{"id":"d8becf04-5b58-423c-9ae0-a90f5f7bcec0","arxiv_id":"2501.06240","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Dynamic routing between capsules is exactly nonlinear gradient descent on the concave objective Ψ(C) = -Σ_j (||U_j C(:,j)|| - arctan ||U_j C(:,j)||), whose value decreases at every routing iteration.","lead":"Capsule networks use a 'routing' loop to decide how strongly lower-level features connect to higher-level objects, but nobody had shown what that loop is actually optimizing. This paper derives the exact mathematical objective the routing step minimizes, shows the routing is a form of nonlinear gradient descent, and proves the objective never increases across iterations.","discovery_kind":"first_principles","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The load-bearing gap is the final inference of Theorem 3.1: monotone decrease of Ψ plus boundedness does not imply convergence of C(r), and the stated convergence to a local optimum is false in a simple zero-gradient interior case.","rationale":"The core algebra checks out: with ψ(z) = z − arctan z, the gradient of ψ(||s||) is the Sabour squash, and the routing update is correctly cast as B(r+1) = B(r) − ∇Ψ(C(r)), C(r) = ∇Φ(B(r)). The monotonicity inequality in Theorem 3.1 follows from Lemma 2.4 if the required conjugacy and convexity conditions are stated precisely. Therefore the objective-identification contribution is genuine and is not the main problem. The weakest point is the final inference from Ψ-decrease to convergence of C(r), exactly as the reader identified. I add a concrete failure mode: a concave minimization problem can have interior stationary points that are maxima, and the routing dynamics can stall on such a maximum. The M = N = 2 example with U_j = [u_j, −u_j] and B(0) = 0 gives s_j = 0, v_j = 0, and B(r+1) = B(r), so the algorithm converges immediately to a global maximum of the objective, not a local minimum. This refutes the unqualified convergence-to-local-optimum statement, although the monotonicity theorem and the objective identification survive. A conditional verdict remains appropriate: the authors should either provide a correct proof with the conclusion weakened to convergence to a critical point, or remove the convergence-to-local-optimum claim. I do not see grounds to reject the paper's main constructive contribution.","tokens_in":10716,"tokens_out":16640,"duration_ms":184841,"concrete_test":"Run the exact routing iteration with M = N = 2, U_1 = [u, −u], U_2 = [v, −v], and B(0) = 0 for any nonzero vectors u and v. The iterates satisfy C(r) = 1/2 in every entry for all r, Ψ(C(r)) = 0, while the feasible set contains points with Ψ < 0; this deterministic check refutes the claim that the limit is a local minimum. To separate the fixable proof gap from the false conclusion, also perturb B(0) by a small ε and observe whether the trajectory leaves the maximum plateau and descends; that outcome supports the corrected statement 'converges to a critical point' rather than 'converges to a local optimum.'","verdict_should_be":"UNCHANGED","load_bearing_attack":"Theorem 3.1 proves only Ψ(C(r)) − Ψ(C(r+1)) ≥ ||C(r) − C(r+1)||²_F, which gives ∑_r ||C(r) − C(r+1)||²_F < ∞. On the compact simplex this implies successive differences tend to zero and subsequential limits exist, but it does not imply convergence of the full sequence. The sentence 'Based on the basic property of the convergence of discrete dynamical system' is not a theorem and is not supplied. The difficulty is not merely technical. The objective in (24) is concave, so interior stationary points are maxima or saddles, not minima. The routing fixed-point condition ∇Ψ(C*) = 0 can occur with s_j = 0. Concretely, take M = N = 2, U_1 = [u, −u], U_2 = [v, −v], and B(0) = 0. Then C(0) has all entries 1/2, every s_j = 0, every v_j = 0, and B(r+1) = B(r) for all r. The algorithm converges immediately to C(0), where Ψ = 0. Since Ψ(C) ≤ 0 for every feasible C, C(0) is a global maximum of the minimization problem, not a local minimum. Thus the conclusion 'converges to the local optimal solution' is false as stated. The boundary/one-hot regime emphasized in the experiments is also unhandled, but the interior maximum example already shows that the missing critical-point analysis changes the theorem's truth value.