{"id":"dc793c9c-efdc-4431-8512-85794f68c5a5","arxiv_id":"2505.19809","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"An equivariant representation learning framework estimates conditional distributions through a block-diagonal conditional expectation operator, with non-asymptotic sample-complexity guarantees that improve with the symmetry group's structure.","lead":"This paper shows how to train neural networks that respect a problem's symmetries while estimating conditional probabilities and uncertainty intervals, not just point predictions. It contributes mathematical learning guarantees and validates the approach on simulated symmetric data and on quadruped robot locomotion tasks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 5.1 is stated without the sub-Gaussian, equal-multiplicity, and anti-symmetry conditions that the appendix proof requires; as stated, the L2 regression bound is false for a heavy-tailed observable, so the advertised guarantees overclaim.","rationale":"The paper makes a coherent methodological contribution: the eNCP architecture, block-diagonal equivariant operator, disentangled isotypic losses, and the experimental validation are all plausible and detailed. The appendices contain substantial derivations, and the empirical results on the quadruped tasks support the practical value of the approach. However, the central theoretical claim is not supported as stated. The proof of Theorem 5.1 in Appendix M requires conditions that are absent from the theorem statement: sub-Gaussian h, equal latent multiplicities mk = m, anti-symmetry of A, and Assumption M.6. The heavy-tailed t3 counterexample shows that the L2 regression bound is false without the moment assumption, not merely unproven, because the sample-mean large deviations are polynomial rather than sub-Gaussian. The rare-event bound similarly depends on the anti-symmetric-A assumption to obtain the clean P[x ∈ G ▷ A] denominator and to drop the ηA term. These gaps directly affect the two quantities advertised as new: the effective-sample-size gain and the rare-event benefit. The right remedy is to restate Theorem 5.1 with the conditions actually used, or to prove the theorem under genuinely weaker assumptions. Since this is fixable by an amendment and the framework and experiments remain valuable, conditional acceptance is the appropriate verdict.","tokens_in":57498,"tokens_out":19800,"duration_ms":226567,"concrete_test":"Re-run the stated theorem in the base case G = {e}: let x and y be independent, y ~ t3, h(y) = y, and set δ = exp(-N^(1/6)). Compute the empirical mean bEy[y] over N samples and compare the failure probability at the threshold c N^(-1/2) log(1/δ) = c N^(-1/3) with δ. The t3 medium-deviation tail gives failure probability ~ N^(-1) for any fixed c, while δ is exponentially smaller, so the claimed bound is violated for large N. This settles that the sub-Gaussian assumption is genuinely load-bearing and must be added to Theorem 5.1.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The main theorem's proof does not support its statement. Appendix Theorem M.7 explicitly assumes that the event A is anti-symmetric for G, that mk = m for all k, that h and f are sub-Gaussian, and that Assumption M.6 holds; Theorem M.8 additionally assumes Y is sub-Gaussian and Eθ ≤ 1. None of these appear in the statement of Theorem 5.1, yet they drive the advertised rates. This is not cosmetic: for G = {e}, choose x and y independent with y ~ t3 (Student t with 3 degrees of freedom) and h(y) = y. Then h ∈ L2, the residual variance is finite, the operator Ey|x is finite-rank (so α = ∞), and all stated hypotheses of Theorem 5.1 are met. The estimator is essentially the sample mean, whose medium-deviation tail is P(|mean − Ey| > t) ≍ (√N t)^(-3). Taking δ = exp(-N^(1/6)) gives log(1/δ) = N^(1/6), so the claimed threshold t_claim ≍ N^(-1/2) log(1/δ) = N^(-1/3). At this threshold the actual failure probability is ≍ N^(-1), which is much larger than δ = exp(-N^(1/6)). Thus bound (19) is false as stated for the trivial-group base case. The rare-event bound (20) has a separate gap: for non-anti-symmetric A, Lemma M.10 introduces an additional relative-error term ηA that is absent from (20), so the effective-sample-size claim for rare events is likewise not proved without the anti-symmetry condition.