Introduction
Let M n h ( C ) M_n^h(\mathbb{C}) M n h ( C ) be the real space of n × n n\times n n × n Hermitian matrices, endowed with the spectral norm ∥ ⋅ ∥ \lVert\cdot\rVert ∥ ⋅ ∥ , and let W ⊆ M n h ( C ) \mathcal{W}\subseteq M_n^h(\mathbb{C}) W ⊆ M n h ( C ) be a real subspace. Given A 0 ∈ M n h ( C ) A_0\in M_n^h(\mathbb{C}) A 0 ∈ M n h ( C ) , we consider the best approximation problem
min W ∈ W ∥ A 0 + W ∥ , \min_{W\in\mathcal{W}}\lVert A_0+W\rVert, W ∈ W min ∥ A 0 + W ∥ ,
and we say that A 0 + W 0 A_0+W_0 A 0 + W 0 is W \mathcal{W} W -minimal if W 0 W_0 W 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 A 0 A_0 A 0 equal to zero, finding a completion of minimal spectral norm is exactly this problem for the subspace W \mathcal{W} 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 ∥ Re ( X )∥ ≤ ∥ X ∥ , the distance from A 0 A_0 A 0 to W \mathcal{W} 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} 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} W is the Hermitian part of a unital C*-subalgebra of M n ( C ) M_n(\mathbb{C}) M n ( 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 ⊆ M n h ( C ) \mathcal{W} \subseteq M_n^h(\mathbb{C}) W ⊆ M n h ( C ) . Neither the identity I I I of M n ( C ) M_n(\mathbb{C}) M n ( 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 { B j } j = 1 r \{B^j\}_{j=1}^r { B j } j = 1 r of W \mathcal{W} W . For a subspace S ⊆ C n S \subseteq\mathbb{C}^n S ⊆ C n , the moment of S S S is the compact convex set
m S = conv { ( ⟨ B 1 v , v ⟩ , … , ⟨ B r v , v ⟩ ) : v ∈ S , ∥ v ∥ = 1 } ⊆ R r , 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, m S = conv { ( ⟨ B 1 v , v ⟩ , … , ⟨ B r v , v ⟩ ) : v ∈ S , ∥ v ∥ = 1 } ⊆ R r ,
that is, the convex hull of the field of { B j } j = 1 r \{B^j\}_{j=1}^r { B j } j = 1 r restricted to S S S in the sense of [10 ] .
The absence of the identity is not a technical detail. If I ∈ W I \in\mathcal{W} I ∈ W , every W \mathcal{W} W -minimal matrix has balanced spectrum, λ max = − λ min \lambda_{\max}=-\lambda_{\min} λ m a x = − λ m i n , and minimality is detected by a witness that splits into two density matrices of equal weight, supported on the eigenspaces of λ max \lambda_{\max} λ m a x and λ min \lambda_{\min} λ m i n 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} W -minimality that accounts for both phenomena, and whose central object is the subdifferential of the norm along W \mathcal{W} W , expressed through moments. More precisely, the following results are new.
∂ ( ∥ A ( ⋅ ) ∥ ) ( x ) = conv ( m S ∥ A ( x ) ∥ ∪ − m S − ∥ A ( x ) ∥ ) , A ( x ) = A 0 + ∑ j = 1 r x j B j , \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, ∂ (∥ A ( ⋅ )∥) ( x ) = conv ( m S ∥ A ( x )∥ ∪ − m S − ∥ A ( x )∥ ) , A ( x ) = A 0 + j = 1 ∑ r x j B j ,
where S ± ∥ A ( x ) ∥ S_{\pm\lVert A(x)\rVert} S ± ∥ A ( x )∥ are the eigenspaces of A ( x ) A(x) A ( x ) for ± ∥ A ( x ) ∥ \pm\lVert A(x)\rVert ± ∥ A ( x )∥ (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 R r \mathbb{R}^r R r .
Characterization of W \mathcal{W} W -minimality without the identity (Theorem 5.9). A 0 + W 0 A_0+\mathcal{W}_0 A 0 + W 0 is W \mathcal{W} W -minimal if and only if 0 ∈ m S M A X 0 \in m_{S_{\mathrm{MAX}}} 0 ∈ m S MAX when the spectrum is not balanced, and if and only if 0 ∈ conv ( m S λ max ∪ − m S λ min ) 0 \in\operatorname{conv}\left(m_{S_{\lambda_{\max}}}\cup-m_{S_{\lambda_{\min}}}\right) 0 ∈ conv ( m S λ m a x ∪ − m S λ m i n ) when it is. We also give the equivalent formulations in terms of finitely many eigenvectors (at most r + 1 r+1 r + 1 ), of the subdifferential, and of functionals in the annihilator of W \mathcal{W} W .
The intersection criterion fails without the identity (Remark 5.10). When I ∈ W I \in\mathcal{W} I ∈ W the condition above reduces to m S λ max ∩ m S λ min ≠ ∅ m_{S_{\lambda_{\max}}}\cap m_{S_{\lambda_{\min}}}\ne\varnothing m S λ m a x ∩ m S λ m i n = ∅ , but we give minimal matrices with balanced spectrum, for subspaces not containing I I I , for which m S λ max ∩ m S λ min = ∅ m_{S_{\lambda_{\max}}}\cap m_{S_{\lambda_{\min}}}=\varnothing m S λ m a x ∩ m S λ m i n = ∅ and m S λ max ∩ − m S λ min = ∅ m_{S_{\lambda_{\max}}}\cap-m_{S_{\lambda_{\min}}}=\varnothing m S λ m a x ∩ − m S λ m i n = ∅ . 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 k k k -dimensional field of [10 ] . For A = A 0 + W 0 A=A_0+\mathcal{W}_0 A = A 0 + W 0 , the field of { A B j } j = 1 r \{AB^j\}_{j=1}^r { A B j } j = 1 r on the subspace where A A A attains its norm is computed in terms of the fields of { B j } j = 1 r \{B^j\}_{j=1}^r { B j } j = 1 r on the eigenspaces of λ max \lambda_{\max} λ m a x and λ min \lambda_{\min} λ m i n .
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 ∈ W I \in\mathcal{W} I ∈ 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 W W W only through its annihilator and through I ∈ W I \in W I ∈ W , so they hold for arbitrary subspaces containing the identity. We also describe the extremal functionals on M n h ( C ) M_n^h(\mathbb{C}) M n h ( 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 W W W , the duality formula and the subspace S M A X S_{\mathrm{MAX}} S 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 W W W -minimality. Section 6 contains the examples and the application to adjacency matrices of graphs.
Hermitian matrix completion and W-minimal matrices
Let A 0 ∈ M n h ( C ) A_0 \in M_n^h(\mathbb{C}) A 0 ∈ M n h ( C ) and I 0 ⊂ { ( i , j ) : 1 ≤ i , j ≤ n } I_0 \subset\{(i,j): 1 \le i,j \le n\} I 0 ⊂ {( i , j ) : 1 ≤ i , j ≤ n } with the following property
( A 0 ) i j = 0 , if ( i , j ) ∈ I 0 . (A_0)_{ij}=0,\quad\text{if }(i,j)\in I_0. ( A 0 ) ij = 0 , if ( i , j ) ∈ I 0 .
