betaFeedback

WorksMathematics

Characterizations of minimal Hermitian completions via moments and subdifferentials

Abstract

Given a Hermitian matrix A0A_0 and a real subspace WW of Hermitian matrices, we study the matrices A0+W0A_0+W_0, W0∈WW_0\in W, of minimal spectral norm, which we call WW-minimal. The main motivation is the Hermitian matrix completion problem, where WW is spanned by the Hermitian matrix units corresponding to the unspecified entries. Our main tool is the moment of a subspace S⊆CnS\subseteq\mathbb{C}^n relative to WW, a compact convex subset of Rr\mathbb{R}^r, r=dim⁡(W)r=\dim(W), which we relate to the joint numerical range and to the kk-dimensional field of a family of matrices. We compute the subdifferential of ∥A0+W(x)∥\|A_0+W(x)\| in terms of the moments of the eigenspaces of the extreme eigenvalues of A0+W(x)A_0+W(x), and we obtain several equivalent characterizations of WW-minimality: in terms of eigenvectors, of moments, of subdifferentials and of norm one functionals that annihilate WW. When the identity belongs to WW, every WW-minimal matrix has balanced spectrum, and minimality is equivalent to the moments of the eigenspaces of the maximum and minimum eigenvalues having nonempty intersection; we show that this characterization fails if I∉WI\notin W. We illustrate the results with several examples, and we prove that the adjacency matrix of a connected graph is minimal for its completion problem if and only if the graph is complete bipartite.

Introduction

Let Mnh(C)M_n^h(\mathbb{C}) be the real space of n×nn\times n Hermitian matrices, endowed with the spectral norm ∥⋅∥\lVert\cdot\rVert, and let W⊆Mnh(C)\mathcal{W}\subseteq M_n^h(\mathbb{C}) be a real subspace. Given A0∈Mnh(C)A_0\in M_n^h(\mathbb{C}), we consider the best approximation problem

min⁡W∈W∥A0+W∥,\min_{W\in\mathcal{W}}\lVert A_0+W\rVert,

and we say that A0+W0A_0+W_0 is W\mathcal{W}-minimal if W0W_0 attains this minimum. Our main motivation is the Hermitian matrix completion problem: if some entries of a Hermitian matrix are specified and the remaining ones are free, then, setting the free entries of A0A_0 equal to zero, finding a completion of minimal spectral norm is exactly this problem for the subspace W\mathcal{W} spanned by the Hermitian matrix units supported on the free positions (see § 2). Since ∥Re⁡(X)∥≤∥X∥\lVert\operatorname{Re}(X)\rVert\leq\lVert X\rVert, the distance from A0A_0 to W\mathcal{W} coincides with the distance to its complex span, so this is also the minimal norm among arbitrary (not necessarily Hermitian) completions.

When W\mathcal{W} is the subspace of real diagonal matrices, minimal Hermitian matrices were characterized in [1] and their supports were studied in [15], where the moment of a subspace was introduced. These ideas were extended to diagonal compact operators in [4, 3, 5], related to the joint numerical range in [13, 5] and to the subdifferential of the maximum eigenvalue and semidefinite programming in [6]. In [7] the case where W\mathcal{W} is the Hermitian part of a unital C*-subalgebra of Mn(C)M_n(\mathbb{C}) was considered; see also [17, 21, 22]. Best approximation in the spectral norm by elements of a general subspace has recently received renewed attention [10, 11]. The classical tools for this problem are the characterization of best approximations by extremal functionals [19, 14] and Watson’s formula for the subdifferential of the spectral norm [20].

In this paper we consider an arbitrary real subspace W⊆Mnh(C)\mathcal{W} \subseteq M_n^h(\mathbb{C}). Neither the identity II of Mn(C)M_n(\mathbb{C}) nor any product structure is assumed, which is the natural setting for completion problems: the subspace spanned by the free positions of a pattern is almost never an algebra, and it contains the identity only when the whole diagonal is free. Fix an orthonormal basis {Bj}j=1r\{B^j\}_{j=1}^r of W\mathcal{W}. For a subspace S⊆CnS \subseteq\mathbb{C}^n, the moment of SS is the compact convex set

mS=conv⁡{(⟨B1v,v⟩,…,⟨Brv,v⟩):v∈S, ∥v∥=1}⊆Rr,m_S = \operatorname{conv}\left\{\left(\langle B^1v,v\rangle,\ldots,\langle B^rv,v\rangle\right): v \in S,\ \lVert v\rVert= 1\right\} \subseteq\mathbb{R}^r,

that is, the convex hull of the field of {Bj}j=1r\{B^j\}_{j=1}^r restricted to SS in the sense of [10].