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript introduces Equivariant Neural Conditional Probability (eNCP), an extension of the Neural Conditional Probability framework of Kostic et al. that incorporates finite group symmetries into representation learning for conditional distributions. The paper argues that, under G-invariance of the joint distribution, the conditional expectation operator decomposes into isotypic components, yielding disentangled equivariant representations and block-diagonal equivariant operator approximations. The central theoretical contribution is Theorem 5.1, which claims non-asymptotic L2 regression bounds with effective sample size amplified by the group structure and a rare-event bound for conditional probabilities controlled by the symmetry index of the conditioning set. The empirical section reports gains over symmetry-agnostic baselines on synthetic conditional Gaussian mixtures, quadruped momentum regression, and uncertainty quantification for ground-reaction forces and mechanical work in legged locomotion.","tokens_in":57908,"tokens_out":5423,"duration_ms":55574,"significance":"The intended contribution is significant: if the theorem is established under the conditions stated in the main text, this would be the first non-asymptotic statistical learning guarantee for equivariant conditional models with neural features, and the rare-event bound would quantify a concrete symmetry-induced variance reduction. The paper is also constructive: it gives a concrete architecture that can wrap any equivariant backbone, proposes a disentangled loss, and ships reproducible experiments with an analytical PMD benchmark that is well suited to isolating approximation error. However, the current version substantially overclaims: the theorem stated in Section 5 is not supported by the appendix proofs without additional assumptions and with a different rate in the regression case. These mismatches are load-bearing because they concern precisely the advertised effective sample size and rare-event improvements. The underlying equivariant operator decomposition and the concentration argument for the symmetry index appear to be a genuine new extension of the prior NCP work, not a circular restatement, but the stated theorem must be corrected before the guarantees can be accepted.","major_comments":[{"comment":"The statement of Theorem 5.1 claims bounds under only G-invariance, a Schatten-class condition, and an equivariant or invariant h, for any measurable A and any subgroup G′. The appendix proof does not establish this: Theorem M.7 additionally assumes that A is anti-symmetric for G, that mk = m for all k, that h and f are sub-Gaussian, and that Assumption M.6 holds; Theorem M.8 additionally assumes that Y is sub-Gaussian and that Eθ ≤ 1. None of these assumptions appear in Theorem 5.1. This is not a cosmetic mismatch: for G = {e}, y ∼ t3, and h(y) = y, all hypotheses stated in Theorem 5.1 are satisfied, but the sample-mean tail is polynomial, so the threshold implied by (19) is false in that base case. The theorem statement must either include the missing regularity and symmetry assumptions or state weaker bounds that follow from the provided proofs.","section":"§5, Thm 5.1 vs App. M, Thm M.7 and Thm M.8"},{"comment":"The rare-event bound (20) omits the relative-error term ηA that the appendix introduces and uses to derive the advertised orbit-probability gain. Theorem M.7's conditional-probability display contains the additive term sqrt(|G| ηA), and Lemma M.10 only produces the clean orbit-probability denominator after assuming g ▷ A ∩ A = ∅ for all nontrivial g. In the main text, ηA is absent and the anti-symmetry condition is not stated, so the bound in (20) is not the one proved in the appendix. The statement must include the anti-symmetry condition and the ηA term, or the rare-event claim must be weakened correspondingly.","section":"§5, Eq. (20) vs App. M, Thm M.7 and Lemma M.10"},{"comment":"The regression rate in Theorem 5.1 is stated with effective sample size disoN, but the appendix theorem that proves the regression bound, Theorem M.8, yields after balancing the dimension m a rate of the form (diso|G|N)^{-2α/(1+2α)} in squared error, i.e., an effective sample size of diso|G|N, not disoN. The main text's statement that symmetries 'boost the effective sample size to disoN' is therefore not supported by the appendix derivation if the appendix rate is the correct one; if the intended rate is instead disoN, a proof of that rate is missing. This discrepancy affects a central advertised contribution and must be resolved.","section":"§5, Thm 5.1 and discussion (item (ii)) vs App. M, Thm M.8"},{"comment":"Theorem M.8 also requires the sample-size condition N ≥ (cu ∨ cv)^2 m log(eδ^{-1}niso) ∨ |G| and the condition Eθ ≤ 1, while Theorem 5.1 states 'with probability at least 1 − δ' with no such restrictions. The missing Eθ ≤ 1 condition is not innocuous, since Eθ is an optimization error that the theorem treats as an arbitrary input. Without these conditions, the claimed high-probability statement (19) is not a consequence of the appendix theorem.","section":"§5, Thm 5.1 vs App. M, Thm M.8"}],"minor_comments":[{"comment":"The same symbol γG′ is used for the symmetry index of a set A and for the symmetry index of an observable h; these are different quantities with different ranges and should be denoted differently to avoid confusion.","section":"§5, Eq. (18) and App. M, Eq. (111)"},{"comment":"The sentence 'For intution refeer to the following blog post' contains typos and is followed by no reference or URL; please complete or delete this sentence.","section":"App. I.2"},{"comment":"The entry for eCQR reads 'Version of eCQR where...' and should read 'Version of CQR where...'