{"id":"f54e93a2-2b8e-42fc-8ac7-d169cf7bf5be","arxiv_id":"2507.05972","paper_version":3,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":8.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"A single efficiently computable function achieves optimal pseudoentropy and optimal predictive hardness simultaneously for a broad family of entropy measures.","lead":"This paper proves a unified theorem connecting computational hardness to pseudoentropy, covering many entropy measures at once with a single witness function. It improves the complexity from exponential to polynomial in the alphabet size, and shows that extra assumptions on the entropy measure are necessary for such efficiency.","discovery_kind":"unification","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The advertised inclusion of Shannon entropy is not formally established: Section 5's transformation σφ leaves an uncontrolled term involving <s−σφ(s), ∇φ(σφ(s))> in identity (9), and the claimed 'exact same way' proof omits the required constant accounting; a fix appears possible but requires a…","rationale":"The core nonuniform theorem (Theorem 1.10) is carefully argued: the mirror-descent boosting proof of Theorem 3.1 is standard and the Fenchel-Young potential accounting is sound; the lower bound (Theorem 1.12) genuinely demonstrates necessity of an efficiency assumption on φ. The concern I identify is not a refutation of Theorem 1.10 itself, but a gap in the paper's advertised extension to Shannon entropy. Section 5's Theorem 5.1 requires nontrivial constant bookkeeping because the transformation σφ shifts the point at which the Bregman divergence is evaluated, introducing a term <s−σφ(s), ∇φ(σφ(s))> that the calibration condition from Theorem 3.1 does not directly control. This is exactly the kind of omitted detail that justifies the reader's CONDITIONAL verdict. I do not think it moves the verdict to ACCEPT or REJECT: if the constant accounting closes (which seems likely), the paper's central claims stand, and if it does not, the Shannon-specific claim would need to be qualified. My agreement with the reader is partial because the reader identified the bounded-subgradient assumption as weakest, whereas I am flagging that the paper's proposed workaround for the unbounded case—the one that substantiates 'including Shannon entropy' in the abstract—is asserted rather than proven at the needed level of precision.","tokens_in":31138,"tokens_out":29420,"duration_ms":277667,"concrete_test":"Write a self-contained proof of Theorem 5.1 for the Shannon entropy φ(v)=Σ v_i ln v_i and σφ(v)=(1−δ)v+δu. Fix B=ln(L/δ)+1, apply Theorem 3.1 to the scaled weight family R={v↦∇φ(σφ(v))/B} with calibration slack ε_cal such that ε_cal+2δB≤ε (e.g. ε_cal=ε/4, δ=ε/(4B)), and verify that the returned s satisfies Hφ(σφ∘s)−Hφ(g*)−Dφ(g*∥σφ∘s)=<g*−s,∇φ(σφ(s))>+<s−σφ(s),∇φ(σφ(s))> ≥ −ε, and that the resulting T' has the form in (40) with only polylog(L/ε)/ε^2 factors. If the constants cannot be closed, the paper's Shannon special case requires a corrected statement or proof.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The reader's weakest assumption is the bounded-subgradient/computability condition in Theorem 1.10. Since Theorem 1.12 shows that assumption is necessary, the remaining load-bearing point is the paper's workaround for Shannon entropy, which the abstract advertises as a special case. Section 5 (Theorem 5.1) allows an arbitrary transformation σφ and guarantees Hφ(σφ∘s)−Hφ(g*) ≥ Dφ(g*∥σφ∘s)−ε with s∈Ind(g*;T,ε)∩Time(T'). The intended proof applies Theorem 3.1 with R={rφ}, where rφ(v) approximates ∇φ(σφ(v)). This yields <s−g*, rφ∘s> ≤ ε/2, i.e. <s−g*, ∇φ(σφ(s))> ≤ ε/2 after scaling. But identity (9), applied to σφ∘s, gives Hφ(σφ∘s)−Hφ(g*)−Dφ(g*∥σφ∘s) = <g*−σφ(s), ∇φ(σφ(s))> = <g*−s, ∇φ(σφ(s))> + <s−σφ(s), ∇φ(σφ(s))>. The calibration condition controls only the first term. For the paper's Shannon choice σφ(v)=(1−δ)v+δu, the second term can be as negative as −2δB, where B≈ln(L/δ). To make (41) hold, one must choose ε_cal and δ so that ε_cal+2δB≤ε, and one must also account for the fact that ∇φ(σφ(v)) has norm B rather than 1. The paper states that Theorem 5.1 is proved 'in the exact same way' as Theorem 1.10 and gives none of this accounting. The gap is plausibly fixable, e.g. by taking ε_cal≈ε/4 and δ≈ε/(4B), with the extra polylog(L/ε) factors absorbed into poly(1/ε, log L); but as written, a central advertised consequence is not formally derived. The uniform version has the same omission: Theorem 4.3 formally requires bounded {∇φ}, and the promised unbounded-subgradient generalization is only asserted.