{"id":"a259f550-b4b6-4b6c-ad30-0bd3bd0aab00","arxiv_id":"2501.15646","paper_version":1,"verdict":"ACCEPT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"For deep feedforward networks with piecewise-smooth activations, the autodiff gradient is shown to be the unique limit of gradients of smoothed activations, a limiting Frechet subgradient, and equal to the true gradient wherever the loss is continuously differentiable.","lead":"This paper gives a precise mathematical definition of the 'gradient' that deep learning libraries actually compute for networks with non-differentiable activations like ReLU: it is the unique limit of gradients of smoothed versions of the network. The result allows rigorous statements about when this generalized gradient equals the true gradient of the loss, and why it behaves like a limiting subgradient.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Lemma 3.10 invokes Lemma 3.8 with the wrong sign, so the approximating sequence approaches singular points from the side where ψ is discontinuous; the proof of Theorem 3.14(iii) is invalid as written.","rationale":"The paper's central contribution divides cleanly into two parts. Theorem 2.15 (uniqueness of G as the pointwise limit of smoothed gradients) is established through dominated convergence and explicit integral representations; I see no load-bearing problem there. The novelty and the integral-representation arguments are credible. The fragile part is the regularity analysis of Section 3, specifically the claim that G(θ) is a limiting Fréchet subgradient (Theorem 3.14(iii)) and hence agrees with the gradient on C^1 open sets (Theorem 3.14(iv)). The reader correctly identifies the one-sided continuity condition as the weakest assumption. My stress-test narrows this further: the proof of Lemma 3.10 uses the sign z from that condition in Lemma 3.8, but Lemma 3.8's inequality points the approximating pre-activations in the opposite direction. For the standard ReLU/left-derivative example, the constructed sequence approaches from the discontinuous side and the claimed convergence of G fails. The reader's ACCEPT is therefore premature as far as the written proof is concerned. The failure is localized and appears repairable by flipping the sign when applying Lemma 3.8; if that repair works, the theorem and the ACCEPT verdict can stand. I recommend CONDITIONAL: require the authors to correct Lemma 3.10 and verify the argument, or provide an alternative proof of Theorem 3.14(iii).","tokens_in":40424,"tokens_out":15400,"duration_ms":130760,"concrete_test":"Re-derive Lemma 3.10 for the minimal case A0 = ReLU, ψ = 1_{(0,∞)} (left derivative), S = {0}, and L(θ) = A0(θ). Follow the proof's instructions: choose z = −1 from the continuity hypothesis and take ϑ_n from U_n ∩ E as constructed via Lemma 3.8. Verify explicitly that the construction forces ϑ_n > 0 for all large n (so N_{k,ϑ_n} ≥ N_{k,θ}), giving G(ϑ_n) = 1 while G(0) = 0, contradicting the lemma. Then replace z by −z in the application of Lemma 3.8 and check that the resulting sequence has ϑ_n < 0, G(ϑ_n) = 0, and all inequalities (155)-(159) go through. This settles whether the current proof has a sign error or a substantive gap.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The most load-bearing defect is in the proof of Lemma 3.10 (Section 3.6), which is the bridge to Proposition 3.12 and hence to Theorem 3.14(iii)-(iv). The one-sided continuity hypothesis fixes a sign z ∈ {−1,1} such that for every x ∈ S, limsup_{h↓0} |ψ(x+zh) − ψ(x)| = 0; for ReLU with the left derivative ψ = 1_{(0,∞)}, the good sign is z = −1. In the proof, Lemma 3.8 is invoked with this same z, yielding open sets U_n on which z(N_{k,ϑ}(x) − N_{k,θ}(x)) ≤ 0. For z = −1 this means N_{k,ϑ}(x) ≥ N_{k,θ}(x), so as ϑ → θ the pre-activations approach each singular point from above. But ψ is left-continuous at 0, not right-continuous, so ψ(N_{k,ϑ_n}(x)) does not converge to ψ(N_{k,θ}(x)); the claimed limsup in (159) is false. In the scalar network L(θ) = A0(θ) with A0 = ReLU and ψ(0)=0, the proof's construction forces ϑ_n > 0, giving G(ϑ_n) = 1 while G(0)=0, so the constructed sequence does not satisfy (153). The fix is to apply Lemma 3.8 with the opposite sign, −z, so the pre-activations approach from the side on which ψ is continuous. This is a proof bug, not a disproof of the theorem, but it must be corrected for the central subgradient claim to be established.