Consider the canonical basis { e i } i = 1 n \{e_i\}_{i=1}^n { e i } i = 1 n of C n \mathbb{C}^n C n and define the following n × n n \times n n × n Hermitian matrices
B j , j = e j e j ∗ , for ( j , j ) ∈ I 0 , B i , j = 1 2 ( e i e j ∗ + e j e i ∗ ) , for ( i , j ) ∈ I 0 with i < j , B i , j = i 2 ( e i e j ∗ − e j e i ∗ ) , for ( i , j ) ∈ I 0 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} B j , j B i , j B i , j = e j e j ∗ , = 2 1 ( e i e j ∗ + e j e i ∗ ) , = 2 i ( e i e j ∗ − e j e i ∗ ) , for ( j , j ) ∈ I 0 , for ( i , j ) ∈ I 0 with i < j , for ( i , j ) ∈ I 0 with j < i , (1)
and the orthonormal set
B = { B i , j : ( i , j ) ∈ I 0 } . (2) \mathbb{B}=\{B^{i,j}:(i,j)\in I_0\}. \tag*{(2)} B = { B i , j : ( i , j ) ∈ I 0 } . (2)
Consider W = span R B W=\operatorname{span}_{\mathbb{R}}\mathbb{B} W = span R B , which is a real subspace of M n h ( C ) M_n^h(\mathbb{C}) M n h ( C ) with dim ( W ) = # I 0 = r \dim(W)=\# I_0=r dim ( W ) = # I 0 = r . Consider an order fixed in I 0 I_0 I 0 , σ : I 0 → { 1 , 2 , … , r } \sigma:I_0\to\{1,2,\ldots,r\} σ : I 0 → { 1 , 2 , … , r } such that B i , j = B σ ( i , j ) = B k B^{i,j}=B^{\sigma(i,j)}=B^k B i , j = B σ ( i , j ) = B k , 1 ≤ k ≤ r 1\le k\le r 1 ≤ k ≤ r . We consider the affine matrix
A ( x ) = A 0 + ∑ k = 1 r x k B k = A 0 + W ( x ) , (3) A(x)=A_0+\sum_{k=1}^{r}x_kB^k=A_0+W(x), \tag*{(3)} A ( x ) = A 0 + k = 1 ∑ r x k B k = A 0 + W ( x ) , (3)
with x = ( x 1 , x 2 , … , x r ) ∈ R r x=(x_1,x_2,\ldots,x_r)\in\mathbb{R}^r x = ( x 1 , x 2 , … , x r ) ∈ R r . We shall also use the linear map Φ B : M n h ( C ) → R r \Phi^{\mathbb{B}}:M_n^h(\mathbb{C})\to\mathbb{R}^r Φ B : M n h ( C ) → R r defined by
Φ B ( R ) = ( tr ( R B 1 ) , tr ( R B 2 ) , … , tr ( R B r ) ) , for every R ∈ M n h ( 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)} Φ B ( R ) = ( tr ( R B 1 ) , tr ( R B 2 ) , … , tr ( R B r ) ) , for every R ∈ M n h ( C ) , (4)
which we will later restrict to the set D S D_S D S of density matrices (see (19) ). In this context, we state the following approximation problem:
min x ∈ R r ∥ A ( x ) ∥ , (5) \min_{x\in\mathbb{R}^r}\|A(x)\|, \tag*{(5)} x ∈ R r min ∥ A ( x ) ∥ , (5)
where ∥ ⋅ ∥ \|\cdot\| ∥ ⋅ ∥ is the spectral norm. Note that (5) is equivalent to finding W 0 = W ( x 0 ) ∈ W W_0=W(x_0)\in W W 0 = W ( x 0 ) ∈ W such that
dist ( A 0 , W ) = ∥ A 0 + W 0 ∥ ≤ ∥ A 0 + W ∥ , for every W ∈ W . \operatorname{dist}(A_0,W)=\|A_0+W_0\|\le\|A_0+W\|,\quad\text{for every }W\in W. dist ( A 0 , W ) = ∥ A 0 + W 0 ∥ ≤ ∥ A 0 + W ∥ , for every W ∈ W .
In this case, we say that A ( x 0 ) = A 0 + W ( x 0 ) A(x_0)=A_0+W(x_0) A ( x 0 ) = A 0 + W ( x 0 ) is a W W W -minimal matrix and − W ( x 0 ) -W(x_0) − W ( x 0 ) is a best approximation of A 0 A_0 A 0 in the subspace W W W . Observe that for every z ∈ C r z\in\mathbb{C}^r z ∈ C r , Re ( A ( z ) ) = A 0 + Re ( W ( z ) ) \operatorname{Re}(A(z))=A_0+\operatorname{Re}(W(z)) Re ( A ( z )) = A 0 + Re ( W ( z )) and
∥ A 0 + Re ( W ( z ) ) ∥ = ∥ Re ( A ( z ) ) ∥ ≤ ∥ A ( z ) ∥ = ∥ A 0 + W ( z ) ∥ . \|A_0+\operatorname{Re}(W(z))\|=\|\operatorname{Re}(A(z))\|\le\|A(z)\|=\|A_0+W(z)\|. ∥ A 0 + Re ( W ( z )) ∥ = ∥ Re ( A ( z )) ∥ ≤ ∥ A ( z ) ∥ = ∥ A 0 + W ( z ) ∥.
This implies that
min x ∈ R r ∥ A ( x ) ∥ ≤ min z ∈ C r ∥ A ( z ) ∥ . \min_{x\in\mathbb{R}^r}\|A(x)\|\le\min_{z\in\mathbb{C}^r}\|A(z)\|. x ∈ R r min ∥ A ( x ) ∥ ≤ z ∈ C r min ∥ A ( z ) ∥.
Since R r ⊂ C r \mathbb{R}^r\subset\mathbb{C}^r R r ⊂ C r , the reverse inequality is trivial, and therefore dist ( A 0 , W ) = dist ( A 0 , span C B ) \operatorname{dist}(A_0,W)=\operatorname{dist}(A_0,\operatorname{span}_{\mathbb{C}}\mathbb{B}) dist ( A 0 , W ) = dist ( A 0 , span C B ) .
Remark 2.1. Let A 0 ∈ M n h ( C ) A_0 \in M_n^h(\mathbb{C}) A 0 ∈ M n h ( C ) and let I 0 I_0 I 0 as defined before. If ( i , j ) ∈ I 0 (i,j) \in I_0 ( i , j ) ∈ I 0 , then ( j , i ) ∈ I 0 (j,i) \in I_0 ( j , i ) ∈ I 0 , since ( A 0 ) j i = ( A 0 ) i j ‾ (A_0)_{ji} = \overline{(A_0)_{ij}} ( A 0 ) j i = ( A 0 ) ij for every i , j i,j i , j .
