{"id":"48606162-fbb9-4679-be58-05910482fa24","arxiv_id":"2502.04495","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"DIF learns and isolates the environment-independent function of a dynamical system from multi-environment trajectories using a causal graph, a hypernetwork, and an adversarial independence constraint.","lead":"This paper introduces DIF, a method that learns the environment-independent part of an ODE from trajectories observed in several environments, recovering for example the undamped pendulum term from damped and powered pendulum data. A generalist may care because this is a step toward extracting physical laws from observational data gathered under varying conditions.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 3.1's uniqueness proof conflates independence from the environment label e with independence from the environment function fe; an e-independent statistic of Xp that is informative about fe suffices to refute the claimed guarantee.","rationale":"The reader correctly identified the gap between the theoretical principle in Eq. 2 and the implemented objective in Eq. 3: Prop 3.3 is shown only for a single predictive model, independence is applied to zc rather than fc, and the final loss is a weighted combination without a proved equivalence. That is a real implementation-level concern. My independent stress-test locates a more fundamental problem upstream of that gap: the theorem itself is not established. Appx C.2's entropy argument uses h(Xp) ⊥⊥ e to conclude h(Xp) ⊥⊥ fe, which is not entailed by the SCM. Since fe may carry environment-independent stochastic components, there exist functions of Xp that are independent of the environment label e but still contain information about fe; such functions would achieve higher I(h(Xp); f | X0) than fc, breaking the claimed uniqueness. The explicit counterexample (fc ≡ 0, fe = e·W, h = W²) shows the theorem is false as stated, not merely hard to implement. Consequently the abstract's guarantee that invariant-function discovery is guaranteed by the information-based principle is unsupported. The empirical results on the three synthetic ODEs may still indicate that DIF works in practice, but they cannot rescue the formal claim, and the paper would need either a substantially stronger identifiability assumption (for example, requiring the extracted function to be independent of fe, not just of e, together with a genuinely informative invariant trajectory target) or a downgrade of the guarantee to a heuristic. The reader's CONDITIONAL verdict understates the severity; the central advertised guarantee is currently refuted, so I recommend REJECT for the present version.","tokens_in":57459,"tokens_out":12550,"duration_ms":143525,"concrete_test":"Analytically instantiate the SCM from Appx C.1 with fc ≡ 0, fe = e·W, e ∈ {±1} equally likely, W ∼ N(0,1), X0 = 0, f = fc + fe, and Xp = [f·t1, ..., f·tTc]. Verify that h(Xp) = (Xp/t)² is a deterministic function of Xp, is independent of e, has I(h(Xp); f) > 0, and is not the true invariant function fc ≡ 0. If this holds, the proof of Theorem 3.1 in Appx C.2 is invalid and the claimed uniqueness of the invariant function is false for the stated assumptions.","verdict_should_be":"REJECT","load_bearing_attack":"The central guarantee is Theorem 3.1/Appx C.2. Its proof needs to show that, among all functions h(Xp) with h(Xp) ⊥⊥ e, the true invariant function fc is the unique maximizer of I(h(Xp); f | X0). The proof instead effectively assumes h(Xp) ⊥⊥ fe. In the existence argument, H(f|fc)=H(fe) and H(f|f'_c) ≥ H(fe|f'_c) lead to H(fe) > H(fe|f'_c), which is said to contradict f'_c ⊥⊥ e; the contradiction only follows if H(fe|f'_c)=H(fe), i.e. if f'_c ⊥⊥ fe. That is not implied by f'_c ⊥⊥ e, because fe = ge(e, εe) can contain environment-independent components εe. A concrete SCM satisfying Fig. 2 is fc ≡ 0, fe = e·W with e ∈ {±1} equally likely and W ∼ N(0,1), f = fc + fe, X0 = 0. Then h(Xp) = (Xp/t)² = W² is independent of e, carries information about f, and is not fc. So the maximizer of Eq. 2 is not uniquely fc, and the theorem's necessity/sufficiency claim fails as stated. Eq. 3 then has no proven connection to fc: penalizing I(e; zc) does not enforce independence from fe, and the MSE term on the full observed derivative cannot by itself separate fc from fe.