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies supervised-learning risk functionals for deep fully-connected feedforward networks with a nonsmooth activation A0 that is C1 away from a finite set S. It fixes a generalized derivative ψ of A0 and considers sequences of C1 activations A_n converging pointwise (eventually exactly) to A0 with derivatives converging to ψ. Under local boundedness and a one-sided continuity condition on ψ, it proves (Theorem 2.15 and Theorem 3.14) that there is a unique G:R^d→R^d to which the gradients ∇L_{A_n} converge pointwise regardless of the approximating sequence; that G(θ) is a limiting Fréchet subgradient of the nonsmooth risk L0 at θ; and that G coincides with ∇L0 on every open set where L0 is C1. The proof uses explicit integral representations for gradient components (Proposition 2.14), dominated convergence, Rademacher's theorem, and a limiting-subgradient density argument.","tokens_in":40714,"tokens_out":9332,"duration_ms":83198,"significance":"If correct, the result is a clean and fairly general mathematical description of the object that automatic differentiation libraries compute for ReLU-type networks. The main conceptual value is the uniqueness statement: unlike set-valued conservative fields, the smoothed-gradient limit is a single-valued function independent of the smoothing mechanism. The paper also extends the ReLU/MSE-specific analysis of prior work to general C1 loss functions and finite-singularity activations with a chosen generalized derivative. The explicit representation formulas and the use of standard analytic tools are strengths. However, the proof of the subgradient property currently contains a localized sign bug that must be corrected before the central claim is established.","major_comments":[{"comment":"The sign used in the application of Lemma 3.8 is inconsistent with the one-sided continuity assumption. The hypothesis before Theorem 3.14 supplies z∈{−1,1} such that ψ is continuous from side z (for ReLU with ψ=1_{(0,∞)}, the good sign is z=−1). Applying Lemma 3.8 with the same z yields z(N_{k,ϑ}(x)−N_{k,θ}(x))≤0; for z=−1 this is N_{k,ϑ}(x)≥N_{k,θ}(x), so the approximating pre-activations approach any singular value from above, i.e. from the side on which ψ is discontinuous. Consequently the asserted limsup in (159) need not hold; in the scalar construction with A0=ReLU and θ=0, the sequence chosen from the U_n produced by Lemma 3.8 with z=−1 has positive pre-activations, so ψ equals 1 there while ψ(0)=0 and (153) fails for the constructed sequence. The fix is to apply Lemma 3.8 with −z, so that z(N_{k,ϑ}(x)−N_{k,θ}(x))≥0 and the pre-activations approach from the side where ψ is continuous. This repairs the proof of Lemma 3.10 and hence of Proposition 3.12 and Theorem 3.14(iii)–(iv). Because this is a proof bug in a load-bearing lemma rather than a counterexample to the theorem, it is correctable, but as written the central subgradient claim is not established.","section":"§3.6, Lemma 3.10, Eq. (155)–(159)"}],"minor_comments":[{"comment":"In item (iii), the statement writes (∇L0)(x)=G(θ) but θ is not quantified in that item; it should be G(x). The proof also contains a typo in Eq. (135), where the integration domain appears as Rθ and the two functions G_k and G_k are conflated.","section":"§3.4, Lemma 3.7"},{"comment":"In Eq. (14), the definition of G_n(v) uses the variable x in the term h(y_{i_v})(x−y_{i_v}), but this should be v; as written the formula is inconsistent with the surrounding definitions.","section":"§2.2, Proposition 2.3"},{"comment":"In the sentence before Eq. (154), the proof states that there exists z∈R satisfying the displayed limit for all x∈R; the correct assertion is that z can be chosen in {−1,1} and that the limit statement holds for all x, with the singular points x∈S covered by the hypothesis and the nonsingular points covered by continuity of ψ away from S.","section":"§3.6, Lemma 3.10"},{"comment":"The boundedness condition in items (ii) contains a malformed expression involving an indicator function of {∞} applied to supp(μ); for example, Theorem 3.14(ii) writes (z+1_{∞}(supp(μ))) without defining this notation. This should be cleaned up for readability.","section":"Theorem 2.15 and Theorem 3.14"},{"comment":"The proof relies on [30, Lemma 3.8] and its