REVIEW 3 major objections 4 minor 5 references
Trace of Multi-variable Matrix Functions and its Application to Functions of Graph Spectrum
T0 review · 3 major / 4 minor · reviewed 2026-08-10 · deepseek-v4-flash
Pith's one-line read For Hermitian matrices, the trace of a multi-variable matrix function inherits the monotonicity and convexity of the scalar function, with applications to graph spectra.
desk verdict A useful multi-variable trace convexity/monotonicity extension with a fixable proof gap; the convexity proof is solid and the graph application is a nice payoff. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The multi-variable matrix extension f(M_1,...,M_m) is built with the Kronecker product: each Hermitian matrix is diagonalized by its eigenvector matrix, and the extended matrix is (P_1⊗...⊗P_m) diag(f(λ_{1j1},...,λ_{mjm})) ($P_1^{{-1}}$⊗...⊗$P_m^{{-1}}$). Its trace is the sum of f over all eigenvalue tuples. The load-bearing identities are the derivative formula d/dt Tr(f(M_1(t),...,M_m(t))) = Σ_k Tr(∂_k f(...) (I⊗...⊗ dM_k/dt ⊗...⊗ I)), which the proof obtains first for monomials, then polynomials, then C^m functions by uniform approximation, and the convexity argument that writes each eigenvalue λ_{lk} as a quadratic form y_{lk}^* M_l y_{lk} and applies the convexity inequality of f. Monotonicity follows from the derivative formula because dM_k/dt in the parameterization is positive semidefinite and therefore a square, making each term in the sum positive semidefinite when f is coordinatewise monotone.
What would settle it
Evaluate the convexity inequality of Proposition 2.4 for f(x,y)=(x+y)^2 on two pairs of 2×2 positive-semidefinite Hermitian matrices that do not commute, at α=1/2; the paper predicts Tr f((M_1+M_1')/2,(M_2+M_2')/2) ≤ (Tr f(M_1,M_2)+Tr f(M_1',M_2'))/2, and a numerical counterexample there would disprove the central convexity claim.
Extended reading notes
Core claim
On the paper's own terms, the central discovery is that the monotonicity and convexity of a scalar function of several variables are preserved by the map M_1,...,M_m ↦ Tr(f(M_1,...,M_m)), where f is evaluated through the Kronecker-product extension. Concretely, Proposition 2.3 shows that when f is monotonically increasing (respectively decreasing) in each coordinate on a product of convex subsets of R, the trace is increasing (respectively decreasing) with respect to the Löwner order on each Hermitian matrix argument. Proposition 2.4 shows that when f is convex (respectively concave), the trace is convex (respectively concave) on the product of Hermitian matrix sets. The derivative identity of Proposition 2.1 is the tool behind the monotonicity statement, while the convexity statement is proved directly from the convexity of f applied to the quadratic forms defining the eigenvalues. Applied to the weighted graph Laplacian, Proposition 3.1 concludes that functions of the multi-spectrum are monotone or convex in edge weights under the same hypotheses on f.
Load-bearing premise
The proof assumes a smooth function and all its partial derivatives can be uniformly approximated by polynomials over the whole domain, but that kind of approximation is only guaranteed on compact sets and the proof does not ensure compactness when the domain is unbounded.
Editorial extensions
If this is right
- For any monotone-in-each-input f, the graph spectral function F(L)=Σ_{k_1,...,k_m} f(λ_{k_1},...,λ_{k_m}) is nondecreasing as any edge weight of the underlying graph increases.
- If f is convex, the same F is convex in the edge weights, so local minima under convex constraints on edge weights are global minima.
- The single-variable result that scalar convexity, not operator convexity, is enough for convexity of the trace now holds in several variables as well.
- When all matrix arguments are equal (M_1=...=M_m=L), the results give a direct criterion for monotonicity and convexity of functions of a single matrix's spectrum, covering sums over repeated eigenvalue tuples.
Reading between the lines
- The convexity proof needs only continuity, while the monotonicity proof needs C^m smoothness; a natural conjecture is that a direct proof could weaken the smoothness requirement for monotonicity, since the derivative route is only one way to obtain the ordering.
- The compactness gap in the Stone-Weierstrass step suggests the derivative-based route is fully justified for compact spectral domains; for unbounded spectra one would need a local approximation argument, and the monotonicity result should be true under that repair.
- For graph optimization, the result opens the door to convex objectives built from pairwise or higher-order eigenvalue interactions (for example, sums of f(λ_i,λ_j) with f coordinatewise monotone); one could test numerically that such objectives are monotone in edge weights on random graphs.
