{"id":"dd5bac05-823c-450a-a7c7-8e31d7c0b949","arxiv_id":"2608.01547","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"A calibrated inverse bound converts small finite-probe drift into a total-variation closeness certificate for distributions in a finite basis, with abstention when the observability margin is nonpositive.","lead":"This paper derives finite-sample total-variation upper bounds for generative drifting models, certifying that a small held-out drift implies the model is close to the target when the distributions live in a known finite basis. A smart generalist would read it to see whether frozen finite probes of a drift field can justify distributional closeness claims, and where the method abstains.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Corollary 4.4's beta_phi is inverted: Eq. (25) divides by ||eta||_2 where the derivation multiplies, so the printed Corollary 5.13 TV certificate is invalid; a disjoint-support m=2 basis yields TV=1 but the formula reports <=1/2.","rationale":"The reader's REJECT is supported. I re-derived both flagged constants. Lemma 3.2's Gaussian bound is too small: for x=0, y_+=u, y_-=-u with ||u||^2 = tau/2, ||K|| = 2||u|| exp(-||u||^2/tau) = sqrt(2 tau/e), exceeding sqrt(tau/e). Lemma 3.3's Laplace bound is also too small: since ell_tau uses squared distance, the correct sup is sqrt(tau/e), not tau/e; the proof's triangle inequality applies to the nonsquared norm. These wrong constants propagate into B_infty, B_N,tau and hence into every radius in Lemma 5.3 and Proposition 5.5. More damaging, Corollary 4.4's beta_phi is literally inverted relative to its own proof; the counterexample above gives a ratio of 2. Since Corollary 5.13's U_TV formula multiplies U_c by beta_phi, the final total-variation upper bound is not a valid 1-(delta_V+delta_M+delta_R) certificate as printed. These are localized arithmetic errors, so the general finite-dimensional inverse framework remains salvageable; the current manuscript's guarantees, however, do not follow. No circularity or data-snooping issue was found: external residual radii are explicit inputs, and the audit uses held-out samples. The reader's stated weakest assumption (external residual radii) is not where the load-bearing failure sits; the failure is in the printed constants, which is why my agreement is partial.","tokens_in":33731,"tokens_out":11196,"duration_ms":97436,"concrete_test":"Run the m=2 disjoint-support check: take phi_1=1_[0,1], phi_2=1_[1,2], a=(1,0), b=(0,1); compute beta via Eq. (25) and via the derivation in Corollary 4.4's proof. The printed value is 1/2 while the derivation requires 1, so Corollary 5.13's U_TV with zero residuals would be 0.5 < TV(p,q)=1, invalidating the certificate as stated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.6, Eq. (25) defines beta_phi = 1/(2||eta||_2) (sum_i ||phi_i||_L1^2)^{1/2}, but the proof of Corollary 4.4 and Lemma 4.3 require beta_phi = (1/2)||eta||_2 (sum_i ||phi_i||_L1^2)^{1/2}. Concrete counterexample: m=2, phi_1=1_[0,1], phi_2=1_[1,2], a=(1,0), b=(0,1). Then c_12=1, ||c||_2=1, p and q are densities with TV(p,q)=1, eta=(1,1), and sum ||phi_i||_L1^2 = 2. The correct beta is 1; the printed beta is 1/2, so the certificate would assert TV <= 1/2. Corollary 5.13 inherits this through U_TV = (rho_p+rho_q)/2 + beta_phi U_c, so the central upper-confidence claim is false as printed. The same certificate also inherits undersized Gaussian/Laplace envelopes: the true Gaussian sup of (7)-(8) is sqrt(2 tau/e) (attained at x=0, y_+=-y_-=sqrt(tau/2)), not sqrt(tau/e), and the squared-exponential Laplace kernel (9) has sup sqrt(tau/e), not tau/e; these feed the Hoeffding and empirical-Bernstein radii of Lemma 5.3. Both defects are arithmetic, not conceptual: replacing the constants restores the framework, but the published bounds do not hold.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops a finite-probe methodology for turning noisy, held-out evaluations of a drifting vector field into finite-sample upper confidence bounds on total variation between a target density p and a model density q, within a declared finite basis or with externally supplied L1 residual radii. The main line is: an exterior-product identity expresses the sampled drift as Mc; a singular-value inverse inequality controls the mismatch c; a bridge lemma converts mismatch to TV; and a composition of drift, operator-calibration, and residual error events yields the held-out certificate of Corollary 5.13, with abstention when the calibrated observability margin is nonpositive. Supporting results include global envelope bounds for Gaussian-RBF and squared-exponential