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces invariant function learning (IFL) for ordinary differential equations observed in multiple environments, with the goal of recovering the shared invariant dynamics while discarding environment-specific function forms. The authors propose a causal SCM in which the full derivative function is f = g_comp(fc, fe), design a hypernetwork-based method called DIF that outputs invariant and environment function embeddings, and state Theorem 3.1 claiming that maximizing I(h(Xp); f | X0) subject to h(Xp) ⊥ e uniquely recovers the true invariant function fc. Proposition 3.3 is used to convert this mutual information objective into an MSE loss, and the final objective in Eq. (3) combines derivative fitting for the full and invariant functions with adversarial environment independence. Experiments on three synthetic multi-environment ODE datasets compare DIF against MAML, CoDA, IRM, and VREx, and report both trajectory NRMSE and symbolic-regression explanations.","tokens_in":57810,"tokens_out":7852,"duration_ms":88522,"significance":"The task formulation and benchmark construction are useful: the paper identifies a genuinely different scenario from coefficient-level environment shifts, releases code, and provides a quantitative evaluation protocol via invariant trajectory ground truths. The DIF architecture and the reference-based hypernetwork implementation are practical contributions, and the empirical comparisons show that the proposed method obtains better invariant trajectory predictions than the adapted baselines. However, the central claimed guarantee is load-bearing and, as stated, fails under the paper's own SCM: the uniqueness proof in Appendix C.2 is invalid, and the implemented loss in Eq. (3) is not shown to be equivalent to the theorem's constrained mutual information objective. The paper's value currently rests on the empirical findings plus a plausible but unproven causal framing, not on the stated theoretical guarantee.","major_comments":[{"comment":"The uniqueness proof is invalid. Equation (8) treats H(fe) > H(fe|f'_c) as contradicting f'_c ⊥ e, but this only follows if H(fe|f'_c) = H(fe), i.e. if f'_c ⊥ fe. The constraint f'_c ⊥ e does not imply f'_c ⊥ fe because fe = ge(e, εe) can contain an environment-independent noise component εe. A concrete SCM satisfying the paper's Fig. 2 is: fc ≡ 0, e ∈ {±1} with equal probability, W ∼ N(0,1), fe = e·W, f = fe, and X0 = 0. Then h(Xp) = (Xp/t)² = W² is independent of e, carries information about f, and satisfies I(h; f | X0) > 0 = I(fc; f | X0). Thus the true invariant function fc is not the maximizer of Eq. (2), and Theorem 3.1 is false as stated. This invalidates the abstract's claim that discovery of invariant functions is guaranteed by the information-based principle.","section":"Appendix C.2 / Theorem 3.1"},{"comment":"Proposition 3.3 only establishes an equivalence between maximizing I(hθc(Xp); f | X0) and minimizing trajectory MSE for the single forecasting model p(X | hθc(Xp), X0). The practical objective in Eq. (3) is not that objective: it minimizes derivative residuals ∥dX/dt − fc(Xt)∥², adds the full-function derivative loss, and adds discriminator and adversarial losses with separate weights. No argument shows that optimizing this weighted multi-loss objective solves the constrained optimization in Eq. (2). In particular, derivative MSE is not equivalent to trajectory MSE for nonlinear ODEs, and the additional λc, λdis, λadv terms are not derived from the theorem. Consequently, the empirical recovery of fc is not guaranteed by Theorem 3.1 even if that theorem were correct.","section":"Section 3.2.1 / Eq. (3) / Proposition 3.3"},{"comment":"The adversarial independence constraint is applied to the hidden embedding zc, while the theorem requires hθc(Xp) ⊥ e. The manuscript does not state or prove that I(e; zc) = 0 implies I(e; fc) = 0; this requires the decoder to be injective and the discriminator to be optimal in a sense that is not established for the finite neural-network implementation. Thus, even under a repaired Theorem 3.1, the practical optimizer could in principle satisfy the constraint on zc while violating it on the function fc that is actually used for invariant forecasting.","section":"Appendix C.5 / Eq. (3)"}],"minor_comments":[{"comment":"The sentence 'an important scenario in scientific discovery has been overlooked, i.e., , which aims' contains a missing phrase and a stray comma; it should be reworded.","section":"Section 1"},{"comment":"Figures 2 and 6 are identical versions of the same SCM; retaining one would reduce redundancy.","section":"Appendix C"},{"comment":"The hyperparameters λc, λdis, and λadv are introduced with a heuristic relationship λadv = λc·λ'adv, but no sensitivity analysis or principled criterion for setting them is provided beyond the random search range in Appendix