{"id":"5399aada-123f-4dc0-800d-336bad84f069","arxiv_id":"2412.09898","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"For convex orthogonally invariant matrix functions, the paper derives explicit second subderivatives and sufficient conditions for twice epi-differentiability, including a closed-form second-order epi-derivative for the nuclear norm.","lead":"This paper develops second-order calculus for matrix functions that depend only on singular values, such as the nuclear norm. It gives explicit formulas for second-order derivatives and optimality conditions that could improve algorithms for matrix completion and rank minimization.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The unproven second-order expansion of singular values (Cor. 2.6) is load-bearing: if it fails at repeated or zero singular values, the chain rules and the explicit nuclear-norm formula collapse.","rationale":"The paper's core mechanism for lifting second-order properties from an absolutely symmetric function f to f∘σ is the singular-value expansion Corollary 2.6. Theorem 5.3 substitutes σ''(X;H,W) into the parabolic subderivative of f to obtain the parabolic subderivative of f∘σ; Theorem 5.6 uses the same expansion to derive the second-subderivative formula and parabolic regularity. If the expansion were false or non-uniform, every one of these conclusions would be unsupported. The reader's weakest assumption flags exactly this, and I agree: the paper neither reproves the expansion nor states the hypotheses from [22] that guarantee it. Formulas (2.11)-(2.13) are delicate precisely at repeated and zero singular values, which are the cases of interest for orthogonally invariant functions. A secondary issue is the self-referential sequence definition in the proof of Theorem 3.5, which is the only proof of the explicit nuclear-norm formula; it is likely a typo, but as written it breaks the argument. Neither issue forces rejection, because the expansion is probably correct and the typo is fixable, but the paper should verify and correct both before publication. Hence the reader's CONDITIONAL verdict stands without change.","tokens_in":37449,"tokens_out":27782,"duration_ms":276042,"concrete_test":"Verify Corollary 2.6 directly: take X0 = diag(1,1,0) (repeated nonzero and zero singular values), choose random H,W, and for t = 10^{-k}, k=3,...,7, compute E(t) = ||σ(X0+tH+1/2t^2W) - [σ(X0)+tσ'(X0;H)+1/2t^2σ''(X0;H,W)]||. If E(t) does not decay as o(t^2) (e.g., t^3 or better), the expansion is false. In parallel, re-derive (2.11)-(2.13) from [22, Thm 3.1] and check its hypotheses; if [22] requires the perturbing direction H to make 1/2(U^T H V + V^T H^T U) have distinct eigenvalues or imposes any generic condition, the result is not established for the repeated/zero cases on which the paper's theorems rely.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's transfer theorems (Thm 5.3, 5.6, 5.7) and the explicit nuclear-norm formula (Cor 3.6) all rest on Corollary 2.6: the parabolic expansion σ(X0+tH+1/2t^2W+o(t^2)) = σ(X0)+tσ'(X0;H)+1/2t^2σ''(X0;H,W)+o(t^2). This is quoted from [22, Thm 3.1] without proof, and the paper does not state the precise hypotheses under which formulas (2.11)–(2.13) hold. The formulas involve eigenvector bases Q that depend on H and on nontrivial spectral decompositions; the cases of repeated nonzero singular values and zero singular values are exactly where the paper needs the result. If [22] assumes distinct singular values, or if the remainder is not uniform in the direction variable, then the equality (5.2) in Thm 5.3 and the second-subderivative formula in Thm 5.6 fail. Theorem 5.3 applies the expansion with W_k→W, requiring uniformity of the o(t^2) remainder along the sequence; Corollary 2.6, as stated, only gives a pointwise expansion for fixed H,W. In addition, the proof of Theorem 3.5 contains the self-referential line 'let us consider H_k = H + t_k H_k V_αΣ^{-1}_α U^T_α H_k', which cannot define a sequence; until corrected, the explicit nuclear-norm formula rests on an invalid proof step. The central claim therefore hinges on an external result that is not verified in the paper and on a fixable but currently broken construction.