{"id":"9096751a-3277-4ac2-b8f9-7954116a827f","arxiv_id":"2509.00538","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"LCUBE picks the causal direction with the shorter MDL description of a cubic spline regression, and the paper claims an asymptotic identifiability guarantee under low-noise assumptions.","lead":"This paper introduces LCUBE, a causal-direction score that compares how compactly a cubic spline can explain Y from X versus X from Y. It targets low-noise nonlinear bivariate data and reports strong benchmark results, especially on the Tuebingen cause-effect pairs.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Proof that LCUBE is an IRSF fails: Eq. (25)'s lambda depends on knot spacings and n, not on parameter count alone, so Lemma 1 does not apply.","rationale":"I read the paper as claiming a proof of identifiability for MDL scores over dense classes, instantiated by LCUBE. The proof relies on Lemma 1, imported from Marx and Vreeken's IRSF framework. For Lemma 1 to apply, the score must split into a strictly increasing function of expected error plus a strictly increasing function of parameter count. Eq. (24) has additional data-dependent and n-dependent terms in the complexity part. The paper's Eq. (25) attempts to identify λ with these terms, but λ then depends on knot positions and n, not only ∥βφ∥0. This is a load-bearing internal inconsistency, not merely a disagreement with external consensus. The reader's weakest_assumption was Assumption 4; that is also a gap, but fixing it would not restore the proof, because the IRSF condition fails before Assumption 4 is used. I therefore agree partially with the reader: the CONDITIONAL verdict is appropriate, but for a different primary reason. The proposed check is a direct, small-scale computation that would confirm or refute whether Eq. (24) is an IRSF. If the check shows the extra terms can be absorbed into γ or λ in a parameter-count-only way, the concern is resolved; otherwise the identifiability theorem for LCUBE is unproven as written. The empirical evaluation is useful independent support, but it does not repair the proof.","tokens_in":17342,"tokens_out":7340,"duration_ms":95378,"concrete_test":"Fix n=100, m=2. Fit two cubic spline models to the same data with the same m (same ∥βφ∥0 = m+4 if all coefficients nonzero) and the same RSS, but different knot sequences, e.g. k=(25,75) and k=(10,90). Compute Eq. (24) for both. If the two code lengths differ, L(Y|X,θ) depends on more than error and parameter count, so it is not an IRSF and Corollary 1 is not derived. To make the asymptotic point, repeat with n=200, n=400 and check whether the difference in Σ log u_j divided by n tends to 0; if it does not, the proof gap affects the claimed limit.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central theoretical claim is that the LCUBE score L(Y|X,θ) from Eq. (24) is an IRSF, so Lemma 1 yields identifiability. Definition 1 requires L = γ(ELSE) + λ(∥βφ∥0) with λ: N→R a strictly increasing function of the parameter count only. But the proof defines λ in Eq. (25) as log m + Σ_{j=1}^m log u_j + (m+4)/2 log n. The sum over log u_j depends on the knot increments u_j defined in Eq. (22), i.e. on the positions of the knots and hence on the data; (m+4)/2 log n depends on sample size. Two models with the same number of parameters and the same RSS can therefore receive different L values, so Eq. (24) is not of the IRSF form. The subsequent sentence 'Function λ is strictly increasing function for n' changes the variable of λ from parameter count to n, which is exactly what Definition 1 does not allow. Thus Lemma 1 cannot be invoked for LCUBE as written. This gap is independent of Assumption 4: even if the simplicity condition is true for cubic splines, the score is not an IRSF in the sense used by the theorem.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an MDL-based bivariate causal score for low-noise additive noise models, using function classes that are dense in C([0,1]). Theorem 1 claims that, under Assumptions 1-4, the ratio of conditional code lengths L(Yn|X,thetaX)/L(X|Yn,thetaY) converges to at most 1, with equality iff the regression function is linear. The method LCUBE instantiates the score with cubic regression splines. Corollary 1 asserts that the LCUBE score is identifiable by showing that the code length in Eq. (24) is an IRSF. The paper reports experiments on synthetic benchmarks and the Tuebingen cause-effect pairs dataset, with competitive or superior AUDRC/accuracy compared to several state-of-the-art methods.","tokens_in":17651,"tokens_out":6416,"duration_ms":80097,"significance":"If the identifiability theorem were correct, the paper would give a clean theoretical justification for a simple, fast, one-hyperparameter MDL score over dense function classes, going beyond the AIC/BIC instantiations of the IRSF framework. The empirical contribution is real: the authors provide public code, evaluate on a wide range of benchmarks, and report strong AUDRC on the Tuebingen pairs. However, the central proof connecting LCUBE to the IRSF framework is invalid as written: the score in Eq. (24) is not an IRSF in the sense of Definition 1, and Assumption 4 is asserted rather than derived for cubic