{"id":"b174d3e9-6f95-4cc9-8e1a-5141f2e8ebf2","arxiv_id":"2601.10950","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A specular derivative built from mirror-weighted one-sided slopes yields weak MVT/Fermat theorems and a subgradient that powers new nonsmooth optimization methods.","lead":"This paper defines a new kind of derivative — the 'specular derivative' — that averages one-sided slopes in a curved way, and extends it to normed vector spaces. The authors prove weak versions of the mean value and Fermat theorems and use the derivative to build three optimization algorithms for nonsmooth convex problems.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The stochastic S-SPEG convergence theorem depends on unbiasedness condition (3.13), which is false: specular differentiation is nonlinear, so E[∇s f_ξ] ≠ ∇s f for natural decompositions.","rationale":"The paper's deterministic core — specular derivatives coinciding with classical derivatives, and the subgradient inclusion for convex functions — is plausible and not undermined by this objection. The weakest point is the stochastic extension: Theorem 3.12's convergence proof depends on the unbiasedness condition (3.13), and that condition is genuinely false for the natural component-decomposition formulation in (3.11). The explicit counterexample is not pathological; it uses convex, specularly differentiable components and a bounded-below objective. This invalidates the advertised S-SPEG and H-SPEG convergence guarantees and the stochastic experiments, while leaving Theorem 3.10 intact. The reader already identified this as the weakest assumption and issued a conditional verdict; my stress-test confirms the same concern, so no verdict change is needed.","tokens_in":32116,"tokens_out":12265,"duration_ms":116444,"concrete_test":"Closed-form evaluation of (3.13) for f_1(x)=|x|+x^2, f_2(x)=2x+x^2, f=(f_1+f_2)/2 at x=0. Using formula (2.5) — or the released specular-differentiation package — compute (∇s f_1(0)+∇s f_2(0))/2 = 1 and ∇s f(0)=tan(½ arctan(3/2)+½ arctan(1/2))≈0.884. If they differ, as this calculation shows, the unbiasedness premise of Theorem 3.12 is false. A further check: run S-SPEG on this f with square-summable step sizes from x_0=0.1 and compare empirical E[f(x^⋄_k)] to the bound claimed in Theorem 3.12; the bound should fail because the conditional expectation term used in the proof is not a subgradient of f.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing assumption in the stochastic part is Definition 3.7 / eq. (3.13): E[∇s f_ξk(x_k)|x_k] = ∇s f(x_k), used verbatim in Theorem 3.12 to reduce the stochastic recurrence to the deterministic subgradient inequality. For an objective written as in (3.11), f = (1/m) Σ f_j, uniform sampling gives E[∇s f_ξ] = (1/m) Σ ∇s f_j. But the specular derivative is not additive: D_s is defined through a weighted average of one-sided slopes followed by a nonlinear tangent/arctangent expression (Theorem 2.5), so ∇s((1/m)Σ f_j) ≠ (1/m)Σ ∇s f_j in general. A concrete failure at x=0 with m=2, f_1(x)=|x|+x^2, f_2(x)=2x+x^2 is: ∇s f_1(0)=0 and ∇s f_2(0)=2, so E=1, while ∇s f(0)=tan(½ arctan 1.5 + ½ arctan 0.5)≈0.884. Thus (3.13) is not a harmless moment condition; it fails at a kink of a bounded-below convex function. Since Theorem 3.12's proof uses (3.13) to replace E[∇s f_ξ(x_ℓ)|x_ℓ] by ∇s f(x_ℓ), the claimed convergence in expectation and in probability for S-SPEG, and for H-SPEG after its switch, currently has no valid proof. The deterministic SPEG convergence (Theorem 3.10) and the subgradient inclusion (Theorem 3.2) are not affected.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces \"specular differentiation\" in normed vector spaces, a nonlinear weighted-average derivative that generalizes Gâteaux and Fréchet differentiability when the latter exist. It proves several structural results: a formula for specular directional derivatives in terms of one-sided derivatives, a Quasi-Mean Value Theorem and Quasi-Fermat Theorem, inclusion of the specular Fréchet differential in the convex subdifferential, and convergence of deterministic and stochastic specular-gradient methods for nonsmooth convex optimization in R^n. The numerical section compares SPEG, S-SPEG, and H-SPEG against GD, Adam, and BFGS on an elastic-net objective, reporting that the proposed methods succeed where classical methods sometimes fail.","tokens_in":32706,"tokens_out":26529,"duration_ms":268493,"significance":"If sound, the deterministic part of the paper would be a useful contribution: it gives an explicit, computable construction of a subgradient for convex functions, a clean arctangent representation, and a public Python package. The convergence proof for the deterministic specular-gradient method is a straightforward but valid adaptation of subgradient-method arguments. However, several load-bearing claims are not correct as stated: the characterization theorem for specular directional derivatives is false in general, the Quasi-Mean Value Theorem is vacuous for codomains of dimension greater than one, and the stochastic convergence theorem does not apply to the implemented component-sampling algorithm. These issues are substantial and require more than cosmetic revision.","major_comments":[{"comment":"The converse direction of Theorem 2.5 is false. Take f(0)=0 and f(x)=|x|(1+sin(1/|x|)) for x≠0. The two weighted terms in (1.2) cancel for every h, so ∂∧_1 f(0)=0 and in fact D_s f(0)=0 (specular Fréchet differentiability holds), but the one-sided derivatives do not exist. Thus (H1) is not necessary for ∂∧_v f(x) to exist. Theorem 2.7, which assumes only existence of ∂∧_v f(x) and then uses ∂±_v f(x), is therefore false as stated. The proof uses a bound |...-c|<h that is not implied by the limit. The forward implication (H1)+(H2)⇒(2.5) appears correct; please restate the theorem as a sufficient condition or add hypotheses that guarantee one-sided differentiability.","section":"§2.1, Theorem 2.5/2.7"},{"comment":"Assumption (2.20) forces Y to be one-dimensional: if φ(y)=0 then ||y||=|φ(y)|=0, so φ is injective from Y to R. Hence the Quasi-Mean Value Theorem does not generalize to vector-valued maps in any nontrivial Banach codomain; it is essentially the one-dimensional result from [5] applied to g=⟨φ,f⟩. The proof also assumes ⟨φ,f(v)-f(u)⟩ = ||f(v)-f(u)||, but a φ satisfying (2.20) may give the opposite sign. Please state the theorem for Y=R (or one-dimensional Y) and explicitly choose the sign of φ.","section":"§2.2, Theorem 2.12"},{"comment":"Condition (3.13) is an unverified and generally false premise for the implemented algorithm. For f=(1/m)Σ f_j and uniform sampling, E[∇s f_ξ(x)]=(1/m)Σ ∇s f_j(x), but the specular derivative is nonlinear in the component slopes (Theorem 2.5), so this average need not equal ∇s f(x). Concrete failure at x=0: f1=|x|+x², f2=2x+x². Then ∇s f1(0)=0, ∇s f2(0)=2, so E=1, while ∇s f(0)=tan(½ arctan(3/2)+½ arctan(1/2))≈0.884. Thus the replacement of E[∇s f_ξ(x_ℓ)|x_ℓ] by ∇s f(x_ℓ) in the proof of Theorem 3.12 is unjustified for the S-SPEG/H-SPEG experiments in §3.2. Additionally, part (a) is false: a zero component specular gradient does not imply a minimizer of f; e.g., f1=(x)_+², f2=(x-1)² at x=0.","section":"§3.2, Definition 3.7/Theorem 3.12"},{"comment":"The assertion that equality in ∇s f(x)·(x−x*)≥0 holds if and only if x=x* is false. For f(x)=max(x,0)^2 on Ω=R and x*=0, every x<0 satisfies ∇s f(x)=0, so the dot product is zero even though x≠x*. The proof only shows that equality implies f(x)=f(x*), not x=x*. The statement should be weakened accordingly.","section":"§3.1, Corollary 3.3(a)"}],"minor_comments":[{"comment":"The break condition appears reversed: the pseudocode stops when ∥∇s f(x)∥ > η, but it should stop when the gradient norm is below tolerance (≤ η). As printed, the algorithms would terminate immediately at most starting points.","section":"Algorithm 3.1 and Algorithm 3.2"},{"comment":"The notation x⋄_k is defined in (3.8) for the deterministic SPEG method only, but Theorem 3.12 uses x⋄_k for S-SPEG without restating the definition. Define it for the stochastic iterates.","section":"§3.2"},{"comment":"The statement that the one-sided limit (h↘0) in (2.4) exists if and only if the limit (h→0) exists is ambiguous, since the definition only uses h>0. Please clarify the intended two-sided statement or remove the remark.","section":"§2.1, Remark 2.4"}],"recommendation":"major_revision","confidential_remarks":"The paper draws heavily on the author's own preprints ([5], [7], [8]); the genuinely new component is the finite-dimensional convex application. The deterministic subgradient-inclusion result is likely correct, but the stochastic section as written is not supported, and the Quasi-MVT overclaims its codomain generality. A careful revision should narrow the claims and repair or remove the stochastic component-sampling theory."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, here's my read of arXiv:2601.10950. The genuinely new and useful core is small but real: for convex functions on R^n, if f is specularly Fréchet differentiable at x then the specular gradient ∇s f(x) is a subgradient (Thm 3.2), and the same idea works in Banach spaces via the Fréchet subdifferential (Thm 2.19). The chain of reasoning from the weighted-slope definition through the A-function representation to the subgradient inclusion is coherent; I don't see a hole in that part. The paper also ships code and is candid that the pure specular method is slower in wall-clock terms than classical methods, which is more honest than most optimization papers.