matrix up to a residual that vanishes when the model fits the data well. In the near-chance regime—where some of the paper's ordering evidence comes from—models do not fit well, so the residual may not vanish, and the deterministic gap may not track the PAC-Bayes bound as tightly as claimed.","fun_headline_variants_meta":{"raw":{"variants":["Entanglement inflates Fisher dimension, worsening quantum model generalization","Entangling gates raise Fisher effective dimension, hurting quantum generalization","Entanglement, not parameter count, governs quantum model generalization","Fisher effective dimension links entanglement to worse quantum generalization","Quantum generalization bound ranks circuits by entanglement, not parameter count","Entanglement worsens quantum generalization by inflating Fisher effective dimension"]},"model":"glm-5.2","effort":"high","cost_usd":0.0,"raw_usage":{"total_tokens":1102,"prompt_tokens":648,"completion_tokens":454,"prompt_tokens_details":null},"tokens_in":648,"tokens_out":454,"duration_ms":8306,"temperature":1.0,"reasoning_tokens":422,"cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-08T12:33:01.622534+00:00","model_set":{"reader":"glm-5.2"},"falsifier":"If two circuits with identical parameter count but different entanglement levels were found to have the same d_eff (e.g., because the extra entangling gates do not enlarge the readout light-cone), and yet still showed different generalization gaps, the central claim that d_eff is the governing complexity measure would be challenged.","supporting_citations":[],"review_version":1}