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes a unified pseudoentropy characterization for arbitrary families of convex entropy functions, aiming to cover Shannon entropy, min-entropy, and Rényi-type entropies in a single framework. The key identity (9) rewrites the difference between the pseudoentropy gap and the Bregman divergence as an inner product involving the subgradient of the entropy function at the simulator; the authors show that enforcing multiaccuracy together with weight-restricted calibration suffices to control this term. They prove a nonuniform regularity lemma (Theorem 3.1) via multiplicative weights, yielding a single low-complexity simulator satisfying both conditions, and then derive nonuniform (Theorem 1.10) and uniform (Theorem 1.11/4.3) pseudoentropy characterizations. They also prove an exponential lower bound (Theorem 1.12) showing that a computational assumption on the subgradient of the entropy function is necessary. The advertised extension to Shannon entropy is routed through Section 5, where a transformation sigma_phi is introduced; it is in this section that the manuscript has a significant proof gap.","tokens_in":31639,"tokens_out":16977,"duration_ms":175674,"significance":"The paper's core algebraic identity and the multiplicative-weights construction are rigorous and elegant, and the lower-bound proof via probabilistic counting over EFF85 designs is convincing. If the Shannon-entropy workaround is made fully rigorous, the paper would constitute a substantial advance: it would provide a single simulator that simultaneously witnesses pseudoentropy and approximability for a broad family of entropies, with polynomial dependence on the alphabet size instead of the doubly exponential cost of multicalibration, and it would prove the necessity of the gradient-complexity assumption. The conceptual reduction to weight-restricted calibration is also an important contribution. However, because the advertised inclusion of Shannon entropy is not formally derived as written, the central claim of the paper is not yet fully established.","major_comments":[{"comment":"The claim that Theorem 5.1 is proved in the exact same way as Theorem 1.10 is not justified for the Shannon-entropy transformation sigma_phi(v) = (1 - delta)v + delta*u. Applying Theorem 3.1 with R = {r_phi} approximating grad(phi) composed with sigma_phi controls <s - g*, r_phi(s)> and hence, after the l_infinity approximation, <s - g*, grad(phi)(sigma_phi(s))>. But identity (9), applied to sigma_phi composed with s, gives <g* - sigma_phi(s), grad(phi)(sigma_phi(s))> = <g* - s, grad(phi)(sigma_phi(s))> + <s - sigma_phi(s), grad(phi)(sigma_phi(s))>, and the second term is not controlled by the regularity lemma. For the Shannon choice, this uncontrolled term can have magnitude Theta(delta log(L/delta)), so inequality (41) does not follow unless the perturbation delta and the calibration error are chosen with explicit parameter accounting (e.g., epsilon_cal + O(delta log(L/delta)) <= epsilon). The subsequent implications in Section 5 address only indistinguishability and circuit complexity of sigma_phi composed with s, not this term. This is a load-bearing gap because the abstract advertises Shannon entropy as a special case; a fix with explicit constants should be supplied.","section":"Section 5 (Theorem 5.1)"},{"comment":"The uniform result Theorem 4.3, which is the formal version of Theorem 1.11, only applies when every phi in Phi has subgradients mapping into [-1,1]^L, because Definition 4.1 and Claim 4.2 concern families R of functions from Delta_L to [-1,1]^L. Shannon entropy does not satisfy this condition, and the final sentence of Section 5 states only that 'similarly, we can prove a generalized version of Theorem 4.3 using Theorem 3.10 to handle unbounded subgradients in the uniform setting' without stating the generalized theorem