\n\nThe soft spots are real. The stochastic section does not hold together. Definition 3.7/eq. (3.13) assumes E[∇s f_ξ(x_k)|x_k] = ∇s f(x_k), but specular differentiation is not additive, so for f = (1/m)Σf_j the conditional expectation is (1/m)Σ∇s f_j, not ∇s f. The paper gives no proof and the condition is false for simple component sums; the stress-test example with f1=|x|, f2=2x at 0 works. Theorem 3.12's convergence in expectation and probability therefore has no valid proof. The deterministic SPEG convergence (Thm 3.10) is fine, but the advertised S-SPEG and H-SPEG results are not backed by the theory.\n\nAlso, Corollary 3.3(a) claims equality ⟺ x=x*; for a convex function constant on an interval, every interior point has zero specular gradient, so equality holds for distinct minimizers. The equality needs strict convexity or an extra argument. And Theorem 2.12's assumption (2.20) essentially forces Y to be one-dimensional, so the normed-space Quasi-MVT reduces to the R case; the generality is decorative rather than substantive.\n\nThe numerical section never compares against the plain subgradient method, which is the natural baseline since SPEG is just a subgradient selection. Showing that GD and Adam fail on a highly nonsmooth elastic-net instance is not the same as showing an advantage over a standard nonsmooth solver.\n\nBottom line: the subgradient-inclusion result and its deterministic method deserve to be known, but the stochastic claims in their current form are not reliable. A referee should ask for a corrected stochastic analysis or a weakened statement, a fix to Cor 3.3(a), and a comparison against SG. That is a substantial revision, not a desk reject.","headline":"Specular gradient is a genuine subgradient for convex functions, but the stochastic convergence theory rests on a false unbiasedness condition and the numerical comparison omits the obvious baseline.","tokens_in":33102,"tokens_out":3534,"would_cite":false,"duration_ms":37088,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["46G05","46T20","49J52","65K05","90C25","90C15"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper introduces specular differentiation, a mirror-inspired generalization of Gâteaux and Fréchet derivatives in normed vector spaces, and shows that for convex functions its gradient is a subgradient, yielding new optimization method","keywords":["specular differentiation","generalized derivative","Quasi-Mean Value Theorem","Quasi-Fermat Theorem","nonsmooth convex optimization","subgradient method","stochastic subgradient method","subdifferential"],"falsifier":"Compute, at x = 0, the specular gradient of f(x) = (|x| + 2x)/2 directly from the half-angle formula (≈0.884) and compare it with the average of the specular gradients of |x| (which is 0) and 2x (which is 2), each sampled with probability 1/2. The two numbers differ (1 vs ≈0.884), so the unbiasedness condition (3.13) is violated, and any convergence theorem for the stochastic method that relies on it does not follow.","tokens_in":32022,"feed_emoji":"🪞","tokens_out":8109,"duration_ms":71529,"temperature":0.7,"pith_summary":"The paper's central project is to define a new kind of derivative, the specular derivative, built from a weighted average of forward and backward difference quotients with weights chosen by distances on the graph of the function. It proves that whenever an ordinary Gâteaux or Fréchet derivative exists, the specular derivative exists and matches it, so the new notion is a true generalization rather than a competing rival. The key payoff is in nonsmooth convex optimization: in R^n, if a convex function is specularly Fréchet differentiable, its specular gradient is a subgradient, so the specular gradient method inherits the convergence guarantees of the classical subgradient method. The paper also establishes weak, inequality-based versions of the Mean Value Theorem and Fermat's Theorem in normed vector spaces, and demonstrates numerically that specular-gradient methods minimize elastic-net problems where gradient descent, Adam, and BFGS fail.","feed_headline":"Specular gradient is a subgradient for convex functions","feed_subtitle":"A mirror-weighted derivative matches classical calculus on smooth functions and tames nonsmooth convex minima.","key_machinery":"The central object is the specular directional derivative and its equivalent half-angle formula in terms of one-sided derivatives. The formula ∂∧_v f(x) = ∥v∥ tan(½ arctan(∂+_v f(x)/∥v∥) + ½ arctan(∂−_v f(x)/∥v∥)) converts left and right derivative information into a single number that always lies between them, and degenerates to the classical derivative when both sides agree. When f is convex, this interpolation guarantees the specular gradient satisfies the subgradient inequality, which is the mechanism that imports all subgradient-method convergence theory. The auxiliary functions A and B encode the same identity in algebraic form for estimates and proofs.","core_discovery":"The discovery at the center of the paper is the specular directional derivative ∂∧_v f(x), formed by reflecting the forward and backward secant lines about a mirror in the graph of f; in R^n it equals ∥v∥ tan(½ arctan(∂+_v f(x)/∥v∥) + ½ arctan(∂−_v f(x)/∥v∥)). When the