Define the norm 1 elements of the annihilator of W \mathcal{W} W in the dual space of M n h ( C ) M_n^h(\mathbb{C}) M n h ( C ) as
W 1 ∘ = { X ∈ M n h ( C ) : tr ( X W ) = 0 , ∀ W ∈ W ∧ ∥ X ∥ 1 = 1 } , (6) \mathcal{W}_1^\circ= \{X \in M_n^h(\mathbb{C}) : \operatorname{tr}(XW) = 0,\ \forall W \in\mathcal{W} \land\lVert X\rVert_1 = 1\}, \tag*{(6)} W 1 ∘ = { X ∈ M n h ( C ) : tr ( X W ) = 0 , ∀ W ∈ W ∧ ∥ X ∥ 1 = 1 } , (6)
where ∥ X ∥ 1 = tr ( ( X ∗ X ) 1 / 2 ) \lVert X\rVert_1 = \operatorname{tr}\left((X^*X)^{1/2}\right) ∥ X ∥ 1 = tr ( ( X ∗ X ) 1/2 ) is the trace norm. Thus W 1 ∘ \mathcal{W}_1^\circ W 1 ∘ is the set of elements of trace norm one of the annihilator W ∘ = { X ∈ M n h ( C ) : tr ( X W ) = 0 , ∀ W ∈ W } \mathcal{W}^\circ= \{X \in M_n^h(\mathbb{C}) : \operatorname{tr}(XW) = 0,\ \forall W \in\mathcal{W}\} W ∘ = { X ∈ M n h ( C ) : tr ( X W ) = 0 , ∀ W ∈ W } . It is evident that X ∈ W ∘ X \in\mathcal{W}^\circ X ∈ W ∘ if and only if tr ( X B i , j ) = 0 \operatorname{tr}(XB^{i,j}) = 0 tr ( X B i , j ) = 0 for every ( i , j ) ∈ I 0 (i,j) \in I_0 ( i , j ) ∈ I 0 . This implies in particular that ker ( Φ B ) = W ∘ \ker(\Phi^{B}) = \mathcal{W}^\circ ker ( Φ B ) = W ∘ and W 1 ∘ = { X ∈ ker ( Φ B ) : ∥ X ∥ 1 = 1 } \mathcal{W}_1^\circ= \{X \in\ker(\Phi^{B}) : \lVert X\rVert_1 = 1\} W 1 ∘ = { X ∈ ker ( Φ B ) : ∥ X ∥ 1 = 1 } , for the map Φ B \Phi^{B} Φ B defined in (2.4). Also, we can deduce that dim R W ∘ = n 2 − dim R W = n 2 − # I 0 = n 2 − r \dim_{\mathbb{R}} \mathcal{W}^\circ= n^2 - \dim_{\mathbb{R}} \mathcal{W} = n^2 - \# I_0 = n^2 - r dim R W ∘ = n 2 − dim R W = n 2 − # I 0 = n 2 − r .
Proposition 2.2. Let A 0 ∈ M n h ( C ) A_0 \in M_n^h(\mathbb{C}) A 0 ∈ M n h ( C ) and let I 0 = { ( i , j ) : ( A 0 ) i j = 0 } I_0 = \{(i,j) : (A_0)_{ij} = 0\} I 0 = {( i , j ) : ( A 0 ) ij = 0 } . For any Y ∈ M n h ( C ) Y \in M_n^h(\mathbb{C}) Y ∈ M n h ( C ) , the following conditions are equivalent:
Y ∈ W 1 ∘ Y \in\mathcal{W}_1^\circ Y ∈ W 1 ∘ .
∥ Y ∥ 1 = 1 \lVert Y\rVert_1 = 1 ∥ Y ∥ 1 = 1 and Y i j = 0 Y_{ij} = 0 Y ij = 0 , for all ( i , j ) ∈ I 0 (i,j) \in I_0 ( i , j ) ∈ I 0 .
Proof. Consider the orthonormal set B = { B i , j : ( i , j ) ∈ I 0 } \mathbb{B} = \{B^{i,j} : (i,j) \in I_0\} B = { B i , j : ( i , j ) ∈ I 0 } with B i , j B^{i,j} B i , j as in (2.1). By Remark 2.1, B i , j ∈ B ⟺ B j , i ∈ B B^{i,j} \in\mathbb{B} \Longleftrightarrow B^{j,i} \in\mathbb{B} B i , j ∈ B ⟺ B j , i ∈ B , so it is sufficient to consider the cases i < j i < j i < j and i = j i = j i = j . Direct calculations show that
tr ( Y B i , j ) = 2 2 Re ( Y i j ) and tr ( Y B j , i ) = − 2 2 Im ( Y i j ) , (7) \operatorname{tr}(YB^{i,j}) = \frac{2}{\sqrt{2}}\operatorname{Re}(Y_{ij}) \quad\text{and} \quad\operatorname{tr}(YB^{j,i}) = -\frac{2}{\sqrt{2}}\operatorname{Im}(Y_{ij}), \tag*{(7)} tr ( Y B i , j ) = 2 2 Re ( Y ij ) and tr ( Y B j , i ) = − 2 2 Im ( Y ij ) , (7)
for i < j i < j i < j , and
tr ( Y B i , i ) = tr ( Y e i e i ∗ ) = ⟨ Y e i , e i ⟩ = Y i i , (8) \operatorname{tr}(YB^{i,i}) = \operatorname{tr}(Ye_i e_i^*) = \langle Ye_i,e_i\rangle= Y_{ii}, \tag*{(8)} tr ( Y B i , i ) = tr ( Y e i e i ∗ ) = ⟨ Y e i , e i ⟩ = Y ii , (8)
for i = j i = j i = j .
(2) ⇒ \Rightarrow ⇒ (1) Since Y i j = 0 Y_{ij} = 0 Y ij = 0 for all ( i , j ) ∈ I 0 (i,j) \in I_0 ( i , j ) ∈ I 0 , it follows from (2.7) and (2.8) that tr ( Y B i , j ) = 0 \operatorname{tr}(YB^{i,j}) = 0 tr ( Y B i , j ) = 0 for every ( i , j ) ∈ I 0 (i,j) \in I_0 ( i , j ) ∈ I 0 , and hence Y ∈ W 1 ∘ Y \in\mathcal{W}_1^\circ Y ∈ W 1 ∘ .
(1) ⇒ \Rightarrow ⇒ (2) Let Y ∈ W 1 ∘ Y \in\mathcal{W}_1^\circ Y ∈ W 1 ∘ . Then
tr ( Y B i , j ) = 0 , for every ( i , j ) ∈ I 0 . \operatorname{tr}(YB^{i,j}) = 0,\quad\text{for every }(i,j) \in I_0. tr ( Y B i , j ) = 0 , for every ( i , j ) ∈ I 0 .
By (2.7) and (2.8), it follows that for every ( i , j ) ∈ I 0 (i,j) \in I_0 ( i , j ) ∈ I 0 :
{ Y i j = Y j i ‾ = 0 , for i < j , Y i i = 0 for i = j \begin{cases}
Y_{ij} = \overline{Y_{ji}} = 0, & \text{for } i < j,\\
Y_{ii} = 0 & \text{for } i = j
\end{cases} { Y ij = Y j i = 0 , Y ii = 0 for i < j , for i = j
Thus, ( i , j ) ∈ I 0 (i,j) \in I_0 ( i , j ) ∈ I 0 implies Y i j = 0 Y_{ij} = 0 Y ij = 0 for all such indices. □ \square □
Remark 2.3 (About the best approximations of real symmetric matrices). Let X ∈ M n ( C ) X \in M_n(\mathbb{C}) X ∈ M n ( C ) , then ℜ ( X ) = X + X ‾ 2 \Re(X) = \dfrac{X+\overline{X}}{2} ℜ ( X ) = 2 X + X is a real matrix and ∥ ℜ ( X ) ∥ ≤ ∥ X ∥ \lVert\Re(X)\rVert\leq\lVert X\rVert ∥ ℜ ( X )∥ ≤ ∥ X ∥ . Suppose that X = A 0 + W 0 X = A_0 + W_0 X = A 0 + W 0 is a W \mathcal{W} W -minimal matrix with ( A 0 ) i j ∈ R (A_0)_{ij} \in\mathbb{R} ( A 0 ) ij ∈ R for all 1 ≤ i , j ≤ n 1 \leq i,j \leq n 1 ≤ i , j ≤ n . Then,
∥ A 0 + ℜ ( W 0 ) ∥ = ∥ ℜ ( A 0 + W 0 ) ∥ ≤ ∥ A 0 + W 0 ∥ . \lVert A_0 + \Re(W_0)\rVert= \lVert\Re(A_0 + W_0)\rVert\leq\lVert A_0 + W_0\rVert. ∥ A 0 + ℜ ( W 0 )∥ = ∥ ℜ ( A 0 + W 0 )∥ ≤ ∥ A 0 + W 0 ∥ .