Signed reviews
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper defines a multi-variable analogue of the matrix functional calculus: for f:S_1×...×S_m→R and diagonalizable matrices M_l with eigenvalues in S_l, it sets f(M_1,...,M_m) to be the tensor product of the diagonalizations with f evaluated on all m-tuples of eigenvalues; its trace is thus the sum of f over all eigenvalue tuples. The main results are a derivative formula for Tr∘f along C^1 paths (Proposition 2.1), a monotonicity statement along positive-semidefinite paths (Corollary 2.2), Loewner-order monotonicity of Tr∘f for coordinatewise monotone f (Proposition 2.3), convexity/concavity inheritance of Tr∘f (Proposition 2.4), and an application to functions of the weighted graph Laplacian spectrum (Proposition 3.1). The proof strategy uses Stone-Weierstrass approximation and eigenvalue perturbation arguments.
Significance. If the results are established rigorously, they give a natural multi-variable generalization of the known fact that Tr∘f inherits scalar convexity and monotonicity from f, without requiring operator convexity or operator monotonicity. The convexity part (Proposition 2.4) is proved by a direct Jensen-argument over Rayleigh quotients and appears correct. The monotonicity claim (Proposition 2.3) is also true and can be proved directly from Weyl's eigenvalue monotonicity theorem. The application to graph Laplacian spectra is a plausible and useful consequence. However, the derivative-based proof route in Section 2.3 has two substantial technical gaps, so the manuscript as written is not fully rigorous. The paper contains no fitted parameters or circular derivations, and the main claims are stated in a falsifiable mathematical form.
major comments (3)
- [Section 2.3, proof of Proposition 2.1] The proof claims that a C^m function on a product of locally compact subsets of R can be uniformly approximated, together with its first partial derivatives, by polynomials on the whole product. This is false when some S_j is unbounded. For example, if S_j=R, no polynomial can approximate exp(x) uniformly on R, and the proposed compactly supported auxiliary function cannot agree with an arbitrary f on all of R. Stone-Weierstrass requires compactness. This gap invalidates the C^m approximation step as stated and is load-bearing because Corollary 2.2 and Proposition 2.3 rely on Proposition 2.1. The argument can be repaired in the context of Proposition 2.3 by working on compact intervals containing the relevant eigenvalue ranges, but this needs to be stated explicitly.
- [Section 2.3, proof of Proposition 2.1] The eigenvalue-differentiability step is not justified. The proof invokes the Implicit Function Theorem to assert that the eigenvalues λ_{lj}(t) vary smoothly with t, but ordered eigenvalues of Hermitian matrices are not differentiable at crossings; for instance, diag(t,-t) has eigenvalues ±|t|. The derivative formula in equation (3) differentiates the individual terms f(λ_{1j_1}(t),...,λ_{m j_m}(t)) and therefore requires a valid choice of differentiable eigenvalue branches, such as those provided by Rellich's theorem for analytic families, or an alternative argument. As written, the proof does not supply this, and the gap affects Corollary 2.2.
- [Section 2.4, Proposition 2.3] Because of the two gaps in Proposition 2.1, the proof of Proposition 2.3 via Corollary 2.2 is not load-bearing as written. The monotonicity claim itself is nevertheless correct and can be proved directly: if M_l ≥ N_l for all l, Weyl's monotonicity theorem gives ordered eigenvalues λ_j(M_l) ≥ λ_j(N_l) for each l, and coordinatewise monotonicity of f yields Tr f((M_l)) ≥ Tr f((N_l)) termwise. The manuscript should either repair Proposition 2.1 or, preferably, replace the monotonicity proof with this direct argument, which also removes the unnecessary C^m assumption in Proposition 2.3.
minor comments (4)
- [Examples 2.2 and 2.3] The function f(x,y)=x^3 y^5 is not monotonically increasing with respect to each input on all of R^2: for fixed y<0 it is decreasing in x, and for y=0 it is constant. The examples should restrict the domain to, say, R_{\ge0}^2 or use a genuinely coordinatewise monotone function on R^2.
- [Example 2.4] The example uses inverses of M^2⊗I and I⊗N^2 but states that M and N are positive semi-definite; these inverses need not exist. The example should be restricted to positive definite matrices, and f should be specified on the positive orthant.
- [Throughout] There are numerous typographical issues, including 'monotonocity' in the abstract, 'diagonaliable', 'Weyly's Perturbation Theorem', and the garbled product symbol 'mą l=1 S_l' in the proof of Proposition 2.4. A careful proofreading pass is needed.