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper studies second-order variational properties of orthogonally invariant matrix functions F = f ∘ σ on M_{m,n}, with n ≤ m. It uses a second-order parabolic expansion of singular values quoted from [22] to prove chain rules for subderivatives and second subderivatives, and it establishes parabolic epi-differentiability and twice epi-differentiability of f ∘ σ when f is convex, lsc, locally Lipschitz, parabolically epi-differentiable, and parabolically regular. The nuclear norm is treated as a special case: Corollary 3.6 gives an explicit second-order epi-derivative under the rank condition n > r, and Corollary 5.9 with Remark 5.10 covers the polyhedral case. The final section derives second-order necessary and sufficient optimality conditions for a class of matrix optimization problems.","tokens_in":37835,"tokens_out":14518,"duration_ms":133054,"significance":"If the proofs are completed, the paper is a useful contribution: it gives computable second-order epi-derivatives for the nuclear norm and for convex orthogonally invariant matrix functions, with no fitted parameters and no circularity. The main theorems are clearly stated, and the application to second-order optimality conditions is natural. The central novelty is the transfer of parabolic regularity and twice epi-differentiability from the absolutely symmetric function f to f ∘ σ, with explicit correction terms expressed through the first- and second-order directional derivatives of singular values.","major_comments":[{"comment":"The sequence H_k is introduced by the self-referential formula H_k = H + t_k H_k V_α Σ_α^{-1} U_α^T H_k. This equation does not define a sequence, and the subsequent convergence claim depends on it. Since this construction provides the recovering sequence needed for the epi-limit upper bound, the formula for d²Ψ_n(X0|Ω)(H) in Theorem 3.5 and the explicit nuclear-norm formula in Corollary 3.6 are not established as written. Please replace the right-hand H_k factors by H if that is the intended argument, or prove existence of the fixed point, and then recompute the limit with the corrected definition.","section":"§3, proof of Theorem 3.5"},{"comment":"Corollary 2.6 quotes from [22] the parabolic expansion σ(X0+tH+1/2t²W+o(t²)) = σ(X0)+tσ'(X0;H)+1/2t²σ''(X0;H,W)+o(t²). The manuscript does not state the precise hypotheses on repeated and zero singular values under which formulas (2.11)–(2.13) hold, and it does not establish uniformity of the o(t²) remainder in the direction variables. This matters because Theorem 5.3 and Propositions 4.5–4.6 apply the expansion along sequences W_k → W and with o(t²) perturbations inside the argument; a pointwise expansion for fixed (H,W) is insufficient for those steps. Please either prove the expansion, cite a uniform version, or supply the missing continuity argument.","section":"§2, Corollary 2.6; §5, Theorem 5.3"},{"comment":"Proposition 5.5(i)–(ii) is stated with σ'(X;Y) in (5.7) and in the proof, while the hypotheses concern H ∈ K_{f∘σ}(X,Y) and the proof of Theorem 5.6 uses σ'(X;H). If σ'(X;Y) is a typo, it should be corrected; as printed, the statement is not the one used later and cannot be verified as a statement about the critical direction H.","section":"§5, Proposition 5.5"}],"minor_comments":[{"comment":"The abstract states without qualification that the nuclear norm of a real m×n matrix is twice epi-differentiable with an explicit second-order epi-derivative, while Corollary 3.6 assumes rank X0 = r and n > r. Please add the rank condition to the abstract or point to Corollary 5.9 and Remark 5.10 for the full-rank case.","section":"Abstract and §3, Corollary 3.6"},{"comment":"The second-order directional derivative is written σ''_s(X0;H) although σ'' is elsewhere a two-direction object; please define the shorthand or write σ''_s(X0;H,H).","section":"§3, Lemma 3.3"},{"comment":"The localization argument 'adding to f the indicator of a polyhedral neighborhood of ¯x' modifies the domain of f and hence the critical cone; since the conclusion is stated for all H, the argument should be made precise.","section":"§5, Corollary 5.9 proof"},{"comment":"The case Ψ'_n(X0;H) < ⟨Ω,H⟩ is not discussed in the formula for d²Ψ_n(X0|Ω)(H); this case cannot occur for Ω ∈ ∂̂Ψ_n(X0), but the statement should say so explicitly.","section":"§3, Theorem 3.5 statement"}],"recommendation":"major_revision","confidential_remarks":"The external reliance on [22] is acceptable if the uniformity and hypothesis issues are addressed. The self-referential H_k in Theorem 3.5 appears to be a fixable typo rather than a substantive gap, but it is load-bearing for Corollary 3.6. The paper is within the journal's scope and the main claims are likely correct, but the proof as printed needs repair."