Therefore, for real symmetric matrices A 0 A_0 A 0 we can consider real symmetric best approximants in W \mathcal{W} W . This reduces the size of the basis B \mathbb{B} B , since it contains only the elements ( B k , j ) (B_{k,j}) ( B k , j ) with ( k ≤ j ) (k \leq j) ( k ≤ j ) , as defined in (2.1). In this case, the annihilator of W \mathcal{W} W in the dual space of M n h ( R ) M_n^h(\mathbb{R}) M n h ( R ) is
W 1 ∘ = { X ∈ M n sim ( R ) : tr ( X W ) = 0 , ∀ W ∈ W ∧ ∥ X ∥ 1 = 1 } . \mathcal{W}_1^\circ= \{X \in M_n^{\operatorname{sim}}(\mathbb{R}) : \operatorname{tr}(XW) = 0,\ \forall W \in\mathcal{W} \land\lVert X\rVert_1 = 1\}. W 1 ∘ = { X ∈ M n sim ( R ) : tr ( X W ) = 0 , ∀ W ∈ W ∧ ∥ X ∥ 1 = 1 } .
We state the well-known Banach duality formula relating the distance from a matrix to a subspace with the norms of functionals in the dual space.
Proposition 2.4 (Banach Duality Formula). Let A 0 ∈ M n h ( C ) A_0 \in M_n^h(\mathbb{C}) A 0 ∈ M n h ( C ) . Then,
max X ∈ W 1 ∘ ∣ tr ( X A 0 ) ∣ = min x ∈ R r ∥ A ( x ) ∥ . (9) \max_{X \in\mathcal{W}_1^\circ} \lvert\operatorname{tr}(XA_0)\rvert= \min_{x \in\mathbb{R}^r} \lVert A(x)\rVert. \tag*{(9)} X ∈ W 1 ∘ max ∣ tr ( X A 0 )∣ = x ∈ R r min ∥ A ( x )∥ . (9)
Since A 0 A_0 A 0 vanishes on I 0 I_0 I 0 , Proposition 2.2 gives A 0 ∥ A 0 ∥ 1 ∈ W 1 ∘ \frac{A_0}{\lVert A_0\rVert_1}\in\mathcal{W}_1^\circ ∥ A 0 ∥ 1 A 0 ∈ W 1 ∘ whenever A 0 ≠ 0 A_0\ne0 A 0 = 0 , and the duality formula (9) immediately yields the following lower bound for the norm of a W \mathcal{W} W -minimal matrix.
Corollary 2.5. For every nonzero A 0 ∈ M n h ( C ) A_0\in M_n^h(\mathbb{C}) A 0 ∈ M n h ( C ) ,
inf W ∈ W ∥ A 0 + W ∥ ≥ ∥ A 0 ∥ F 2 ∥ A 0 ∥ 1 , \inf_{W\in\mathcal{W}}\lVert A_0+W\rVert\geq\frac{\lVert A_0\rVert_F^2}{\lVert A_0\rVert_1}, W ∈ W inf ∥ A 0 + W ∥ ≥ ∥ A 0 ∥ 1 ∥ A 0 ∥ F 2 , where ∥ ⋅ ∥ F \lVert\cdot\rVert_F ∥ ⋅ ∥ F denotes the Frobenius norm, that is ∥ X ∥ F = tr ( X ∗ X ) \lVert X\rVert_F=\sqrt{\operatorname{tr}(X^*X)} ∥ X ∥ F = tr ( X ∗ X ) , for X ∈ M n ( C ) X\in M_n(\mathbb{C}) X ∈ M n ( C ) .
Proof. By (9) ,
inf W ∈ W ∥ A 0 + W ∥ = max X ∈ W 1 ∘ ∣ tr ( X A 0 ) ∣ ≥ ∣ tr ( A 0 ∥ A 0 ∥ 1 A 0 ) ∣ = tr ( A 0 2 ) ∥ A 0 ∥ 1 = ∥ A 0 ∥ F 2 ∥ A 0 ∥ 1 . \inf_{W\in\mathcal{W}}\lVert A_0+W\rVert=\max_{X\in\mathcal{W}_1^\circ}\lvert\operatorname{tr}(XA_0)\rvert\geq\left\lvert\operatorname{tr}\left(\frac{A_0}{\lVert A_0\rVert_1}A_0\right)\right\rvert=\frac{\operatorname{tr}(A_0^2)}{\lVert A_0\rVert_1}=\frac{\lVert A_0\rVert_F^2}{\lVert A_0\rVert_1}. W ∈ W inf ∥ A 0 + W ∥ = X ∈ W 1 ∘ max ∣ tr ( X A 0 )∣ ≥ tr ( ∥ A 0 ∥ 1 A 0 A 0 ) = ∥ A 0 ∥ 1 tr ( A 0 2 ) = ∥ A 0 ∥ 1 ∥ A 0 ∥ F 2 .
We now introduce a subspace that plays a central role, and establish some of its properties.
Proposition 2.6 (Properties of S M A X S_{\mathrm{MAX}} S MAX for Hermitian matrices). Let A ∈ M n h ( C ) A\in M_n^h(\mathbb{C}) A ∈ M n h ( C ) , A ≠ 0 A\ne0 A = 0 , and for λ ∈ R \lambda\in\mathbb{R} λ ∈ R let S λ = { v ∈ C n : A v = λ v } S_\lambda=\{v\in\mathbb{C}^n:Av=\lambda v\} S λ = { v ∈ C n : A v = λ v } (so S λ = { 0 } S_\lambda=\{0\} S λ = { 0 } if λ ∉ σ ( A ) \lambda\notin\sigma(A) λ ∈ / σ ( A ) ). Define the subspace
S M A X ( A ) = span { v ∈ C n : ∃ u ∈ C n such that A v = ∥ A ∥ u , ∥ v ∥ = ∥ u ∥ = 1 } . (10) S_{\mathrm{MAX}}(A)=\operatorname{span}\{v\in\mathbb{C}^n:\exists\,u\in\mathbb{C}^n\text{ such that }Av=\lVert A\rVert u,\ \lVert v\rVert=\lVert u\rVert=1\}. \tag*{(10)} S MAX ( A ) = span { v ∈ C n : ∃ u ∈ C n such that A v = ∥ A ∥ u , ∥ v ∥ = ∥ u ∥ = 1 } . (10) Then,
(1) S M A X ( A ) = span { v ∈ C n : A v = ∥ A ∥ v or A v = − ∥ A ∥ v } S_{\mathrm{MAX}}(A)=\operatorname{span}\{v\in\mathbb{C}^n:Av=\lVert A\rVert v\text{ or }Av=-\lVert A\rVert v\} S MAX ( A ) = span { v ∈ C n : A v = ∥ A ∥ v or A v = − ∥ A ∥ v }
= S ∥ A ∥ ⊕ S − ∥ A ∥ . =S_{\lVert A\rVert}\oplus S_{-\lVert A\rVert}. = S ∥ A ∥ ⊕ S − ∥ A ∥ . (2) Exactly one of the following conditions holds
S − ∥ A ∥ = { 0 } S_{-\lVert A\rVert}=\{0\} S − ∥ A ∥ = { 0 } and S M A X ( A ) = S ∥ A ∥ S_{\mathrm{MAX}}(A)=S_{\lVert A\rVert} S MAX ( A ) = S ∥ A ∥ ,
S ∥ A ∥ = { 0 } S_{\lVert A\rVert}=\{0\} S ∥ A ∥ = { 0 } and S M A X ( A ) = S − ∥ A ∥ S_{\mathrm{MAX}}(A)=S_{-\lVert A\rVert} S MAX ( A ) = S − ∥ A ∥ ,
S ∥ A ∥ ≠ { 0 } ≠ S − ∥ A ∥ S_{\lVert A\rVert}\ne\{0\}\ne S_{-\lVert A\rVert} S ∥ A ∥ = { 0 } = S − ∥ A ∥ and S M A X ( A ) = S ∥ A ∥ ⊕ S − ∥ A ∥ S_{\mathrm{MAX}}(A)=S_{\lVert A\rVert}\oplus S_{-\lVert A\rVert} S MAX ( A ) = S ∥ A ∥ ⊕ S − ∥ A ∥ .