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the dynamic routing algorithm of capsule networks (Sabour et al., 2017). Its central claims are that the routing procedure minimizes the concave objective Ψ(C) = -Σ_j ψ(||U_j C(:,j)||) with ψ(z) = z - arctan(z) under the constraint that each row of C is a probability vector, that the routing updates constitute a nonlinear gradient method, and that the algorithm converges, with a rigorous proof given in Theorem 3.1. The paper also reports numerical experiments visualizing the monotone behavior of Ψ and the polarization of coupling coefficients.","tokens_in":10879,"tokens_out":3094,"duration_ms":30851,"significance":"If correct, the paper would give the first closed-form objective for dynamic routing, resolve the tension between routing's empirical success and the absence of a known objective, and explain polarization as a consequence of minimizing a concave function over the simplex. The identification of Ψ and the derivation of the monotonicity inequality in Section III.B are genuine contributions: the algebraic equivalence between Algorithm 1 and the matrix-form update is sound, and the inequality Ψ(C(r)) - Ψ(C(r+1)) ≥ ||C(r) - C(r+1)||²_F is a useful quantitative statement. However, the headline convergence claim is not established, and the stated conclusion that C(r) converges to a local optimum is false as written. The paper's main theorem therefore does not deliver what the abstract promises.","major_comments":[{"comment":"The proof's final inference is invalid. The paper shows Ψ(C(r)) - Ψ(C(r+1)) ≥ ||C(r) - C(r+1)||²_F and then asserts: 'Based on the basic property of the convergence of discrete dynamical system, we can get the convergence of C(r).' This is not a theorem and is not proved. Monotone decrease of a bounded scalar sequence implies convergence of Ψ(C(r)), and the square-summability of successive differences gives ||C(r+1) - C(r)|| → 0, but on a compact set this only yields subsequential convergence, not convergence of the full iterate sequence. The missing step is load-bearing because it is the bridge from monotonicity to the abstract's claim of a 'strictly proved' convergence.","section":"§III.C, Theorem 3.1"},{"comment":"The stated conclusion that C(r) converges to 'the local optimal solution' is false. Since Ψ is concave, interior stationary points are maxima or saddle points of the minimization problem, not minima. A concrete counterexample is M = N = 2, U_1 = [u, -u], U_2 = [v, -v], and B(0) = 0. Then C(0) has all entries 1/2, every s_j = 0, every v_j = 0, and B(r+1) = B(r) for all r. The algorithm is fixed at C(0), where Ψ = 0. Because ψ(z) ≥ 0 for all z, Ψ(C) ≤ 0 for every feasible C, so C(0) is a global maximum of the minimization objective, not a local minimum. The boundary or one-hot regime emphasized by the experiments is likewise not analyzed; the fixed-point argument used in the interior does not extend to it.","section":"§III.C, Theorem 3.1 / Theorem 3.2"},{"comment":"The convexity assertions in Corollary 3.2 are stated without a domain. From Corollary 2.2, φ*(y) - (1/2)y^T y is convex only on the probability simplex {y ≥ 0, Σ y_i = 1}, not on all of R^N. The proof of Theorem 3.1 applies Lemma 2.4 with F(C) = Φ*(C), and the omission of the domain leaves the application unjustified as written. The inequality can be recovered by restricting to the feasible set of C, but the manuscript must state this domain explicitly and verify that the updates remain in it.","section":"§III.C, Corollary 3.2 and Lemma 2.4"}],"minor_comments":[{"comment":"The text says 'the line above in blue color represents the value of C in each iteration,' but Equation (16), which the experiment reports, is the energy function Ψ(C); this should be clarified to avoid confusing the matrix C with the scalar objective.","section":"§IV.B, Figure 1"},{"comment":"The phrase 'strictly monotone decreasing' is stronger than what the inequality proves; the inequality Ψ(C(r)) - Ψ(C(r+1)) ≥ ||C(r) - C(r+1)||²_F gives non-strict monotonicity unless consecutive iterates are distinct.","section":"§III.C, Theorem 3.1"},{"comment":"The