Pith. sign in

REVIEW 3 major objections 4 minor 22 references

Twice Epi-Differentiability of Orthogonally Invariant Matrix Functions and Application

T0 review · 3 major / 4 minor · reviewed 2026-08-11 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Twice epi-differentiability transfers from an absolutely symmetric function to its singular-value composite, and is made explicit for the nuclear norm.

desk verdict 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. read the letter →

arxiv 2412.09898 v2 pith:XYQM46SU submitted 2024-12-13 math.OC

Jiahuan He , Chao Kan , Wen Song This is my paper · ORCID
classification math.OC MSC 15A1849J5249J5394A11
keywords orthogonallyinvariantmatrixfunctionssecondsubderivativestwiceepi-differentiabilitynuclearnormparabolicsingularvaluesecond-orderoptimalityconditions
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

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.

What carries the argument

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.

What would settle it

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.

Watch

Extended reading notes

Core claim

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.

Load-bearing premise

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.

Editorial extensions

If this is right

  • 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.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • 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.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 4 minor

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.

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 (3)
  1. [§3, proof of Theorem 3.5] 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.
  2. [§2, Corollary 2.6; §5, Theorem 5.3] 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.
  3. [§5, Proposition 5.5] 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.
minor comments (4)
  1. [Abstract and §3, Corollary 3.6] 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.
  2. [§3, Lemma 3.3] 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).
  3. [§5, Corollary 5.9 proof] 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.
  4. [§3, Theorem 3.5 statement] 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.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper's second-order formulas are derived from an external second-order directional-derivative expansion of singular values [22] plus standard composite-function calculus; there are no fitted parameters, no self-citation chains, and no quantity defined in terms of the claimed conclusion.

full rationale

Score 0. The derivation chain is: (i) import the second-order directional derivatives of singular values, formulas (2.11)-(2.13) and the parabolic expansion (2.14), as Theorem 3.1 of the external paper [22] by Zhang, Zhang and Xiao; (ii) apply these together with standard composite-function calculus [13,15] to obtain the subderivative chain rule (Thm 4.3), the tangent and second-order tangent set representations (Cor 4.4, Prop 4.6), the parabolic subderivative chain rule (Thm 5.3), and the second subderivative formula (Thm 5.6); (iii) specialize to the nuclear norm by writing it as the sum of the C2 function Phi_r and the tail function Psi_n (Thm 3.5, Cor 3.6). No parameter is fitted, no prediction is generated from a subset of data, and no function is defined in terms of the claimed conclusion. The reference list contains no self-citations by the present authors, and the load-bearing expansion of singular values is attributed to an independent external source rather than to an unverified prior claim of the same authors. The main robustness concern is that Corollary 2.6 is quoted rather than reproved, and the proof of Theorem 3.5 contains the malformed construction H_k = H + t_k H_k V_alpha Sigma_alpha^{-1} U_alpha^T H_k, which cannot define a sequence as written; the uniformity of the o(t^2) remainder in (2.14) is also not established inside the paper. However, reliance on an external theorem and a fixable proof gap are correctness risks, not circularity under the stated rules: the formulas do not reduce to their inputs by construction, and there is no fitted input renamed as a prediction.

Assumptions & free parameters 0 free parameters · 7 assumptions · 0 invented entities

No free parameters or invented entities. The central results rely on standard spectral inequalities (Fan, von Neumann), the second-order directional derivative formulas for singular values from Zhang, Zhang and Xiao [22], and the composite function theory of Mohammadi, Mordukhovich and Sarabi [13]. All are external benchmarks, and the paper's new chain rules are proved rather than assumed.