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: this is a serious paper that does what it claims. It extends the Mohammadi–Sarabi spectral-function program to singular values, gives an explicit second-order epi-derivative for the nuclear norm, and computes second subderivatives for convex orthogonally invariant functions. The formulas are genuinely new and the proofs are detailed. I think it deserves peer review.\n\nWhat is good: the transfer theorems (Thm 5.3, 5.6) are not just restatements of the spectral case. The extra terms involving Σ^{-1}_α are where the singular-value difficulty lives, and the authors handle them explicitly. The application to second-order optimality conditions for matrix optimization is a natural and useful payoff. The paper also correctly credits the second-order expansion of singular values to Zhang, Zhang and Xiao [22] rather than claiming to reprove it.\n\nSoft spots, in order of importance. First, the proof of Theorem 3.5 contains the line \"let us consider H_k = H + t_k H_k V_α Σ^{-1}_α U^T_α H_k\" — H_k is defined in terms of itself. That is clearly a typo, but as written the construction is invalid, and Corollary 3.6 depends on that step. It must be fixed before publication. Second, the abstract claims the nuclear norm is twice epi-differentiable for any real m×n matrix, but Corollary 3.6 requires n > r. That is an overstatement; the rank-deficient case may behave differently. Third, the edifice rests on Corollary 2.6, quoted from [22] without proof. The paper does not state the precise hypotheses under which formulas (2.11)–(2.13) hold, and the applications use the expansion with sequences W_k→W, so a pointwise expansion is not obviously enough. The stress-test concern about uniformity is legitimate. I am not convinced it is a real gap — [22] is peer-reviewed and the formulas are written to handle repeated singular values — but the authors should state the hypotheses or give a proof sketch.\n\nThe citation pattern is clean. There are no fitted parameters, no self-citations beyond reasonable credit, and the math looks mostly right. Who is this for? People working on second-order conditions for nuclear norm, Ky Fan k-norm, or general convex matrix optimization, and researchers in nonsmooth variational analysis. I would send it out for review, with a request to fix the H_k typo, correct the abstract's rank claim, and clarify the status of Corollary 2.6.","headline":"A solid second-order variational analysis paper that extends spectral-function results to singular values and delivers explicit nuclear-norm formulas, but it has a few fixable typos and one load-bearing external expansion that should be checked.","tokens_in":38316,"tokens_out":2575,"would_cite":true,"duration_ms":27343,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["15A18","49J52","49J53","94A11"],"pacs":[],"model":"deepseek-v4-flash","headline":"Twice epi-differentiability transfers from an absolutely symmetric function to its singular-value composite, and is made explicit for the nuclear norm.","keywords":["orthogonally invariant matrix functions","second subderivatives","twice epi-differentiability","nuclear norm","parabolic epi-differentiability","singular value functions","second-order optimality conditions"],"falsifier":"A concrete check: take a rank-deficient matrix with a repeated zero singular value, for instance $X=\\operatorname{diag}(1,0,0)$, choose $H$ with nonzero off-diagonal blocks, and compare the formula in Corollary 3.6 with the direct liminf of the second-order difference quotients of the nuclear norm; any mismatch, or any dependence of the expression on the chosen SVD, would refute the claimed formula.","tokens_in":37279,"feed_emoji":"🧮","tokens_out":11932,"duration_ms":102018,"temperature":0.7,"pith_summary":"The paper studies convex functions on real matrices that depend only on singular values, written as $F(X)=f(\\sigma(X))$, where $\\sigma(X)$ lists the singular values in decreasing order and $f$ is absolutely symmetric (unchanged by signed permutations of its arguments). It aims to show that when $f$ is convex, locally Lipschitz, parabolically regular, and parabolically epi-differentiable, these second-order variational properties lift to $F$. The main result is that the nuclear norm $\\|X\\|_*$ is twice epi-differentiable—its second-order difference quotients have a well-defined epigraphical limit—with an explicit formula for its second-order epi-derivative. This gives second-order optimality conditions for matrix optimization problems such as matrix