(3)
S M A X ( A ) ⊥ = ⨁ λ ∈ σ ( A ) ∣ λ ∣ < ∥ A ∥ S λ . S_{\mathrm{MAX}}(A)^\perp=\bigoplus_{\substack{\lambda\in\sigma(A)\\|\lambda|<\lVert A\rVert}}S_\lambda. S MAX ( A ) ⊥ = λ ∈ σ ( A ) ∣ λ ∣ < ∥ A ∥ ⨁ S λ . (4) If M A M_A M A denotes the set of unit vectors at which A A A attains its norm, that is,
M A = { v ∈ C n : ∥ v ∥ = 1 , ∥ A v ∥ = ∥ A ∥ } , M_A=\{v\in\mathbb{C}^n:\lVert v\rVert=1,\ \lVert Av\rVert=\lVert A\rVert\}, M A = { v ∈ C n : ∥ v ∥ = 1 , ∥ A v ∥ = ∥ A ∥} , then
M A = S M A X ( A ) ∩ { x ∈ C n : ∥ x ∥ = 1 } . (11) M_A=S_{\mathrm{MAX}}(A)\cap\{x\in\mathbb{C}^n:\lVert x\rVert=1\}. \tag*{(11)} M A = S MAX ( A ) ∩ { x ∈ C n : ∥ x ∥ = 1 } . (11) (5) Every unit vector v ∈ S M A X ( A ) v\in S_{\mathrm{MAX}}(A) v ∈ S MAX ( A ) , written as v = v + + v − v=v_++v_- v = v + + v − with v ± ∈ S ± ∥ A ∥ v_\pm\in S_{\pm\lVert A\rVert} v ± ∈ S ± ∥ A ∥ , satisfies A v = ∥ A ∥ u Av=\lVert A\rVert u A v = ∥ A ∥ u with
u = v + − v − ∈ S M A X ( A ) and ∥ u ∥ = 1. u=v_+-v_-\in S_{\mathrm{MAX}}(A)\quad\text{and}\quad\lVert u\rVert=1. u = v + − v − ∈ S MAX ( A ) and ∥ u ∥ = 1.
Proof. For λ ∈ σ ( A ) \lambda\in\sigma(A) λ ∈ σ ( A ) , let P λ P_\lambda P λ be the orthogonal projection onto S λ S_\lambda S λ . By the spectral theorem, C n = ⨁ λ ∈ σ ( A ) S λ \mathbb{C}^n=\bigoplus_{\lambda\in\sigma(A)}S_\lambda C n = ⨁ λ ∈ σ ( A ) S λ is an orthogonal decomposition and ∥ A ∥ = max { ∣ λ ∣ : λ ∈ σ ( A ) } \lVert A\rVert=\max\{|\lambda|:\lambda\in\sigma(A)\} ∥ A ∥ = max { ∣ λ ∣ : λ ∈ σ ( A )} . Hence, for every v ∈ C n v\in\mathbb{C}^n v ∈ C n ,
∥ A v ∥ 2 = ∑ λ ∈ σ ( A ) λ 2 ∥ P λ v ∥ 2 ≤ ∥ A ∥ 2 ∑ λ ∈ σ ( A ) ∥ P λ v ∥ 2 = ∥ A ∥ 2 ∥ v ∥ 2 . (12) \lVert Av\rVert^2=\sum_{\lambda\in\sigma(A)}\lambda^2\lVert P_\lambda v\rVert^2\leq\lVert A\rVert^2\sum_{\lambda\in\sigma(A)}\lVert P_\lambda v\rVert^2=\lVert A\rVert^2\lVert v\rVert^2. \tag*{(12)} ∥ A v ∥ 2 = λ ∈ σ ( A ) ∑ λ 2 ∥ P λ v ∥ 2 ≤ ∥ A ∥ 2 λ ∈ σ ( A ) ∑ ∥ P λ v ∥ 2 = ∥ A ∥ 2 ∥ v ∥ 2 . (12)
with equality if and only if P λ v = 0 P_\lambda v=0 P λ v = 0 whenever ∣ λ ∣ < ∥ A ∥ |\lambda|<\lVert A\rVert ∣ λ ∣ < ∥ A ∥ , that is, if and only if v ∈ S ∥ A ∥ + S − ∥ A ∥ v\in S_{\lVert A\rVert}+S_{-\lVert A\rVert} v ∈ S ∥ A ∥ + S − ∥ A ∥ . This sum is direct (and orthogonal) because A ≠ 0 A\ne0 A = 0 implies ∥ A ∥ ≠ − ∥ A ∥ \lVert A\rVert\ne-\lVert A\rVert ∥ A ∥ = − ∥ A ∥ .
(1) The second equality holds because the span of the union of two subspaces is their sum, which is direct here, as noted above. If ∥ v ∥ = 1 \lVert v\rVert=1 ∥ v ∥ = 1 and v ∈ S s ∥ A ∥ v\in S_{s\lVert A\rVert} v ∈ S s ∥ A ∥ with s ∈ { 1 , − 1 } s\in\{1,-1\} s ∈ { 1 , − 1 } , then A v = ∥ A ∥ u Av=\lVert A\rVert u A v = ∥ A ∥ u with u = s v u=sv u = s v and ∥ u ∥ = 1 \lVert u\rVert=1 ∥ u ∥ = 1 , so v v v is one of the generators in (2.10). Hence S ∥ A ∥ ⊕ S − ∥ A ∥ ⊆ S M A X ( A ) S_{\lVert A\rVert}\oplus S_{-\lVert A\rVert}\subseteq S_{\mathrm{MAX}}(A) S ∥ A ∥ ⊕ S − ∥ A ∥ ⊆ S MAX ( A ) . Conversely, if ∥ v ∥ = ∥ u ∥ = 1 \lVert v\rVert=\lVert u\rVert=1 ∥ v ∥ = ∥ u ∥ = 1 and A v = ∥ A ∥ u Av=\lVert A\rVert u A v = ∥ A ∥ u , then ∥ A v ∥ = \lVert Av\rVert= ∥ A v ∥ = ∥ A ∥ = ∥ A ∥ ∥ v ∥ \lVert A\rVert= \lVert A\rVert\lVert v\rVert ∥ A ∥ = ∥ A ∥ ∥ v ∥ , so equality holds in (12) and v ∈ S ∥ A ∥ ⊕ S − ∥ A ∥ v \in S_{\lVert A\rVert} \oplus S_{-\lVert A\rVert} v ∈ S ∥ A ∥ ⊕ S − ∥ A ∥ . Since the latter is a subspace, S M A X ( A ) ⊆ S ∥ A ∥ ⊕ S − ∥ A ∥ S_{\mathrm{MAX}}(A) \subseteq S_{\lVert A\rVert} \oplus S_{-\lVert A\rVert} S MAX ( A ) ⊆ S ∥ A ∥ ⊕ S − ∥ A ∥ .
(2) Since ∥ A ∥ = max { ∣ λ ∣ : λ ∈ σ ( A ) } \lVert A\rVert= \max\{|\lambda| : \lambda\in\sigma(A)\} ∥ A ∥ = max { ∣ λ ∣ : λ ∈ σ ( A )} , at least one of ± ∥ A ∥ \pm\lVert A\rVert ± ∥ A ∥ is an eigenvalue of A A A , so S ∥ A ∥ S_{\lVert A\rVert} S ∥ A ∥ and S − ∥ A ∥ S_{-\lVert A\rVert} S − ∥ A ∥ are not both zero. Hence exactly one of the three listed cases occurs, and in each of them the stated description of S M A X ( A ) S_{\mathrm{MAX}}(A) S MAX ( A ) follows from item 1.
