[ { "slug": "big-bird-infinity-norm", "title": "Big Bird: Transformers for Longer Sequences", "issueTitle": "Failure of the L∞ approximation lemma", "year": 2020, "category": "Transformers & in-context learning", "status": "counterexample", "version": "v2", "location": "Appendix A.2.3, Lemma 5: arXiv v2 PDF page 21; published NeurIPS supplement PDF page 5 (printed page 20).", "originalClaim": "Every network in the modified hardmax/piecewise-linear class can be approximated arbitrarily accurately by an ordinary softmax/ReLU transformer, including at p=infinity.", "formula": "\\forall\\bar g\\in\\overline{\\mathcal T}^{2,1,1},\\quad 1\\le p\\le\\infty:\\quad\\exists g\\in\\mathcal T^{2,1,4},\\quad d_p(g,\\bar g)\\le\\varepsilon/3.", "assumptions": [ "The modified class permits hardmax and piecewise-linear activations with discontinuities, inherited from the cited Yun et al. construction.", "The ordinary network has finite weights and finitely many softmax/ReLU layers, so its output is continuous on the compact input cube.", "The selected claim includes p=infinity. The main-text Theorem 1 instead states 11/2.", "formula": "\\bar g_1(X)=X_{11}+\\mathbf 1\\{X_{11}>1/2\\}." }, { "title": "Use continuity of the approximant", "text": "Fix the other input coordinates in the interior of the cube. The first coordinate of any finite ordinary transformer has the same limit from the left and right at x=1/2. Call this common value c. The target limits are 1/2 and 3/2.", "formula": "\\max\\{|c-1/2|,|c-3/2|\\}\\ge 1/2." }, { "title": "The essential supremum also fails", "text": "For any proposed error strictly below 1/2, one side of the jump has error above that threshold. Continuity on that open side gives an open set of inputs with the same failure. Thus the obstruction has positive measure; it is not just the value assigned at the discontinuity.", "formula": "d_\\infty(g,\\bar g)\\ge 1/2." }, { "title": "Separate the endpoint from finite p", "text": "A transition region of width delta can have finite-p error tending to zero as delta shrinks. That argument does not make the maximum error vanish.", "formula": "\\|e_\\delta\\|_p=O(\\delta^{1/p})\\ \\ (p<\\infty),\\qquad\\|e_\\delta\\|_\\infty\\ge 1/2." } ], "counterexample": "For ε=1, Lemma 5 requires an approximation error at most 1/3. The discontinuous target has distance at least 1/2 from every continuous approximant in the uniform norm.", "witnessFormula": "\\bar g(x)=x+\\mathbf1\\{x>1/2\\},\\qquad\\inf_{g\\in C([0,1])}\\|g-\\bar g\\|_\\infty\\ge\\frac12>\\frac13.", "impact": "Refutes the infinity endpoint of Lemma 5 and invalidates that intermediate-network approximation route to the appendix endpoint theorem. It does not refute uniform universality for continuous target functions proved by a different construction.", "practical": "No experimental result, training run, or implementation benchmark was re-evaluated by this counterexample.", "repair": "Restrict the modified-to-ordinary approximation lemma to finite p. For a uniform approximation theorem on continuous targets, supply a separate construction that controls the transition regions in the uniform norm. The cited Yun et al. Lemma 9 explicitly has p=2||w*||, eta<=2/L_f, both initial iterates are zero, each additive perturbation has norm at most epsilon, and L<=R/(2epsilon)." ], "plainExplanation": "The local-smoothness argument identifies a bound relative to the minimizer with a bound relative to the origin. Starting from zero, nonexpansiveness yields an origin-radius bound involving 2‖w*‖, rather than ‖w*‖. Under the stated radius condition, an admissible perturbation can consequently leave the smoothness domain. A one-dimensional convex C³ objective gives an accumulated error of 2.1 after two steps, exceeding the claimed bound of 0.2.", "steps": [ { "title": "Choose a locally smooth convex objective", "text": "Take the quadratic centred at 1 and add a positive quartic term only to the right of 2. The function is globally convex and C3; on [-2,2] it is exactly the quadratic, with gradient Lipschitz constant 1. Its unique minimizer is 1.", "formula": "f(w)=\\tfrac12(w-1)^2+250(w-2)_+^4,\\quad f\\prime(w)=w-1+1000(w-2)_+^3." }, { "title": "Check every numerical hypothesis", "text": "Use R=2, L_f=1, eta=2, epsilon=0.1, and two steps. R=2|w*|, eta=2/L_f, and 2<=R/(2epsilon)=10. The perturbations are (0.1,0).", "formula": "(R,L_f,\\eta,\\varepsilon,L)=(2,1,2,0.1,2),\\quad(e^0,e^1)=(0.1,0)." }, { "title": "Compute the exact trajectory", "text": "The first gradient is -1, so exact gradient descent moves from 0 to 2. At 2 the gradient is 1, so the next exact iterate returns to 0.", "formula": "w^0_{\\rm GD}=0,\\quad w^1_{\\rm GD}=2,\\quad w^2_{\\rm GD}=0." }, { "title": "Compute the perturbed trajectory", "text": "The first perturbed iterate is 2.1, already outside B(2). There the gradient is 1.1+1=2.1. The next perturbed iterate is therefore -2.1.", "formula": "\\widehat w^1=2.1,\\quad f\\prime(2.1)=2.1,\\quad\\widehat w^2=2.1-2(2.1)=-2.1." }, { "title": "Compare with both claimed bounds", "text": "The final error is 2.1 while the claimed accumulated-error bound is 0.2. Already at the first step, the claimed radius bound is 1.1 while the iterate is 2.1. The proof’s final induction instead bounds distance to w*, which does not establish membership in the required ball about zero.", "formula": "|\\widehat w^2-w^2_{\\rm GD}|=2.1>0.2;\\quad|\\widehat w^1|=2.1>1.1." } ], "counterexample": "For the specified convex objective, the exact iterates are (0, 2, 0), whereas the perturbed iterates are (0, 2.1, −2.1). Both the radius assertion and the two-step error bound fail.", "witnessFormula": "f(w)=\\frac12(w-1)^2+250(w-2)_+^4,\\quad 0\\to2\\to0,\\quad0\\to2.1\\to-2.1,\\quad2.1>2\\varepsilon=0.2.", "impact": "Lemma D.1 is false under its stated local-smoothness assumptions. The published text explicitly uses it to justify linear error accumulation in Theorem D.1; that use needs a valid domain argument. This is not a counterexample to every transformer construction or to the broad claim that transformers can implement learning algorithms.", "practical": "The counterexample does not evaluate the paper’s empirical in-context learning results.", "repair": "For a fixed horizon L, a sufficient domain condition is 2||w*||+L epsilon<=R. Replace the origin-radius assertion by ||w_hat^l||<=2||w*||+l epsilon, and retain the local nonexpansiveness induction within that ball. Alternatively assume global L_f-smoothness for the accumulated-error claim; that alone does not restore the erroneous R/2 origin bound. From initialization zero, nonexpansiveness gives ||w_hat^l-w*||<=||w*||+l epsilon. The triangle inequality then gives the stated corrected origin bound. Under the stronger headroom condition all update inputs remain in the smoothness region, making the induction legitimate.", "repairFormula": "2\\|w^*\\|+L\\varepsilon\\le R\\quad\\Longrightarrow\\quad\\|\\widehat w^\\ell\\|\\le2\\|w^*\\|+\\ell\\varepsilon\\le R.", "repairProofFormula": null, "implementation": "Enlarge the smoothness and update-approximation region, or use a separately analysed projected algorithm. Projection changes the update rule and is not a silent repair of the original construction.", "caveats": [ "The witness is for the supporting stability lemma, not a complete refutation of Theorem D.1’s existential network claim.", "Global smoothness would restore the linear accumulated-error estimate but not the stated origin-radius estimate.", "The downstream ICLR 2026 paper is cited only as an observed reuse; no finding about its full theorems is made." ], "correctionSearch": "2026-09-09 Fresh arXiv history confirms latest v2, 6 July 2023. Re-read exact arXiv and published NeurIPS statements and complete short proof; visually checked the published statement. Exact-title searches with correction, erratum, mistake, Lemma 14, Lemma D.1, and the descriptive lemma name. Fresh all-state author repository issue listing contained zero issues. Author publication page checked. Indexed OpenReview versions and a later ICLR 2026 lemma repeat the same assumptions and bound; no repair identified. No explicit qualification, replacement lemma, or addressing discussion for this domain escape was found in the inspected material. OpenReview direct PDF and API access returned 403; its indexed snippets cannot certify the full discussion history. No exhaustive audit of all later applications was performed.", "sources": [ { "label": "Published NeurIPS paper", "url": "https://papers.neurips.cc/paper_files/paper/2023/file/b2e63e36c57e153b9015fece2352a9f9-Paper-Conference.pdf#page=28", "location": "Lemma D.1 and its use after Theorem D.1, p.28; proof p.34" }, { "label": "arXiv v2", "url": "https://arxiv.org/pdf/2306.04637v2#page=11", "location": "Lemma 14, pp.11–12; proof Appendix C.4, p.40" }, { "label": "Author publication page", "url": "https://yubai.org/", "location": "Publication listing checked for a revised or correction link" }, { "label": "Author code repository", "url": "https://github.com/allenbai01/transformers-as-statisticians/issues?q=is%3Aissue", "location": "All-state issue listing was empty on retrieval" }, { "label": "Later reuse: In-Context Algorithm Emulation in Fixed-Weight Transformers", "url": "https://openreview.net/pdf?id=BC7YLA0zJ0", "location": "Lemma D.4, p.34, reproduces the same local-smoothness statement; only this reuse was checked" } ], "auditId": "IARC-073", "primaryUrl": "https://papers.neurips.cc/paper_files/paper/2023/file/b2e63e36c57e153b9015fece2352a9f9-Paper-Conference.pdf#page=28", "citationCount": 348, "citationUrl": "https://www.semanticscholar.org/paper/70c3d5ab03a54281be91709b19e3f50a2e4be0e3", "citationDate": "2026-09-09", "paperSummary": "This paper constructs transformers that carry out familiar learning algorithms, such as regression and gradient descent, using examples in a prompt without updating the transformer’s weights. It also shows how a single transformer can select an algorithm or task from the supplied data. The analysis gives quantitative prediction and pretraining guarantees for specified statistical settings, alongside experiments on algorithm selection.", "theoryRole": "The selected lemma controls how errors accumulate when transformer layers approximate successive steps of convex gradient descent, supporting quantitative guarantees for that construction. Its failure means the argument needs a valid bound keeping the iterates inside the region where the update is controlled; it does not invalidate every construction in the paper." }, { "slug": "implicit-bayesian-icl-prior-shift", "title": "An Explanation of In-context Learning as Implicit Bayesian Inference", "issueTitle": "Prior total variation does not control conditional label agreement", "year": 2021, "category": "Transformers & in-context learning", "status": "counterexample", "version": "v6", "location": "Assumption 3, PDF p.5; Lemma 1 and proof, Appendix C, pp.17–18, equations (31)–(47), arXiv v6.", "originalClaim": "A total-variation bound on the prompt’s hidden start distribution ensures that its most likely label agrees with the model’s prediction conditioned on the correct concept and the observed prompt.", "formula": "\\max_{d\\in\\mathcal D}\\operatorname{TV}(u,T_d)<\\Delta/4\\quad\\Longrightarrow\\quad\\arg\\max_y p(y\\mid S_n,x,\\theta^*)=\\arg\\max_y p_{\\rm prompt}(y\\mid x).", "assumptions": [ "The prompt and model use the same HMM transition and emission rules under the correct concept; they differ in the hidden-state prior after a delimiter.", "Delta is the prompt predictive margin after observing the test input, whereas the TV assumption concerns priors before observing that input.", "The lemma invokes Assumption 3. The explicit witness also uses positive transitions and a delimiter emitted by a unique hidden state." ], "plainExplanation": "Assumption 3 bounds the total-variation distance between hidden-state priors before conditioning on the test input. Equations (37)–(38) subsequently omit the distinct evidence normalizers and apply this bound to predictive distributions after conditioning. The five-state HMM below satisfies the stated prior-TV condition, with strictly positive transitions, but reverses the predictive argmax. A posterior-TV condition or an evidence-dependent bound is needed for the label-agreement step.", "steps": [ { "title": "Expose the missing normalization", "text": "The proof defines W with entries p(y|x,h)p(x|h). Its columns sum to p(x|h), not to one. The conditional prediction for a start prior u is Wu divided by its total mass. Equations (37)–(38) omit that denominator separately for each prior.", "formula": "q_u(y\\mid x)=\\frac{(Wu)_y}{\\mathbf1^\\top Wu},\\qquad\\sum_y W_{yh}=p(x\\mid h)." }, { "title": "See the mechanism in two states", "text": "Take priors u=(0.04,0.96), v=(0.005,0.995), input likelihoods (0.9,0.01), and labels that identify the hidden branch. Prior TV is 0.035, but conditioning makes the first branch most likely under u and the second branch most likely under v.", "formula": "\\frac{0.04(0.9)}{0.04(0.9)+0.96(0.01)}=0.78947,\\qquad\\frac{0.005(0.9)}{0.005(0.9)+0.995(0.01)}=0.31142." }, { "title": "Embed the example in a positive-transition HMM", "text": "Use hidden states Ax,Bx,Ay,By,D. Ax and Bx emit x with probabilities 0.9 and 0.01; otherwise they emit z. Ay, By and D emit a, b and the delimiter respectively. T0 sends Ax to Ay, Bx to By, Ay and By to D, and D to 0.005 Ax+0.995 Bx. Set T=0.999 T0+0.0002 times the all-ones matrix. Every transition is positive. Choose a positive-probability prompt ending in the delimiter, such as [x,a,delimiter]. Only D emits that symbol, so the next hidden-state prior is exactly the D row of T, independent of the earlier prompt.", "formula": "T=0.999T_0+0.0002\\mathbf1\\mathbf1^\\top,\\qquad\\min_{i,j}T_{ij}=0.0002." }, { "title": "Verify Assumption 3", "text": "Set the prompt prior to 0.999(0.04,0.96,0,0,0)+0.0002 times the all-ones vector. The model prior after any delimiter is the D row of T. Their TV distance is 0.034965. For one-token test inputs x,z,a,b,delimiter, every prompt predictive margin is at least 0.57995488, so the required TV inequality holds for every possible input.", "formula": "0.034965<0.57995488/4=0.14498872." }, { "title": "Compute the conflicting labels", "text": "For input x, the prompt predicts a with probability 0.78967744 versus b with 0.20972256. After the delimiter, the model predicts b with 0.67966461 versus a with 0.31973539. The same concept is used throughout; the discrepancy comes from conditioning the two close start priors.", "formula": "\\arg\\max q_u(\\cdot\\mid x)=a\\ne b=\\arg\\max q_{T_D}(\\cdot\\mid x)." } ], "counterexample": "The priors have TV distance 0.034965, below one quarter of the minimum prompt margin, 0.57995488. Conditional on x, the prompt predictor assigns approximately 0.78968 to a, whereas the delimiter-conditioned predictor assigns approximately 0.67966 to b.", "witnessFormula": "\\operatorname{TV}(u,T_D)=0.034965<\\Delta/4,\\quad q_u(a\\mid x)=0.78967744,\\quad q_{T_D}(b\\mid x)=0.67966461.", "impact": "Lemma 1 fails under its stated prior-TV assumption. Its application in the main convergence proof consequently requires a stronger conditioning argument. The construction here establishes the lemma-level failure; it does not verify a counterexample to the complete main theorem.", "practical": "The GINC experiments and the general usefulness of a Bayesian account of in-context learning are not refuted by this lemma-level witness.", "repair": "Retain the normalizers and assume a bound on the hidden-state posterior after conditioning, or derive such a bound using a positive lower bound on the evidence probability and an appropriately smaller prior-TV tolerance. If the two conditioned hidden-state distributions have TV distance below Delta/2, the common conditional label kernel cannot increase TV. Each label probability then changes by less than Delta/2, which preserves the unique argmax. Prior TV alone does not imply this condition.", "repairFormula": "\\operatorname{TV}(u(\\cdot\\mid x),v(\\cdot\\mid x))<\\Delta/2\\quad\\Longrightarrow\\quad\\arg\\max q_u=\\arg\\max q_v.", "repairProofFormula": null, "implementation": "Use separate likelihood normalizers for each prior. A theorem based on prior shift must include evidence-sensitive constants or directly control the conditional distribution.", "caveats": [ "The two-state calculation illustrates the mechanism; the five-state construction is the HMM witness used for the lemma.", "The public claim is restricted to Lemma 1, even though the main proof invokes it.", "Correctly applying Bayes’ rule does not itself guarantee the claimed label agreement when the two priors differ." ], "correctionSearch": "2026-09-09 Fresh arXiv history confirms latest v6, 21 July 2022. Re-read Assumption 3 and the complete Lemma 1 proof; visually inspected equations (37)–(47). Exact-title searches with correction, erratum, mistake, normalization, posterior, Assumption 3, Lemma 1, and total variation. Fresh author repository listing: all five issue bodies inspected. Comment retrieval then hit the GitHub rate limit. Read the author explanation and relevant theoretical/related-work portions of three later primary analyses listed in Sources. These discuss Bayesian inference, distribution mismatch, or biased inputs, but no correction of the specific prior-to-posterior bound was identified. No addressing source for this exact missing-normalization implication was found in the inspected material. OpenReview direct PDF/API returned 403 and repository comments could not be refreshed. Related-paper inspection was targeted, not an exhaustive citation-graph search. No priority claim is made.", "sources": [ { "label": "arXiv v6", "url": "https://arxiv.org/pdf/2111.02080v6#page=17", "location": "Assumption 3, p.5; Lemma 1, pp.17–18; equations (33)–(38) and (44)" }, { "label": "Author code repository", "url": "https://github.com/p-lambda/incontext-learning/issues?q=is%3Aissue", "location": "All five issue bodies inspected; none addressed this lemma" }, { "label": "Author explanation", "url": "https://ai.stanford.edu/blog/understanding-incontext/", "location": "2022 explanation and limitations; no correction to the prior-TV lemma identified" }, { "label": "Later Bayesian model-averaging analysis", "url": "https://arxiv.org/html/2305.19420", "location": "Section 4.1, Proposition 4.1 and discussion of Xie et al.; no repair of the selected prior-TV implication identified" }, { "label": "Later kernel-regression analysis", "url": "https://arxiv.org/html/2305.12766", "location": "Sections 3.2–3.3 and Appendix A; different covariance-based assumptions, not a correction of Lemma 1" }, { "label": "Dual Operating Modes of In-Context Learning", "url": "https://arxiv.org/html/2402.18819", "location": "Sections 2 and 5.3 discuss biased inputs and task retrieval; no statement addressing the selected hidden-start-prior bound identified" } ], "auditId": "IARC-072", "primaryUrl": "https://arxiv.org/pdf/2111.02080v6#page=17", "citationCount": 1152, "citationUrl": "https://www.semanticscholar.org/paper/10bd4160b44803ada6a3d2e366c44b7e2a4ffe90", "citationDate": "2026-09-09", "paperSummary": "This paper explains in-context learning as inference about a hidden concept shared by examples in a prompt. In a model where documents come from hidden Markov models, it analyzes when an ideal predictor can recover the relevant concept despite differences between prompts and pretraining data. Experiments on synthetic data explore related behaviour in transformers and recurrent models.", "theoryRole": "Lemma 1 connects knowing the correct concept to predicting the prompt’s most likely label, providing a bridge in the main convergence argument. The counterexample shows that its assumption on starting distributions does not justify this bridge after conditioning on the input; it does not establish a failure of the main theorem under all of its assumptions." }, { "slug": "cql-linear-projection", "title": "Conservative Q-Learning for Offline Reinforcement Learning", "issueTitle": "Sign reversal of the projected CQL penalty", "year": 2020, "category": "Reinforcement learning", "status": "counterexample", "version": "arXiv:2006.04779v3, 19 August 2020; also present in the NeurIPS 2020 supplement", "location": "Theorem D.1, printed pp20–22; projected-penalty argument immediately after equation (16), p21.", "originalClaim": "The linear-function extension claims that, with an invertible feature Gram matrix and its stated choice of penalty weight, the learned policy value stays below the corresponding tabular value. Its proof asserts that the projected conservative penalty is nonnegative.", "formula": "P_F=F(F^\\top D F)^{-1}F^\\top D,\\qquad \\pi^\\top P_F\\!\\left(\\frac{\\pi-\\beta}{\\beta}\\right)\\ge 0", "assumptions": [ "One discounted state, two actions, deterministic return to the same state; discount γ=1/2.", "Behaviour β=(1/2,1/2) and evaluated policy π=(9/10,1/10), so both actions have positive data support.", "Linear critic Qw=(w,2w); D=diag(1/2,1/2) and FᵀDF=5/2, hence the required inverse exists.", "Exact population backup, penalty α=1, rewards r=(9/20,29/20), and initial critic Q=(1,2), equal to the true Qπ." ], "plainExplanation": "Theorem D.1 extends conservatism to linear critics through a feature-space projection. The proof requires the policy expectation of the projected penalty to be nonnegative, which does not follow from invertibility of the feature Gram matrix. With features F=(1,2)ᵀ, this expectation is −22/125. An exact update initialized at a representable true Q-function therefore increases the estimated policy value from 1.1 to 1.276 when α=1.", "steps": [ { "title": "", "text": "The true policy value is (9/10)·1+(1/10)·2=11/10. The Bellman targets are r+γ·11/10=(1,2), so there is no backup or representation error.", "formula": null }, { "title": "", "text": "For Qw=(w,2w), the conservative term is α(11w/10−3w/2)=−2αw/5. Half the behaviour-weighted squared Bellman error is 5(w−1)²/4.", "formula": null }, { "title": "", "text": "The strictly convex objective is minimized at w=1+4α/25. At α=1 this gives w=29/25 and estimated policy value (11/10)(29/25)=319/250=1.276.", "formula": null }, { "title": "", "text": "The correct value is 11/10=1.1, so the estimate exceeds it by 22/125=0.176.", "formula": null }, { "title": "", "text": "For comparison, the exact tabular regularized update at α=1 is (1,2)−(4/5,−4/5)=(1/5,14/5), with policy value 23/50=0.46. The projected update also exceeds that corresponding tabular iterate.", "formula": null }, { "title": "", "text": "The D-orthogonal projector is [[1/5,2/5],[2/5,4/5]]. Applying it to (π−β)/β=(4/5,−4/5) gives (−4/25,−8/25); its π-average is −22/125.", "formula": null }, { "title": "", "text": "The approximation-error numerator of the stated penalty threshold is zero, while its penalty denominator is nonzero. The paper's own representable-target interpretation therefore admits this positive α.", "formula": null } ], "counterexample": "The objective has minimizer w=29/25 at α=1. Its policy value is 319/250, exceeding the true value 11/10 by 22/125 despite exact Bellman targets and a realizable critic.", "witnessFormula": "J_\\alpha(w)=\\frac54(w-1)^2-\\frac{2\\alpha}{5}w,\\quad w_\\alpha=1+\\frac{4\\alpha}{25},\\quad \\widehat V_\\alpha-V^\\pi=\\frac{22\\alpha}{125}>0", "impact": "Refutes the stated extension to arbitrary linear feature spaces, including a realizable true-value example. It also identifies the failed positivity step.", "practical": "Does not refute the tabular CQL result, the empirical benchmark results, or every neural CQL implementation.", "repair": "For a fixed backup target, write the estimated value as A−αB. A lower-bound argument must establish the sign of B and require αB to cover any positive projection bias. If B=0, the unpenalized approximation must already satisfy the desired bound. Representing the constant function is a sufficient condition for nonnegativity of the globally averaged projected penalty: with h=π/β−1 and P_F1=1, (1+h)ᵀDP_Fh=hᵀDP_Fh≥0. It does not by itself guarantee that a zero penalty can correct a positive projection bias.", "repairFormula": null, "repairProofFormula": null, "implementation": "A certificate for a shared linear critic must control both the projected penalty and the projection error. This requires an explicit condition on the feature space or direct verification of these quantities for the fitted critic.", "caveats": [ "The feature space is deliberately small; the theorem explicitly permits it.", "The optional value clip discussed elsewhere is inactive: 2Rmax/(1−γ)=5.8 exceeds both updated values.", "The adjacent nonlinear theorem has additional issues and an incomplete displayed condition; it is not bundled into this counterexample.", "Adding a constant feature repairs this particular sign-reversal example. This is not a proof that adding a bias repairs every linear or neural statement in the paper." ], "correctionSearch": "2026-09-09 Rechecked the latest arXiv version and the exact printed linear argument; inspected the rendered equation page. Revisited the earlier same-day audit of the published supplement, four public reviews, all 28 repository issue/PR entries, and selected relevant comments. Repeated searches for the exact title, Theorem D.1, projection, linear features, counterexample, correction and error. No retrieved source addresses this specific projected-feature failure. This is a dated, bounded search, not proof of undiscovered status. The live GitHub HTML did not expose all issue comments; the distinction from issue #21 relies on the same-day archived API discussion.", "sources": [ { "label": "Current primary paper", "url": "https://arxiv.org/pdf/2006.04779v3", "location": "Theorem D.1 and proof, pp20–22" }, { "label": "Published supplement", "url": "https://papers.nips.cc/paper/2020/file/0d2b2061826a5df3221116a5085a6052-Supplemental.pdf", "location": "Theorem D.1; PDF pp7–9, printed pp20–22" }, { "label": "Related public discussion", "url": "https://github.com/aviralkumar2907/CQL/issues/21", "location": "Tabular nonnegativity discussion; distinct from projection through arbitrary features" } ], "auditId": "E-RL-05", "primaryUrl": "https://arxiv.org/pdf/2006.04779v3", "citationCount": 2786, "citationUrl": "https://www.semanticscholar.org/paper/28db20a81eec74a50204686c3cf796c42a020d2e", "citationDate": "2026-09-09", "paperSummary": "CQL learns decision policies from a fixed collection of past experience, without collecting new interactions. It adds a penalty to the value-learning objective to discourage optimistic estimates for actions that the data do not support well, then uses those estimates to guide policy improvement.", "theoryRole": "The selected theorem extends the lower-value guarantee from separate tabular values to critics that share parameters through linear features. That extension is an important bridge toward the neural critics used in practice; the tabular analysis is a separate result." }, { "slug": "bear-suboptimality-constant", "title": "Stabilizing Off-Policy Q-Learning via Bootstrapping Error Reduction", "issueTitle": "The Q-backup discrepancy omits immediate-action suboptimality", "year": 2019, "category": "Reinforcement learning", "status": "counterexample", "version": "arXiv:1906.00949v2, 25 November 2019; NeurIPS 2019 main and supplement", "location": "Definitions 4.1–4.2, Assumption 4.1 and Theorem 4.1, p4; proof in Appendix B.1, pp12–13.", "originalClaim": "The performance guarantee bounds the constrained policy's state-value loss using approximation error and a suboptimality constant defined by the difference between two Q-backup operators at Q*.", "formula": "\\alpha_Q(\\Pi)=\\|\\mathcal T^\\Pi Q^*-\\mathcal T Q^*\\|_\\infty,\\qquad \\delta=0\\Longrightarrow \\lim_k\\mathbb E_{\\rho_0}|V^{\\pi_k}-V^*|\\le \\frac{\\alpha_Q(\\Pi)}{1-\\gamma}", "assumptions": [ "Finite discounted MDP with γ=1/2, initial state s0 and absorbing state s1.", "At s0, actions good and bad both go to s1 and give rewards 1 and 0; every reward at s1 is zero.", "The allowed policy set is the singleton that always chooses bad; the behaviour policy is this same policy.", "Use the discounted data-state marginal μ(s0)=μ(s1)=1/2. For every positive time k, c(k)=2 satisfies the stated concentrability condition, so C(Π)=2.", "Initialize exact constrained Q-iteration at Q(s0,·)=(1,0), Q(s1,·)=(0,0). All approximation errors δ are zero." ], "plainExplanation": "Theorem 4.1 uses a discrepancy between constrained and unconstrained Q-backup operators to bound state-value suboptimality. This discrepancy evaluates continuation values after the current action has been fixed. It can vanish even when the policy restriction excludes an action with strictly larger immediate reward. In the two-state construction, approximation error and Q-backup discrepancy are both zero, while the constrained policy has value loss one at the initial state.", "steps": [ { "title": "", "text": "Because s1 has no rewards, the exact optimal Q-table is Q*(s0,good)=1, Q*(s0,bad)=0, and Q*(s1,·)=0.", "formula": null }, { "title": "", "text": "At every state-action pair, both backup operators add the same immediate reward to zero continuation value. Thus TΠQ*=TQ*=Q*, and αQ=0.", "formula": null }, { "title": "", "text": "The same Q-table is a fixed point of the constrained iteration. Initializing there makes δ(s,a)=0 for every iteration and action, including the action absent from the behaviour data.", "formula": null }, { "title": "", "text": "The only allowed policy chooses bad at s0, so its initial value is 0. The optimal initial value is 1.", "formula": null }, { "title": "", "text": "Concentrability is finite and its error multiplier is multiplied by zero. The displayed guarantee consequently reduces to 1≤0.", "formula": null }, { "title": "", "text": "Appendix B.1 switches to value-function backup reasoning for its suboptimality term. That reasoning needs an immediate-action/value discrepancy, not the Q-backup discrepancy defined in the theorem.", "formula": null } ], "counterexample": "At the initial state, both actions transition to a zero-reward absorbing state. Their immediate rewards are 1 and 0. The two Q-backup operators agree exactly, but the constrained policy selects the zero-reward action and loses one unit of value.", "witnessFormula": "\\alpha_Q=0,\\quad \\delta=0,\\quad C(\\Pi)=2,\\qquad V^*(s_0)-V^\\Pi(s_0)=1>0", "impact": "Refutes the state-value guarantee with the printed Q-based suboptimality constant. No finite rescaling of this zero constant repairs the example.", "practical": "Does not refute BEAR's empirical results or the motivation for avoiding unsupported actions.", "repair": "Measure the missing immediate choice using ΔΠ=||V*−maxπ∈Π EπQ*||∞, or an equivalent discrepancy between the value Bellman operators. For rectangular policy restrictions, contraction yields ||VΠ−V*||∞≤ΔΠ/(1−γ). Let TΠV(s)=maxπ∈Π Eπ[r(s,a)+γPV(s,a)]. Then ||VΠ−V*||∞≤γ||VΠ−V*||∞+||TΠV*−TV*||∞; the last term is ΔΠ, equal to 1 here.", "repairFormula": null, "repairProofFormula": null, "implementation": "Performance certificates for support-restricted policies must account for the value of excluded current actions. Small Bellman residuals and benign state-distribution shift alone do not establish near-optimality.", "caveats": [ "The theorem explicitly allows a singleton policy set and assumes a residual bound; it does not require the data to identify every action value.", "This is an exact mathematical counterexample, not a report of a failed benchmark run.", "This repairs the exact-iteration suboptimality component. It is not a fresh proof of every constant in the paper's approximate-iteration bound or of the practical MMD policy constraint." ], "correctionSearch": "2026-09-09 Rechecked the current arXiv definitions, theorem and value-backup proof; visually inspected p4. Revisited the earlier same-day published supplement, three reviews, author feedback, author BEAR repository history and the recommended evaluation repository discussion. Checked related tutorial and DOGE suboptimality treatments; repeated title/theorem/suboptimality correction searches. No retrieved source corrects the Q-versus-value mismatch in this guarantee. The search is bounded and does not establish priority. A related bound for Q-function performance is not automatically a corrected state-value bound.", "sources": [ { "label": "Current primary paper", "url": "https://arxiv.org/pdf/1906.00949v2", "location": "p4; Appendix B.1, pp12–13" }, { "label": "Published main paper", "url": "https://papers.neurips.cc/paper/9349-stabilizing-off-policy-q-learning-via-bootstrapping-error-reduction.pdf", "location": "Definitions 4.1–4.2 and Theorem 4.1, p4" }, { "label": "Related same-author tutorial", "url": "https://arxiv.org/abs/2005.01643", "location": "Offline RL policy-constraint discussion; no located repair of this constant" } ], "auditId": "IR009", "primaryUrl": "https://arxiv.org/pdf/1906.00949v2", "citationCount": 1344, "citationUrl": "https://www.semanticscholar.org/paper/82b4b03a4659d6e04bd7cbf51d6e08fde1348dbd", "citationDate": "2026-09-09", "paperSummary": "BEAR studies why learning from a fixed dataset can become unstable when value updates rely on actions scarcely represented in that data. Its algorithm constrains the learned policy toward the data's action support while still trying to select rewarding actions within that support.", "theoryRole": "The selected theorem formalizes the proposed trade-off between avoiding distribution shift and excluding good actions. Its suboptimality term is meant to measure the performance cost of the policy restriction." }, { "slug": "combo-zero-penalty", "title": "COMBO: Conservative Offline Model-Based Policy Optimization", "issueTitle": "A zero occupancy penalty cannot compensate for model bias", "year": 2021, "category": "Reinforcement learning", "status": "counterexample", "version": "arXiv:2102.08363v2, 27 January 2022; NeurIPS 2021", "location": "Proposition 4.1, p5; Corollary A.1, pp17–18, especially equation (14). Lemma A.1, p15, permits a zero expected penalty.", "originalClaim": "The expected learned value is claimed to become a lower bound for sufficiently large conservative weight, including when model or sampling error is present.", "formula": "\\widehat Q_{k+1}=\\widehat{\\mathcal B}^{\\pi}\\widehat Q_k-\\beta\\frac{\\rho-d}{d_f},\\qquad \\mathbb E_{\\mu_0,\\pi}\\widehat Q^\\pi\\le\\mathbb E_{\\mu_0,\\pi}Q^\\pi\\quad\\text{for sufficiently large }\\beta", "assumptions": [ "One state, one action, deterministic self-loop, discount γ=1/2 and interpolation f=1/2.", "True and empirical rewards are zero. The learned reward is 1/10; learned and true transitions coincide.", "Model and dataset occupancies satisfy ρ=d=df=1, so every denominator has positive support.", "Population data remove sampling error; choose a valid reward bound Rmax=1. The resulting iterates remain within the bounded range used in the argument." ], "plainExplanation": "The sufficiently-large-β conclusion requires a strictly positive effective conservative penalty whenever the residual model bias is positive. The stated assumptions allow model and data occupancies to coincide, making the penalty identically zero. In a one-state, one-action MDP with a positively biased learned reward, the interpolated Bellman fixed point exceeds the true value for every β. The error-dependent bound preceding the corollary is consistent with this construction.", "steps": [ { "title": "", "text": "With one state and one action, the only possible normalized occupancy is 1. Hence (ρ−d)/df=0 and ν(ρ,f)=0.", "formula": null }, { "title": "", "text": "The interpolated backup uses reward f·0+(1−f)·(1/10)=1/20, plus γQ.", "formula": null }, { "title": "", "text": "The update is therefore Qk+1=1/20+Qk/2 for every β. It converges to Qhat=1/10.", "formula": null }, { "title": "", "text": "The true value is zero because every true reward is zero. Thus Qhat exceeds the true value by 1/10 for every finite penalty coefficient.", "formula": null }, { "title": "", "text": "Lemma A.1 already characterizes equality ν=0. The later corollary labels the same quantity strictly positive and uses it to dominate positive error.", "formula": null }, { "title": "", "text": "The valid nonnegative identity is ν=f Σ(ρ−d)²/df, for positive denominators and normalized distributions. Nonnegativity does not imply the strict positivity needed by the final step.", "formula": null } ], "counterexample": "With identical occupancies, f=1/2, discount 1/2, true reward zero and learned reward 1/10, the conservative term vanishes. The estimated fixed point is 1/10 for every β, while the true value is zero.", "witnessFormula": "\\rho=d=d_f=1,\\quad \\nu=0,\\quad \\widehat Q_{k+1}=\\frac1{20}+\\frac12\\widehat Q_k,\\quad \\widehat Q^\\pi=\\frac1{10}>0=Q^\\pi", "impact": "Refutes the unrestricted sufficiently-large-β lower-bound statement. The preceding bound that retains an explicit model-error term is not refuted by this example.", "practical": "Does not show that COMBO fails its benchmarks, or that it fails when the model and data estimates are exact.", "repair": "Keep the error term in the conclusion. A corollary may remove it only if its effective expected conservative penalty is strictly positive and large enough to cover the error. If the penalty is zero, require zero or favourably signed bias instead. An inequality of the form Vhat≤V+E−βν only implies Vhat≤V when βν≥E. For ν=0 and E>0 there is no admissible finite β; that is exactly this example.", "repairFormula": null, "repairProofFormula": null, "implementation": "If a certified lower bound is required, occupancy matching cannot replace a bound on reward/model error. One would need an independently justified error penalty or another assumption that prevents this zero-penalty case.", "caveats": [ "The example preserves full support; it does not rely on division by zero.", "The learned reward error can be arbitrarily small but positive.", "The equality case in Lemma A.1 is acknowledged internally, but the later unconditional corollary does not incorporate it.", "Restricting to ν>0 repairs the displayed division step, not every other assumption or occupancy normalization in the full analysis. Adding a practical uncertainty penalty would be an algorithm change requiring separate evaluation." ], "correctionSearch": "2026-09-09 Rechecked current v2, the interpolation lemma and the rendered equation (14). Revisited the earlier same-day official NeurIPS supplement, ICLR/workshop versions and search-accessible thesis theorem. Repeated exact-title, Proposition 4.1, Corollary A.1, zero/strictly-positive penalty and correction searches. No retrieved source supplies the missing qualification for the zero-penalty case. Public OpenReview reviews were unavailable behind verification/403 responses in the documented search. The full thesis appendix was not successfully downloaded in that search; only accessible relevant text is claimed inspected. The history search is therefore less complete than for the other five cases.", "sources": [ { "label": "Current primary paper", "url": "https://arxiv.org/pdf/2102.08363v2", "location": "Proposition 4.1 p5; Lemma A.1 p15; Corollary A.1 pp17–18" }, { "label": "OpenReview version", "url": "https://openreview.net/pdf?id=dUEpGV2mhf", "location": "Section 4.1 and Appendix A" }, { "label": "Later same-author thesis", "url": "https://www2.eecs.berkeley.edu/Pubs/TechRpts/2023/EECS-2023-223.pdf", "location": "Theorem 5.2.1 repeats the sufficiently-large-penalty statement in search-accessible text" } ], "auditId": "IR012", "primaryUrl": "https://arxiv.org/pdf/2102.08363v2", "citationCount": 544, "citationUrl": "https://www.semanticscholar.org/paper/245682e8b3fa76f4a3e2991b5497577af95cbb3f", "citationDate": "2026-09-09", "paperSummary": "COMBO learns a model of the environment and uses simulated transitions alongside an offline dataset to train a policy. Its critic regularization is designed to make value estimates conservative on model-generated state-action pairs, without requiring a separate explicit uncertainty estimator.", "theoryRole": "The selected proposition provides the expected lower-value guarantee that motivates this form of pessimism. It supports the paper's policy-improvement argument; it is distinct from the empirical evidence for the algorithm." }, { "slug": "pac-bayes-finite-symmetry-mixture", "title": "Computing Nonvacuous Generalization Bounds for Deep (Stochastic) Neural Networks with Many More Parameters than Training Data", "issueTitle": "The KL identity requires more than a finite symmetry set", "year": 2017, "category": "Optimization & generalization", "status": "counterexample", "version": "arXiv:1703.11008v2, 19 October 2017; UAI 2017 supplement", "location": "Appendix B.1, Lemma B.1, physical p11; official supplement p1.", "originalClaim": "For any finite set of parameter symmetries preserving the prior, Lemma B.1 assigns an exact KL reduction to the equally weighted mixture of transformed posteriors.", "formula": "Q^S=\\frac1{|S|}\\sum_{\\sigma\\in S}Q_\\sigma,\\qquad \\mathrm{KL}(Q^S\\|P)=\\mathrm{KL}(Q\\|P)-\\mathrm{KL}(Q\\|Q^S)", "assumptions": [ "Use a three-parameter network whose output is invariant to permuting three hidden units; for example h_w(x)=sign(Σi ReLU(w_i x)−1).", "The prior is the full-support Gaussian P=N(0,I3). Let Ei be the sector where coordinate i is largest; ties have probability zero and P(Ei)=1/3.", "Set dQ/dP=3qi on Ei, with q=(0.6,0.3,0.1). This positive bounded density ratio gives an absolutely continuous Q with finite differential entropy.", "S contains the identity and one coordinate 3-cycle. It is a finite set of genuine network symmetries, but it is not a group." ], "plainExplanation": "Lemma B.1 asserts an exact KL reduction for averaging an arbitrary finite set of prior-preserving network symmetries. The valid mixture identity contains the average component-to-mixture divergence; it generally cannot be replaced by the divergence of one distinguished component. The replacement is justified for a complete finite-group average, whose mixture is invariant. A partial cyclic average violates the printed identity and can underestimate the resulting KL complexity.", "steps": [ { "title": "", "text": "Choose the cycle so its sector masses are qσ=(0.1,0.6,0.3). The mixture has masses qS=(0.35,0.45,0.2).", "formula": null }, { "title": "", "text": "Because the density ratios are constant on sectors, every continuous KL integral equals its three-point categorical calculation.", "formula": null }, { "title": "", "text": "KL(Q||P)=0.20066656381, KL(QS||P)=0.04995849931 and KL(Q||QS)=0.13244364995.", "formula": null }, { "title": "", "text": "The claimed right side is 0.06822291386, not the actual 0.04995849931. The exact left-minus-right residual is log(11664/16807)/20.", "formula": null }, { "title": "", "text": "A second admissible ordering q=(0.6,0.1,0.3) gives mixture (0.45,0.35,0.2). The actual KL remains 0.04995849931, but the printed right side is 0.03169408476: it can also understate complexity.", "formula": null }, { "title": "", "text": "The correct mixture decomposition subtracts the average of KL(Qσ||QS) over all selected transformations. Under full finite-group averaging these terms coincide; an arbitrary subset does not ensure that equality.", "formula": null } ], "counterexample": "For q=(0.6,0.1,0.3), its cyclic transform (0.3,0.6,0.1), and a uniform prior, the mixture KL is approximately 0.04995850. The claimed difference of KL terms is approximately 0.03169408. Equal-probability Gaussian sectors embed this calculation in absolutely continuous distributions.", "witnessFormula": "\\mathrm{KL}(Q^S\\|P)=0.04995849931\\ne0.06822291386=\\mathrm{KL}(Q\\|P)-\\mathrm{KL}(Q\\|Q^S)", "impact": "Refutes the exact identity for arbitrary finite symmetry sets. The weaker claim that such invertible prior-preserving symmetrization does not increase KL still follows from convexity.", "practical": "The reported numerical PAC-Bayes bounds did not use this symmetry reduction, so this finding does not invalidate those reported bounds.", "repair": "For a finite collection of invertible prior-preserving maps, use KL(QS||P)=KL(Q||P)−|S|⁻¹Σσ KL(Qσ||QS). Alternatively, require S to be a finite group and average over the whole group. Expand the average KL(Qσ||P), insert log(qσ/p)=log(qσ/qS)+log(qS/p), and average. The second integral becomes KL(QS||P). Invariance of a full-group mixture makes the remaining KL terms equal.", "repairFormula": null, "repairProofFormula": null, "implementation": "A symmetry-based certificate should evaluate the actual mixture complexity or the average component-to-mixture divergence. Subtracting only KL(Q||QS) after sampling an arbitrary list of permutations is not an exact certificate.", "caveats": [ "The finite set is intentionally not closed under composition; the printed lemma does not require closure.", "The Gaussian construction avoids relying on a discrete posterior or a compactly supported prior.", "The finding concerns an auxiliary identity, not the validity of PAC-Bayes theory.", "The simplified group formula requires the usual measurable bijection and prior-invariance conditions. No claim is made that full-group averaging is computationally feasible for a large network." ], "correctionSearch": "2026-09-09 Rechecked the current lemma and its explicit finite-set wording in the rendered source. Revisited the official UAI supplement and the relevant parts of Lyle et al. and Beck–Ochs. Repeated exact-title/arXiv, finite-set, mixture, symmetry, KL and correction searches. No retrieved source explicitly repairs this arbitrary-finite-set identity. The later group/pushforward results concern different operations. This does not prove the mixture identity is unknown; the mathematical repair itself is elementary. The correction search is selective and bounded, not an inventory of all PAC-Bayes literature.", "sources": [ { "label": "Current primary paper", "url": "https://arxiv.org/pdf/1703.11008v2", "location": "Appendix B.1, Lemma B.1, p11" }, { "label": "Later invariance analysis", "url": "https://arxiv.org/abs/2005.00178", "location": "Group-based function averaging; a distinct operation from arbitrary finite mixtures of parameter distributions" }, { "label": "Later PAC-Bayes symmetry analysis", "url": "https://arxiv.org/pdf/2510.17303v2", "location": "Lemma 1.4 and averaging-operator construction; general deterministic-pushforward decomposition" } ], "auditId": "IO-1703.11008", "primaryUrl": "https://arxiv.org/pdf/1703.11008v2", "citationCount": 975, "citationUrl": "https://www.semanticscholar.org/paper/540c226fdf7047ac602c7cb05a18db19ee595df0", "citationDate": "2026-09-09", "paperSummary": "This paper computes informative numerical generalization bounds for stochastic neural networks, including networks with many more parameters than training examples. It optimizes a PAC-Bayes objective that balances measured training error against the complexity of a distribution over network weights.", "theoryRole": "The selected appendix lemma explores whether averaging equivalent parameter settings can further reduce the complexity term. It is an auxiliary route to tighter bounds, not a step used to obtain the paper's reported numerical bounds." }, { "slug": "batchnorm-weight-gradient-bound", "title": "How Does Batch Normalization Help Optimization?", "issueTitle": "A negative upper bound on the squared BatchNorm weight gradient", "year": 2018, "category": "Optimization & generalization", "status": "counterexample", "version": "arXiv:1805.11604v5, 15 April 2019; NeurIPS 2018 main and supplement", "location": "Theorem 4.4, p8; Appendix C.2, pp20–21, especially the final step after equation (17).", "originalClaim": "Theorem 4.4 gives a worst-case squared weight-gradient bound for BatchNorm by subtracting mean-gradient and activation-correlation terms from the corresponding ordinary-network quantity.", "formula": "g_j=\\max_{\\|X\\|\\le\\lambda}\\|\\nabla_W L\\|^2,\\quad \\widehat g_j=\\max_{\\|X\\|\\le\\lambda}\\|\\nabla_W\\widehat L\\|^2,\\quad \\widehat g_j\\le\\frac{\\gamma^2}{\\sigma_j^2}\\left(g_j^2-m\\mu_{g_j}^2-\\lambda^2\\langle\\nabla_{y_j}L,\\widehat y_j\\rangle^2\\right)", "assumptions": [ "Batch size m=2, two input coordinates, W=(1,−1), operator-norm bound λ=1, fixed BN scale γ=1 and shift 0.", "Use the same nonnegative differentiable downstream mean Huber loss in both networks: L(y)=[Huber1(y1+3)+Huber1(y2−3)]/2.", "BN is the paper's idealized formula without numerical epsilon; restrict to batches with positive variance.", "The downstream gradient is a=(1/2,−1/2) on every ordinary activation with ||X||≤1 and on both possible normalized BN outputs." ], "plainExplanation": "Theorem 4.4 subtracts an activation-correlation term without the normalization required by the BatchNorm Jacobian. For a batch of size two, the centered and normalized activation directions span the batch space, so the idealized BatchNorm projector is zero at every nondegenerate batch. The construction below nevertheless gives a negative right-hand side for the stated squared-gradient bound. The fixed-batch projection identity retains factors of 1/m in both subtracted terms.", "steps": [ { "title": "", "text": "For ||X||op≤1, ordinary activations y=XW have norm at most √2. Both Huber residuals stay in their linear regions, so the downstream gradient is always a=(1/2,−1/2).", "formula": null }, { "title": "", "text": "The ordinary squared weight-gradient maximum is ||a||²=1/2, attained at X=I.", "formula": null }, { "title": "", "text": "Every nonconstant two-example batch normalizes to yhat=±(1,−1). The BN derivative contains P=I−11ᵀ/2−yhat yhatᵀ/2, which is the zero matrix.", "formula": null }, { "title": "", "text": "Hence the BN weight-gradient maximum is zero. At X=I, variance σ²=1, mean(a)=0 and ²=1.", "formula": null }, { "title": "", "text": "Using the printed definition g_j=max||∇WL||², its right side is (1/2)²−1=−3/4. Even reading g_j as an unsquared norm gives 1/2−1=−1/2.", "formula": null }, { "title": "", "text": "The fixed-activation projection calculation needs the factor 1/m on the squared correlation. That normalization appears earlier in the appendix but disappears in the final weight-space expression.", "formula": null } ], "counterexample": "For m=d=2, X=I, W=(1,−1), γ=1 and the specified Huber loss, the BatchNorm weight gradient is zero. The printed right-hand side is −3/4 under its squared-norm notation, yielding the contradiction 0≤−3/4.", "witnessFormula": "\\widehat g=0,\\quad g=\\frac12,\\quad \\langle a,\\widehat y\\rangle^2=1,\\qquad 0\\not\\le-\\frac34\\quad(\\text{or }-\\tfrac12\\text{ under the alternate norm convention})", "impact": "Refutes the quantitative weight-space bound as printed, under either natural reading of its extra squared notation.", "practical": "Does not refute the empirical BN observations, every smoothing result, or the separate activation-space projection identity.", "repair": "Use the exact fixed-batch chain rule ∇W LBN=Xᵀ(γ/σ)Pa, with a evaluated at the BN output. It implies ||∇W LBN||²≤λ²γ²/σ²[||a||²−<1,a>²/m−²/m]. The vectors 1/√m and yhat/√m are orthonormal. P projects away their two directions, so the Pythagorean identity supplies both factors 1/m.", "repairFormula": null, "repairProofFormula": null, "implementation": "Standard BN backpropagation need not be changed because of this theorem error. Any certificate or learning-rate argument based on the printed worst-case formula must use the normalized Jacobian and track how a and σ vary with the batch.", "caveats": [ "The example uses m=2, which is not excluded by the printed theorem.", "Zero-variance batches are outside the paper's epsilon-free BN definition.", "Real implementations usually add epsilon; this is a counterexample to the idealized theorem, not a claim that practical BN has an exactly zero gradient at batch size two.", "Restoring 1/m is not a complete proof of a corrected minimax theorem: the downstream gradient, normalized activation and variance all depend on X. The stated repair is a fixed-batch bound." ], "correctionSearch": "2026-09-09 Rechecked current v5, the printed theorem and the derivation; inspected the theorem image. Revisited the NeurIPS supplement, public reviews and the 2025 initialization counterexample. Repeated exact-title, Theorem 4.4, minimax, normalization, correction and error searches. The located 2025 criticism addresses Lemma 4.5. No retrieved source corrects this specific Theorem 4.4 bound. The same paper already contains the valid activation-space normalization; that does not by itself supply a corrected weight-space maximization theorem. No claim of exhaustive absence of later discussion or of priority.", "sources": [ { "label": "Current primary paper", "url": "https://arxiv.org/pdf/1805.11604v5", "location": "Theorem 4.4 p8; Appendix C.2 pp20–21" }, { "label": "Distinct published criticism", "url": "https://arxiv.org/pdf/2502.17913v1", "location": "Dannemann–Junike, Sections 1 and 3: counterexample to initialization Lemma 4.5, not weight-space Theorem 4.4" } ], "auditId": "IO-1805.11604", "primaryUrl": "https://arxiv.org/pdf/1805.11604v5", "citationCount": 1763, "citationUrl": "https://www.semanticscholar.org/paper/94be567c32ae76bdaadabd4975807a94181e39b3", "citationDate": "2026-09-09", "paperSummary": "This paper investigates why batch normalization helps neural-network training. It challenges the explanation based solely on stabilizing intermediate activation distributions and argues, through experiments and mathematical analysis, that normalization can make the optimization landscape and gradients better behaved.", "theoryRole": "The selected theorem tries to translate an effect on activation gradients into a worst-case bound on weight gradients. That translation matters because training updates the weights; it complements the paper's other theoretical statements and experiments." }, { "slug": "trades-loss-domination", "title": "Theoretically Principled Trade-off between Robustness and Accuracy", "issueTitle": "Calibration is insufficient for the TRADES surrogate bound", "year": 2019, "category": "Optimization & generalization", "status": "counterexample", "version": "arXiv:1901.08573v3, 24 June 2019; ICML 2019 paper and supplement", "location": "Theorem 3.1, p6; Appendix B.1, p20. Instance space and relative adversarial balls are defined in Section 2.1.", "originalClaim": "Theorem 3.1 upper-bounds robust excess error for any nonnegative classification-calibrated loss with φ(0)≥1, any measurable predictor and the stated instance space.", "formula": "R_{\\rm rob}(f)-R_{\\rm nat}^*\\le\\psi^{-1}(R_\\phi(f)-R_\\phi^*)+\\mathbb E\\max_{x'\\in B(X,\\epsilon)}\\phi\\!\\left(\\frac{f(x')f(X)}{\\lambda}\\right)", "assumptions": [ "Instance space X={−1,+1}⊂R, uniform data and deterministic label Y=X; f(x)=x.", "Adversarial radius ε=2, with balls intersected with X as the paper defines; λ=1.", "Define S(u)=0 for u≤0, S(u)=6u⁵−15u⁴+10u³ for 0H(η), so the loss satisfies the printed calibration assumption. Its calibration transform is ψ(θ)=θ/2.", "formula": null }, { "title": "", "text": "The predictor classifies both points correctly with zero surrogate risk. The optimal natural risk and surrogate excess are both zero.", "formula": null }, { "title": "", "text": "Every adversarial ball contains the opposite point, so robust error is one. The boundary surrogate is max{φ(1),φ(−1)}=1/2.", "formula": null }, { "title": "", "text": "The claimed upper bound is consequently 1≤1/2. The proof's step from a wrong-sign indicator to φ(u) fails at u=−1.", "formula": null } ], "counterexample": "The two-point construction has zero surrogate excess risk, robust error one and boundary surrogate 1/2. The second inequality therefore reduces to 1≤1/2.", "witnessFormula": "\\phi(-1)=\\frac12,\\quad \\phi(0)=1,\\quad \\phi(1)=0,\\quad \\psi(\\theta)=\\frac{\\theta}{2},\\qquad 1=R_{\\rm rob}-R_{\\rm nat}^*>\\frac12", "impact": "Refutes the universal second inequality of Theorem 3.1 under its printed loss and instance-space assumptions. The first risk decomposition/calibration inequality is not refuted.", "practical": "Does not refute the usual TRADES algorithm or its benchmark results. Common non-increasing margin losses, with the stated normalization, satisfy the extra domination property.", "repair": "Add φ(u)≥1 for every nonpositive u. A sufficient, stronger condition is that φ is non-increasing and φ(0)≥1. Then the wrong-sign indicator is pointwise bounded by the surrogate, restoring the disputed step. For u≤0 the added condition supplies the required unit penalty; for u>0 the indicator is zero and nonnegativity suffices. Taking the same maxima and expectation preserves the inequality.", "repairFormula": null, "repairProofFormula": null, "implementation": "When adapting the binary TRADES guarantee to a new loss or a discrete input domain, verify wrong-sign domination rather than calibration alone. No change to a standard loss that already satisfies this condition is implied.", "caveats": [ "On a connected perturbation domain with a continuous predictor, an intermediate zero score may restore the boundary surrogate. The theorem explicitly permits the two-point domain used here.", "The witness loss is deliberately non-monotone; the printed theorem does not require monotonicity.", "The scope is binary population theory, not an empirical multiclass robustness certificate.", "This repairs the selected upper-bound step. It does not independently establish the paper's lower-bound tightness theorem or its multiclass practical objective." ], "correctionSearch": "2026-09-09 Rechecked current v3, the relative-ball domain and the rendered theorem hypotheses. Revisited the official PMLR proof and earlier same-day inventory of 33 repository issues/PRs with available comments. Repeated theorem/calibration/monotonicity/domination/counterexample correction searches. No retrieved source adds the needed wrong-sign domination condition to this printed universal statement. The