.","section":"App. B"},{"comment":"The appendix overview says 'App. A summarizes the notations used, while App. A provides a glossary'; the second reference should presumably be to another appendix or should be removed.","section":"App. A"},{"comment":"The code reproducibility statement mentions an open-access repository and package 'symm_rep_learn', but no URL or persistent identifier is provided; please include one.","section":"App. G and §6"}],"recommendation":"major_revision","confidential_remarks":"The novelty of the equivariant extension and the independent concentration argument for the symmetry index are not in question; the problem is that the main theorem is stated more broadly than the appendix proves. This is correctable by aligning the statement with the appendix assumptions and rates. I would not recommend rejection because the central derivation appears sound under the appendix hypotheses, but the advertised guarantees cannot be accepted in their current form."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear [Colleague],\n\nThe paper's core idea is genuinely new: combining spectral contrastive learning (NCP) with the isotypic decomposition of G-symmetric function spaces, plus a block-diagonal equivariant matrix parameterization. That part is well done and the empirical results on robotics are plausible. But the main theorem as stated is not supported by the proof. The appendix requires extra conditions—A anti-symmetric, m_k = m for all k, sub-Gaussian observables, and Assumption M.6—none of which appear in Theorem 5.1. This is not cosmetic. For G = {e}, independent x and y with y ~ t_3, h(y) = y, the theorem's hypotheses hold (h ∈ L2, finite-rank operator), but the claimed Gaussian-type concentration bound (19) fails: the sample mean has medium-deviation tails of order (√N t)^{-3}, so the claimed threshold with δ = exp(-N^{1/6}) gives failure probability ~ N^{-1/2}, vastly larger than δ. So the advertised guarantees overclaim.\n\nThe rare-event bound (20) has a separate gap: Lemma M.10 introduces an extra relative-error term η_A that is absent from (20), so the effective-sample-size claim for non-anti-symmetric A is not proved.\n\nWhat the paper does well: the isotypic decomposition of the conditional expectation operator, the singular-space dimension argument (Prop K.3), and the block-diagonal architecture are solid and genuinely useful. The empirical section shows eNCP matches or beats baselines on regression and gives calibrated coverage on a robot locomotion task. The appendices are substantial.\n\nWhere it needs work: restate Theorem 5.1 with the assumptions actually used, or weaken the conclusions to match the proof. The current statement is too broad. Also, the main text advertises effective sample size diso N while the appendix uses different rates—clarify.\n\nWho is this for: researchers in equivariant representation learning and uncertainty quantification will find the architecture and empirical pipeline useful, but should not rely on the stated guarantees until the theorem is fixed.\n\nRecommendation: send to peer review, but flag the theorem-proof mismatch to the authors and require them to either add the missing assumptions to the statement or prove a bound that holds under the stated hypotheses. The core idea deserves referee time; the current draft is not ready for acceptance.","headline":"Novel equivariant contrastive architecture with a main theorem that overclaims; the proof requires extra conditions that are not stated, so the advertised statistical guarantees fail for heavy-tailed observables.","tokens_in":58394,"tokens_out":4422,"would_cite":true,"duration_ms":40792,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T07","62G05"],"pacs":[],"model":"deepseek-v4-flash","headline":"Symmetry-aware spectral contrastive learning yields equivariant conditional models with non-asymptotic error bounds and an effective sample size of $d_{\\mathrm{iso}}N$.","keywords":["equivariant representation learning","conditional expectation operator","spectral contrastive learning","isotypic decomposition","non-asymptotic statistical guarantees","uncertainty quantification","rare events","group symmetry"],"falsifier":"Run eNCP on a symmetric cGMM where the conditioning set $A$ is neither invariant nor anti-symmetric, so $0<\\gamma_{G'}(A)<1$, and compare the empirical error rate against $1/\\sqrt{|G'|\\,P(x\\in A)}$ with the rate $1/\\sqrt{P(x\\in G'\\triangleright A)}$; whichever denominator governs the observed rate settles which rare-event bound is real. A second decisive check is a target whose isotypic multiplicities $m_k$ differ across blocks, since the proof's balancing step assumes they are equal.","tokens_in":57305,"feed_emoji":"🔄","tokens_out":9873,"duration_ms":80212,"temperature":0.7,"pith_summary":"The paper aims to show that a known symmetry prior on the data-generating process