(3) By item 1 and the orthogonal decomposition C n = ⨁ λ ∈ σ ( A ) S λ \mathbb{C}^{n} = \bigoplus_{\lambda\in\sigma(A)} S_{\lambda} C n = ⨁ λ ∈ σ ( A ) S λ , the orthogonal complement of S M A X ( A ) S_{\mathrm{MAX}}(A) S MAX ( A ) is the sum of the eigenspaces S λ S_{\lambda} S λ with λ ∈ σ ( A ) ∖ { ∥ A ∥ , − ∥ A ∥ } \lambda\in\sigma(A) \setminus\{\lVert A\rVert,-\lVert A\rVert\} λ ∈ σ ( A ) ∖ {∥ A ∥ , − ∥ A ∥} , that is, with ∣ λ ∣ < ∥ A ∥ |\lambda| < \lVert A\rVert ∣ λ ∣ < ∥ A ∥ .
(4) If v ∈ S M A X ( A ) v \in S_{\mathrm{MAX}}(A) v ∈ S MAX ( A ) and ∥ v ∥ = 1 \lVert v\rVert= 1 ∥ v ∥ = 1 , then by item 1 equality holds in (12) , so ∥ A v ∥ = ∥ A ∥ \lVert Av\rVert= \lVert A\rVert ∥ A v ∥ = ∥ A ∥ and v ∈ M A v \in M_A v ∈ M A . Conversely, if v ∈ M A v \in M_A v ∈ M A , then u = A v / ∥ A ∥ u = Av/\lVert A\rVert u = A v / ∥ A ∥ satisfies ∥ u ∥ = 1 \lVert u\rVert= 1 ∥ u ∥ = 1 and A v = ∥ A ∥ u Av = \lVert A\rVert u A v = ∥ A ∥ u , so v ∈ S M A X ( A ) v \in S_{\mathrm{MAX}}(A) v ∈ S MAX ( A ) by definition.
(5) Let v ∈ S M A X ( A ) v \in S_{\mathrm{MAX}}(A) v ∈ S MAX ( A ) with ∥ v ∥ = 1 \lVert v\rVert= 1 ∥ v ∥ = 1 . By item 1, v = v + + v − v = v_{+} + v_{-} v = v + + v − with v ± ∈ S ± ∥ A ∥ v_{\pm} \in S_{\pm\lVert A\rVert} v ± ∈ S ± ∥ A ∥ , and ∥ v + ∥ 2 + ∥ v − ∥ 2 = ∥ v ∥ 2 = 1 \lVert v_{+}\rVert^{2} + \lVert v_{-}\rVert^{2} = \lVert v\rVert^{2} = 1 ∥ v + ∥ 2 + ∥ v − ∥ 2 = ∥ v ∥ 2 = 1 because v + ⊥ v − v_{+} \perp v_{-} v + ⊥ v − . Then
A v = ∥ A ∥ v + − ∥ A ∥ v − = ∥ A ∥ u , u = v + − v − ∈ S M A X ( A ) , Av = \lVert A\rVert v_{+} - \lVert A\rVert v_{-} = \lVert A\rVert u,\qquad u = v_{+} - v_{-} \in S_{\mathrm{MAX}}(A), A v = ∥ A ∥ v + − ∥ A ∥ v − = ∥ A ∥ u , u = v + − v − ∈ S MAX ( A ) ,
and ∥ u ∥ 2 = ∥ v + ∥ 2 + ∥ v − ∥ 2 = 1 \lVert u\rVert^{2} = \lVert v_{+}\rVert^{2} + \lVert v_{-}\rVert^{2} = 1 ∥ u ∥ 2 = ∥ v + ∥ 2 + ∥ v − ∥ 2 = 1 .
Moment and joint numerical range
Let
D = { ρ ∈ M n h ( C ) : ρ ≥ 0 and tr ( ρ ) = 1 } (13) \mathcal{D} = \{\rho\in M_n^h(\mathbb{C}) : \rho\ge0 \text{ and } \operatorname{tr}(\rho) = 1\} \tag*{(13)} D = { ρ ∈ M n h ( C ) : ρ ≥ 0 and tr ( ρ ) = 1 } (13)
be the set of density matrices. We now define two sets relevant to our analysis.
Definition 3.1. Let S S S be a subspace of C n \mathbb{C}^{n} C n and P S P_S P S be the orthogonal projection onto S S S . Consider an order fixed in I 0 I_0 I 0 , σ : I 0 → { 1 , 2 , … , r } \sigma: I_0 \to\{1,2,\ldots,r\} σ : I 0 → { 1 , 2 , … , r } such that B i , j = B σ ( i , j ) = B k B^{i,j} = B^{\sigma(i,j)} = B^k B i , j = B σ ( i , j ) = B k , 1 ≤ k ≤ r 1 \le k \le r 1 ≤ k ≤ r , where B i , j B^{i,j} B i , j are the matrices described in (1) . Define the following family of matrices
P S B P S = { P S B 1 P S , P S B 2 P S , … , P S B r P S } . P_S\mathbb{B}P_S = \{P_SB^1P_S,P_SB^2P_S,\ldots,P_SB^rP_S\}. P S B P S = { P S B 1 P S , P S B 2 P S , … , P S B r P S } .
The joint (algebraic) numerical range with respect to the family P S B P S P_S\mathbb{B}P_S P S B P S is
W ( P S B P S ) = { ( tr ( ρ P S B 1 P S ) , … , tr ( ρ P S B r P S ) ) ∈ R r : ρ ∈ D } (14) W(P_S\mathbb{B}P_S) = \{(\operatorname{tr}(\rho P_SB^1P_S),\ldots,\operatorname{tr}(\rho P_SB^rP_S)) \in\mathbb{R}^{r} : \rho\in\mathcal{D}\} \tag*{(14)} W ( P S B P S ) = {( tr ( ρ P S B 1 P S ) , … , tr ( ρ P S B r P S )) ∈ R r : ρ ∈ D } (14)
Definition 3.2. Given a subspace S S S of C n \mathbb{C}^{n} C n , we define the moment of S S S relative to W \mathcal{W} W in the basis B \mathbb{B} B as
m S B = m S = conv { ( ⟨ B 1 v , v ⟩ , … , ⟨ B r v , v ⟩ ) : v ∈ S , ∥ v ∥ = 1 } . (15) m^{\mathbb{B}}_{S} = m_S = \operatorname{conv}\{(\langle B^1v,v\rangle,\ldots,\langle B^rv,v\rangle) : v \in S,\lVert v\rVert= 1\}. \tag*{(15)} m S B = m S = conv {(⟨ B 1 v , v ⟩ , … , ⟨ B r v , v ⟩) : v ∈ S , ∥ v ∥ = 1 } . (15)
Some properties of m S m_S m S are summarized below.
Remark 3.3.
m S ⊆ R r m_S \subseteq\mathbb{R}^{r} m S ⊆ R r , since B i B^i B i is Hermitian for every 1 ≤ i ≤ r 1 \le i \le r 1 ≤ i ≤ r .
The moment of a subspace was first defined in [15 ] for W \mathcal{W} W as the subspace of diagonal real matrices; the name comes from the fact that it is the image of a certain moment map. In [5 ] it was generalized to compact operators. Recently, in [7 ] , this was generalized by considering W \mathcal{W} W to be any unital C ∗ C^* C ∗ subalgebra of M n ( C ) M_n(\mathbb{C}) M n ( C ) , with identity I = I n I = I_n I = I n .
If W = D n ( R ) \mathcal{W} = D_n(\mathbb{R}) W = D n ( R ) the subspace of n × n n \times n n × n real diagonal matrices and B = E = { E i = e i e i ∗ : 1 ≤ i ≤ n } \mathbb{B} = \mathbb{E} = \{E_i = e_i e_i^* : 1 \le i \le n\} B = E = { E i = e i e i ∗ : 1 ≤ i ≤ n } is the diagonal canonical basis, then
m S E = conv { ( ⟨ E 1 v , v ⟩ , … , ⟨ E n v , v ⟩ ) : v ∈ S , ∥ v ∥ = 1 } = conv { ( ∣ v 1 ∣ 2 , … , ∣ v n ∣ 2 ) : v = ( v 1 , … , v n ) ∈ S , ∥ v ∥ = 1 } = conv { ∣ v ∣ 2 : v ∈ S , ∥ v ∥ = 1 } , \begin{aligned}
m_S^{\mathbb{E}} = \operatorname{conv}\{(\langle E_1v,v\rangle,\ldots,\langle E_nv,v\rangle) : v \in S,\lVert v\rVert= 1\} \\
&= \operatorname{conv}\{(|v_1|^2,\ldots,|v_n|^2) : v = (v_1,\ldots,v_n) \in S,\lVert v\rVert= 1\} \\
&= \operatorname{conv}\{|v|^2 : v \in S,\lVert v\rVert= 1\},
\end{aligned} m S E = conv {(⟨ E 1 v , v ⟩ , … , ⟨ E n v , v ⟩) : v ∈ S , ∥ v ∥ = 1 } = conv {( ∣ v 1 ∣ 2 , … , ∣ v n ∣ 2 ) : v = ( v 1 , … , v n ) ∈ S , ∥ v ∥ = 1 } = conv { ∣ v ∣ 2 : v ∈ S , ∥ v ∥ = 1 } ,
which is exactly the definition of moment in [13 , 5 ] .
Recall that, for a family of matrices A 1 , … , A k ∈ M n ( C ) A^1,\ldots,A^k \in M_n(\mathbb{C}) A 1 , … , A k ∈ M n ( C ) and a subspace S ⊆ C n S \subseteq\mathbb{C}^n S ⊆ C n , the k k k -dimensional field of the family restricted to S S S was defined in [10 Definition 10] as
F ( { A j } j = 1 k , S ) = { ( ⟨ A 1 v , v ⟩ , … , ⟨ A k v , v ⟩ ) : v ∈ S , ∥ v ∥ = 1 } ⊆ C k . (16) \mathcal{F}(\{A^j\}_{j=1}^k,S)=\{(\langle A^1v,v\rangle,\ldots,\langle A^kv,v\rangle):v\in S,\ \|v\|=1\}\subseteq\mathbb{C}^k. \tag*{(16)} F ({ A j } j = 1 k , S ) = {(⟨ A 1 v , v ⟩ , … , ⟨ A k v , v ⟩) : v ∈ S , ∥ v ∥ = 1 } ⊆ C k . (16)
In particular, since the matrices B j B^j B j are Hermitian,
F ( { B j } j = 1 r , S ) = { ( ⟨ B 1 v , v ⟩ , … , ⟨ B r v , v ⟩ ) : v ∈ S , ∥ v ∥ = 1 } ⊆ R r , \mathcal{F}(\{B^j\}_{j=1}^r,S)=\{(\langle B^1v,v\rangle,\ldots,\langle B^rv,v\rangle):v\in S,\ \|v\|=1\}\subseteq\mathbb{R}^r, F ({ B j } j = 1 r , S ) = {(⟨ B 1 v , v ⟩ , … , ⟨ B r v , v ⟩) : v ∈ S , ∥ v ∥ = 1 } ⊆ R r ,
which is exactly the set whose convex hull defines m S m_S m S in (15) . Hence
m S = conv F ( { B j } j = 1 r , S ) . m_S=\operatorname{conv}\mathcal{F}(\{B^j\}_{j=1}^r,S). m S = conv F ({ B j } j = 1 r , S ) .
If we consider the joint numerical range with respect to the family P S B 1 P S , … , P S B r P S P_SB^1P_S,\ldots,P_SB^rP_S P S B 1 P S , … , P S B r P S (14) , then m S ⊆ W ( P S B 1 P S , … , P S B r P S ) m_S\subseteq W(P_SB^1P_S,\ldots,P_SB^rP_S) m S ⊆ W ( P S B 1 P S , … , P S B r P S ) , since
⟨ B i v , v ⟩ = tr ( v v ∗ B i ) = tr ( P S v v ∗ P S B i ) \langle B^iv,v\rangle=\operatorname{tr}(vv^*B^i)=\operatorname{tr}(P_Svv^*P_SB^i) ⟨ B i v , v ⟩ = tr ( v v ∗ B i ) = tr ( P S v v ∗ P S B i )
and v v ∗ ∈ D vv^*\in\mathcal{D} v v ∗ ∈ D for every 1 ≤ i ≤ r 1\leq i\leq r 1 ≤ i ≤ r and v ∈ S v\in S v ∈ S .
Let us now define the subset of density matrices relative to the subspace S ⊂ C n S\subset\mathbb{C}^n S ⊂ C n
D S = { ρ ∈ D : im ( ρ ) ⊂ S } ⊆ D . (17) \mathcal{D}_S=\{\rho\in\mathcal{D}:\operatorname{im}(\rho)\subset S\}\subseteq\mathcal{D}. \tag*{(17)} D S = { ρ ∈ D : im ( ρ ) ⊂ S } ⊆ D . (17)
Then every ρ ∈ D S \rho\in\mathcal{D}_S ρ ∈ D S can be written as
ρ = ∑ i = 1 dim ( S ) λ i u i u i ∗ (18) \rho=\sum_{i=1}^{\dim(S)}\lambda_i u_i u_i^* \tag*{(18)} ρ = i = 1 ∑ d i m ( S ) λ i u i u i ∗ (18)
for λ i ∈ R ≥ 0 \lambda_i\in\mathbb{R}_{\geq0} λ i ∈ R ≥ 0 , with ∑ i = 1 dim ( S ) λ i = 1 \sum_{i=1}^{\dim(S)}\lambda_i=1 ∑ i = 1 d i m ( S ) λ i = 1 and { u i } i = 1 dim ( S ) \{u_i\}_{i=1}^{\dim(S)} { u i } i = 1 d i m ( S ) an orthonormal basis of S S S .
We now restrict Φ B \Phi^{\mathbb{B}} Φ B (defined in (4) ) to D S \mathcal{D}_S D S ; that is, for every ρ ∈ D S \rho\in\mathcal{D}_S ρ ∈ D S ,
Φ B ( ρ ) = ( ⟨ ρ , B 1 ⟩ , ⟨ ρ , B 2 ⟩ , … , ⟨ ρ , B r ⟩ ) = ( tr ( ρ B 1 ) , tr ( ρ B 2 ) , … , tr ( ρ B r ) ) , (19) \begin{aligned}
\Phi^{\mathbb{B}}(\rho)=(\langle\rho,B^1\rangle,\langle\rho,B^2\rangle,\ldots,\langle\rho,B^r\rangle) \\
&=(\operatorname{tr}(\rho B^1),\operatorname{tr}(\rho B^2),\ldots,\operatorname{tr}(\rho B^r)), \tag*{(19)}
\end{aligned} Φ B ( ρ ) = (⟨ ρ , B 1 ⟩ , ⟨ ρ , B 2 ⟩ , … , ⟨ ρ , B r ⟩) = ( tr ( ρ B 1 ) , tr ( ρ B 2 ) , … , tr ( ρ B r )) , (19)
where the k k k -th coordinate is the coefficient of B k B^k B k in the orthogonal projection of ρ \rho ρ onto W \mathcal{W} W . Additionally,
ker ( Φ B ∣ D S ) = { ρ ∈ D S : ρ ∈ W 1 ∘ } = D S ∩ W 1 ∘ . \operatorname{ker}(\Phi^{\mathbb{B}}|_{\mathcal{D}_S})=\{\rho\in\mathcal{D}_S:\rho\in\mathcal{W}_1^\circ\}=\mathcal{D}_S\cap\mathcal{W}_1^\circ. ker ( Φ B ∣ D S ) = { ρ ∈ D S : ρ ∈ W 1 ∘ } = D S ∩ W 1 ∘ .