or giving any proof. Since the paper claims that the uniform characterization encompasses the uniform versions of Theorems 1.2 and 1.3 of Vadhan and Zheng, which include Shannon entropy, this omission is load-bearing and must be filled in, together with the nonuniform parameter accounting described in the previous comment.","section":"Section 4.2 and Section 5 (Theorem 4.3)"}],"minor_comments":[{"comment":"Item 3 states that ||g_hat_k - g_k||_infty <= epsilon/10, but Algorithm 1 (Lines 3, 10, and 15) requires the l1 bound ||g_hat_k - g_k||_1 <= epsilon/10, and the proof of Lemma 3.2 uses the l1 bound. The coordinatewise precision supplied by Lemma 3.4 does give the l1 bound, but the statement of Lemma 3.3 should be corrected.","section":"Lemma 3.3, item 3"},{"comment":"The proof of Lemma 3.11 omits the analysis of the running time and evaluation time of the constructed function s, saying only that this follows 'similarly to Lemma 3.3'. In the uniform setting the construction involves sample access and oracle invocations, so the details are not literally identical; a short argument should be included.","section":"Lemma 3.11"},{"comment":"The same symbol epsilon is used both for the error parameter in Theorem 5.1 and for the perturbation parameter in sigma_phi(v) = (1 - epsilon)v + epsilon*u. Since the two parameters must be chosen with different scales to close the argument, using distinct names (e.g., delta for the perturbation) is necessary for clarity.","section":"Section 5"},{"comment":"There are several typographical errors, including 'addtionally' in Theorem 3.10, 'strenthening' in Section 1.5, 'diference' in Section 1.5, and 'un-transformeds' in Section 5. These should be corrected.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The paper is technically strong and the core results are likely publishable after the Section 5 gap is repaired. The self-citation pattern is not problematic given the direct prior work, and the paper is self-contained against external benchmarks. I would support acceptance once the authors provide the missing parameter accounting for the Shannon-entropy transformation and the corresponding uniform generalization."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, here's my read on arXiv:2507.05972 (Hu–Vadhan). The headline result is genuinely new: one theorem that simultaneously characterizes pseudoentropy for a whole family of convex entropy functions with a single simulator, using multiaccuracy plus weight-restricted calibration instead of full multicalibration. The improvement from doubly exponential in ℓ to polynomial in L is real, and the lower bound (Theorem 1.12) shows the efficiency assumption on the entropy family is necessary. I checked the core algebra: identity (9) is correct, the multiplicative-weights analysis in Lemma 3.2 is sound, and the lower bound uses a clean counting argument. For entropy functions with uniformly bounded computable subgradients—min-entropy and Rényi-2 collision entropy included—Theorem 1.10 appears to go through as stated.\n\nThe soft spot is Section 5's handling of Shannon entropy, whose gradient is unbounded near the simplex boundary. The paper introduces a transformation σφ(v) = (1-ε)v + εu and claims Theorem 5.1 can be proved 'in the exact same way' as Theorem 1.10. That claim doesn't hold as written. Applying the regularity lemma to the composed gradient ∇φ(σφ(v)) controls ⟨s-g*, ∇φ(σφ(s))⟩, but identity (9) applied to σφ∘s includes an extra term ⟨s-σφ(s), ∇φ(σφ(s))⟩. For the chosen σφ, this term can be as negative as about ε ln(L/ε), which swamps the claimed ε error. Choosing σφ with a smaller mixing parameter, say δ ≈ ε/ln(L/ε), should fix it, at the cost of an extra log factor in the complexity. The uniform version, Theorem 4.3, has the same omission: the bounded-gradient condition is assumed, and the unbounded generalization is only asserted.\n\nSo: the main machine is in good shape, but a central advertised special case is currently not formally derived. This is a gap a referee would want closed, not a fatal flaw. I'd send it out—the core theorems deserve referee time, and the fix is likely a few lines of constants.\n\nWho should read it: anyone working on hardness-randomness equivalences, pseudoentropy, or the regularity-lemma/calibration literature. I'd cite the main theorem once the Shannon case is repaired.","headline":"Strong unified pseudoentropy result with a real but fixable gap in the Shannon-entropy case.","tokens_in":32173,"tokens_out":5528,"would_cite":true,"duration_ms":54235,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68Q15","68Q17","94A17"],"pacs":[],"model":"deepseek-v4-flash","headline":"A single efficiently computable simulator is simultaneously entropy-optimal and divergence-optimal for every entropy notion in a family with efficiently computable bounded subgradients, unifying hardness–randomness equivalences and…","keywords":["pseudoentropy","hardness–randomness equivalence","computational indistinguishability","weight-restricted calibration","multiaccuracy","Bregman divergence","multiplicative weights","exponential lower bound"],"falsifier":"An explicit counterexample would settle the matter: exhibit a single convex function $\\varphi : \\Delta_L \\to \\mathbb{R}$ whose subgradient is computable by a $\\operatorname{poly}(L)$-size circuit and a distribution pair $(\\mu, g^*)$ such that every function $s$ that is $0.05$-indistinguishable from $g^*$ against the identity distinguisher and satisfies $H_\\varphi(s) \\geq H_\\varphi(g^*) - 0.05$ requires circuit size $\\exp(\\Omega(L))$; that would refute the claimed polynomial bound of Theorem 1.10. A direct computational check on small instances (for example $L = 2, 4$ with $\\varphi(v) = \\lVert v \\rVert_2^2$) of whether the paper's multiplicative-weights algorithm satisfies $H_\\varphi(s) - H_\\varphi(g^*) \\geq D_\\varphi(g^*\\|s) - \\varepsilon$ at the stated circuit bound would expose any hidden constant or $\\varepsilon$-dependence not captured by the theorem.","tokens_in":30934,"feed_emoji":"🎲","tokens_out":16163,"duration_ms":163194,"temperature":0.7,"pith_summary":"The paper sets out to prove that the many known hardness–randomness equivalences — theorems asserting that a distribution which is hard to predict or approximate must have an efficiently indistinguishable copy with higher entropy — are all instances of one statement. The central claim is that for any family $\\Phi$ of convex entropy notions whose subgradients are bounded and computable to $\\ell^\\infty$ accuracy $\\varepsilon/4$ in nonuniform time $T_\\Phi$, a single simulator $s$ exists that is simultaneously $\\varepsilon$-indistinguishable from the true distribution $g^*$, entropy-optimal among indistinguishable simulators, and divergence-optimal among efficient approximators, for every entropy in the family at once. This unifies the characterizations for average min-entropy, Shannon entropy, and collision probability under one theorem and one proof, and it reduces the dependence on the alphabet size from doubly exponential in $\\ell$ to polynomial in $L = 2^\\ell$. The paper also claims this unification is essentially optimal: without the subgradient-efficiency assumption, even a single trivial distinguisher forces the simulator to have exponential circuit size.","feed_headline":"One simulator settles every pseudoentropy equivalence","feed_subtitle":"Fairness-style calibration plus indistinguishability cuts alphabet cost to polynomial and unites Shannon and min entropy.","key_machinery":"The load-bearing object is the identity $H_\\varphi(s) - H_\\varphi(g^*) - D_\\varphi(g^*\\|s) = \\langle g^* - s, \\nabla\\varphi \\circ s\\rangle$ (equation (9) of the paper), which converts the desired coincidence of entropy gap and divergence gap into two inner-product conditions: multiaccuracy, meaning $|\\langle s - g^*, f\\rangle| \\leq \\varepsilon$ against every low-complexity distinguisher $f$, and weight-restricted calibration, meaning $\\langle s - g^*, r_\\varphi \\circ s\\rangle \\leq \\varepsilon$ for weight functions $r_\\varphi$ that approximate the subgradient $\\nabla\\varphi$ to accuracy $\\varepsilon/4$. The paper proves an enhanced regularity lemma — extending the Complexity-Theoretic Regularity Lemma and the Leakage Simulation Lemma from boolean to $L$-ary outputs — by multiplicative weights / mirror descent: starting from the uniform distribution, each violated condition triggers an additive shift $h \\leftarrow h - \\varepsilon f$ or $h \\leftarrow h - \\varepsilon (r \\circ \\hat{g})$, and a Fenchel–Young divergence potential with the softmax map bounds the number of updates by $O((\\log L)/\\varepsilon^2)$, yielding the stated circuit bound. Unbounded subgradients, as with Shannon entropy, are handled by perturbing distributions away from the simplex boundary before evaluation. The lower-bound half uses a combinatorial design of exponentially many near-disjoint subsets together with a probabilistic counting argument over a random masking function, forcing the simulator to encode $\\exp(\\Omega(L))$ distinct behaviors.","core_discovery":"The paper's discovery is that the equivalence between computational hardness and pseudoentropy can be witnessed by one function at once for an entire family of entropy measures. Formally, for every family $\\Phi$ of convex functions $\\varphi : \\Delta_L \\to \\mathbb{R}$ whose subgradients lie in $[-1,1]^L$ and can be computed to $\\ell^\\infty$ accuracy $\\varepsilon/4$ in nonuniform time $T_\\Phi$, and for every target $g^* : \\{0,1\\}^n \\to \\Delta_L$, there is a single simulator $s$ that is $(T,\\varepsilon)$-indistinguishable from $g^*$, computable in time $T' = O((T+T_\\Phi)\\log L / \\varepsilon^2 + L\\cdot\\operatorname{poly}(1/\\varepsilon, \\log L))$, and satisfies $H_\\varphi(s) - H_\\varphi(g^*) \\geq D_\\varphi(g^*\\|s) - \\varepsilon$ for every $\\varphi \\in \\Phi$. Thus the pseudoentropy gap and the Bregman-divergence hardness gap coincide up to $\\varepsilon$, achieved by the same function $s$, for every entropy notion in the family simultaneously. The matching converse shows every achievable pseudoentropy gap is at most the best efficient divergence approximation plus $\\varepsilon$. The paper further establishes that the efficiency hypothesis is load-bearing: an exponential-in-$L$ lower bound holds even against a single trivial distinguisher when the subgradient of one convex function is not efficiently computable, and this forces the doubly exponential dependence on $\\ell$ for multicalibration and calibrated multiaccuracy.","pith_inferences":["A trade-off the paper leaves implicit is the role of the $\\varepsilon/4$ subgradient accuracy: treating it as a free parameter $t$ suggests a conjecture that $T'$ grows polynomially in $1/t$, which could be tested numerically for collision probability on small $L$.","Equation (9) reads as a design principle: any pair of conditions that controls $\\langle s - g^*, f\\rangle$ and $\\langle s - g^*, r \\circ s\\rangle$ yields a pseudoentropy characterization, so other fairness-inspired conditions could generate new characterizations with different complexity profiles.","The lower bound's convex function is defined through a maximum over a huge design set and depends on a hidden random mask, so it is not explicitly presentable; finding an explicit, small-circuit-presentable $\\varphi$ with the same $\\exp(\\Omega(L))$ effect would pin down exactly what subgradient efficiency buys.","Cryptographic extraction steps are currently analyzed against a single entropy measure at a time, such as min-entropy for hardcore bits or Shannon entropy for pseudorandom-generator constructions; the universal simulator suggests one extraction step could satisfy all such analyses simultaneously whenever the relevant subgradients are efficiently computable."],"forward_implications":["Shannon entropy, average min-entropy, and collision probability become corollaries of one theorem, with the very same simulator witnessing the pseudoentropy gap for all of them at once.","The single function $s$ is simultaneously the entropy-optimal indistinguishable simulator and the divergence-optimal low-complexity approximator of $g^*$ — the omnipredictor phenomenon — for every $\\varphi$ in the family.","The alphabet-size dependence improves from doubly exponential in $\\ell$ to polynomial in $L = 2^\\ell$, and the paper proves that doubly exponential dependence is unavoidable for multicalibration and even for the weaker calibrated multiaccuracy.","When the subgradient family is weakly agnostically learnable, the characterization holds in the uniform model of computation with one algorithm for all input lengths, encompassing the uniform min-entropy and Shannon-entropy