statement reads 'ψ′(z) > 0, ψ′(z) ≥ 0'; the second condition should be ψ''(z) ≥ 0, as used in the proof.","section":"§II.A, Theorem 2.1"},{"comment":"Equation (21) contains a typo: 'ϕ(B(i,ˆi))' should be 'φ(B(i,:))'.","section":"§III.C, Corollary 3.2 proof"}],"recommendation":"reject","confidential_remarks":"The core algebraic identity and the monotonicity inequality are correct and could be the basis of a useful paper, but the stated convergence theorem is false as written, and the counterexample is a fixed point of the algorithm in which the algorithm converges to a global maximum of the minimization objective. Since the paper's announced contribution is the rigorous proof of convergence, this is a load-bearing error that cannot be repaired by a local edit; a substantial reformulation (e.g., replacing convergence claims with monotonicity and subsequential-limit analysis under additional nondegeneracy assumptions) would be needed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The one thing to know: the core mathematical contribution here is real. The paper constructs Ψ(C) = -Σ_j (||U_j C(:,j)|| - arctan ||U_j C(:,j)||) and shows that the Sabour squash is exactly the gradient of ψ(z)=z-arctan z. I checked the rewriting of dynamic routing into B(r+1)=B(r)-∇Ψ(C(r)), C(r)=∇Φ(B(r)), and it works. The monotonicity inequality Ψ(C(r))-Ψ(C(r+1)) ≥ ||C(r)-C(r+1)||²_F is correctly derived from the convexity lemmas, and this is a real advance: for the first time dynamic routing is shown to have a closed-form objective, contradicting the belief attributed to [1].\n\nWhere it falls apart: the paper then asserts “based on the basic property of the convergence of discrete dynamical system” that C(r) converges to a local optimum. That is not a theorem, and the stress-test counterexample is decisive. For M=N=2, U_1=[u,-u], U_2=[v,-v], B(0)=0, every s_j=0, every v_j=0, the algorithm stays at C(0) with all entries 1/2. Ψ(C(0))=0, but Ψ(C) ≤ 0 everywhere on the feasible set, so C(0) is a global maximum of the minimization problem, not a local minimum. The conclusion “converges to the local optimal solution” is false as stated. The proof only establishes monotone decrease and boundedness of Ψ, not convergence of the iterate sequence. This is a load-bearing gap between the abstract’s “strictly proved convergence” and what is actually shown.\n\nOther soft spots, in proportion: Theorem 2.2 states a condition (2A1 - A2 PSD) that does not match the lemma it cites (A1 + A2 PSD); likely a typo, but it is in the theorem statement. The experiments are synthetic, unseeded, and the Figure 1 caption says C is strictly increasing while the text says the objective is decreasing—internally inconsistent. The boundary/one-hot regime is unhandled, though the interior counterexample already suffices to falsify the convergence claim.\n\nWhat I’d tell you: this paper deserves a serious referee because the objective identification is a genuine, verifiable contribution that will be useful to people working on capsule routing. But the convergence claims need major revision—trim the abstract, prove or drop the local-optimality statement, fix the theorem condition, and clean up the experiments. If the authors resubmit with the overclaims removed and the counterexample addressed, the monotonicity result and the objective are worth keeping.\n\nI’d bring it to a reading group, mostly to show the gap between monotone decrease and convergence. I wouldn’t cite it in its current form.\n\nRecommendation: send it to peer review, but only with the understanding that the convergence-to-local-optimum claim must be corrected—not just patched.","headline":"The objective identification for dynamic routing is real and checkable, but the advertised convergence to a local optimum is asserted, not proved, and is false in a simple interior case.","tokens_in":11621,"tokens_out":1965,"would_cite":false,"duration_ms":17494,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90C26","90C30","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"Dynamic routing between capsules has a hidden concave objective.","keywords":["capsule networks","dynamic routing","convergence analysis","concave objective function","nonlinear