assumptions (7)
  • standard math Fan's inequality ⟨A,B⟩ ≤ ⟨λ(A),λ(B)⟩ for symmetric matrices
    Used in Theorem 4.3 and Proposition 5.2 to relate traces of matrix products to eigenvalue inner products.
  • standard math Von Neumann's trace theorem ⟨X,Y⟩ ≤ ⟨σ(X),σ(Y)⟩ for arbitrary matrices
    Used throughout Sections 4 and 5 to transfer inner products to singular value vectors.
  • domain assumption Second-order directional derivative formulas for singular values (Propositions 2.4 and 2.5, from [22])
    The paper's entire second-order analysis rests on these quoted formulas, especially Corollary 2.6. They are not reproved.
  • domain assumption Composite function chain rule from Mohammadi, Mordukhovich and Sarabi [13, Theorem 3.4]
    Used in Theorem 4.8 to obtain subderivative chain rules for absolutely symmetric functions.
  • domain assumption f is proper, lsc, convex, absolutely symmetric, and locally Lipschitz relative to its domain
    These assumptions on the absolutely symmetric function f are stated in Theorem 5.6 and Corollary 5.9 and are needed for the second subderivative formulas.
  • domain assumption f is parabolically epi-differentiable at σ(X) and parabolically regular at σ(X) for σ(Y)
    These are the key inherited properties that Theorem 5.6 and Proposition 5.7 require of f to conclude about f∘σ.
  • standard math Lipschitz property of singular values: ||σ(X)-σ(Y)|| ≤ ||X-Y||
    Used in Proposition 4.1 to prove distance preservation for orthogonally invariant sets.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Twice Epi-Differentiability of Orthogonally Invariant Matrix Functions and Application." pith.science (2026). https://pith.science/paper/XYQM46SU

@misc{pith2026241209898,
  author       = {Pith},
  title        = {Pith review of: Twice Epi-Differentiability of Orthogonally Invariant Matrix Functions and Application},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XYQM46SU}},
  note         = {Machine review of arXiv:2412.09898}
}
abstract

In this paper, our focus lies on the study of the second-order variational analysis of orthogonally invariant matrix functions. It is well-known that an orthogonally invariant matrix function is an extended-real-value function defined on ${\mathbb M}_{m,n}\,(n \leqslant m)$ of the form $f \circ \sigma$ for an absolutely symmetric function $f \colon \R^n \rightarrow [-\infty,+\infty]$ and the singular values $\sigma \colon {\mathbb M}_{m,n} \rightarrow \R^{n}$. We establish several second-order properties of orthogonally invariant matrix functions, such as parabolic epi-differentiability, parabolic regularity, and twice epi-differentiability when their associated absolutely symmetric functions enjoy some properties. Specifically, we show that the nuclear norm of a real $m \times n$ matrix is twice epi-differentiable and we derive an explicit expression of its second-order epi-derivative. Moreover, for a convex orthogonally invariant matrix function, we calculate its second subderivative and present sufficient conditions for twice epi-differentiability. This enables us to establish second-order optimality conditions for a class of matrix optimization problems.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

22 extracted references · 21 canonical work pages

  1. [22]

    and Xiao, X.: On the Second-order Directio nal Derivatives of Singular Values of Matrices and Symmetric Matrix-valued Functions

    Zhang, L., Zhang, N. and Xiao, X.: On the Second-order Directio nal Derivatives of Singular Values of Matrices and Symmetric Matrix-valued Functions. Set-Va lued Var. Anal. 21(3): 557-586 (2013) 30

    work page 2013

  2. [1]

    Springer, New York (2000)

    Bonnans, J.F., Shapiro, A.: Perturbation Analysis of Optimization P roblems. Springer, New York (2000)

    work page 2000

  3. [2]

    and Zhao, X.: Quadratic Growth Conditions for Con vex Matrix Optimiza- tion Problems Associated with Spectral Functions

    Cui, Y., Ding, C. and Zhao, X.: Quadratic Growth Conditions for Con vex Matrix Optimiza- tion Problems Associated with Spectral Functions. SIAM J. Optim. 2 7(4): 2332-2355 (2017)

    work page 2017

  4. [3]

    and Sendov, H.: Prox-Regularity of Spectral Functions and Spectral Sets

    Daniilidis, A., Lewis, A., Malick, J. and Sendov, H.: Prox-Regularity of Spectral Functions and Spectral Sets. J. Convex Anal. 15(3): 547-560 (2008)

    work page 2008

  5. [4]