results.","The reverse direction yields a general upper bound: for every $\\varphi$, the best efficient approximation of $g^*$ under $D_\\varphi$ is within $\\varepsilon$ of the largest pseudoentropy gap any indistinguishable simulator can achieve."],"supporting_citations":[{"why":"supplies the Complexity-Theoretic Regularity Lemma and its boosting framework, which the paper extends to larger alphabets.","marker":"[TTV09]"},{"why":"the Leakage Simulation Lemma for $L > 2$ outputs that the paper's enhanced regularity lemma builds on and strengthens.","marker":"[JP14]"},{"why":"the multicalibration-based pseudoentropy characterizations that this work generalizes, and whose doubly exponential alphabet cost is proved inherent for multicalibration.","marker":"[CDV24]"},{"why":"established the min-entropy and Shannon-entropy pseudoentropy characterizations, including uniform versions, that Theorem 1.10 unifies.","marker":"[VZ12]"},{"why":"Theorem 1.4's characterization for differentiable strictly convex entropy functions, which this work extends to non-differentiable $\\varphi$ and whole families.","marker":"[Zhe14]"},{"why":"introduced weight-restricted (weighted) calibration, the fairness notion combined with multiaccuracy in the proof.","marker":"[GKSZ22]"},{"why":"introduced multicalibration in the fairness literature, the stronger condition shown to be overkill for pseudoentropy.","marker":"[HKRR18]"},{"why":"constructed omnipredictors from calibrated multiaccuracy, a consequence recovered and sharpened by the universal simulator.","marker":"[GHK+23]"},{"why":"supplies the combinatorial design of exponentially many low-intersection subsets used in the exponential lower bound.","marker":"[EFF85]"}],"fun_headline_variants":["Unified pseudoentropy: one simulator for every entropy notion","Fairness tool unifies hardness-randomness equivalences","Single function achieves all pseudoentropy gaps at once","Exponential alphabet cost shown inevitable for multicalibration","Enhanced leakage simulation yields exponential improvement"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The polynomial-in-$L$ construction rests entirely on assuming that for every entropy $\\varphi$ in the family, a subgradient of $\\varphi$ bounded in $[-1,1]^L$ can be computed to $\\ell^\\infty$ accuracy $\\varepsilon/4$ by a nonuniform circuit of size $T_\\Phi$, and the paper proves that if this assumption fails, even a single trivial distinguisher forces the simulator to exponential circuit size.","fun_headline_variants_meta":{"raw":{"variants":["Unified pseudoentropy: one simulator for every entropy notion","Fairness tool unifies hardness-randomness equivalences","Single function achieves all pseudoentropy gaps at once","Exponential alphabet cost shown inevitable for multicalibration","Enhanced leakage simulation yields exponential improvement"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000278,"raw_usage":{"total_tokens":1758,"prompt_tokens":1152,"completion_tokens":606,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":768,"completion_tokens_details":{"reasoning_tokens":532}},"tokens_in":768,"tokens_out":606,"duration_ms":7210,"temperature":1.0,"reasoning_tokens":532,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T19:14:51.052418+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"An explicit counterexample would settle the matter: exhibit a single convex function $\\varphi : \\Delta_L \\to \\mathbb{R}$ whose subgradient is computable by a $\\operatorname{poly}(L)$-size circuit and a distribution pair $(\\mu, g^*)$ such that every function $s$ that is $0.05$-indistinguishable from $g^*$ against the identity distinguisher and satisfies $H_\\varphi(s) \\geq H_\\varphi(g^*) - 0.05$ requires circuit size $\\exp(\\Omega(L))$; that would refute the claimed polynomial bound of Theorem 1.10. A direct computational check on small instances (for example $L = 2, 4$ with $\\varphi(v) = \\lVert v \\rVert_2^2$) of whether the paper's multiplicative-weights algorithm satisfies $H_\\varphi(s) - H_\\varphi(g^*) \\geq D_\\varphi(g^*\\|s) - \\varepsilon$ at the stated circuit bound would expose any hidden constant or $\\varepsilon$-dependence not captured by the theorem.","supporting_citations":[],"review_version":1}