interactions, random-probe conditioning through a population Gram matrix, and a large-bandwidth collapse to mean matching. The paper is explicit about its scope: certificates are conditional on a finite basis or valid external residual radii, on fresh held-out audit data, and on the unnormalized drift numerator.","tokens_in":34001,"tokens_out":7981,"duration_ms":71857,"significance":"If the printed constants are corrected, the paper is a useful contribution. It composes standard concentration and matrix-perturbation tools into an end-to-end TV certificate with a clearly stated abstention rule, and it is unusually careful about the distinction between a plug-in diagnostic and a valid upper confidence bound. The treatment of external residual radii, the separation of formal versus feasible mismatch directions, the population Gram characterization of probe design, and the explicit large-bandwidth failure mode are all valuable and clearly explained. The paper also ships reproducible scripts and distinguishes machine-verified quantities from descriptive numerical evaluations. However, the manuscript as printed contains load-bearing arithmetic errors that make the central certificate invalid: the mismatch-to-TV constant in Eq. (25) is inverted, and the two global interaction envelopes in Lemmas 3.2 and 3.3 are too small. These are local and correctable, but they must be fixed and the affected numerical statements re-evaluated before the central claim can be accepted.","major_comments":[{"comment":"The definition of beta_phi is inverted. The proof of Corollary 4.4 correctly gives ||p-q||_L1 <= ||a-b||_2 (sum_i ||phi_i||_L1^2)^{1/2} <= ||eta||_2 ||c||_2 (sum_i ||phi_i||_L1^2)^{1/2}, so the correct constant is beta_phi = (1/2) ||eta||_2 (sum_i ||phi_i||_L1^2)^{1/2}. As printed, Eq. (25) divides by ||eta||_2 instead of multiplying. Concretely, for m=2 with phi_1=1_[0,1], phi_2=1_[1,2], a=(1,0), b=(0,1), one has c_12=1, eta=(1,1), TV(p,q)=1, and the correct beta_phi is 1, while the printed formula gives beta_phi=1/2. Thus Corollary 4.4 as printed would assert TV(p,q) <= 1/2 for this example. Since Corollary 5.13 inherits this via U_TV = (rho_p+rho_q)/2 + beta_phi U_c, the held-out TV certificate is not valid as printed. The intended statement 'TV(p,q) <= (m/2) ||c||_2' in the density-basis case is consistent with the corrected formula and should replace Eq. (25).","section":"§4.6, Eq. (25)"},{"comment":"The Gaussian-RBF envelope constant is wrong by a factor of sqrt(2). For K_tau from (7)-(8), writing u=y_+-x, v=y_- -x, and s=||u-v||_2, the proof correctly obtains ||K_tau|| <= s exp(-(||u||^2+||v||^2)/(2 tau)) <= s exp(-s^2/(4 tau)). The maximum of s exp(-s^2/(4 tau)) is sqrt(2 tau/e), attained at s=sqrt(2 tau), not sqrt(tau/e) as stated in Eq. (10) and used throughout (A5), (A7), Table 2, Lemma 5.3, Corollary 5.10, and the audit defaults in Corollary 5.13. In particular, the claimed coordinate bound B_infty = sqrt(tau/e) is not an almost-sure envelope for the Gaussian interaction.","section":"§3.4, Lemma 3.2 and §B.1"},{"comment":"The envelope for the squared-exponential interaction (9) is also too small. With s=||y_+-y_-||_2, the proof gives ||K_tau|| <= s exp(-s^2/(2 tau)), whose maximum is sqrt(tau/e) at s=sqrt(tau), not tau/e as stated in Eq. (12). The value tau/e is the maximum of s exp(-s/tau), i.e. the true Laplace kernel exp(-||x-y||_2/tau), which is not the kernel (9). This error propagates into B_infty and B_X for the Laplace benchmark and into the empirical-Bernstein and bounded-vector radii of Lemma 5.3 and Proposition 5.5, so the explicit numerical radii in the experiments are understated.","section":"§3.4, Lemma 3.3 and §B.2"}],"minor_comments":[{"comment":"The component study reports U_TV = min{1, U_c} in the exact-basis m=2 case. This coincides with the corrected density-case constant m/2 = 1, but contradicts the printed Eq. (25), which would give min{1, U_c/2}. Please reconcile the notation so that the displayed formula, the theorem, and the experiment use the same constant.","section":"§7.2"},{"comment":"The entries for B_N,tau and the stated 'rigorous value' B_infty = sqrt(tau/e) = 0.857764 in Section 7.2 inherit the wrong envelope constants. After replacing them with sqrt(2 tau/e) for the Gaussian interaction and sqrt(tau/e) for (9), the numerical medians and informative-certificate rates in Sections 7.2 and 7.3 should be recomputed, since the printed radii are not valid envelopes.","section":"Table 2 and §7.2"},{"comment":"Calling (9) the 'Laplace similarity' is potentially confusing: it is a squared-exponential kernel exp(-||x-y||^2/tau). If the original drifting objective indeed