is enough to turn spectral representation learning for conditional distributions into a provably sample-efficient method. It introduces an equivariant version of Neural Conditional Probability (eNCP): a G-equivariant, block-diagonal bilinear model that approximates the truncated singular value decomposition of the conditional expectation operator, with one block per isotypic component of the group action. The central claim is that this structural constraint increases the effective sample size from $N$ to $d_{\\mathrm{iso}}N$, and that for rare-event conditioning the relevant probability denominator shrinks from $\\sqrt{P(x\\in A)}$ to $\\sqrt{|G'|\\,P(x\\in A)}$. If correct, the result provides the first non-asymptotic statistical learning guarantees for equivariant conditional inference with learned neural features, and it makes uncertainty quantification a byproduct of regression in a single model.","feed_headline":"Symmetries enlarge effective sample size in conditional learning","feed_subtitle":"An equivariant contrastive model learns conditional distributions and regression with error rates that improve with the symmetry group's…","key_machinery":"The load-bearing object is the conditional expectation operator viewed as a G-equivariant linear integral operator between symmetric Hilbert spaces. Its isotypic decomposition $L^2=\\bigoplus_{k=1}^{n_{\\mathrm{iso}}} L^2_{(k)}$ splits the operator into blocks, and the group action forces each block to have the Kronecker form $\\Theta^{(k)}\\otimes I_{d_k}$, where $d_k$ is the dimension of the smallest faithful representation of the relevant subgroup; this is the mechanism that injects extra sample efficiency. The symmetry index $\\gamma_{G'}(A)$ measures how much the conditioning set overlaps its own orbit, and group-averaged U-statistics provide unbiased estimates of the contrastive loss with effective sample size $|G|N$ and $d_kN^2$.","core_discovery":"Under the assumptions that the joint distribution and the marginal of $x$ are G-invariant, the conditional expectation operator $E_{y|x}:L^2_y\\to L^2_x$ is G-equivariant and its kernel is the G-invariant pointwise mutual dependency. By the isotypic decomposition of the symmetric Hilbert spaces, the operator disentangles into blocks $E^{(k)}_{y|x}$; each block is parametrized as $\\Theta^{(k)}\\otimes I_{d_k}$, so its singular spaces have dimension at least $d_k$, the dimension of the corresponding irreducible representation. The paper defines eNCP as the model that learns these blocks with an equivariant bilinear contrastive loss, then reads off regression and conditional probabilities from the same learned operator. Theorem 5.1 states that, with probability at least $1-\\delta$, the $L^2$ error of the estimated conditional expectation $\\hat z_\\theta$ is bounded by $\\sqrt{\\mathrm{Var}[\\|h\\|]}$ times $[E^r_\\theta + (d_{\\mathrm{iso}}N)^{-\\alpha/(1+2\\alpha)}\\log(n_{\\mathrm{iso}}/\\delta)]$, while the set-conditioned error has the same factor divided by $\\sqrt{|G'|\\,P(x\\in A)}$ and multiplied by a symmetry-index term $\\gamma_{G'}(A)$.","pith_inferences":["A testable consequence not drawn in the paper: because the learned operator is a surrogate for the regression loss, the same machinery should provide a surrogate-gap bound that quantifies how much the contrastive training loss overestimates the equivariant regression MSE.","The rare-event result suggests a data-augmentation strategy: for a rare asymmetric event $A$, conditioning on the group orbit of $A$ is statistically as informative as observing roughly $|G'|$ independent samples of the event; this could be verified by comparing eNCP to a model trained with standard orbit averaging.","The equal-multiplicity assumption is likely a proof artifact rather than a structural requirement; if true in practice, allowing $m_k$ to differ would let the framework adapt to groups whose irreps appear with very different frequencies in a target.","One could turn the framework into a symmetry test: eNCP's validation performance versus its symmetry-agnostic counterpart measures how much of the conditional structure is actually explained by $G$, suggesting a score for detecting symmetry misspecification."],"forward_implications":["The same trained model yields G-equivariant regression, conditional probabilities, and confidence intervals at any coverage level without retraining, because it regresses the full conditional CDF.","Sample complexity of equivariant conditional inference improves from $N^{-\\alpha/(1+2\\alpha)}$ to $(d_{\\mathrm{iso}}N)^{-\\alpha/(1+2\\alpha)}$, a gain that grows with both the number of isotypic components and the dimensions of the faithful representations.","For an asymmetric conditioning set $A$, the rare-event probability denominator is multiplied by $|G'|$, so events with probability $P(x\\in A)$ can be conditioned on as if they had probability $|G'|P(x\\in A)$.","The framework is architecture-agnostic: any G-equivariant backbone can be used, and the