The absence of the identity is not a technical detail. If I∈WI \in\mathcal{W}, every W\mathcal{W}-minimal matrix has balanced spectrum, λmax⁡=−λmin⁡\lambda_{\max}=-\lambda_{\min}, and minimality is detected by a witness that splits into two density matrices of equal weight, supported on the eigenspaces of λmax⁡\lambda_{\max} and λmin⁡\lambda_{\min} and with the same moment [7]. For a general subspace neither property holds: the spectrum of a minimal matrix need not be balanced, and even when it is, the two parts of a witness may carry different weights. The main contribution of this paper is a description of W\mathcal{W}-minimality that accounts for both phenomena, and whose central object is the subdifferential of the norm along W\mathcal{W}, expressed through moments. More precisely, the following results are new.

  • Subdifferential of the norm along W\mathcal{W} (Theorem 4.5 and Corollary 4.6). For every real subspace W\mathcal{W} and every x∈Rrx \in\mathbb{R}^r,

∂(∥A(⋅)∥)(x)=conv⁡(mS∥A(x)∥∪−mS−∥A(x)∥),A(x)=A0+∑j=1rxjBj,\partial(\lVert A(\cdot)\rVert)(x)=\operatorname{conv}\left(m_{S_{\lVert A(x)\rVert}}\cup-m_{S_{-\lVert A(x)\rVert}}\right), \qquad A(x)=A_0+\sum_{j=1}^r x_jB^j,

where S±∥A(x)∥S_{\pm\lVert A(x)\rVert} are the eigenspaces of A(x)A(x) for ±∥A(x)∥\pm\lVert A(x)\rVert (one of them may be trivial). This single formula covers the balanced and the non-balanced cases, and reduces the computation of the subdifferential of the spectral norm, restricted to an affine subspace of Hermitian matrices, to the geometry of two compact convex subsets of Rr\mathbb{R}^r.

  • Characterization of W\mathcal{W}-minimality without the identity (Theorem 5.9). A0+W0A_0+\mathcal{W}_0 is W\mathcal{W}-minimal if and only if 0∈mSMAX0 \in m_{S_{\mathrm{MAX}}} when the spectrum is not balanced, and if and only if 0∈conv⁡(mSλmax⁡∪−mSλmin⁡)0 \in\operatorname{conv}\left(m_{S_{\lambda_{\max}}}\cup-m_{S_{\lambda_{\min}}}\right) when it is. We also give the equivalent formulations in terms of finitely many eigenvectors (at most r+1r+1), of the subdifferential, and of functionals in the annihilator of W\mathcal{W}.

  • The intersection criterion fails without the identity (Remark 5.10). When I∈WI \in\mathcal{W} the condition above reduces to mSλmax⁡∩mSλmin⁡≠∅m_{S_{\lambda_{\max}}}\cap m_{S_{\lambda_{\min}}}\ne\varnothing, but we give minimal matrices with balanced spectrum, for subspaces not containing II, for which mSλmax⁡∩mSλmin⁡=∅m_{S_{\lambda_{\max}}}\cap m_{S_{\lambda_{\min}}}=\varnothing and mSλmax⁡∩−mSλmin⁡=∅m_{S_{\lambda_{\max}}}\cap-m_{S_{\lambda_{\min}}}=\varnothing. One of them comes from a completion pattern (Example 6.1).

  • Moments and fields (Proposition 3.8 and Corollary 3.9). We relate the moments to the kk-dimensional field of [10]. For A=A0+W0A=A_0+\mathcal{W}_0, the field of {ABj}j=1r\{AB^j\}_{j=1}^r on the subspace where AA attains its norm is computed in terms of the fields of {Bj}j=1r\{B^j\}_{j=1}^r on the eigenspaces of λmax⁡\lambda_{\max} and λmin⁡\lambda_{\min}.

  • Adjacency matrices of graphs (Proposition 6.4). For the completion problem in which the edges of a connected graph are specified and the non-edges and the diagonal are free, the adjacency matrix is minimal if and only if the graph is complete bipartite.