search is bounded and does not establish priority. Generic analyses of adversarial calibration or a different loss do not automatically correct this exact statement.", "sources": [ { "label": "Current primary paper", "url": "https://arxiv.org/pdf/1901.08573v3", "location": "Section 2.1; Theorem 3.1 p6; Appendix B.1 p20" }, { "label": "Official ICML version", "url": "https://proceedings.mlr.press/v97/zhang19p.html", "location": "Main paper and supplementary proof of Theorem 3.1" }, { "label": "Author repository discussion", "url": "https://github.com/yaodongyu/TRADES/issues", "location": "Earlier same-day issue/comment inventory; binary/multiclass and loss-implementation discussions distinguished" } ], "auditId": "IO-1901.08573", "primaryUrl": "https://arxiv.org/pdf/1901.08573v3", "citationCount": 3147, "citationUrl": "https://www.semanticscholar.org/paper/6c405d4b5dc41a86be05acd59c06ed19daf01d14", "citationDate": "2026-09-09", "paperSummary": "TRADES studies the tension between ordinary prediction accuracy and resistance to adversarial perturbations. It decomposes robust error into ordinary error and a boundary-related term, then uses surrogate losses to motivate a training objective balancing accurate predictions with stability under perturbation.", "theoryRole": "The selected upper bound is part of the theoretical justification for replacing the difficult robust-error objective with a tractable surrogate. Its assumptions determine how broadly that justification applies to alternative losses and input spaces." }, { "slug": "infovae-divergence-sign", "title": "InfoVAE: Balancing Learning and Inference in Variational Autoencoders", "issueTitle": "A missing coefficient condition in InfoVAE Proposition 2", "year": 2019, "category": "Generative models", "status": "counterexample", "version": "arXiv:1706.02262v3, 30 May 2018; published AAAI 2019 version also inspected", "location": "arXiv v3, Proposition 2 and Eq. (6), PDF p. 4; proof pp. 11–12. Published AAAI version: Proposition 2 and Eq. (6), printed p. 5888.", "originalClaim": "Proposition 2 states that, at a fixed encoder mutual information, replacing the marginal KL term by any strict divergence preserves global optimality of a model that matches the data and has exact amortized inference, assuming α<1 and λ>0.", "formula": "\\widehat{\\mathcal L}=\\mathbb E_{p_D(x)q(z\\mid x)}\\log p_\\theta(x\\mid z)-(1-\\alpha)\\mathbb E_{p_D(x)}D_{\\mathrm{KL}}(q(z\\mid x)\\|p(z))-(\\alpha+\\lambda-1)D(q(z)\\|p(z)).", "assumptions": [ "Both data and latent spaces are continuous.", "The proposition permits α<1 and λ>0, without requiring α+λ−1≥0.", "The marginal divergence is nonnegative and vanishes exactly when its two distributions agree.", "Mutual information Iq(X;Z) is held fixed; the witness uses Iq=0 and finite, unit encoder variance.", "Optimization is considered over a family containing the stated Gaussian encoders and an independent Gaussian decoder." ], "plainExplanation": "The assumptions α<1 and λ>0 permit α+λ−1<0. In this region, substituting an arbitrary strict divergence into the marginal regularizer can reverse its intended effect. Set α=0, λ=1/2 and D=3 KL. An independent Gaussian encoder with mean m then has fixed mutual information zero and objective −H(pD)+m²/4. The matched configuration m=0 is not a global maximizer, and the objective is unbounded above along this family.", "steps": [ { "title": "Choose an allowed parameter pair", "text": "Set α=0 and λ=1/2. The coefficient α+λ−1 is −1/2, although the appendix treats it as positive.", "formula": "\\alpha=0,\\quad\\lambda=\\tfrac12,\\quad\\alpha+\\lambda-1=-\\tfrac12." }, { "title": "Use ordinary Gaussian distributions", "text": "Let the data and prior both be standard normal. Keep the decoder equal to the data distribution and let the encoder mean m vary independently of x.", "formula": "p_D(x)=p_\\theta(x\\mid z)=\\mathcal N(0,1),\\quad p(z)=\\mathcal N(0,1),\\quad q_m(z\\mid x)=\\mathcal N(m,1)." }, { "title": "Check mutual information and the divergence", "text": "The encoder ignores x, so Iq(X;Z)=0 for every m. Three times KL is still a strict divergence, and the relevant Gaussian KL equals m²/2.", "formula": "D=3D_{\\mathrm{KL}},\\quad I_{q_m}(X;Z)=0,\\quad D_{\\mathrm{KL}}(\\mathcal N(m,1)\\|\\mathcal N(0,1))=\\tfrac{m^2}{2}." }, { "title": "Evaluate the complete objective", "text": "The reconstruction term stays fixed. The negative KL contribution is outweighed by the positively signed substituted divergence.", "formula": "\\widehat{\\mathcal L}(m)=-H(p_D)-\\tfrac{m^2}{2}+\\tfrac12\\left(3\\tfrac{m^2}{2}\\right)=-H(p_D)+\\tfrac{m^2}{4}." }, { "title": "Compare with the claimed optimum", "text": "At m=0 the generative model matches the data and its posterior equals the encoder. Nevertheless, every nonzero m increases the objective, and the objective is unbounded above along this family.", "formula": "\\widehat{\\mathcal L}(2)-\\widehat{\\mathcal L}(0)=1,\\qquad \\widehat{\\mathcal L}(m)\\longrightarrow+\\infty\\quad\\text{as }|m|\\to\\infty." } ], "counterexample": "At α=0, λ=1/2 and D=3 KL, shifting the encoder mean from zero to two increases the objective by one. The decoder, encoder variance and mutual information remain fixed.", "witnessFormula": "\\alpha=0,\\ \\lambda=\\tfrac12,\\ D=3D_{\\mathrm{KL}}\\quad\\Longrightarrow\\quad\\widehat{\\mathcal L}(m)=-H(\\mathcal N(0,1))+\\tfrac14m^2.", "impact": "The arbitrary-divergence global-optimality statement fails in part of its printed parameter range. The KL-only objective is not refuted by this example.", "practical": "The example does not test the paper's experiments or show failure of its recommended positive-penalty settings.", "repair": "Adding α+λ−1≥0 removes the displayed reward-for-mismatch mechanism. Under finite entropy/KL quantities and availability of an exact model with the prescribed mutual information, it also gives the following sufficient argument: E_q log pθ(X|Z)≤−H_q(X|Z)=−H(pD)+Iq(X;Z), so the objective is bounded above by −H(pD)+αIq minus two nonnegative marginal penalties; an exact matched joint distribution attains that bound.", "repairFormula": "1-\\alpha>0,\\qquad \\alpha+\\lambda-1\\ge0.", "repairProofFormula": "\\widehat{\\mathcal L}\\le-H(p_D)+\\alpha I_q-(1-\\alpha)D_{\\mathrm{KL}}(q_Z\\|p_Z)-(\\alpha+\\lambda-1)D(q_Z\\|p_Z)\\le-H(p_D)+\\alpha I_q.", "implementation": "A general implementation that exposes arbitrary divergence choices should validate the coefficient range, or document an appropriate divergence-specific condition. Existing configurations with a nonnegative marginal penalty do not need a new training algorithm.", "caveats": [ "The witness uses a valid rescaling of KL because the proposition permits any strict divergence.", "All objective values are finite for each finite m; the construction does not rely on a vanishing encoder variance.", "The repair stated here is sufficient; it is not claimed to characterize every admissible divergence and parameter combination.", "The sufficient optimality repair assumes the entropy/KL manipulations are well-defined and that the prescribed mutual information is attainable by an exact matched joint model." ], "correctionSearch": "2026-09-09 The primary proposition and appendix were reread; the published proposition was visually checked. Current arXiv history and targeted web searches for a correction to Proposition 2 and the negative marginal-divergence coefficient were checked. The later Information Autoencoding Family and Coupled Variational Autoencoder formulations were inspected at the relevant equations. \"InfoVAE\" \"proposition\" \"correction\" \"InfoVAE\" \"alpha\" \"lambda\" \"negative\" divergence InfoVAE alpha lambda divergence coefficient negative condition theorem No source addressing this exact failing parameter range was found in the bounded search. The Information Autoencoding Family uses nonnegative dual multipliers in a different constrained formulation. Coupled Variational Autoencoder restates the KL objective and uses α=0 with λ=2 or 3 in the inspected experiments. These are relevant valid formulations, but neither inspected passage qualifies the original unrestricted arbitrary-divergence proposition. The author-linked InfoVAE repository was unavailable (404). The OpenReview attachment could not be fully inspected because of access/verification failures. This is not an exhaustive search of all literature or private correspondence, and it does not establish novelty.", "sources": [ { "label": "InfoVAE, exact arXiv version", "url": "https://arxiv.org/pdf/1706.02262v3", "location": "Proposition 2; Eq. (6); proof of Proposition 2, pp. 4, 11–12" }, { "label": "Published AAAI paper", "url": "https://ojs.aaai.org/index.php/AAAI/article/download/4538/4416", "location": "Proposition 2 and Eq. (6), printed p. 5888" }, { "label": "Version history", "url": "https://arxiv.org/abs/1706.02262", "location": "Latest listed revision v3, 30 May 2018" }, { "label": "The Information Autoencoding Family: A Lagrangian Perspective", "url": "https://arxiv.org/pdf/1806.06514", "location": "Eq. (18), Section 6, and the InfoVAE conversion" }, { "label": "Coupled Variational Autoencoder", "url": "https://proceedings.mlr.press/v202/hao23b/hao23b.pdf", "location": "Section 2, Eq. (4); experimental parameter footnote on p. 6" } ], "auditId": "IG007", "primaryUrl": "https://arxiv.org/pdf/1706.02262v3", "citationCount": 495, "citationUrl": "https://www.semanticscholar.org/paper/6f88397a7df1861d18303b19a77d6453ed32621e", "citationDate": "2026-09-09", "paperSummary": "A variational autoencoder learns an encoder from data to hidden variables and a decoder back to data. InfoVAE changes its training objective to balance generation, inference, and the information retained in those hidden variables. It introduces adjustable weights and alternative ways to compare the aggregate latent distribution with the model's prior.", "theoryRole": "Proposition 2 supports the use of alternative divergence measures by asserting that exact data and inference matching remain globally optimal at a fixed mutual information. It provides an idealized optimization justification for the broader objective family." }, { "slug": "ebgan-off-support-equilibrium", "title": "Energy-based Generative Adversarial Network", "issueTitle": "An omitted off-support condition in the EBGAN equilibrium characterization", "year": 2017, "category": "Generative models", "status": "counterexample", "version": "arXiv:1609.03126v4, 6 March 2017; ICLR 2017", "location": "Theorem 2 and footnote 1, PDF p. 3; Appendix A, proof of Theorem 2, PDF p. 11.", "originalClaim": "Theorem 2 characterizes equilibria by a matching generator and a constant discriminator energy γ on the data distribution. Footnote 1 extends the statement to regions where the data density is zero by allowing arbitrary discriminator values between zero and the margin there.", "formula": "p_G=p_D,\\qquad D(x)=\\gamma\\text{ on }\\operatorname{supp}(p_D),\\qquad 0\\le\\gamma\\le m,\\qquad D(x)\\in[0,m]\\text{ off support}.", "assumptions": [ "The analysis is the paper's nonparametric game with unrestricted generator and discriminator capacity.", "Both players minimize their respective objectives; the generator minimizes its expected discriminator energy.", "The discriminator energy is nonnegative and the hinge margin m is positive.", "The selected extension concerns data distributions with a region of zero density." ], "plainExplanation": "The off-support extension of Theorem 2 permits discriminator energies below the constant energy attained on the data support. Such a discriminator is a best response to a matching generator, but the generator can reduce its objective by moving to an accessible lower-energy region. The extension therefore fails as a sufficient Nash-equilibrium condition. The necessary distribution-matching result in Theorem 1 is unaffected.", "steps": [ { "title": "Specify the losses", "text": "For the matched generator, the discriminator sees the same distribution in its two terms. The generator minimizes energy under its own distribution.", "formula": "V(G,D)=\\mathbb E_{p_D}D(X)+\\mathbb E_{p_G}[m-D(X)]_+,\\qquad U(G,D)=\\mathbb E_{p_G}D(X)." }, { "title": "Construct a pair allowed by the footnote", "text": "Set m=1 and pD=pG=Uniform[0,1]. Choose a continuous nonnegative energy equal to 1/2 on [0,1], decreasing to zero on [1,2], and zero from 2 onward.", "formula": "D(x)=\\tfrac12\\min\\{1,\\max\\{0,2-x\\}\\},\\qquad \\gamma=\\tfrac12." }, { "title": "Verify the discriminator is a best response", "text": "On the common data/generated support, D+[1−D]+=1. This is the minimum of the pointwise sum, so no discriminator deviation lowers V.", "formula": "V(G,D)=1,\\qquad d+[1-d]_+\\ge1\\quad(d\\ge0)." }, { "title": "Exhibit the generator's profitable deviation", "text": "Keep D fixed and change the generator distribution to Uniform[2,3]. Its energy is now zero instead of one half.", "formula": "U(G,D)=\\tfrac12,\\qquad U(G',D)=0,\\qquad p_{G'}=\\operatorname{Uniform}[2,3]." }, { "title": "Apply the equilibrium definition", "text": "The required inequality U(G,D)≤U(G′,D) is false. The claimed sufficient characterization, including its off-support footnote, therefore admits a pair that is not an equilibrium.", "formula": "\\tfrac12\\not\\le0." } ], "counterexample": "Let data and generator be uniform on [0,1], with margin one and discriminator energy 1/2 on that interval. If the energy is zero on [2,3], the generator can reduce its objective from 1/2 to zero by changing support.", "witnessFormula": "p_G=p_D=\\operatorname{Unif}[0,1],\\quad D|_{[0,1]}=\\tfrac12,\\quad D|_{[2,3]}=0:\\quad U(G',D)=0<\\tfrac12=U(G,D).", "impact": "This refutes the sufficiency of the off-support extension. It does not refute existence of equilibria, Theorem 1's necessary distribution-matching conclusion, or Theorem 2 under its full-support assumption.", "practical": "This is a counterexample to the abstract game characterization. It is not a training experiment or a claim that a particular finite autoencoder architecture exhibits the deviation.", "repair": "For a matched generator and constant energy γ on its support, require D(x)≥γ almost everywhere in every off-support region reachable by an allowed generator. Together with 0≤γ≤m, this supplies the missing best-response condition. A global constant discriminator is one sufficient special case.", "repairFormula": "D(x)\\ge\\gamma\\quad\\text{on generator-reachable off-support regions}.", "repairProofFormula": null, "implementation": "Matching generated and data distributions is not, by itself, an equilibrium certificate for a fixed learned energy function. A certificate must also rule out lower-energy reachable regions. This correction does not require replacing the paper's loss functions.", "caveats": [ "The selected claim is the footnote read as part of Theorem 2's sufficient characterization.", "The example uses intervals of positive measure, so it does not depend on changing a density at a single point.", "The proposed energy can be continuous; discontinuity is not needed." ], "correctionSearch": "2026-09-09 The original equilibrium definition, theorem, footnote, and appendix proof were reread and the theorem page visually checked. Current arXiv history and targeted searches for support-related corrections were checked. Calibrating EBGAN Appendix A.2 was reread through Proposition A.1. \"Energy-based Generative Adversarial Network\" \"off-support\" \"Energy-based Generative Adversarial Network\" \"constant\" \"support\" equilibrium correction \"Energy-based\" \"GAN\" \"off-support\" equilibrium No source supplying the missing fixed-discriminator off-support lower bound was found in the bounded search. Calibrating EBGAN explicitly analyzes discriminator best responses and then profiles the discriminator as a function of the generator distribution. Its Eq. (21) also includes a data-expectation term. That optimization is not the fixed-discriminator generator deviation required by the original Nash definition, so its inspected result does not repair this sufficient characterization. The original OpenReview discussion and notes endpoints were not fully accessible. The search is bounded and does not establish that the observation is previously unknown.", "sources": [ { "label": "EBGAN, exact version", "url": "https://arxiv.org/pdf/1609.03126v4", "location": "Section 2.2, Theorem 2 and footnote 1, p. 3; Appendix A, p. 11" }, { "label": "Version history", "url": "https://arxiv.org/abs/1609.03126", "location": "Latest listed revision v4, 6 March 2017" }, { "label": "Calibrating Energy-based Generative Adversarial Networks", "url": "https://arxiv.org/pdf/1702.01691", "location": "Appendix A.2, Eqs. (19)–(23) and Proposition A.1, pp. 11–12" }, { "label": "ICLR discussion page", "url": "https://openreview.net/forum?id=ryh9pmcee", "location": "Discussion access was incomplete in this audit" } ], "auditId": "IG016", "primaryUrl": "https://arxiv.org/pdf/1609.03126v4", "citationCount": 1152, "citationUrl": "https://www.semanticscholar.org/paper/2ba23d9b46027e47b4483243871760e315213ffe", "citationDate": "2026-09-09", "paperSummary": "EBGAN trains a discriminator that assigns an energy score to each sample, with lower energy favored for real data and a margin-based penalty for generated samples. The generator learns to produce samples with low discriminator energy. The paper uses an autoencoder's reconstruction error as one way to define that energy.", "theoryRole": "The equilibrium analysis explains what the two-player objective should recover when both networks have unrestricted capacity. Theorem 2 characterizes pairs of generators and energy functions that are claimed to be equilibria, extending the distribution-matching conclusion of Theorem 1." }, { "slug": "numerics-gan-skew-scale", "title": "The Numerics of GANs", "issueTitle": "An omitted skew-part scale in the consensus damping bound", "year": 2017, "category": "Generative models", "status": "counterexample", "version": "arXiv:1705.10461v3, 11 June 2018; NeurIPS 2017 main paper and author-hosted supplement", "location": "Lemma 9, Eqs. (17)–(18), PDF p. 6; proof, Eq. (32), PDF p. 14 of the combined arXiv version.", "originalClaim": "Lemma 9 bounds the largest imaginary-to-real eigenvalue ratio of A−γAᵀA using a ratio c formed from the symmetric and skew parts of A, the smallest-singular-value lower bound ρ, and γ.", "formula": "q(\\gamma)=\\max_{\\lambda\\in\\operatorname{spec}(A-\\gamma A^\\top A)}\\frac{|\\operatorname{Im}\\lambda|}{|\\operatorname{Re}\\lambda|}\\ \\le\\ \\frac{1}{c+2\\rho^2\\gamma},\\qquad c=\\inf_{\\|v\\|=1}\\frac{|v^*(A+A^\\top)v|}{|v^*(A-A^\\top)v|}.", "assumptions": [ "A is a real, invertible matrix whose symmetric part is negative semidefinite; the paper explicitly permits nonsymmetric A.", "||Av||≥ρ||v|| for a positive ρ, and γ>0.", "The ratio defining c is evaluated on complex unit vectors; zero denominators are interpreted as an infinite ratio when the numerator is positive.", "The counterexample satisfies the stronger condition A+Aᵀ=−2I." ], "plainExplanation": "Lemma 9 bounds the imaginary-to-real eigenvalue ratio after replacing A by A−γAᵀA. The derivation drops the denominator associated with the skew quadratic form when introducing the singular-value lower bound. For A=[(−1,−1),(1,−1)] and γ=1, the transformed eigenvalues are −3±i: the ratio is 1/3, exceeding the printed upper bound 1/5. Retaining ‖A−Aᵀ‖₂ yields a valid sufficient bound and preserves the qualitative damping conclusion.", "steps": [ { "title": "Choose an admissible matrix", "text": "Use A=−I+J, where J rotates by ninety degrees. Its symmetric part is strictly negative and AᵀA=2I.", "formula": "A=\\begin{pmatrix}-1&-1\\\\1&-1\\end{pmatrix},\\quad J=\\begin{pmatrix}0&-1\\\\1&0\\end{pmatrix},\\quad A^\\top A=2I,\\quad\\rho=\\sqrt2." }, { "title": "Compute c", "text": "For every complex unit vector the numerator is 2 and the denominator is at most 2. A complex eigenvector of J attains the upper denominator, so c=1.", "formula": "|v^*(A+A^\\top)v|=2,\\quad |v^*(A-A^\\top)v|\\le2,\\quad v=\\tfrac1{\\sqrt2}(1,-i)^\\top\\Longrightarrow c=1." }, { "title": "Compute the damped eigenvalues exactly", "text": "At γ=1 the transformed matrix is −3I+J, with eigenvalues −3±i.", "formula": "B=A-A^\\top A=\\begin{pmatrix}-3&-1\\\\1&-3\\end{pmatrix},\\qquad \\operatorname{spec}(B)=\\{-3+i,-3-i\\}." }, { "title": "Compare the two sides", "text": "The actual ratio is 1/3. The printed right-hand side is 1/(1+4)=1/5.", "formula": "q(1)=\\tfrac13>\\tfrac15=\\frac{1}{c+2\\rho^2}." }, { "title": "Locate the missing factor", "text": "For a unit eigenvector, write a=|v*(A+Aᵀ)v| and b=|v*(A−Aᵀ)v|. The exact ratio is b/(a+2γ||Av||²). Dividing by b retains the factor 1/b on the regularization term.", "formula": "\\frac{|\\operatorname{Im}\\lambda|}{|\\operatorname{Re}\\lambda|}=\\frac{b}{a+2\\gamma\\|Av\\|^2}=\\frac{1}{a/b+2\\gamma\\|Av\\|^2/b}." }, { "title": "Check compatibility with a smooth game", "text": "The same A is the Jacobian of the ascent–descent field for a smooth concave–convex quadratic payoff, so the example is compatible with the paper's game framework.", "formula": "f(x,y)=-\\tfrac12x^2-xy+\\tfrac12y^2,\\qquad(\\partial_xf,-\\partial_yf)=A(x,y)^\\top." } ], "counterexample": "The matrix has AᵀA=2I, c=1 and ρ=√2. With γ=1, its regularized eigenvalues are −3±i, so the observed ratio 1/3 exceeds the claimed bound 1/5.", "witnessFormula": "A=-I+J,\\quad \\gamma=1,\\quad c=1,\\quad\\rho^2=2:\\qquad q(1)=\\tfrac13>\\tfrac15.", "impact": "Lemma 9's quantitative bound is false. The example and the repaired bound both retain the qualitative conclusion that sufficiently strong consensus damping reduces this eigenvalue ratio.", "practical": "This does not disprove the reported experiments or establish that consensus optimization diverges. It is a failure of the stated numerical guarantee.", "repair": "Write s=‖A−Aᵀ‖₂. For s>0, bounding the skew quadratic form by s gives the following sufficient estimate. If s=0, the regularized matrix is symmetric and the imaginary-to-real eigenvalue ratio is zero.", "repairFormula": "q(\\gamma)\\le\\frac{1}{c+2\\gamma\\rho^2/\\|A-A^\\top\\|_2}\\qquad(\\|A-A^\\top\\|_2>0).", "repairProofFormula": null, "implementation": "Any step-size or damping certificate derived from this particular quantitative bound must retain the skew scale or compute the relevant spectrum directly. The consensus update A−γAᵀA itself does not need to change.", "caveats": [ "The negative-semidefinite convention is the one explicitly used by the paper for nonsymmetric game Jacobians.", "The repaired estimate is a sufficient general bound and is exact for the displayed witness.", "Only Lemma 9's quantitative claim is evaluated here." ], "correctionSearch": "2026-09-09 The lemma, matrix assumptions, and proof were reread and the theorem page visually checked. Current arXiv history, the NeurIPS main paper, the author-hosted supplement, public reviews, and available author repository issues were checked. The relevant qualitative discussion in the author's dissertation was also inspected during the audit. \"The Numerics of GANs\" \"Lemma 9\" correction \"Numerics of GANs\" erratum correction No source repairing the missing skew scale in