splines. The theoretical claim is therefore not currently established, although it may be repairable with a substantial revision.","major_comments":[{"comment":"The proof of Corollary 1 defines lambda(||beta_phi||_0) := log m + sum_j log u_j + ((m+4)/2) log n. This depends on the data-derived knot spacings u_j defined in Eq. (22) and on the sample size n, not solely on the parameter count ||beta_phi||_0 as Definition 1 requires. Two spline models with the same number of parameters and the same RSS can therefore receive different values of L(Y|X,theta). Consequently Eq. (24) is not an IRSF, and Lemma 1 cannot be invoked for LCUBE. This is load-bearing for both Theorem 1 and Corollary 1.","section":"Section V-H, Eq. (25)"},{"comment":"The verification of the gamma condition is also not in IRSF form. The authors write gamma := (n/2) log(RSS(theta)/n) and say it is strictly increasing 'for n'. Definition 1 requires gamma: R_{>=0} -> R to be strictly increasing in the expected squared error only. The factor n and the use of the finite-sample RSS mean that the score is a function of (n, RSS, knot positions), not of (ELSE, parameter count). The asymptotic convergence RSS/n -> ELSE does not make the finite-sample score an IRSF.","section":"Section IV, proof of Theorem 1"},{"comment":"Assumption 4 is load-bearing for Lemma 1 and Corollary 1. For cubic splines it is asserted rather than derived. The text says 'we can therefore assume' and cites Kilbertus et al. [29] for low-degree polynomials, but no argument is given for splines with variable or equidistant knots. The intuitive smoothness discussion in Section VI-A does not establish ||beta_phi||_0 <= ||beta_psi||_0. If the anti-causal spline can be represented with fewer effective parameters, the ratio inequality in Lemma 1 can fail.","section":"Section V-H and Section VI-A, Assumption 4"}],"minor_comments":[{"comment":"Typo: 'numerator and denominator' is written as 'nomonator an denominator'. Also, the sentence claiming that fixed parameterizations 'do not affect the limit' is unclear and should be expanded or removed.","section":"Section IV, proof of Theorem 1"},{"comment":"The transition from Eq. (2) to Eq. (3) is inconsistent: the text mentions approximation error 1/(2n)N and then writes Y_n ≈ h(X) + (1/n)N. Please clarify the noise scaling.","section":"Section III-A, Eq. (3)"},{"comment":"The claim that the product m * prod_j u_j * n^{(m+4)/2} is a natural number for even m does not address the more serious issue that lambda must depend only on the parameter count. Also, if any u_j = 0, log u_j is undefined; the Schoenberg-Whitney conditions should be explicitly imposed in Algorithm 1 and checked in practice.","section":"Section V-H, Eq. (25)"},{"comment":"References [8] and [43] are the same paper (Galkin), and [2] and [25] are the same paper (Janzing and Schoelkopf). Please consolidate.","section":"References"},{"comment":"The column headings are confusing: '1. column', '2. column', and '2. column' are repeated. Clarify which columns correspond to the 10-dataset versus 13-dataset averages.","section":"Table I"}],"recommendation":"major_revision","confidential_remarks":"The empirical work is solid and the paper is clearly written, but the theoretical identifiability claim is the main advertised contribution and it currently rests on an incorrect IRSF verification. A revision that either repairs the proof by introducing a generalized score framework or modifies the MDL encoding so that Eq. (24) is genuinely an IRSF would be needed. The paper may be better suited to a venue that values the empirical comparison even if the theory remains incomplete."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The method is plausible and the empirical results are genuinely striking, but the central proof does not survive contact with its own definitions. The stress-test note is correct: Eq. (25) defines λ as log m + Σ log u_j + (m+4)/2 log n. That depends on the sample size n and on the data-derived knots u_j, not solely on the parameter count ||βφ||0 as Definition 1 requires. Lemma 1 from [30] only applies to scoring functions of the form γ(ELSE) + λ(||βφ||0), so the invocation fails. The sentence “Function λ is strictly increasing for n” changes the argument of λ from parameter count to sample size, which is exactly what the definition forbids. This is a load-bearing flaw, independent of Assumption 4.\n\nThere is also a second gap: Assumption 4 is asserted for cubic splines by analogy with low-degree polynomials, not derived. And the code-length derivation in Appendix B assumes knots are a subsequence of the data, while Algorithm 1 uses equidistant knots. The u_j counts then no longer correspond to the index differences the code length is built on. These are all fixable in principle, but as written the theorem is not established.\n\nTo give credit where it is due: LCUBE as a concrete method—MDL with cubic splines, one hyperparameter, fast, publicly available code—is new. The empirical results are strong: AUDRC 87 on Tuebingen pairs, perfect or near-perfect on several low-noise synthetic benchmarks, and the runtime comparison with Sloppy and ROCHE is concrete and useful. The paper is clearly written and situates itself honestly against Slope and Sloppy.\n\nThe soft spots are not cosmetic. The abstract and title promise identifiability, and that claim is currently unsupported. But the empirical work may still stand on its own, and the method is plausible enough to warrant a serious referee.\n\nMy recommendation: send it to peer review. A good referee can push for a corrected proof, a weakened claim, or a clearer separation of the empirical contribution from the theoretical one. I would not take the identifiability theorem at face value, but the paper deserves engagement rather than a desk rejection.","headline":"Plausible method, strong empirical numbers, but the identifiability proof does not go through as written: the score is not an IRSF because λ depends on n and knot spacings, not just parameter count.","tokens_in":18154,"tokens_out":2244,"would_cite":false,"duration_ms":26734,"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":"A minimum description length ratio over dense functional classes asymptotically identifies causal direction in low-noise bivariate additive noise models; the cubic-spline instantiation LCUBE reports high precision on benchmarks.","keywords":["causal direction","minimum description length","dense functional class","cubic splines","identifiability","additive noise model","low noise","LCUBE"],"falsifier":"Simulate X → Y with a nonlinear spline function f and low noise, then choose a function whose inverse is a spline with fewer knots; compute the LCUBE ratio for increasing n. If the ratio eventually exceeds 1, Assumption 4 is violated and Theorem 1's conclusion fails; a direct linearity check of f should likewise predict whether the method goes undecided.","tokens_in":17250,"feed_emoji":"↔️","tokens_out":9490,"duration_ms":100432,"temperature":0.7,"pith_summary":"The paper aims to prove that a Minimum Description Length causal score, defined through any functional class dense in the continuous functions on a compact interval, can separate cause from effect in bivariate additive noise models with low noise and no hidden confounders. Its main theorem states that the MDL score in the true causal direction is asymptotically at most the score in the anti-causal direction, with equality only for linear functions. The paper instantiates this with cubic regression splines in a method called LCUBE, which it shows is identifiable, has one hyperparameter, and runs quickly; it reports an AUDRC of 87 on the Tuebingen pairs and perfect AUDRC on most low-noise synthetic benchmarks. If correct, this gives practitioners a fast, principled way to decide causal direction under testable conditions and without assuming Gaussian noise.","feed_headline":"Cubic-spline MDL score identifies cause-effect direction in low noise","feed_subtitle":"The proof applies to any dense function class and needs no Gaussian-noise assumption; LCUBE tests it with splines.","key_machinery":"The load-bearing object is the score ratio L(Y|X, θ_X) / L(X|Y, θ_Y), where each code is a two-part MDL description: a parameter part log m + Σ log u_j + ((m+4)/2) log n and a goodness-of-fit part (n/2) log(RSS/n). The proof works by showing this score is an IRSF—a scoring function that adds a strictly increasing penalty for error to a strictly increasing penalty for parameter count—so the ratio inherits the regression-error ratio of Assumptions 1–3. The density property of the function class lets the approximation error be absorbed into the low-noise term, and Assumption 4 keeps the anti-causal model from being simpler than the causal one. For LCUBE, cubic splines supply the dense class and","core_discovery":"The paper proves that for a bivariate additive noise model Y = f(X) + (1/n)N, with compactly supported cause and low noise, the direction X → Y is identifiable from a ratio of two-part MDL code lengths. Theorem 1 states that if the regression functions come from any dense functional class and Assumptions 1–4 hold, then L(Y|X, θ_X) / L(X|Y, θ_Y) converges to at most 1 as n grows, with equality only when f is linear. The instantiation LCUBE uses cubic regression splines, which are dense in C([0,1]); its score is an IRSF, so the theorem applies. The paper also reports that LCUBE achieves AUDRC 87 on the Tuebingen pairs and perfect AUDRC on several low-noise synthetic benchmarks, with no Gaussia","pith_inferences":["Editorial inference: Assumption 4 is checkable on each dataset by comparing fitted parameter counts in both directions; if the anti-causal spline is systematically more compact, the claimed limit can fail and the score should not be trusted.","Editorial inference: because the theorem is stated for any dense class, the same proof strategy could be instantiated with other dense bases, but the paper only works out cubic splines; extending to another class would require a new practical MDL encoding and a new check of Assumption 4.","Editorial inference: the strong empirical numbers come mostly from low-noise synthetic benchmarks and one real-world collection; a wider set of real cause-effect pairs with known ground truth would test how the asymptotic guarantee holds at finite sample sizes."],"forward_implications":["Under the theorem, any dense functional class satisfying Assumptions 1–4 inherits the same asymptotic