guarantees connect the quality of the learned representation $E^r_\\theta=\\|E_{y|x}-E_\\theta\\|_{\\mathrm{op}}$ directly to the downstream error.","Setting $G$ to the trivial group recovers the symmetry-agnostic NCP baseline, so the guarantees include the earlier method as a special case."],"supporting_citations":[{"why":"Defines the neural conditional probability framework and the contrastive loss that eNCP generalizes, and supplies the baseline guarantees recovered when G is trivial.","marker":"[37]"},{"why":"Introduces operator-theoretic modeling of conditional expectations and U-statistic estimation that the paper builds on.","marker":"[36]"},{"why":"Provides the spectral contrastive loss perspective that justifies learning representations from the truncated singular functions.","marker":"[26]"},{"why":"Gives the optimal rank-r approximation result that makes the truncated SVD of the conditional expectation operator the reference solution.","marker":"[16]"},{"why":"Supplies the harmonic-analysis isotypic decomposition of symmetric function spaces used to disentangle the operator.","marker":"[43]"},{"why":"Shows the block-diagonal structure of G-equivariant operators used to justify the architecture.","marker":"[53]"},{"why":"Provides the conditional Gaussian mixture model used to generate symmetric synthetic data with an analytic kernel.","marker":"[23]"},{"why":"Provides the concentration inequalities used for the finite-sample bounds in the appendix.","marker":"[4]"}],"fun_headline_variants":["Equivariant inference: non-asymptotic guarantees for regression","Symmetry-enriched learning with provable uncertainty","Guaranteed equivariant learning via isotypic decomposition","Symmetry boosts sample efficiency: provable error bounds","Provable equivariant representation for conditional models"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is unstated in Theorem 5.1 but used in its proof: the conditioning set $A$ must be anti-symmetric, all isotypic multiplicities must be equal, the observable $h$ must be sub-Gaussian, and the symmetry-index bound of Assumption M.6 must hold; if any of these fails, the advertised effective-sample-size and rare-event gains are not proven.","fun_headline_variants_meta":{"raw":{"variants":["Equivariant inference: non-asymptotic guarantees for regression","Symmetry-enriched learning with provable uncertainty","Guaranteed equivariant learning via isotypic decomposition","Symmetry boosts sample efficiency: provable error bounds","Provable equivariant representation for conditional models"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000365,"raw_usage":{"total_tokens":1976,"prompt_tokens":966,"completion_tokens":1010,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":582,"completion_tokens_details":{"reasoning_tokens":935}},"tokens_in":582,"tokens_out":1010,"duration_ms":30652,"temperature":1.0,"reasoning_tokens":935,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T14:06:07.453501+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run eNCP on a symmetric cGMM where the conditioning set $A$ is neither invariant nor anti-symmetric, so $0<\\gamma_{G'}(A)<1$, and compare the empirical error rate against $1/\\sqrt{|G'|\\,P(x\\in A)}$ with the rate $1/\\sqrt{P(x\\in G'\\triangleright A)}$; whichever denominator governs the observed rate settles which rare-event bound is real. A second decisive check is a target whose isotypic multiplicities $m_k$ differ across blocks, since the proof's balancing step assumes they are equal.","supporting_citations":[{"cited_title":"Neural conditional probability for uncertainty quantification","cited_arxiv_id":null,"evidence_quote":"Defines the neural conditional probability framework and the contrastive loss that eNCP generalizes, and supplies the baseline guarantees recovered when G is trivial."},{"cited_title":"Learning invariant representations of time-homogeneous stochastic dynamical systems","cited_arxiv_id":null,"evidence_quote":"Introduces operator-theoretic modeling of conditional expectations and U-statistic estimation that the paper builds on."},{"cited_title":"HaoChen, Colin Wei, Adrien Gaidon, and Tengyu Ma","cited_arxiv_id":null,"evidence_quote":"Provides the spectral contrastive loss perspective that justifies learning representations from the truncated singular functions."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the harmonic-analysis isotypic decomposition of symmetric function spaces used to disentangle the operator."},{"cited_title":"Dynamics harmonic analysis of robotic systems: Appli- cation in data-driven koopman modelling","cited_arxiv_id":null,"evidence_quote":"Shows the block-diagonal structure of G-equivariant operators used to justify the architecture."},{"cited_title":"Conditional gaussian mixture models for environmental risk mapping","cited_arxiv_id":null,"evidence_quote":"Provides the conditional Gaussian mixture model used to generate symmetric synthetic data with an analytic kernel."}],"review_version":1}