classical directional derivative exists, the specular derivative agrees with it, and in normed vector spaces specular Gâteaux and Fréchet differentiability generalize the classical notions. The load-bearing result is Theorem 3.2: for a convex function on R^n that is specularly Fréchet differentiable, the specular gradient ∇s f(x) is a genuine subgradient, i.e., f(w) ≥ f(x) + ∇s f(x)·(w−x) for all w. Combined with the Quasi-Fer","pith_inferences":["A caution that goes beyond the paper's claims: the unbiasedness condition E[∇s f_ξ(x)|x_k] = ∇s f(x_k) assumed in Definition 3.7 is not established in the paper, and because the specular derivative is a nonlinear function of function values it can fail for natural decompositions; for f = (|x| + 2x)/2 at x = 0, sampling f1 = |x| and f2 = 2x with equal probability gives E[∇s f_ξ] = 1 while ∇s f ≈ 0.","A testable design consequence: the half-angle formula suggests specular gradients are a one-parameter family of secant-like interpolants; one could generalize by replacing the arithmetic mean of arctangents with other symmetric means, producing a continuum of specular derivatives with adjustable bias between left and right slopes.","Because Theorem 2.12 requires a norm-preserving linear functional on Y, an assumption that fails for many infinite-dimensional spaces, the Quasi-Mean Value Theorem is most naturally stated for real-valued or Hilbert-space-valued maps; extending it to general Banach targets would require a different mechanism.","The equivalence in Corollary 2.8 between existence of one-sided derivatives and specular differentiability suggests a fast differentiability check: a function is specularly differentiable exactly when its left and right slopes exist and are not both +∞ or both −∞, which may ease automatic differentiation of nonsmooth programs."],"forward_implications":["For any convex, specularly Fréchet differentiable function on R^n, the specular gradient gives a deterministic rule for choosing a subgradient, so subgradient-method convergence theorems apply without separately computing the subdifferential.","In normed vector spaces, local extrema of specularly Gâteaux differentiable functionals satisfy the explicit necessary condition |∂∧_v f(x*)| ≤ ∥v∥, a quantitative generalization of Fermat's theorem to nondifferentiable functions.","The Quasi-Mean Value Theorem bounds the change in f between two points by the supremum and infimum of specular directional derivatives along the segment, giving a substitute for the classical mean value theorem when f is not differentiable.","The specular gradient method with square-summable, non-summable step sizes converges to a minimizer, matching subgradient-method worst-case guarantees and, per the numerical experiments, outperforming gradient descent, Adam, and BFGS on elastic-net instances where those methods fail.","The hybrid and stochastic variants improve wall-clock time while retaining convergence, provided the stochastic unbiasedness assumption holds."],"fun_headline_variants":["Specular gradient: a subgradient classical derivatives miss","Mirror secants craft a derivative that doubles as subgradient","Specular differentiation: subgradient for convex functions","New specular calculus yields subgradients at nonsmooth points","Reflect the secants, get a subgradient: specular calculus"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The stochastic specular gradient method rests on the unproven assumption that averaging specular gradients of randomly chosen component functions gives the specular gradient of the whole function, an equality that is not proven and is false for simple convex decompositions.","fun_headline_variants_meta":{"raw":{"variants":["Specular gradient: a subgradient classical derivatives miss","Mirror secants craft a derivative that doubles as subgradient","Specular differentiation: subgradient for convex functions","New specular calculus yields subgradients at nonsmooth points","Reflect the secants, get a subgradient: specular calculus"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000963,"raw_usage":{"total_tokens":3883,"prompt_tokens":638,"completion_tokens":3245,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":382,"completion_tokens_details":{"reasoning_tokens":3158}},"tokens_in":382,"tokens_out":3245,"duration_ms":22453,"temperature":1.0,"reasoning_tokens":3158,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T10:10:56.835276+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute, at x = 0, the specular gradient of f(x) = (|x| + 2x)/2 directly from the half-angle formula (≈0.884) and compare it with the average of the specular gradients of |x| (which is 0) and 2x (which is 2), each sampled with probability 1/2. The two numbers differ (1 vs ≈0.884), so the unbiasedness condition (3.13) is violated, and any convergence theorem for the stochastic method that relies on it does not follow.","supporting_citations":[],"review_version":2}