identifiability result, so the method is a template rather than a spline-only trick.","The score remains valid for non-Gaussian low-noise errors, extending regression-error causal inference beyond Gaussian likelihood settings.","For linear f, the score ratio converges to 1, so the algorithm should output 'undecided' rather than a direction; users should expect ties exactly in the linear case.","LCUBE's one hyperparameter and low computational cost make it practical for large collections of bivariate pairs, where neural approaches need hours.","Because the assumptions are testable, practitioners can check a given dataset for compact support and low noise before trusting the direction output."],"supporting_citations":[{"why":"Supplies Assumptions 1–3 and the low-noise regression-error inequality that Lemma 1 is built on.","marker":"[31]"},{"why":"Supplies Assumption 4, the IRSF definition, and the earlier identifiability theorem this proof adapts.","marker":"[30]"},{"why":"Used to argue the simplicity assumption holds because inverse polynomials cannot use fewer parameters.","marker":"[29]"},{"why":"Gives the geometric conditions under which a unique interpolating cubic spline exists.","marker":"[40]"},{"why":"Used to justify the density property and approximation order for cubic splines with many knots.","marker":"[42]"},{"why":"Provides the MDL coding of knot positions through successive index differences used in the LCUBE code length.","marker":"[27]"},{"why":"Gives the stochastic-complexity encoding of estimated parameters (half log n per parameter) in the MDL code.","marker":"[38]"},{"why":"Supplies the AUDRC evaluation measure and the LOCI baselines used in the empirical comparison.","marker":"[23]"},{"why":"Provides the Tuebingen cause-effect pairs benchmark and the forced-decision accuracy protocol.","marker":"[32]"},{"why":"Provides the AN, AN-s, LS, LS-s, and MN-U synthetic low-noise benchmarks used in evaluation.","marker":"[44]"}],"fun_headline_variants":["MDL score identifies causal direction without Gaussian noise","Cubic splines prove causal direction identifiable in low noise","Dense function classes make causality testable via MDL","LCUBE: fast MDL score for cause-effect direction detection","Causal direction inference from dense functional classes"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The load-bearing premise is Assumption 4: the best-fitting anti-causal function must need at least as many parameters as the best-fitting causal function; if a more compact inverse representation exists, the inequality behind the theorem can reverse.","fun_headline_variants_meta":{"raw":{"variants":["MDL score identifies causal direction without Gaussian noise","Cubic splines prove causal direction identifiable in low noise","Dense function classes make causality testable via MDL","LCUBE: fast MDL score for cause-effect direction detection","Causal direction inference from dense functional classes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000176,"raw_usage":{"total_tokens":1150,"prompt_tokens":795,"completion_tokens":355,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":539,"completion_tokens_details":{"reasoning_tokens":277}},"tokens_in":539,"tokens_out":355,"duration_ms":4724,"temperature":1.0,"reasoning_tokens":277,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T13:29:55.171098+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate X → Y with a nonlinear spline function f and low noise, then choose a function whose inverse is a spline with fewer knots; compute the LCUBE ratio for increasing n. If the ratio eventually exceeds 1, Assumption 4 is violated and Theorem 1's conclusion fails; a direct linearity check of f should likewise predict whether the method goes undecided.","supporting_citations":[{"cited_title":"Bl ¨obaum, D","cited_arxiv_id":null,"evidence_quote":"Supplies Assumptions 1–3 and the low-noise regression-error inequality that Lemma 1 is built on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies Assumption 4, the IRSF definition, and the earlier identifiability theorem this proof adapts."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the geometric conditions under which a unique interpolating cubic spline exists."},{"cited_title":"Subbotin, N.I","cited_arxiv_id":null,"evidence_quote":"Used to justify the density property and approximation order for cubic splines with many knots."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the MDL coding of knot positions through successive index differences used in the LCUBE code length."},{"cited_title":"Rissanen","cited_arxiv_id":null,"evidence_quote":"Gives the stochastic-complexity encoding of estimated parameters (half log n per parameter) in the MDL code."},{"cited_title":"Immer, C","cited_arxiv_id":null,"evidence_quote":"Supplies the AUDRC evaluation measure and the LOCI baselines used in the empirical comparison."},{"cited_title":"Mooij, J","cited_arxiv_id":null,"evidence_quote":"Provides the Tuebingen cause-effect pairs benchmark and the forced-decision accuracy protocol."},{"cited_title":"Tagasovska, V","cited_arxiv_id":null,"evidence_quote":"Provides the AN, AN-s, LS, LS-s, and MN-U synthetic low-noise benchmarks used in evaluation."}],"review_version":1}