completion and rank minimization.","feed_headline":"Nuclear norm is twice epi-differentiable, with an explicit formula","feed_subtitle":"For convex orthogonally invariant matrix functions, this yields explicit second-order optimality conditions.","key_machinery":"The workhorse is the singular-value map $\\sigma:\\mathbb{M}_{m,n}\\to\\mathbb{R}^n$ together with its second-order Taylor expansion along parabolic arcs, $\\sigma\\bigl(X+tH+\\tfrac12 t^2W\\bigr)=\\sigma(X)+t\\sigma'(X;H)+\\tfrac12 t^2\\sigma''(X;H,W)+o(t^2)$, quoted from Zhang, Zhang and Xiao. This expansion is what lets the authors transfer parabolic properties of $f$ to the matrix function $f\\circ\\sigma$. A second device is the symmetric embedding $B(X)=\\begin{bmatrix}0&X\\\\ X^T&0\\end{bmatrix}$, which turns singular values of $X$ into eigenvalues of $B(X)$ and thereby brings known eigenvalue perturbation formulas into play. Together these two objects carry the chain rules, the second-subderivative formula, and the explicit nuclear-norm computation.","core_discovery":"The central claim is a transfer principle: every well-behaved absolutely symmetric function $f$ produces an orthogonally invariant matrix function $f\\circ\\sigma$ with the same second-order variational properties. Concretely, if $f$ is lower semicontinuous, convex, locally Lipschitz relative to its domain, parabolically epi-differentiable at $\\sigma(X)$, and parabolically regular at $\\sigma(X)$ for $\\sigma(Y)$, then $f\\circ\\sigma$ is parabolically regular at $X$ for $Y$ and twice epi-differentiable at $X$ for $Y$. The proof computes the second subderivative exactly: for $Y\\in\\partial(f\\circ\\sigma)(X)$ and $H$ in the critical cone, $d^2(f\\circ\\sigma)(X|Y)(H)$ equals a base term $d^2f(\\sigma(X)|\\sigma(Y))(\\sigma'(X;H))$ plus two curvature corrections determined by the cross blocks of the linearization $B(H)$ and by the coupling between the zero and the nonzero singular spaces. Specializing to $f(x)=\\|x\\|_1$ gives the nuclear norm statement, and specializing to polyhedral $f$ gives a no-gap quadratic growth condition.","pith_inferences":["A natural extension the paper does not pursue is the nonconvex case: the same transfer argument should work for prox-regular absolutely symmetric functions, since most of the proof's ingredients are variational rather than convexity-specific.","The explicit nuclear-norm second-order epi-derivative could feed directly into second-order methods for low-rank matrix recovery, replacing the common habit of approximating the regularizer's Hessian.","Because the chain rules depend on the quoted expansion (2.14), a direct numerical check of that expansion at repeated singular values would independently test the entire transfer principle."],"forward_implications":["For the nuclear norm, Corollary 3.6 provides a closed-form second-order epi-derivative at any real matrix, including rank-deficient and repeated-singular-value points.","For any convex orthogonally invariant matrix function built from a parabolically regular absolutely symmetric function, the second subderivative is computable from Theorem 5.6, so second-order necessary and sufficient conditions for problem (P) can be written explicitly.","When the absolutely symmetric function is polyhedral, such as the $\\ell^1$-norm, the second subderivative collapses to an indicator of a critical cone plus two curvature terms, yielding no-gap quadratic growth conditions.","Theorem 5.8 extends second-order optimality conditions from eigenvalue-based spectral models to singular-value-based models, covering the nuclear norm, the spectral norm, and the Ky Fan $k$-norm regularizers."],"supporting_citations":[{"why":"supplies the second-order directional derivative formulas for singular values and the parabolic expansion (2.14) that the transfer results build on.","marker":"[22]"},{"why":"provides the composite-function framework for parabolic epi-differentiability and twice epi-differentiability used in Theorem 5.6 and Proposition 5.7.","marker":"[15]"},{"why":"establishes parabolic regularity for spectral functions, the model adapted here to singular-value functions.","marker":"[14]"},{"why":"gives the foundational variational theory of singular-value functions and the first-order directional derivatives used in Proposition 2.4.","marker":"[10]"},{"why":"proves that convexity and lower semicontinuity transfer from $f$ to $f\\circ\\sigma$ and gives the subdifferential characterization used in Proposition 4.2.","marker":"[8]"},{"why":"supplies the definitions of second subderivatives, epi-convergence, and the optimality conditions invoked in Theorem 5.8.","marker":"[18]"},{"why":"gives the first-order directional derivative formulas for singular values used in the subderivative chain rule of Theorem 4.3.","marker":"[5]"},{"why":"provides the epi-differentiability result used in Corollary 3.6 to pass from the tail function $\\Psi_n$ to the nuclear norm.","marker":"[17]"},{"why":"supplies the chain rule for subderivatives of composite functions used in Theorem 4.8.","marker":"[13]"}],"fun_headline_variants":["Nuclear norm twice epi-differentiable: explicit formula","Explicit second epi-derivative for nuclear norm","Twice epi-differentiability for orthogonally invariant matrix functions","Absolutely symmetric f gives twice epi-differentiable matrix functions","Nuclear norm's second-order structure made explicit"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that, along parabolic arcs $X+tH+\\tfrac12 t^2W$, the singular-value vector obeys the quoted second-order expansion with remainder $o(t^2)$ uniformly in the directions; the paper does not reprove this expansion, so if it fails at repeated singular values the chain rules and the nuclear-norm formula collapse.","fun_headline_variants_meta":{"raw":{"variants":["Nuclear norm twice epi-differentiable: explicit formula","Explicit second epi-derivative for nuclear norm","Twice epi-differentiability for orthogonally invariant matrix functions","Absolutely symmetric f gives twice epi-differentiable matrix functions","Nuclear norm's second-order structure made explicit"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000883,"raw_usage":{"total_tokens":3850,"prompt_tokens":1019,"completion_tokens":2831,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":635,"completion_tokens_details":{"reasoning_tokens":2760}},"tokens_in":635,"tokens_out":2831,"duration_ms":19209,"temperature":1.0,"reasoning_tokens":2760,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T16:36:36.191947+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A concrete check: take a rank-deficient matrix with a repeated zero singular value, for instance $X=\\operatorname{diag}(1,0,0)$, choose $H$ with nonzero off-diagonal blocks, and compare the formula in Corollary 3.6 with the direct liminf of the second-order difference quotients of the nuclear norm; any mismatch, or any dependence of the expression on the chosen SVD, would refute the claimed formula.","supporting_citations":[{"cited_title":"and Xiao, X.: On the Second-order Directio nal Derivatives of Singular Values of Matrices and Symmetric Matrix-valued Functions","cited_arxiv_id":null,"evidence_quote":"supplies the second-order directional derivative formulas for singular values and the parabolic expansion (2.14) that the transfer results build on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"provides the composite-function framework for parabolic epi-differentiability and twice epi-differentiability used in Theorem 5.6 and Proposition 5.7."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"establishes parabolic regularity for spectral functions, the model adapted here to singular-value functions."},{"cited_title":"P art I: Theory","cited_arxiv_id":null,"evidence_quote":"gives the foundational variational theory of singular-value functions and the first-order directional derivatives used in Proposition 2.4."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"proves that convexity and lower semicontinuity transfer from $f$ to $f\\circ\\sigma$ and gives the subdifferential characterization used in Proposition 4.2."},{"cited_title":"Springer, B erlin (1998)","cited_arxiv_id":null,"evidence_quote":"supplies the definitions of second subderivatives, epi-convergence, and the optimality conditions invoked in Theorem 5.8."},{"cited_title":"and Toh, K.C.: An Introduction to a Class of Matrix Cone Programming","cited_arxiv_id":null,"evidence_quote":"gives the first-order directional derivative formulas for singular values used in the subderivative chain rule of Theorem 4.3."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"provides the epi-differentiability result used in Corollary 3.6 to pass from the tail function $\\Psi_n$ to the nuclear norm."},{"cited_title":"and Sarabi, M.E.: Variational A nalysis of Composite Models with Applications to Continuous Optimization","cited_arxiv_id":null,"evidence_quote":"supplies the chain rule for subderivatives of composite functions used in Theorem 4.8."}],"review_version":1}