consequence for C1 domains without restating them; since these results are central to the argument, the authors should either state the exact versions used or give a short proof, so that the paper is more self-contained.","section":"§3.7, Proposition 3.12 and Corollary 3.13"}],"recommendation":"major_revision","confidential_remarks":"The sign error in Lemma 3.10 is localized and almost certainly fixable by applying Lemma 3.8 with the opposite sign. Assuming the authors make that correction and verify the surrounding argument, I would be comfortable with acceptance. I do not see grounds for rejection: the theorem statements appear sound, the representation formulas are explicit, and the remaining issues are presentation and proof-detail matters."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The main new result here is the uniqueness theorem: for any continuous activation that is C^1 off a finite set, and any admissible smoothing sequence, the gradients of the smoothed loss converge to a single function G, and G agrees with the ordinary gradient wherever the loss is C^1. That is a genuine extension of the earlier ReLU-only results from [30], and the integral representations in Proposition 2.14 are explicit and useful. The approximation lemmas in Section 2 look solid, and the paper is carefully written.\n\nThe problem is in Section 3. The proof of Proposition 3.12, which gives the limiting Fréchet subgradient property, relies on Lemma 3.10. There, the one-sided continuity assumption fixes a sign z ∈ {−1,1} such that ψ is continuous at each singular point from the side z. But the proof invokes Lemma 3.8 with that same z, which forces the approximating parameters ϑ_n to approach the singular pre-activations from the opposite side. For ReLU with left-continuous derivative, this means the pre-activations approach 0 from above, where ψ is discontinuous, and the claimed convergence in (159) is false. The fix is straightforward: apply Lemma 3.8 with −z instead, so the sequence approaches from the continuous side. I did not find a way to make the current argument go through, and this affects the central subgradient theorem and the C^1-agreement conclusion.\n\nThe one-sided continuity assumption itself is a genuine restriction, and the paper states it clearly. There are also minor typos in Lemma 3.7 and Lemma 3.8, but those are not serious.\n\nSo: the uniqueness part is a real contribution and deserves to be published. The subgradient part is not proven as written. I would send this to a serious referee, but the authors need to repair Lemma 3.10 before acceptance. If the sign fix works, the paper should be acceptable.","headline":"Generalizes the ReLU smoothing result to arbitrary continuous activations, but the proof of the subgradient claim has a sign error in Lemma 3.10 that must be fixed.","tokens_in":41320,"tokens_out":3881,"would_cite":false,"duration_ms":34761,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T07","49J52"],"pacs":[],"model":"deepseek-v4-flash","headline":"There is a unique generalized gradient for training ReLU networks.","keywords":["generalized gradients","ReLU activation","nonsmooth analysis","limiting Frechet subgradient","backpropagation","automatic differentiation","stochastic gradient descent","deep neural networks"],"falsifier":"Implement two different $C^1$ smoothings of ReLU whose derivatives converge pointwise to the left derivative and that are uniformly bounded on compact sets, then evaluate the smoothed-loss gradients at a parameter vector where some pre-activation equals zero. The theorem predicts that the two gradient vectors converge to the same limit; observing a nonzero difference between the limits would disprove the uniqueness claim.","tokens_in":40156,"feed_emoji":"🧠","tokens_out":8388,"duration_ms":66726,"temperature":0.7,"pith_summary":"Deep learning libraries routinely compute gradients through non-differentiable activations such as ReLU by using a chosen one-sided derivative. This paper asks whether that choice is a harmless implementation detail or a well-defined mathematical object. It establishes that, for fully-connected feedforward networks, every reasonable smoothing of the activation produces gradients that converge pointwise to the same function $G: \\mathbb{R}^d \\to \\mathbb{R}^d$, so the generalized gradient is unique. It then proves that $G(\\theta)$ is a limiting Fréchet subgradient of the nonsmooth loss at every parameter vector $\\theta$, and that on every open set where the loss is continuously differentiable, $G$ coincides with the ordinary gradient. The result gives a precise target for theoretical analyses of stochastic gradient descent.","feed_headline":"One gradient vector governs ReLU network training","feed_subtitle":"Smoothing ReLU any way you like, backprop converges to one vector—the true gradient where the loss is smooth.","key_machinery":"The load-bearing mechanism is the admissible smoothing sequence: a sequence of $C^1$ activations $A_n$ that eventually agrees pointwise with the original activation and with its generalized derivative, while staying uniformly bounded on compact intervals. Representation formulas in Proposition 2.14 write each coordinate of $G$ as an explicit integral over paths through the network, with products of weights and values of the generalized derivative at pre-activation values. Because every admissible sequence eventually evaluates the same quantities at the same finitely many pre-activation values, all smoothed-loss gradients share the same pointwise limit. A second mechanism is the one-sided continuity condition on the generalized derivative at the kink set $S$, which lets the proof approximate an arbitrary parameter by nearby differentiability points and pass $G(\\theta)$ through as a limiting Fréchet subgradient.","core_discovery":"The central claim is that the gradient returned by backpropagation through a ReLU-like network is not implementation-dependent. For any continuous activation that is $C^1$ off a finite set $S$ and any generalized derivative $\\psi$ that agrees with the true derivative off $S$, the paper considers sequences of $C^1$ activations $A_n$ that converge pointwise to the original activation while their derivatives converge pointwise to $\\psi$, with uniform boundedness on compact intervals. Theorems 2.15 and 3.14 show that the gradients of the smoothed losses converge pointwise to a single function $G$ that depends only on the network, the loss, the data measure, and the choice of $\\psi$, not on the particular smoothing sequence. Moreover, $G(\\theta)$ is a limiting Fréchet subgradient of the nonsmooth loss for every $\\theta$, and $G$ agrees with the classical gradient of the loss on every open set where that loss is $C^1$. For the ReLU activation with the left derivative, this is an exact description of the vector that automatic differentiation in standard deep-learning libraries produces.","pith_inferences":["A natural next question the paper leaves open is whether uniqueness survives for activations whose generalized derivative oscillates on both sides of a kink; two smoothings with different one-sided limits there would likely produce different generalized gradients.","The single-valued uniqueness result suggests that stochastic gradient descent analyses could treat this $G$ as the canonical descent field, and it would be worth checking whether convergence guarantees known for conservative set-valued fields transfer to this smaller object.","One could numerically probe the theorem: at parameters where a hidden pre-activation lands exactly on a kink, compare the gradients produced by two different smoothing widths; the theorem predicts identical limiting vectors."],"forward_implications":["Backpropagation through a ReLU kink is deterministic: every smoothing family that respects the generalized derivative converges to the same parameter vector field, so convergence analyses can be written against a single function rather than a set-valued subgradient.","The generalized gradient is always a limiting Fréchet subgradient, giving a nonsmooth variational certificate at every point of parameter space.","On any open region where the loss is continuously differentiable, the generalized gradient equals the ordinary gradient, so smoothing-based training and gradient-flow dynamics coincide on smooth parts of the landscape.","The theorems extend prior ReLU-specific analyses to arbitrary piecewise-smooth activations with finitely many kinks, general $C^1$ loss functions, and compactly supported data measures."],"supporting_citations":[{"why":"supplies the explicit smoothing-sequence construction for ReLU networks that this paper generalizes to arbitrary activations.","marker":"[10]"},{"why":"uses pointwise