E.2.2.","section":"Section 3.2.1 / Eq. (3)"},{"comment":"The invariant ground truth Xc is generated by removing environment terms from the same SCM used in the paper's causal model; this is a reasonable synthetic benchmark, but the text should state explicitly that this evaluation measures recovery of the assumed causal decomposition rather than model-free discovery from raw observations.","section":"Section 5.2"}],"recommendation":"major_revision","confidential_remarks":"The paper's empirical framework and benchmarks are potentially salvageable, but the central theoretical claim is false as stated and the practical loss is not derived from the theorem. I would not reject the paper outright because a strengthened assumption on fe (e.g., no environment-independent noise) and a separate proof connecting Eq. (3) to the principle could repair the contribution. However, the submission should not be accepted until those load-bearing gaps are addressed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague—\n\nKnow this: the paper's central guarantee, Theorem 3.1, is not proven as stated. The proof in Appx C.2 conflates independence from the environment label e with independence from the environment function fe. A concrete SCM shows the claimed uniqueness fails: take fc ≡ 0, fe = e·W with e ∈ {±1} equiprobable, W ∼ N(0,1), X0 = 0. Then (Xp/t)² = W² is independent of e but carries information about f, and is not fc. So the maximizer of Eq. 2 is not uniquely the true invariant function. This is load-bearing: the abstract's \"guaranteed\" discovery claim does not follow.\n\nWhat the paper does well: it identifies a genuinely new task—invariant function learning with environments differing in function form, not just coefficients—that CoDA, LEAD, and MetaphysiCa don't cover. The DIF architecture (transformer encoder, hypernetwork decoder, adversarial independence) is a reasonable first method. The three synthetic ODE benchmarks and released code are useful, and the symbolic regression explanations give a nice sanity check. The empirical advantage over adapted baselines is credible.\n\nOther soft spots are secondary. Prop 3.3 shows the MI-MSE equivalence for a single predictive model, not the actual weighted multi-loss objective, and the independence constraint is applied to the embedding zc rather than the function fc. Evaluation is on simulated data with ground truth generated from the same causal decomposition the method assumes, and there are no comparisons to standard equation discovery baselines. These are addressable; the theorem gap is not a mere technicality.\n\nWho should read this: researchers in scientific ML and invariant learning who want to think about what \"invariant function\" should mean for dynamical systems. It deserves a serious referee: the task is new, the method is a sensible first attempt, and the empirical story is coherent. But the theoretical claims need to be substantially softened or the proof repaired before the paper can be trusted as a guarantee. I'd recommend sending it to peer review with the expectation of major revision.\n\nBest,","headline":"New task with a plausible first method, but the central identifiability theorem is not proven as stated—e-independence does not imply independence from the environment function, so the paper's guarantee overreaches.","tokens_in":58354,"tokens_out":2387,"would_cite":false,"duration_ms":24722,"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 claims that shared ODE dynamics can be recovered from environment-biased trajectories by maximizing information under an independence constraint, implemented as the hypernetwork method DIF.","keywords":["invariant function learning","dynamical systems","ordinary differential equations","causal disentanglement","hypernetwork","environment independence","scientific equation discovery","multi-environment ODE datasets"],"falsifier":"Generate a multi-environment ODE dataset in which the environment-specific term is functionally correlated with the invariant term — for example, one environment adds a force proportional to $\\sin(\\theta_t)$ with an environment-varying coefficient. The theory predicts DIF still recovers exactly the shared invariant function; if the extracted function instead varies across environments or includes the spurious term, the independence constraint is not sufficient for identification.","tokens_in":57197,"feed_emoji":"🎯","tokens_out":9585,"duration_ms":87636,"temperature":0.7,"pith_summary":"The paper introduces invariant function learning: given trajectories of the same ODE system observed in environments that differ not just in coefficients but in whole function forms, recover the dynamics shared by all of them — for a pendulum, the frictionless law $-\\alpha^2\\sin\\theta_t$ from damped and powered observations. It argues this recovery is identifiable, not heuristic: in a structural causal model where an invariant function $f_c$ and an environment-specific function $f_e$ compose into the full derivative $f$, the invariant function is the one that maximizes mutual information with $f$ while being independent of the environment. The proposed method, Disentanglement of Invariant Functions (DIF), implements this with a transformer encoder, two disentangling MLPs, a decoder hypernetwork that outputs the predicted derivative function, and an adversarial discriminator that strips environment information from the invariant embedding. On three multi-environment ODE datasets (pendulum, Lotka–Volterra, SIR-epidemic) the recovered invariant functions beat meta-learning and invariant-learning baselines, and symbolic regression converts them into equations close to the true laws. The reason to care is concrete: if correct, it is a route from raw, environment-entangled observations to the underlying physical law.","feed_headline":"Neural net recovers the shared physics law from environment-biased data","feed_subtitle":"DIF extracts the frictionless pendulum equation from damped and powered trajectories without ever seeing the pure form.","key_machinery":"The load-bearing object is the invariant function random variable $f_c$, defined by a structural causal model in which $f_c$ and an environment-specific function $f_e$ compose into the observable full dynamics $f$; because $f$ is a collider, conditioning on the observed trajectory entangles the two components and blocks direct fitting. The load-bearing identity is Theorem 3.1, which says $f_c$ is the unique function maximizing $I(h_c(X_p); f \\mid X_0)$ under the constraint $h_c(X_p) \\perp e$, with Proposition 3.3 converting that information objective into an MSE loss on trajectories. The implementation machinery is a hypernetwork: a transformer encoder maps a trajectory to an embedding, two MLPs separate the invariant and environment parts, a decoder maps the invariant embedding to the parameters of a neural-network derivative function, and an adversarial discriminator enforces the independence constraint. A reference-based hypernetwork implementation that stores parameters by pointer instead of per-sample assignment gives a 16.8x speedup over a naive implementation.","core_discovery":"The paper's central claim is that the invariant function $f_c$ is identifiable from observed trajectories $X_p$ alone, without seeing the environment-free system. Theorem 3.1 states that the true $f_c$ is obtained by maximizing the conditional mutual information $I(h_c(X_p); f \\mid X_0)$ subject to $h_c(X_p) \\perp e$, and Proposition 3.3 shows this maximization is equivalent to minimizing the mean-square error between the predicted and true invariant trajectories. DIF realizes the principle by splitting the encoded function into an invariant embedding and an environment embedding, decoding the invariant part into a neural-network derivative function, and training adversarially so that a discriminator cannot predict the environment from the invariant embedding. In the reported experiments the extracted function predicts the environment-free trajectory with lower error than the baselines, and symbolic regression on the recovered function yields expressions such as $0.99\\,\\omega_t$ and $-0.97\\,\\alpha^2\\sin\\theta_t$ for the pendulum's natural motion.","pith_inferences":["If the identification theorem is right, 'maximum information under environment-independence' is a general definition of an invariant mechanism that could be transferred to partial differential equations or stochastic dynamics once multi-environment datasets for those domains exist.","The sharpest stress test would be environments whose extra terms are functionally correlated with the invariant term — for example, an environment force proportional to $\\sin\\theta_t$ with an environment-varying coefficient — which the current datasets may not include; a failure there would bound the identifiability regime.","The framework suggests a measurable signature of non-identifiability: if the invariant embedding's environment-conditional variance stays high, practitioners should distrust the recovered law before symbolic explanation.","The recovered invariant function could serve as an inductive bias for symbolic regression, potentially improving closed-form discovery in systems where direct sparse-regression methods fail because environments entangle the terms."],"forward_implications":["Scientific discovery becomes a single forward pass: a trained DIF network reads one biased trajectory and outputs the shared derivative function, with no per-system fitting or test-time adaptation.","The principle applies to environment changes of whole function form, not just coefficient shifts, so it targets systems where earlier invariant-learning and meta-learning methods were not designed to work.","The recovered invariant function can be handed to symbolic regression, turning a black-box neural predictor into closed-form equations such as $-\\alpha^2\\sin\\theta_t$.","A correct invariant function should make the environment unpredictable, giving a practical diagnostic: if the discriminator can still identify the environment from $\\hat{z}_c$, the candidate is spurious.","Because training fits derivatives rather than rolling out an integrator, the method scales to longer trajectories and larger state spaces at training time."],"supporting_citations":[{"why":"Defines invariant risk minimization, the principle the paper extends to function spaces and one of its baselines.","marker":"Arjovsky et al., 2019"},{"why":"Establishes invariant causal prediction, the theoretical basis for requiring invariance across environments.","marker":"Peters et al., 2016"},{"why":"Introduces hypernetworks, the architecture used to predict the neural-network derivative function.","marker":"Ha et al., 2016"},{"why":"CoDA, the coefficient-environment dynamics model adapted as a baseline and comparison for hypernetwork design.","marker":"Kirchmeyer et al., 2022"},{"why":"Supplies the DampedPendulum system that ME-Pendulum extends to function-form environments.","marker":"Yin et al., 2021b"},{"why":"Metaphysica, the source of the derivative-fitting training scheme and an OOD physics baseline.","marker":"Mouli et al., 2024"},{"why":"Provides PySR, used to turn extracted functions into closed-form symbolic equations for evaluation.","marker":"Cranmer, 2023"},{"why":"Supplies the adversarial training logic used to implement the environment-independence constraint.","marker":"Goodfellow et al., 2020"},{"why":"Provides the structural causal model and d-separation framework underpinning the causal graph and Theorem 3.1.","marker":"Pearl, 2009"},{"why":"MAML, the meta-learning baseline contrasted with invariant function discovery.","marker":"Finn et al., 2017"}],"fun_headline_variants":["DIF learns the true ODE from environment-tinted trajectories","Adversarial decoder separates invariant physics from environment drift","Causal hypernetwork extracts the frictionless law from biased data","Invariant function learning reveals shared ODE without pure observations","DIF disentangles cause from effect to recover the pendulum equation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The guarantee rests on the assumption that the practical training objective of DIF really enforces the theoretical condition — that the recovered function carries maximum information about the full dynamics while being independent of the environment — since the paper proves that equivalence only in a simplified single-loss case.","fun_headline_variants_meta":{"raw":{"variants":["DIF learns the true ODE from environment-tinted trajectories","Adversarial decoder separates invariant physics from environment drift","Causal hypernetwork extracts the frictionless law from biased data","Invariant function learning reveals shared ODE without pure observations","DIF disentangles cause from effect to recover the pendulum equation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000873,"raw_usage":{"total_tokens":3833,"prompt_tokens":1055,"completion_tokens":2778,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":671,"completion_tokens_details":{"reasoning_tokens":2695}},"tokens_in":671,"tokens_out":2778,"duration_ms":18535,"temperature":1.0,"reasoning_tokens":2695,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T22:32:18.506565+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate a multi-environment ODE dataset in which the environment-specific term is functionally correlated with the invariant term — for example, one environment adds a force proportional to $\\sin(\\theta_t)$ with an environment-varying coefficient. The theory predicts DIF still recovers exactly the shared invariant function; if the extracted function instead varies across environments or includes the spurious term, the independence constraint is not sufficient for identification.","supporting_citations":[],"review_version":1}