Some of the remaining results extend known ones to the present setting, and we include them, with proofs when these are short, to keep the paper self-contained. The description of the moment through density matrices and the joint numerical range (Proposition 3.4) follows [7]. The characterization by a single witness (Theorem 5.1) is [21 Theorem 2.6] and when I∈WI \in\mathcal{W}, the equivalence between minimality and the intersection of the moments (Theorem 5.2) was proved in [7] for C*-subalgebras.

Moreover, the support conditions of the witness come from [1]. We observe that these proofs use WW only through its annihilator and through I∈WI \in W, so they hold for arbitrary subspaces containing the identity. We also describe the extremal functionals on Mnh(C)M_n^h(\mathbb{C}) (Proposition 5.7), a fact that is probably known but for which we could not find a reference.

The paper is organized as follows. In Section 2 we set up the completion problem, the annihilator of WW, the duality formula and the subspace SMAXS_{\mathrm{MAX}} where a Hermitian matrix attains its norm. Section 3 is devoted to the moment of a subspace, the joint numerical range and the field. In Section 4 we compute the subdifferentials, and in Section 5 we prove the characterizations of WW-minimality. Section 6 contains the examples and the application to adjacency matrices of graphs.

Hermitian matrix completion and W-minimal matrices

Let A0∈Mnh(C)A_0 \in M_n^h(\mathbb{C}) and I0⊂{(i,j):1≤i,j≤n}I_0 \subset\{(i,j): 1 \le i,j \le n\} with the following property

(A0)ij=0,if (i,j)∈I0.(A_0)_{ij}=0,\quad\text{if }(i,j)\in I_0.

Consider the canonical basis {ei}i=1n\{e_i\}_{i=1}^n of Cn\mathbb{C}^n and define the following n×nn \times n Hermitian matrices

Bj,j=ejej∗,for (j,j)∈I0,Bi,j=12(eiej∗+ejei∗),for (i,j)∈I0 with i<j,Bi,j=i2(eiej∗−ejei∗),for (i,j)∈I0 with j<i,(1)\begin{aligned} B^{j,j} &= e_j e_j^*, &&\text{for }(j,j)\in I_0,\\ B^{i,j} &= \frac{1}{\sqrt{2}}\left(e_i e_j^*+e_j e_i^*\right), &&\text{for }(i,j)\in I_0\text{ with }i<j,\\ B^{i,j} &= \frac{\mathrm{i}}{\sqrt{2}}\left(e_i e_j^*-e_j e_i^*\right), &&\text{for }(i,j)\in I_0\text{ with }j<i, \tag*{(1)} \end{aligned}

and the orthonormal set

B={Bi,j:(i,j)∈I0}.(2)\mathbb{B}=\{B^{i,j}:(i,j)\in I_0\}. \tag*{(2)}

Consider W=span⁡RBW=\operatorname{span}_{\mathbb{R}}\mathbb{B}, which is a real subspace of Mnh(C)M_n^h(\mathbb{C}) with dim⁡(W)=#I0=r\dim(W)=\# I_0=r. Consider an order fixed in I0I_0, σ:I0→{1,2,…,r}\sigma:I_0\to\{1,2,\ldots,r\} such that Bi,j=Bσ(i,j)=BkB^{i,j}=B^{\sigma(i,j)}=B^k, 1≤k≤r1\le k\le r. We consider the affine matrix

A(x)=A0+∑k=1rxkBk=A0+W(x),(3)A(x)=A_0+\sum_{k=1}^{r}x_kB^k=A_0+W(x), \tag*{(3)}

with x=(x1,x2,…,xr)∈Rrx=(x_1,x_2,\ldots,x_r)\in\mathbb{R}^r. We shall also use the linear map ΦB:Mnh(C)→Rr\Phi^{\mathbb{B}}:M_n^h(\mathbb{C})\to\mathbb{R}^r defined by

ΦB(R)=(tr⁡(RB1),tr⁡(RB2),…,tr⁡(RBr)),for every R∈Mnh(C),(4)\Phi^{\mathbb{B}}(R)=\left(\operatorname{tr}(RB^1),\operatorname{tr}(RB^2),\ldots,\operatorname{tr}(RB^r)\right),\quad\text{for every }R\in M_n^h(\mathbb{C}), \tag*{(4)}