Lemma 9 was found in the bounded search. The inspected later qualitative convergence discussion does not give this quantitative inequality. Previously discussed minibatch-gradient bias concerns a different mechanism. The public issue inventory contained two records; absence of a correction there is not evidence about private correspondence. This search does not establish novelty or exhaust every citing paper.", "sources": [ { "label": "The Numerics of GANs, exact combined version", "url": "https://arxiv.org/pdf/1705.10461v3", "location": "Lemma 9, Eqs. (17)–(18), p. 6; proof Eq. (32), p. 14" }, { "label": "NeurIPS main paper", "url": "https://papers.neurips.cc/paper/6779-the-numerics-of-gans.pdf", "location": "Lemma 9, p. 6" }, { "label": "Author-hosted supplementary material", "url": "https://www.cvlibs.net/publications/Mescheder2017NIPS_supplementary.pdf", "location": "Lemma 9 and its proof" }, { "label": "Version history", "url": "https://arxiv.org/abs/1705.10461", "location": "Latest listed revision v3, 11 June 2018" }, { "label": "Official reviews", "url": "https://proceedings.neurips.cc/paper_files/paper/2017/file/4588e674d3f0faf985047d4c3f13ed0d-Reviews.html", "location": "Three public review reports" }, { "label": "Author code issues", "url": "https://github.com/LMescheder/TheNumericsOfGANs/issues", "location": "Open and closed issue records checked" } ], "auditId": "IG023", "primaryUrl": "https://arxiv.org/pdf/1705.10461v3", "citationCount": 490, "citationUrl": "https://www.semanticscholar.org/paper/459fbc416eb9a55920645c741b1e4cce95f39786", "citationDate": "2026-09-09", "paperSummary": "The Numerics of GANs studies training as a dynamical system in which two networks continually respond to one another. It connects poor local convergence to properties of the combined gradient's Jacobian and proposes consensus optimization, which adds a penalty based on the size of that gradient. The aim is to reduce unstable or strongly oscillatory behavior near an equilibrium.", "theoryRole": "Lemma 9 quantifies how consensus regularization changes the ratio of oscillation to decay in the local linear dynamics. This numerical estimate supports the paper's explanation of improved damping; it is distinct from the preceding qualitative local-convergence argument." }, { "slug": "dps-gaussian-noise-scaling", "title": "Diffusion Posterior Sampling for General Noisy Inverse Problems", "issueTitle": "Incorrect Gaussian noise scaling in the DPS Jensen-gap bound", "year": 2023, "category": "Generative models", "status": "counterexample", "version": "arXiv:2209.14687v4, 20 May 2024; ICLR 2023", "location": "Definition 1, Eq. (12), PDF p. 4; Theorem 1, Eq. (14), p. 5; Lemmas 2–3 and proof of Theorem 1, pp. 14–15.", "originalClaim": "Theorem 1 bounds the Jensen gap between the conditional likelihood and its posterior-mean approximation by a Gaussian-noise-dependent constant times the forward operator's derivative bound and the posterior's first centered absolute moment.", "formula": "\\mathcal J=\\mathbb E[p(y\\mid X_0)\\mid X_t]-p(y\\mid\\mathbb E[X_0\\mid X_t]),\\qquad \\mathcal J\\le\\frac{d}{\\sqrt{2\\pi\\sigma^2}}e^{-1/(2\\sigma^2)}\\|\\nabla\\mathcal A\\|m_1.", "assumptions": [ "The measurement noise is Gaussian with standard deviation σ>0.", "The forward operator has a finite derivative bound.", "The conditional distribution of X0 given Xt has a finite first centered absolute moment m1.", "The witness is one-dimensional and comes from an ordinary variance-preserving Gaussian forward diffusion with an exact conditional mean.", "The Jensen gap is signed as defined in Eq. (12); the witness below makes it positive." ], "plainExplanation": "The Gaussian likelihood constant in Lemma 2, propagated into Theorem 1, has incorrect dependence on the noise standard deviation. For identity observation, σ=0.1, y=0.2 and conditional state distribution N(0,0.01), the signed Jensen gap is approximately 0.497859. Equation (14) gives an upper bound of approximately 6.14×10⁻²³. The univariate likelihood has Lipschitz constant exp(−1/2)/(√(2π)σ²); substituting this constant repairs the density-bound argument.", "steps": [ { "title": "Use a realizable Gaussian conditional distribution", "text": "Take independent standard normals X0 and Z and Xt=√0.99 X0+√0.01 Z. At Xt=0, Gaussian conditioning gives X0|Xt=0 distributed as N(0,0.01), with conditional mean zero.", "formula": "X_0,Z\\stackrel{\\mathrm{ind}}{\\sim}\\mathcal N(0,1),\\quad X_t=\\sqrt{0.99}X_0+\\sqrt{0.01}Z,\\quad X_0\\mid X_t=0\\sim\\mathcal N(0,0.01)." }, { "title": "Set the measurement model", "text": "Use the identity forward operator, measurement standard deviation 0.1, and observed value y=0.2. The operator bound is 1 and the centered first moment is explicit.", "formula": "\\mathcal A(x)=x,\\quad \\sigma=0.1,\\quad y=0.2,\\quad m_1=0.1\\sqrt{2/\\pi}." }, { "title": "Integrate the likelihood exactly", "text": "Convolving the conditional N(0,0.01) distribution with N(0,0.01) measurement noise produces an N(0,0.02) likelihood for y. The plug-in approximation instead evaluates the N(0,0.01) density at y.", "formula": "p(0.2\\mid X_t=0)=\\frac{e^{-1}}{\\sqrt{0.04\\pi}},\\qquad p(0.2\\mid\\widehat X_0=0)=\\frac{e^{-2}}{\\sqrt{0.02\\pi}}." }, { "title": "Compare the signed gap with the printed bound", "text": "The gap is positive, approximately 0.497859. The theorem's right-hand side is approximately 6.14×10⁻²³.", "formula": "\\mathcal J=\\frac{e^{-1}}{\\sqrt{0.04\\pi}}-\\frac{e^{-2}}{\\sqrt{0.02\\pi}}\\approx0.4978590784>\\frac{e^{-50}}{\\pi}\\approx6.139401446\\times10^{-23}." }, { "title": "Differentiate the Gaussian density", "text": "For φσ(r)=(√(2π)σ)⁻¹exp(−r²/(2σ²)), the magnitude of the derivative is maximized at |r|=σ. This gives the correct one-dimensional Lipschitz constant.", "formula": "|\\phi_\\sigma'(r)|=\\frac{|r|}{\\sqrt{2\\pi}\\sigma^3}e^{-r^2/(2\\sigma^2)},\\qquad \\sup_r|\\phi_\\sigma'(r)|=\\frac{e^{-1/2}}{\\sqrt{2\\pi}\\sigma^2}." }, { "title": "Recover a valid first-moment estimate", "text": "Apply the corrected Lipschitz bound to the likelihood composed with the forward operator, then average the distance from the conditional mean. This bounds the absolute gap and therefore also its signed value.", "formula": "|\\mathcal J|\\le\\frac{e^{-1/2}}{\\sqrt{2\\pi}\\sigma^2}\\|\\nabla\\mathcal A\\|m_1." } ], "counterexample": "Under the stated Gaussian construction, E[p(y|X₀)]−p(y|E[X₀])=0.49785908>6.14×10⁻²³. The gap uses the signed convention in Definition 1; its positivity is essential to this counterexample.", "witnessFormula": "\\sigma=0.1,\\ y=0.2,\\ X_0\\mid X_t=0\\sim\\mathcal N(0,0.01):\\qquad \\mathcal J\\approx0.497859>6.1394\\times10^{-23}.", "impact": "The printed Eq. (14) fails even for a one-dimensional Gaussian forward diffusion and a linear observation operator. The univariate Gaussian constant in Lemma 2 is also incorrect.", "practical": "This is not an empirical test of DPS reconstructions. It does not show that the sampler fails on the reported tasks. The separate limitation that a density approximation does not by itself control the score has already been discussed in later work and is not presented here as an unresolved finding.", "repair": "Replace the univariate likelihood Lipschitz constant by e^(−1/2)/(√(2π)σ²). In n measurement dimensions with Euclidean norms, the isotropic Gaussian density has gradient supremum e^(−1/2)/((2π)^(n/2)σ^(n+1)); combine it with an explicit operator Lipschitz constant and m1.", "repairFormula": "|\\mathcal J|\\le\\frac{e^{-1/2}}{(2\\pi)^{n/2}\\sigma^{n+1}}\\operatorname{Lip}(\\mathcal A)\\,m_1.", "repairProofFormula": null, "implementation": "A certificate or error tolerance based on the printed noise constant needs correction. The mathematical repair does not change the DPS denoiser or guidance update. A guarantee on the score, which the algorithm uses, still requires a separate argument.", "caveats": [ "Likelihood densities can exceed one; the displayed likelihood values are densities, not event probabilities.", "The positive signed-gap witness uses y=0.2. At y=0, the corresponding signed gap is negative and would not by itself refute Eq. (14).", "The conditional Gaussian is the standard regular conditional distribution, and the violation persists for a neighborhood of the displayed observation and diffusion state." ], "correctionSearch": "2026-09-09 The signed-gap definition, theorem, Gaussian lemmas, and proof were reread; the theorem page was visually checked. Current arXiv history and targeted searches for the Gaussian Lipschitz constant and theorem correction were checked. Available author repository issues and comments and the indicated later Gaussian/score analyses were inspected during the audit. \"Diffusion Posterior Sampling\" \"Lipschitz constant\" error \"Diffusion Posterior Sampling\" \"Lemma 2\" correction \"Diffusion Posterior Sampling\" \"Gaussian\" \"bound\" \"incorrect\" No source correcting this exact Gaussian constant was found in the bounded search. Later work discusses Gaussian conditional covariance and the inability to derive a score bound directly from a density bound. Those concerns are treated as related, already discussed limitations and are excluded from this finding. The original OpenReview discussion and notes endpoints were not fully accessible. An indexed later document reproduced the disputed constant, but its complete identity and text were not verified; no downstream-propagation claim is made from that result. The literature search is bounded and does not establish novelty.", "sources": [ { "label": "DPS, exact version", "url": "https://arxiv.org/pdf/2209.14687v4", "location": "Definition 1/Eq. (12), p. 4; Theorem 1/Eq. (14), p. 5; Lemmas 2–3 and theorem proof, pp. 14–15" }, { "label": "Version history", "url": "https://arxiv.org/abs/2209.14687", "location": "Latest listed revision v4, 20 May 2024" }, { "label": "Author repository discussions", "url": "https://github.com/DPS2022/diffusion-posterior-sampling/issues", "location": "Available open and closed issues and comments checked during the audit" }, { "label": "Frequency-Guided Posterior Sampling for Diffusion-Based Image Restoration", "url": "https://arxiv.org/abs/2411.15295", "location": "Gaussian conditional/score analysis, Section 5 and Appendix B" }, { "label": "On the Guidance of Flow Matching", "url": "https://proceedings.mlr.press/v267/feng25s.html", "location": "Section 3.2, discussion following the local guidance error bound; density error versus guidance-vector-field error" }, { "label": "ICLR discussion page", "url": "https://openreview.net/forum?id=OnD9zGAGT0k", "location": "Discussion access was incomplete in this audit" } ], "auditId": "IG045", "primaryUrl": "https://arxiv.org/pdf/2209.14687v4", "citationCount": 1867, "citationUrl": "https://www.semanticscholar.org/paper/61e46884567be7cad12e999365b16a8d3414b678", "citationDate": "2026-09-09", "paperSummary": "Diffusion Posterior Sampling uses a pretrained diffusion model to reconstruct signals or images from noisy measurements, including nonlinear measurement processes. It combines diffusion sampling with measurement guidance computed through a denoised estimate of the unknown clean input. A key approximation evaluates the measurement likelihood at that estimate instead of averaging it over the entire conditional distribution.", "theoryRole": "Theorem 1 is intended to quantify the likelihood-level error introduced by the posterior-mean approximation under Gaussian measurement noise. Its bound is part of the justification for that approximation, but it does not directly bound the guidance-gradient error used by the sampler." } ]