convergent smoothings for ReLU stochastic gradient descent, a specialized precursor of the approximation scheme.","marker":"[26]"},{"why":"establishes the ReLU, mean-squared-error special case of the limiting-Frechet-subgradient and C^1-agreement conclusions.","marker":"[30]"},{"why":"provides a convergence analysis for ReLU smoothing that motivates describing the limiting gradient.","marker":"[31]"},{"why":"introduces conservative set-valued fields as the standard model of automatic differentiation outputs, the set-valued framework the uniqueness result refines.","marker":"[7]"},{"why":"proves that limits of conservative fields of definable functions remain conservative, a closely related uniqueness-type statement.","marker":"[49]"},{"why":"supplies the variational-analysis definition of Frechet and limiting subgradients used in Theorem 3.14.","marker":"[46]"}],"fun_headline_variants":["Backprop gradient for ReLU nets is unique","Smoothing ReLU always yields same gradient","One true gradient emerges in ReLU training","ReLU backprop gradient: path-independent"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument rests on the generalized derivative approaching its value at each non-differentiability point from at least one side; if it oscillates on both sides of a kink, the proof that the generalized gradient is a limiting Fréchet subgradient, and with it the agreement with the true gradient on smooth regions, no longer goes through.","fun_headline_variants_meta":{"raw":{"variants":["Backprop gradient for ReLU nets is unique","Smoothing ReLU always yields same gradient","One true gradient emerges in ReLU training","ReLU backprop gradient: path-independent"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000436,"raw_usage":{"total_tokens":2271,"prompt_tokens":1051,"completion_tokens":1220,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":667,"completion_tokens_details":{"reasoning_tokens":1163}},"tokens_in":667,"tokens_out":1220,"duration_ms":8241,"temperature":1.0,"reasoning_tokens":1163,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T14:04:52.476208+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Implement two different $C^1$ smoothings of ReLU whose derivatives converge pointwise to the left derivative and that are uniformly bounded on compact sets, then evaluate the smoothed-loss gradients at a parameter vector where some pre-activation equals zero. The theorem predicts that the two gradient vectors converge to the same limit; observing a nonzero difference between the limits would disprove the uniqueness claim.","supporting_citations":[{"cited_title":"A proof of conver- gence for gradient descent in the training of artificial neur al networks for constant target functions","cited_arxiv_id":null,"evidence_quote":"supplies the explicit smoothing-sequence construction for ReLU networks that this paper generalizes to arbitrary activations."},{"cited_title":"Convergence proof for stochastic gradient descent in the training of deep neural networks with ReLU activation for constant target functions","cited_arxiv_id":"2112.07369","evidence_quote":"uses pointwise convergent smoothings for ReLU stochastic gradient descent, a specialized precursor of the approximation scheme."},{"cited_title":"On the existence of global minima and convergence analyses for gradient descent methods in the training of dee p neural networks","cited_arxiv_id":null,"evidence_quote":"establishes the ReLU, mean-squared-error special case of the limiting-Frechet-subgradient and C^1-agreement conclusions."},{"cited_title":"A proof of convergence for stochastic gradient descent in the training of artificial neural networks with ReLU activation for constant target functions","cited_arxiv_id":null,"evidence_quote":"provides a convergence analysis for ReLU smoothing that motivates describing the limiting gradient."},{"cited_title":"Conservative set valued fields, automatic differentiation, stochastic gradient methods and deep learning","cited_arxiv_id":null,"evidence_quote":"introduces conservative set-valued fields as the standard model of automatic differentiation outputs, the set-valued framework the uniqueness result refines."},{"cited_title":"T., and Wets, R","cited_arxiv_id":null,"evidence_quote":"supplies the variational-analysis definition of Frechet and limiting subgradients used in Theorem 3.14."}],"review_version":1}