    Set-Valued Var

    Ding, C.: Variational Analysis of the Ky Fan k-norm. Set-Valued Var. Anal. 25: 265-296 (2017)

    work page 2017

  6. [5]

    and Toh, K.C.: An Introduction to a Class of Matrix Cone Programming

    Ding, C., Sun, D. and Toh, K.C.: An Introduction to a Class of Matrix Cone Programming. Math. Program. 144(1-2): 141-179 (2014)

    work page 2014

  7. [6]

    Springer, Berlin (2001)

    Hiriart-Urruty, J.-B., Lemar´ echal, C.: Fundamentals of Convex Analysis. Springer, Berlin (2001)

    work page 2001

  8. [7]

    Cambridge univ ersity press, Cambridge (2012)

    Horn, R.A., Johnson, C.R.: Matrix Analysis (2nd ed.). Cambridge univ ersity press, Cambridge (2012)

    work page 2012

Show all 22 references
  1. [8]

    Lewis, A.S.: The Convex Analysis of Unitarily Invariant Matrix Funct ions. J. Convex Anal. 2(1–2): 173-183 (1995)

    1995

  2. [9]

    Lewis, A.S.: Nonsmooth Analysis of Eigenvalues. Math. Program. 8 4(1): 1-24 (1999)

    1999

  3. [10]

    P art I: Theory

    Lewis, A.S., Sendov, H.S.: Nonsmooth Analysis of Singular Values. P art I: Theory. Set-Valued Anal. 13: 213-241 (2005)

    2005

  4. [11]

    P art II: Applications

    Lewis, A.S., Sendov, H.S.: Nonsmooth Analysis of Singular Values. P art II: Applications. Set-Valued Anal. 13: 243-264 (2005)

    2005

  5. [12]

    Linear Algebra Appl

    Mathias, R.: The Spectral Norm of a Nonnegative Matrix. Linear Algebra Appl. 139: 269-284 (1990)

    1990

  6. [13]

    and Sarabi, M.E.: Variational A nalysis of Composite Models with Applications to Continuous Optimization

    Mohammadi, A., Mordukhovich, B.S. and Sarabi, M.E.: Variational A nalysis of Composite Models with Applications to Continuous Optimization. Math. Oper. Res . 47(1): 397–426 (2021)

    2021

  7. [14]

    Mohammadi, A., Sarabi, E.: Parabolic Regularity of Spectral Func tions. Math. Oper. Res. (2024) https://doi.org/10.1287/moor.2023.0010

    2024

  8. [15]

    Mohammadi, A., Sarabi, M.E.: Twice Epi-Differentiability of Extended -Real-Valued Func- tions with Applications in Composite Optimization. SIAM J. Optim. 30: 23 79–2409 (2020)

    2020

  9. [16]

    Princeton University Press, Princeton (1970)

    Rockafellar, R.T.: Convex Analysis. Princeton University Press, Princeton (1970)

    1970

  10. [17]

    Rockafellar, R.T.: First- and Second-Order Epi-Differentiability in Nonlinear Programming. Trans. Amer. Math. Soc. 307(1): 75-108 (1988)

    1988

  11. [18]

    Springer, B erlin (1998)

    Rockafellar, R.T., Wets, R.J.-B.: Variational Analysis. Springer, B erlin (1998)

    1998

  12. [19]

    Nonlinear Anal

    Torki, M.: Second-Order Directional Derivatives of All Eigenvalu es of a Symmetric Matrix. Nonlinear Anal. 46(8): 1133-1150 (2001)

    2001

  13. [20]

    SIAM Re v

    Vandenberghe, L., Boyd, S.: Semidefinite Programming. SIAM Re v. 38(1): 49-95 (1996)

    1996

  14. [21]

    Linear Algebra Appl

    Watson, G.A.: Characterization of the Subdifferential of Some M atrix Norms. Linear Algebra Appl. 170(1): 33-45 (1992)

    1992

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.