References

  1. [1]Andruchow, Esteban; Larotonda, Gabriel; Recht, Lázaro; Varela, Alejandro. A characterization of minimal Hermitian matrices. Linear Algebra Appl. 436 (2012), 2366–2374.DOI
  2. [2]Bhatia, Rajendra. Matrix analysis. Graduate Texts in Mathematics, vol. 169, Springer-Verlag, New York, 1997.
  3. [3]Bottazzi, Tamara. Best approximation by diagonal operators in Schatten ideals. Linear Algebra Appl. 620 (2021), 1–26.DOI
  4. [4]Bottazzi, Tamara; Varela, Alejandro. Best approximation by diagonal compact operators. Linear Algebra Appl. 439 (2013), no. 10, 3044–3056.arxiv.org/abs/1303.0739
  5. [5]Bottazzi, Tamara; Varela, Alejandro. Minimal self-adjoint compact operators, moment of a subspace and joint numerical range. J. Math. Anal. Appl. 528 (2023), no. 2, Paper No. 127552, 22 pp.arxiv.org/abs/2303.04128
  6. [6]Bottazzi, Tamara; Varela, Alejandro. Minimal compact operators, subdifferential of the maximum eigenvalue and semi-definite programming. Linear Algebra Appl. 716 (2025), 1–31.arxiv.org/abs/2405.08923
  7. [7]Bottazzi, Tamara; Varela, Alejandro. Best approximants relative to a C*-subalgebra, joint numerical range and subdifferentials. Preprint (2026), arXiv:2604.16041.arxiv.org/abs/2604.16041
  8. [8]Brouwer, Andries E.; Haemers, Willem H. Spectra of graphs. Universitext, Springer, New York, 2012.
  9. [9]Clarke, Frank H. Optimization and nonsmooth analysis. Second edition, Classics in Applied Mathematics, vol. 5, SIAM, Philadelphia, PA, 1990.
  10. [10]Faber, Vance; Liesen, Jörg; Tichý, Petr. Matrix best approximation in the spectral norm. Linear Algebra Appl. 733 (2026), 178–204.
  11. [11]Grover, Priyanka; Gupta, Krishna Kumar. Properties of best approximations with respect to the Ky Fan p-k norm, and the strict spectral approximant of a matrix. Linear Algebra Appl. 743 (2026), 102–122.
  12. [12]Horn, Roger A.; Johnson, Charles R. Matrix analysis. Second edition, Cambridge University Press, Cambridge, 2013.
  13. [13]Klobouk, Abel H.; Varela, Alejandro. Moment of a subspace and joint numerical range. Linear Multilinear Algebra 71 (2023), no. 9, 1470–1503.arxiv.org/abs/2110.10584
  14. [14]Lau, K. K.; Riha, W. O. J. Characterization of best approximations in normed linear spaces of matrices by elements of finite-dimensional linear subspaces. Linear Algebra Appl. 35 (1981), 109–120.DOI
  15. [15]Mendoza, Alberto; Recht, Lázaro; Varela, Alejandro. Supports for minimal Hermitian matrices. Linear Algebra Appl. 584 (2020), 458–482.arxiv.org/abs/1906.06417
  16. [16]Overton, Michael L. Large-scale optimization of eigenvalues. SIAM J. Optim. 2 (1992), no. 1, 88–120.DOI
  17. [17]Rieffel, Marc A. Leibniz seminorms and best approximation from C*-subalgebras. Sci. China Math. 54 (2011), 2259–2274.DOI
  18. [18]Rockafellar, R. Tyrrell. Convex analysis. Princeton Mathematical Series, vol. 28, Princeton University Press, Princeton, NJ, 1970.
  19. [19]Singer, Ivan. Best approximation in normed linear spaces by elements of linear subspaces. Springer-Verlag, New York-Berlin, 1970.
  20. [20]Watson, G. A. Characterization of the subdifferential of some matrix norms. Linear Algebra Appl. 170 (1992), 33–45.DOI
  21. [21]Zhang, Ying; Jiang, Lining; Han, Yongheng. Constructions of minimal Hermitian matrices related to a C*-subalgebra of M_n(C). Proc. Amer. Math. Soc. 151 (2023), no. 1, 73–84.DOI
  22. [22]Zhang, Ying; Jiang, Lining. Minimal Hermitian compact operators related to a C*-subalgebra of K(H). J. Math. Anal. Appl. 506 (2022), no. 2, Paper No. 125649.DOI
  23. [23]Ziętak, Krystyna. On approximation problems with zero-trace matrices. Linear Algebra Appl. 247 (1996), 169–183.DOI

Paper details

vlnc:W_4ku5y4dv
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 1

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents