{"id":"c74c878b-53b7-4b86-a925-34234afe7996","arxiv_id":"2602.00797","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A rectified flow has zero midpoint velocity iff its source and target distributions match; the paper turns this into a loss for sufficient encoders.","lead":"This paper shows that a rectified flow's velocity field is exactly zero at the midpoint of its path when the source and target distributions are identical, and that this zero-flow condition can be used as a training loss for learning sufficient representations and Markov blankets. A generalist should read it because it offers a non-parametric alternative to contrastive and reconstruction-based methods for representation learning, with a claim of robustness to shortcut feature","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 3.1's 'only if' direction is false as stated: zero midpoint velocity does not imply distributional equality unless characteristic functions are non-vanishing, a condition omitted from the main text.","rationale":"The reader's weakest assumption was the expected-norm relaxation of the zero-flow loss. That is a genuine gap, but the more fundamental issue is that the theoretical equivalence itself is not true without a condition not stated in the main text. The counterexample with a characteristic function vanishing on an interval shows that the 'only if' direction of Theorem 3.1 can fail, invalidating the pure mathematical statement. This does not depend on training or loss relaxation; it is a failure of the claimed necessary and sufficient criterion. The paper's own Lemma B.1 includes the non-vanishing assumption, but Theorem 3.1 and Theorem 3.3 omit it. The experiments use finite-support empirical distributions or bounded data, whose characteristic functions are trigonometric polynomials with at most isolated zeros, so the theorem may be repairable by adding the missing condition or a finite-support/analyticity restriction. The verdict therefore remains CONDITIONAL: the central claim is not sound as written, but the defect is localized and fixable, and the empirical evidence may still support the method once the condition is stated and verified.","tokens_in":21443,"tokens_out":33924,"duration_ms":390778,"concrete_test":"Verify the counterexample analytically: define ϕ0(t)=(1−|t|)_+ and choose a Schwartz function s with ∫s=0 and Fourier transform ŝ supported in |t|>1; set ϕ1(t)=ϕ0(t)+εŝ(t) with ε small enough that the inverse Fourier transform of ϕ1 is a nonnegative density. Show that ϕ1'ϕ0−ϕ0'ϕ1≡0. By the Fourier argument, this implies the midpoint conditional expectation E[X1−X0|X0+X1=s] is identically zero, while the two marginal densities differ. This settles whether Theorem 3.1 requires a non-vanishing characteristic-function assumption.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The load-bearing concern is not the finite-sample heuristic (which the authors admit) but a gap in the theoretical criterion itself. The proof of Lemma B.1 (Appendix B.2) derives ϕ1'ϕ0 = ϕ0'ϕ1 from v_{0.5}=0 and then divides by ϕ0(t)ϕ1(t) to conclude log ϕ1 = log ϕ0. This requires both characteristic functions to be non-vanishing for all t. Theorem 3.1 states the if-and-only-if without this assumption. The assumption is essential: take X0 with density f(x)=(1−cos x)/(π x²), whose characteristic function is ϕ0(t)=(1−|t|)_+ (identically 0 for |t|>1). Let X1 have density f(x)+εs(x), where s is a Schwartz function with ∫s=0 and Fourier transform supported in |t|>1, with ε>0 small enough for the density to be nonnegative. Then ϕ1=ϕ0 on [−1,1], and on |t|>1 both ϕ0 and ϕ0' vanish, so ϕ1'ϕ0−ϕ0'ϕ1≡0. Fourier inversion of the signed measure E[(X1−X0)1_{X0+X1∈·}] gives E[X1−X0|X0+X1=s]=0 for all s, yet p_X0≠p_X1. Thus Theorem 3.1 is false as stated; Theorem 3.3 inherits the defect for conditional distributions. Lemma B.1 contains the missing non-vanishing assumption, but the main theorems do not.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a representation-learning framework based on a 'zero-flow' criterion for rectified flows. It claims that for independent coupling, the optimal rectified-flow velocity at t=0.5 vanishes everywhere if and only if the source and target distributions are equal (Theorem 3.1), and that the conditional analogue certifies that an encoder f is sufficient for predicting X, i.e. X⊥⊥Y|f(Y) (Theorem 3.3). Based on this criterion the authors introduce a simulation-free loss (Eq. 8) and apply it to amortized Markov blanket discovery and to self-supervised learning, with experiments on synthetic graphical models, S&P 500 time series, and image datasets. The paper explicitly acknowledges in Section 7 that the loss is a heuristic relaxation of the pointwise zero-flow condition.","tokens_in":21849,"tokens_out":3659,"duration_ms":47526,"significance":"If the central theorems were correct as stated, the zero-flow criterion would be an elegant and useful bridge between flow matching and conditional-independence-driven representation learning. The paper has notable strengths: the mathematical development is substantial and includes proofs in appendices; the method is simulation-free; code is provided; and the experiments demonstrate practical promise, especially the non-parametric Markov blanket recovery and robustness to shortcut features. However, the main theoretical claim is not correct as stated, and the empirical objective is an acknowledged relaxation of the criterion that the theory supports. The significance of the paper therefore depends on whether the theoretical gap can be repaired without undermining the applicability of the method.","major_comments":[{"comment":"Theorem 3.1 states an if-and-only-if without any assumption on characteristic functions, but the 'only if' proof relies on Lemma B.1, which explicitly assumes non-vanishing characteristic functions. The division by ϕ0(t)ϕ1(t) in Eq. (22) is invalid when either characteristic function has zeros. This is not a technicality: a counterexample exists with X0 having density proportional to (1−cos x)/x² and X1 obtained by adding a small-ε Schwartz perturbation with Fourier support outside the support of ϕ0. Then ϕ1′ϕ0−ϕ0′ϕ1≡0, so v_{0.5}=0, while p_X0≠p_X1. Thus Theorem 3.1 is false as stated. Theorem 3.3 inherits the same defect in its 'only if' direction, which is the direction used to certify sufficiency of the learned encoder. Theorem C.1 also inherits the issue.","section":"Section 3.1 / Appendix B.2"},{"comment":"The theoretical criterion is v_{t=0.5}(z)=0 for every z, but the training objective minimizes an expectation of squared velocity evaluated only at the interpolated points X_t. Even if Theorem 3.3 were corrected, it would not imply that minimizing L(u,f) yields an encoder satisfying X⊥⊥Y|f(Y). The paper admits this in Section 7 ('we used heuristics to derive zero-flow loss'), but the abstract and introduction present the method as grounded in a rigorous criterion. This gap should be addressed explicitly, either by proving a finite-sample/expected-norm analogue, or by clearly stating the conditions under which the heuristic loss is justified and reframing the theoretical contribution accordingly.","section":"Section 3.3, Eq. (8), and Section 7"},{"comment":"The statement of Theorem 3.3 is ambiguous: it says 'For any pair (ξ,η), if f(ξ)=η, the velocity field ... =0 iff pX|Y=pX|f(Y).' If the zero-flow condition is verified for only one pair (ξ,η), the conclusion does not follow; one would need the condition for all ξ with f(ξ)=η. The proof in Appendix B.4 argues using U=X|Y=ξ and V=X′|f(Y′)=η and obtains p_U=p_V, i.e. p_{X|Y=ξ}=p_{X|f(Y)=η}, not the global equality p_{X|Y}=p_{X|f(Y)}. The theorem should be reformulated with the quantifiers made precise, or the conclusion weakened. This is a load-bearing point for the claimed certification of sufficiency.","section":"Section 3.2, Theorem 3.2 and 3.3"}],"minor_comments":[{"comment":"Several references are incomplete in the running text: 'Li et al.; Beizaee et al.', 'He et al.', and 'Park et al.' appear without publication years in the introduction (pages 1-2).","section":"Introduction"},{"comment":"Typo: 'Zeors' should be 'Zeros'; also 'Zeor-flow encoder' appears in the same paragraph.","section":"Appendix F.2.4"},{"comment":"The notation for the decoder input is inconsistent: the text says 'the latent vector reshaped to match the spatial dimensions' for both Y and D(f(Y)), but the equation (39) uses Y directly. Clarify the conditioning variables.","section":"Appendix F.2.3"},{"comment":"The S&P 500 experiment treats each stock as a sample and each trading day as a feature. This yields 500 samples and 252 features, which is a very small sample-to-dimension ratio; the interpretation of the learned Markov blanket as 'trading days most predictive of a five-day window' should be discussed in light of this transposed setup.","section":"Section 6.1.2"},{"comment":"The proof of Theorem 3.2 assumes the continuity equation with the given drift has a unique solution. This assumption is stated but not discussed; uniqueness for such conditional velocity fields is not automatic and should be commented on.","section":"Appendix B.3"}],"recommendation":"major_revision","confidential_remarks":"The core theoretical claim is false as stated because of the omitted non-vanishing characteristic-function assumption. This is repairable in principle, but the repair is nontrivial: it restricts the theorem to a class of distributions that excludes many common cases (e.g., compactly supported distributions can have zeroes in their characteristic functions), and the same restriction propagates to the conditional sufficiency guarantee. The empirical heuristic gap is also substantial. The paper has interesting ideas and promising experiments, but the theoretical foundation needs careful reworking before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things you should know. The central theoretical claim is overstated. The practical idea is still worth a look.\n\nThe observation itself—that when source and target are identical, the optimal rectified-flow velocity at t=0.5 vanishes pointwise—is correct and nicely demonstrated. The conditional transport statement (Theorem 3.2) is a standard flow-matching result, cleanly proved. Packaging zero-flow as a squared-loss objective for sufficiency is new to me, and the amortized Markov blanket experiments are solid: the method beats GLasso and PC on non-Gaussian graphs by large margins, and the LSTM inductive-bias version gets near-perfect AUC. Code is available.\n\nThe soft spots are in the theory. Theorem 3.1 claims an if-and-only-if: zero midpoint velocity for all z implies equality of distributions. The proof in Lemma B.1 divides by the characteristic functions φ0 and φ1. That requires them to be non-vanishing everywhere. The lemma states this condition; the main theorem does not. And it is essential, not a technicality. Here is the counterexample: let X0 have density proportional to (sin x / x)^4, whose characteristic function is a B-spline supported on a bounded interval (or use the simpler (1-cos x)/(πx^2) density with triangular characteristic function). Add a small mean-zero perturbation with spectral support outside that interval. Then φ0 and φ1 agree on the support of φ0, and both vanish on the complementary interval, so the equation φ1'φ0=φ0'φ1 holds everywhere, the midpoint conditional expectation E[X1-X0 | X0+X1=s] is zero for all s, yet the distributions differ. This example can be made with finite first moment, so it is not just a heavy-tail pathology. Theorem 3.3 inherits the defect. The authors need to either state the non-vanishing assumption prominently, or prove a different condition that applies to the distributions they actually use.\n\nAlso, the practical loss (8) enforces the criterion only in expectation over sampled Xt; the authors admit this is a heuristic. That is acceptable as a heuristic, but it means the paper does not deliver the certificate it advertises.\n\nWho should read this? People interested in flow-based criteria for sufficiency and invariance. It is a promising direction with a fixable flaw. I would send it to peer review; the reviewers should push on the assumptions and ask the authors to characterize when the equivalence holds.","headline":"The zero-flow criterion is an elegant idea, but the main theorem's converse depends on a missing non-vanishing characteristic-function assumption; the paper is promising, and the gap is fixable.","tokens_in":22271,"tokens_out":12949,"would_cite":false,"duration_ms":147389,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper proves that the velocity of a rectified flow at its temporal midpoint is zero exactly when the source and target distributions coincide, and that the same zero-flow condition, applied conditionally, certifies when a learned encod","keywords":["rectified flow","zero-flow criterion","conditional independence","sufficient representation","Markov blanket","self-supervised learning","shortcut problem","flow matching"],"falsifier":"Search for two distinct distributions whose characteristic functions vanish on a set (for example, distributions with zeros in their characteristic functions) and check whether the midpoint velocity of the rectified flow with independent coupling is identically zero; the paper's necessity proof assumes non-vanishing characteristic functions, so such a counterexample would bound the theorem's validity.","tokens_in":21363,"feed_emoji":"🎯","tokens_out":8432,"duration_ms":93668,"temperature":0.7,"pith_summary":"The paper tries to establish that the midpoint velocity of a rectified flow carries exact information about whether two distributions—or two conditional distributions—are the same. It proves a zero-flow criterion: the velocity at t=0.5 vanishes everywhere if and only if the distributions match, and in the conditional setting this vanishing is equivalent to X being conditionally independent of Y given f(Y). That equivalence gives a new way to learn sufficient representations: train an encoder so that the conditioning flow becomes still at its midpoint. The authors show this criterion yields a tractable, simulation-free loss, and they demonstrate it on two tasks—recovering Markov blankets in graphical models without parametric assumptions, and learning image representations that ignore superficial watermarks that derail contrastive methods.","feed_headline":"Zero velocity at flow's midpoint certifies a sufficient encoder","feed_subtitle":"A flow-based test for conditional independence lets encoders learn without labels or shortcuts.","key_machinery":"The central object is the optimal velocity field of a rectified flow: a time-indexed vector field v_t learned by regressing the straight-line displacement X' - X against the interpolation point X_t = tX' + (1-t)X, with X and X' sampled independently. The paper proves that evaluating this field at t=0.5 gives a zero vector for every z exactly when the two endpoint distributions are equal, and that conditioning the field on Y and f(Y') extends this to a test for p_{X|Y} = p_{X|f(Y)}. This zero-flow criterion does the work of the argument: it converts the abstract conditional-independence requirement into a regression target that can be minimized without simulating the ODE, and it supplies the","core_discovery":"The paper establishes a precise equivalence between a geometric property of a rectified flow and a statistical property of a learned representation. For a rectified flow trained on independent pairs (X, X'), the learned velocity field at the temporal midpoint t=0.5 is zero everywhere if and only if X and X' have the same distribution (Theorem 3.1). When the flow is conditioned on a candidate encoder f, the same midpoint condition holds if and only if the conditional distributions p_{X|Y} and p_{X|f(Y)} coincide (Theorem 3.3), which is exactly the conditional independence X ⊥⊥ Y | f(Y). Thus a vanishing velocity at the midpoint serves as a certificate that f is a sufficient representation for","pith_inferences":["The zero-flow criterion suggests a non-parametric two-sample test that could complement existing kernel-based tests, since it requires only training a flow and checking the midpoint norm.","The anti-symmetry theorem (Theorem C.1) implies the entire time-indexed velocity field, not just the midpoint, encodes distributional mismatch; this might support structural diagnostics, such as localizing where two distributions differ most.","The paper restricts to continuous variables; a natural testable extension is to embed discrete data or use stochastic interpolants to bring discrete Markov blanket learning under the same criterion.","The gap between pointwise zero and expected-norm minimization could be probed empirically by checking whether the learned encoder passes an independent conditional-independence test on holdout data; failures would motivate a stronger loss that samples more interpolation points."],"forward_implications":["If the zero-flow loss reaches zero, the encoder f provably satisfies X ⊥⊥ Y | f(Y), so f is a sufficient statistic for predicting X from Y.","The loss is simulation-free: it requires only samples from the joint distribution and a shuffled copy, not iterative ODE solving.","Markov blankets can be learned non-parametrically, without estimating a joint density or a partition function, and the amortized version generalizes to target feature sets unseen at training time.","Applying the criterion to two augmented views of an image enforces the multi-view assumption directly, so the learned representation does not collapse onto shortcut features such as watermarks or colors."],"fun_headline_variants":["Zero-flow criterion certifies sufficiency in encoders","Midpoint flow zero proves conditional independence","Flow's zero velocity at t=0.5 guarantees sufficient features","Encoders certified by zero-flow check","Zero-flow test for shortcut-free representation learning"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that the zero-flow criterion holds pointwise for all z, but the training objective only enforces it at sampled interpolation points; the paper does not prove that minimizing the empirical loss produces a representation satisfying the required conditional independence.","fun_headline_variants_meta":{"raw":{"variants":["Zero-flow criterion certifies sufficiency in encoders","Midpoint flow zero proves conditional independence","Flow's zero velocity at t=0.5 guarantees sufficient features","Encoders certified by zero-flow check","Zero-flow test for shortcut-free representation learning"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000648,"raw_usage":{"total_tokens":2786,"prompt_tokens":693,"completion_tokens":2093,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":437,"completion_tokens_details":{"reasoning_tokens":2022}},"tokens_in":437,"tokens_out":2093,"duration_ms":16055,"temperature":1.0,"reasoning_tokens":2022,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T05:55:01.343911+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Search for two distinct distributions whose characteristic functions vanish on a set (for example, distributions with zeros in their characteristic functions) and check whether the midpoint velocity of the rectified flow with independent coupling is identically zero; the paper's necessity proof assumes non-vanishing characteristic functions, so such a counterexample would bound the theorem's validity.","supporting_citations":[],"review_version":1}