gradient descent","convex optimization","polarization"],"falsifier":"Run Algorithm 2 on a random fixed set of prediction vectors $\\hat{u}_{j|i}$ and record $C(r)$ and $\\Psi(C(r))$ for many iterations; if $C(r)$ ever enters a non-convergent cycle or drifts forever while $\\Psi(C(r))-\\Psi(C(r+1)) \\ge \\|C(r)-C(r+1)\\|_F^2$ holds in floating point, the convergence-to-a-local-optimum claim fails even though the descent inequality holds.","tokens_in":10250,"feed_emoji":"🧠","tokens_out":6479,"duration_ms":60160,"temperature":0.7,"pith_summary":"This paper tries to settle a question that has lingered around capsule networks: does the iterative routing procedure actually optimize something? It answers yes. The dynamic routing algorithm is cast as a nonlinear gradient method minimizing a specific concave energy over matrices whose rows are probability distributions, and the paper argues this yields a rigorous convergence proof. If the argument is right, the routing's well-known tendency to polarize link strengths is not a side effect but a mathematical necessity, and the practical problem becomes redesigning the objective's constraints rather than tuning iteration counts.","feed_headline":"Capsule routing is secretly minimizing a concave function","feed_subtitle":"Routing's notorious polarization is what you get when a concave function is minimized over a simplex.","key_machinery":"The engine of the paper is a descent inequality for a concave energy built from $\\psi(z)=z-\\arctan z$. The routing couplings $C$ are row-stochastic, and the energy $\\Psi(C)=-\\sum_{j=1}^N \\psi(\\|\\hat{U}_j C(:,j)\\|)$ is shown to be concave by composing this increasing convex $\\psi$ with the norm. The update is rewritten as a nonlinear gradient step on the logit matrix $B$: $B(r+1)=B(r)-\\nabla_C\\Psi(C(r))$, then $C(r+1)=\\nabla_B\\Phi(B(r+1))$, with $\\Phi(B)=\\sum_i \\phi(B(i,:))$ the sum of log-sum-exp functions. The proof uses a decomposition $\\Psi=F-G$ with $F=\\Phi^*$ and $G=\\Phi^* - \\Psi$; Lemma 2.4 converts convexity of $F-\\frac12\\|C\\|_F^2$ and $G-\\frac12\\|C\\|_F^2$ into $\\Psi(C(r))-\\Psi(C(r+1))\\ge\\|C(r)-C(r+1)\\|_F^2$. This inequality is what carries the monotonicity and the claimed convergence.","core_discovery":"The central claim is that the dynamic routing algorithm from the original capsule architecture is the realization of a nonlinear gradient method on a constrained optimization problem with an explicit closed-form objective. The objective is $E(C) = -\\sum_{j=1}^{N} \\psi(\\|\\hat{U}_j C(:,j)\\|)$ with $\\psi(z) = z - \\arctan z$, minimized over coupling matrices $C$ whose rows satisfy $C(i,:)e=1$ and $C(i,j)\\ge 0$. In matrix form the routing update becomes $B(r+1)=B(r)-\\nabla_C \\Psi(C(r))$ followed by $C(r+1)=\\nabla_B \\Phi(B(r+1))$, where $\\Phi$ is the sum of log-sum-exp row functions; Theorem 3.1 establishes the energy decrease $\\Psi(C(r))-\\Psi(C(r+1)) \\ge \\|C(r)-C(r+1)\\|_F^2$, so the energy is strictly monotone and bounded. The paper concludes that the iterates converge to a local optimum of this concave problem, giving dynamic routing the mathematical convergence proof it previously lacked.","pith_inferences":["An immediate corollary not developed in the paper: because a concave function over a polytope attains its minimum at an extreme point, the observed one-hot polarization of couplings is the expected outcome of running the routing to convergence; any fix must change the objective or constraints, not the iteration count.","If the missing convergence step cannot be supplied, the paper's theorem should be read as a monotonicity result rather than a full convergence proof, since a bounded monotone energy alone does not force the iterates themselves to converge.","A testable extension: add a small entropy regularizer to the objective or soften the softmax temperature, and measure whether the energy landscape's boundary minima become interior, reducing the polarization the experiments visualize."],"forward_implications":["Dynamic routing is not an unsupervised heuristic: it is minimizing $E(C)$ under row-stochastic constraints, so its behavior can be analyzed with existing optimization theory.","Each routing iteration strictly decreases the energy by at least $\\|C(r)-C(r+1)\\|_F^2$, so the algorithm cannot cycle while the descent inequality holds.","Polarization is intrinsic: minimizing a concave function over the simplex pushes couplings toward the boundary, which explains why extra iterations typically do not improve and can hurt classification.","The proof machinery generalizes to a broad class of capsule routing algorithms that share the same energy form, giving them a common convergence framework.","Changing the constraints can split the optimization into separable subproblems, potentially reducing routing's computational cost."],"supporting_citations":[{"why":"Supplies the dynamic routing algorithm whose objective and convergence are analyzed.","marker":"[1]"},{"why":"States the belief the paper challenges: that no sensible objective function underlies routing.","marker":"[2]"},{"why":"Documents the empirical behavior that motivates the convergence analysis.","marker":"[19]"},{"why":"Proposes an alternative optimization view of routing that the paper contrasts with its closed-form objective.","marker":"[22]"},{"why":"Reports the polarization of link strengths that the paper explains through concave minimization.","marker":"[23]"},{"why":"Provides the convex-analysis facts, including conjugate functions and log-sum-exp properties, used in the proof.","marker":"[26]"},{"why":"Gives the nonlinear gradient descent framework and convergence conditions the paper applies.","marker":"[27]"}],"fun_headline_variants":["Capsule routing: gradient descent on a concave objective","Dynamic routing converges—here's the proof","Routing's polarization explained by concave minimization","Capsule routing proven to be a nonlinear gradient method","The concave function behind capsule routing"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a bounded, monotonically decreasing energy $\\Psi$ together with the routing update forces the coupling matrices themselves to converge to a local optimum; the paper states this as a consequence of a basic property of discrete dynamical systems but does not prove the limit exists at the boundary where couplings become one-hot.","fun_headline_variants_meta":{"raw":{"variants":["Capsule routing: gradient descent on a concave objective","Dynamic routing converges—here's the proof","Routing's polarization explained by concave minimization","Capsule routing proven to be a nonlinear gradient method","The concave function behind capsule routing"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000252,"raw_usage":{"total_tokens":1584,"prompt_tokens":991,"completion_tokens":593,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":607,"completion_tokens_details":{"reasoning_tokens":524}},"tokens_in":607,"tokens_out":593,"duration_ms":6215,"temperature":1.0,"reasoning_tokens":524,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:34:39.093437+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run Algorithm 2 on a random fixed set of prediction vectors $\\hat{u}_{j|i}$ and record $C(r)$ and $\\Psi(C(r))$ for many iterations; if $C(r)$ ever enters a non-convergent cycle or drifts forever while $\\Psi(C(r))-\\Psi(C(r+1)) \\ge \\|C(r)-C(r+1)\\|_F^2$ holds in floating point, the convergence-to-a-local-optimum claim fails even though the descent inequality holds.","supporting_citations":[{"cited_title":"Matrix capsules with em routing,","cited_arxiv_id":null,"evidence_quote":"States the belief the paper challenges: that no sensible objective function underlies routing."},{"cited_title":"An optimization view on dynamic routing between capsules,","cited_arxiv_id":null,"evidence_quote":"Proposes an alternative optimization view of routing that the paper contrasts with its closed-form objective."},{"cited_title":"Capsule networks need an improved routing algorithm,","cited_arxiv_id":null,"evidence_quote":"Reports the polarization of link strengths that the paper explains through concave minimization."},{"cited_title":"Improved concave-convex procedure and its application to analysis for the stability of hopfield neural network,","cited_arxiv_id":null,"evidence_quote":"Gives the nonlinear gradient descent framework and convergence conditions the paper applies."}],"review_version":1}