- [Proof of Proposition 2.4] The notation for eigenvectors and eigenvalue indices is sometimes confusing, particularly the use of y_{l j_l}, m_{lk}, and n_{lk} in the same expression. The double-stochasticity of the coefficients |y^*_{lj} m_{lk}|^2 is used but not explicitly named; stating it would improve readability.
Circularity Check
No significant circularity: the paper's trace inequalities follow from its tensor-product definition and standard external results (Jensen, Weyl, Bhatia, Carlen), with no fitted quantity presented as a prediction.
full rationale
The paper's central results, Propositions 2.3 and 2.4, are derived directly from the definition of the multi-variable matrix extension and from standard external theorems. Proposition 2.4's convexity proof is self-contained: it expresses the trace as a sum of f evaluated at Rayleigh quotients, applies scalar convexity, and uses the resolution-of-identity property of Hermitian eigenbases. Proposition 2.3's monotonicity proof is based on Proposition 2.1 and Corollary 2.2, which in turn use standard tools (Stone-Weierstrass approximation, Weyl's perturbation bound, and the implicit function theorem) attributed to external references (Abbott, Bhatia, Rahman, Carlen). No parameter is fitted, no quantity is defined in terms of the result it is used to prove, and no load-bearing claim rests on the authors' own prior work. The only self-citation, Sahin et al. (2024), is cited as an example of an application area and is not used to justify any theorem. A reviewer might question the Stone-Weierstrass step in Proposition 2.1 for unbounded S_j, but that is a proof gap or correctness concern, not circularity; the monotonicity claim is independently supported by Weyl's min-max principle. Accordingly, the circularity score is 0.
Assumptions & free parameters
assumptions (5)
- standard math Stone-Weierstrass approximation theorem provides uniform polynomial approximation of continuous functions on compact Hausdorff spaces.
- standard math Weyl's perturbation theorem gives |λ_j(t+Δt)-λ_j(t)| ≤ ||M(t+Δt)-M(t)||_2 for Hermitian matrices.
- standard math Implicit Function Theorem ensures smooth variation of eigenvalues of smoothly varying diagonalizable matrices.
- standard math Jensen's inequality for convex functions on product convex sets.
- domain assumption Spectral theorem for diagonalizable matrices with bounded inverse modal matrices.
Cite this review
Pith. "Pith review of Trace of Multi-variable Matrix Functions and its Application to Functions of Graph Spectrum." pith.science (2026). https://pith.science/paper/RZUE2COQ
@misc{pith2026250114515,
author = {Pith},
title = {Pith review of: Trace of Multi-variable Matrix Functions and its Application to Functions of Graph Spectrum},
year = {2026},
howpublished = {\url{https://pith.science/paper/RZUE2COQ}},
note = {Machine review of arXiv:2501.14515}
}
abstract
Matrix extension of a scalar function of a single variable is well-studied in literature. Of particular interest is the trace of such functions. It is known that for diagonalizable matrices, $M$, the function $g(M) = \text{Tr}(f(M)) = \sum_{j=1}^n f(\mu_j)$ (where $\{\mu_j\}_{j=1,2,\cdots,n}$ are the eigenvalues of $M$) inherits the monotonocity and convexity properties of $f$ (i.e., for $g$ to be convex, $f$ need not be operator convex -- convexity is sufficient). In this paper we formalize the idea of matrix extension of a function of multiple variables, study the monotonicity and convexity properties of the trace, and thus show that a function of form $g(M) = \sum_{j_1=1}^n \sum_{j_2=1}^n \cdots \sum_{j_m=1}^n f(\mu_{j_1}, \mu_{j_2},\cdots, \mu_{j_m})$ also inherits the monotonocity and convexity properties of the multi-variable function, $f$. We apply these results to functions of the spectrum of the weighted Laplacian matrix of undirected, simple graphs.
Reference graph
Works this paper leans on
-
[1]
Abbott, S. et al. (2001). Understanding analysis , volume 2. Springer
work page 2001
-
[2]
Bhatia, R. (2013). Matrix analysis , volume 169. Springer Science & Business Media
work page 2013
-
[3]
Carlen, E. (2010). Trace inequalities and quantum entropy: an introductory course. Entropy and the quantum , 529:73--140
work page 2010
-
[4]
Rahman, Q. (2002). Analytic theory of polynomials . Oxford University Press
work page 2002
-
[5]
Sahin, A., Kozachuk, N., Blum, R. S., and Bhattacharya, S. (2024). Resonance reduction against adversarial attacks in dynamic networks via eigenspectrum optimization. arXiv preprint arXiv:2410.00126
work page Pith review arXiv 2024
Reviewed August 10, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.