uses this kernel, consider naming it 'squared-exponential' to distinguish it from exp(-||x-y||_2/tau), whose envelope is tau/e.","section":"§3.4"}],"recommendation":"major_revision","confidential_remarks":"The two arithmetic errors in Eq. (25) and Lemmas 3.2-3.3 are real and invalidate the printed central certificate, but they are local constant errors: the surrounding derivation is coherent, the proofs in B.1 and B.2 already contain the correct inequalities, and Corollary 5.10 already uses the correct density-case constant m/2. I therefore do not recommend rejection; the authors should correct the constants, update all dependent displayed radii, and re-run the affected experiments to confirm that the reported numerical claims still hold with valid envelopes."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things you should know. The paper has a genuinely useful idea: turn a finite-probe drift measurement into a finite-sample TV upper bound with abstention, under a declared finite basis or external residual radii. The composition—inverse inequality, concentration, operator calibration, Gram-matrix probe design, large-bandwidth collapse—is clearly argued and honestly scoped. But as printed, the central certificate is not true. Corollary 4.4 defines beta_phi with 1/(2||eta||_2) where the proof gives (1/2)||eta||_2, so the TV bound is too small by a factor of ||eta||_2^2. The concrete m=2 example (two unit intervals, swapped coefficients) gives TV=1 while the formula reports <=1/2. Corollary 5.13 inherits this, so the headline TV upper confidence bound is invalid. Also, Lemma 3.2's Gaussian envelope is wrong: the correct sup over the interaction (7)-(8) is sqrt(2 tau/e), not sqrt(tau/e), so the Hoeffding and bounded-vector radii in Lemma 5.3 are understated by sqrt(2). The stress-test also claimed the Laplace constant was wrong, but that one doesn't land: for the Laplacian kernel in (9), the paper's tau/e bound is correct.\n\nWhat's actually new and good: the calibrated end-to-end bound with explicit abstention, the population Gram-matrix condition for random probes, and the O(1/tau) large-bandwidth collapse are not in the prior literature, and the paper credits the underlying finite-basis identity to Deng et al. The experiments are thorough and reproducible, with prespecified radii, held-out batches, external residual radii, and an outward-rounded Gram bound. No circularity: the residual radii are declared inputs, not fitted from the audit batch.\n\nWhere the soft spots are: the two arithmetic errors are load-bearing, not cosmetic. The beta_phi inversion can be fixed by changing one denominator to a multiplication; the Gaussian constant by replacing tau/e with 2tau/e. The high-level structure survives. But the paper as it stands advertises rigorous radii it does not have. The \"invented entities\" and \"free parameters\" flags are not issues; those are honest declarations.\n\nWho it's for: someone working on drift/generative modeling or finite-probe distributional inference will want this after a revision. The current version should not be published as is. A serious referee should engage with it—the claims are checkable, the mistakes are identifiable, and the design is worth salvaging. I'd reject this version but invite a resubmission.","headline":"The framework is real, but the printed TV certificate is invalid: Corollary 4.4 inverts a factor and the Gaussian envelope is off by sqrt(2); both are fixable.","tokens_in":34609,"tokens_out":5793,"would_cite":false,"duration_ms":47080,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62G05","62G20","62F25"],"pacs":[],"model":"deepseek-v4-flash","headline":"A finite noisy drift field can certify total-variation closeness within a declared finite basis when the measurement operator is well conditioned.","keywords":["total variation certificate","drifting objectives","finite-basis identifiability","observability margin","random probes","Gram matrix","large-bandwidth collapse","empirical Bernstein radii"],"falsifier":"Take the m=2 Gaussian-basis example of Section 7.2 with n=256 and the empirical-Bernstein radius, draw many fresh audit batches, and check whether the true total variation 0.47725 lies below U_TV in at least 95% of runs; a coverage rate below the nominal level would falsify the certificate's probability statement.","tokens_in":33430,"feed_emoji":"🎯","tokens_out":8392,"duration_ms":72553,"temperature":0.7,"pith_summary":"The paper asks what can be concluded about two probability laws when the vector field of a drifting objective is observed noisily at finitely many probe locations. Its answer is that, within a declared finite density basis—or for normalized finite-basis approximants with externally validated $L^1$ residual radii—the sampled drift obeys a finite linear identity, and that identity can be inverted into an a posteriori upper confidence bound on total variation. The bound explicitly accounts for held-out field noise, operator-estimation error, and representation residuals, and it abstains by returning the trivial bound $1$ when the calibrated observability margin is not positive. The result matters because it turns small held-out drift into a conditional distributional guarantee, while making the conditions for that guarantee explicit.","feed_headline":"Small drift can certify closeness if probes are well conditioned","feed_subtitle":"A held-out audit returns a TV upper confidence bound—or abstains when the observability margin fails.","key_machinery":"The load-bearing object is the observation matrix $M\\in\\mathbb{R}^{dN\\times r}$, whose columns are the probe-stacked pair responses $U_{ij}=\\int\\!\\int K(x_\\ell,y^+,y^-)\\phi_i(y^+)\\phi_j(y^-)\\,dy^+dy^-$; the identity $\\operatorname{vec}(V_X)=Mc$ reduces drift identifiability to a finite-dimensional inverse problem. The argument's backbone is the inverse inequality of Theorem 5.1, the mismatch-to-coefficient bridge $\\|a-b\\|_2\\le\\|\\eta\\|_2\\|c\\|_2$ (a Lagrange-identity/operator-norm identity), and the conversion constant $\\beta_\\phi$ that maps $\\|c\\|$ to total variation. For random probes, the population quantity $\\gamma(\\nu)=\\lambda_{\\min}(\\Gamma(\\nu))$ with $\\Gamma(\\nu)=\\mathbb{E}[G(X)^\\top G(X)]$ controls conditioning: positive $\\gamma$ plus a probe-count condition yields high-probability full column rank, and a nonpositive calibrated margin $\\sigma-\\varepsilon_M$ is the explicit abstention trigger.","core_discovery":"Within the declared scope, every antisymmetric interaction-kernel drift can be written at the probes as $\\operatorname{vec}(V_X)=Mc$, where $c=a\\wedge b$ is the antisymmetric coefficient mismatch and $M$ is a probe-dependent observation matrix; full column rank of $M$ makes zero drift imply $p=q$ inside the basis, and a smallest-singular-value margin converts approximate zero drift into a norm bound on $c$. The paper's central certificate (Corollary 5.13) bounds the mismatch as $\\|c_m\\|_2\\le (\\|\\operatorname{vec}(\\hat V_X)\\|_2+\\varepsilon_V+\\|R_m\\|_2)/(\\sigma-\\varepsilon_M)$ when $\\sigma-\\varepsilon_M>0$, then bridges that bound to $\\operatorname{TV}(p,q)$ through an exact exterior-product coefficient identity; with the three declared error budgets it yields $P\\{\\operatorname{TV}(p,q)\\le U_{\\operatorname{TV}}\\mid F_0\\}\\ge 1-\\delta_V-\\delta_M-\\delta_R$ almost surely, and $U_{\\operatorname{TV}}=1$ with abstention otherwise. It also shows that random probes achieve full rank with high probability governed by the population Gram matrix $\\Gamma(\\nu)$, and that large bandwidth collapses the Gaussian and Laplace fields to mean matching at rate $O(1/\\tau)$.","pith_inferences":["A user monitoring a black-box generator would need an independent approximation certificate for the residual radii; otherwise the audit's honest output is only the conditional sensitivity curve, not a full-distribution total-variation bound.","The same inverse bound can be specialized to bounded integral probability metrics: by the paper's Remark 4.5, any bounded test-function class inherits the total-variation certificate with a constant factor, suggesting a cheap way to certify bounded-kernel MMD without a new argument.","Repeated audits over time would require a simultaneous-confidence or confidence-sequence correction; the paper flags this as future work, so naively re-running Corollary 5.13 on fresh batches would inflate the error rate."],"forward_implications":["If the margin $\\sigma-\\varepsilon_M$ is positive, the audit's $U_{\\operatorname{TV}}$ is a valid one-sided $1-(\\delta_V+\\delta_M+\\delta_R)$ upper confidence bound on the true total variation, conditional on the declared basis and residual structure.","The induced equivalence test has controlled type-I error: it rejects a prespecified null $\\operatorname{TV}(p,q)\\ge\\epsilon$ only when $U_{\\operatorname{TV}}<\\epsilon$, and a failure to certify remains abstention rather than evidence that $p\\neq q$.","For i.i.d. probes from a law $\\nu$, the sample size condition $N\\gtrsim (8L^2/\\gamma)\\log(r/\\delta')$ makes the observation matrix full-rank and well conditioned with high probability, so maximizing $\\lambda_{\\min}(\\Gamma(\\nu))$ is a concrete probe-design rule.","Degenerate or overestimated measurement systems are honestly handled: when $dN