The following result is based on ideas from [7 Proposition 1] . For completeness, we include the proofs.
Proposition 3.4. Let S ⊆ C n S\subseteq\mathbb{C}^n S ⊆ C n be a subspace with dim ( S ) > 0 \dim(S)>0 dim ( S ) > 0 , B = { B k } k = 1 r \mathbb{B}=\{B^k\}_{k=1}^r B = { B k } k = 1 r be the fixed orthonormal basis of Hermitian matrices of a subspace W ⊂ M n h ( C ) \mathcal{W}\subset\mathcal{M}_n^h(\mathbb{C}) W ⊂ M n h ( C ) and Φ B : D S → R r \Phi^{\mathbb{B}}:\mathcal{D}_S\to\mathbb{R}^r Φ B : D S → R r defined as in (19) . Then the moment m S m_S m S defined in (15) admits the following descriptions.
(1) m S = Φ B ( D S ) m_S=\Phi^{\mathbb{B}}(\mathcal{D}_S) m S = Φ B ( D S ) .
(2) m S = { ( tr ( μ B 1 ) , … , tr ( μ B r ) ) : μ ∈ D S } m_S=\{(\operatorname{tr}(\mu B^1),\ldots,\operatorname{tr}(\mu B^r)):\mu\in\mathcal{D}_S\} m S = {( tr ( μ B 1 ) , … , tr ( μ B r )) : μ ∈ D S } (see (17) )
(3) m S = ⋃ { z k } k = 1 dim ( S ) o.n. basis of S conv { ( ⟨ B 1 z k , z k ⟩ , … , ⟨ B r z k , z k ⟩ ) } k = 1 dim ( S ) m_S=\bigcup_{\substack{\{z_k\}_{k=1}^{\dim(S)}\\ \text{o.n. basis}\\ \text{of }S}}\operatorname{conv}\{(\langle B^1z_k,z_k\rangle,\ldots,\langle B^rz_k,z_k\rangle)\}_{k=1}^{\dim(S)} m S = ⋃ { z k } k = 1 d i m ( S ) o.n. basis of S conv {(⟨ B 1 z k , z k ⟩ , … , ⟨ B r z k , z k ⟩) } k = 1 d i m ( S ) .
(4) If S ≠ C n S\ne\mathbb{C}^n S = C n , then W ( P S B 1 P S , … , P S B r P S ) = { ε x ∈ R r : ε ∈ [ 0 , 1 ] , x ∈ m S } W(P_SB^1P_S,\ldots,P_SB^rP_S)=\{\varepsilon x\in\mathbb{R}^r:\varepsilon\in[0,1],x\in m_S\} W ( P S B 1 P S , … , P S B r P S ) = { ε x ∈ R r : ε ∈ [ 0 , 1 ] , x ∈ m S } (see (14) ). If S = C n S=\mathbb{C}^n S = C n , then W ( B 1 , … , B r ) = Φ B ( D ) = m C n W(B^1,\ldots,B^r)=\Phi^{\mathbb{B}}(\mathcal{D})=m_{\mathbb{C}^n} W ( B 1 , … , B r ) = Φ B ( D ) = m C n .
Proof. (1) Recall that any μ ∈ D S \mu\in\mathcal{D}_S μ ∈ D S is a convex combination of rank-one orthogonal projections. That means that it can be described as μ = ∑ k = 1 dim ( S ) t k z k z k ∗ \mu=\sum_{k=1}^{\dim(S)}t_kz_kz_k^* μ = ∑ k = 1 d i m ( S ) t k z k z k ∗ for t k ≥ 0 t_k\geq0 t k ≥ 0 , ∑ k = 1 dim ( S ) t k = 1 \sum_{k=1}^{\dim(S)}t_k=1 ∑ k = 1 d i m ( S ) t k = 1 and { z k } k = 1 dim ( S ) \{z_k\}_{k=1}^{\dim(S)} { z k } k = 1 d i m ( S ) an orthonormal basis of S S S . Then
tr ( μ B i ) = ∑ k = 1 dim ( S ) t k tr ( B i z k z k ∗ ) = ∑ k = 1 dim ( S ) t k ⟨ B i z k , z k ⟩ , for i = 1 , … , r . \operatorname{tr}(\mu B^i)=\sum_{k=1}^{\dim(S)}t_k\operatorname{tr}(B^i z_kz_k^*)=\sum_{k=1}^{\dim(S)}t_k\langle B^iz_k,z_k\rangle,\qquad\text{for }i=1,\ldots,r. tr ( μ B i ) = k = 1 ∑ d i m ( S ) t k tr ( B i z k z k ∗ ) = k = 1 ∑ d i m ( S ) t k ⟨ B i z k , z k ⟩ , for i = 1 , … , r .
Therefore
Φ B ( μ ) = ( tr ( B 1 μ ) , … , tr ( B r μ ) ) = ∑ k = 1 dim ( S ) t k ( ⟨ B 1 z k , z k ⟩ , … , ⟨ B r z k , z k ⟩ ) ∈ m S , \Phi^{\mathbb{B}}(\mu)=\left(\operatorname{tr}(B^{1}\mu),\ldots,\operatorname{tr}(B^{r}\mu)\right)
=\sum_{k=1}^{\dim(S)}t_k\left(\langle B^{1}z_k,z_k\rangle,\ldots,\langle B^{r}z_k,z_k\rangle\right)\in m_S, Φ B ( μ ) = ( tr ( B 1 μ ) , … , tr ( B r μ ) ) = k = 1 ∑ d i m ( S ) t k ( ⟨ B 1 z k , z k ⟩ , … , ⟨ B r z k , z k ⟩ ) ∈ m S ,
which proves one inclusion.
Now, for every v ∈ S v\in S v ∈ S with ∥ v ∥ = 1 \lVert v\rVert=1 ∥ v ∥ = 1 , we have that
( ⟨ B 1 v , v ⟩ , … , ⟨ B r v , v ⟩ ) = Φ B ( v v ∗ ) , v v ∗ ∈ D S . \left(\langle B^{1}v,v\rangle,\ldots,\langle B^{r}v,v\rangle\right)=\Phi^{\mathbb{B}}(vv^*),\qquad vv^*\in\mathcal{D}_S. ( ⟨ B 1 v , v ⟩ , … , ⟨ B r v , v ⟩ ) = Φ B ( v v ∗ ) , v v ∗ ∈ D S .
Since D S \mathcal{D}_S D S is convex and Φ B \Phi^{\mathbb{B}} Φ B is linear, Φ B ( D S ) \Phi^{\mathbb{B}}(\mathcal{D}_S) Φ B ( D S ) is convex, and therefore
m S = conv { ( ⟨ B 1 v , v ⟩ , … , ⟨ B r v , v ⟩ ) : v ∈ S , ∥ v ∥ = 1 } ⊆ Φ B ( D S ) , m_S=\operatorname{conv}\left\{\left(\langle B^{1}v,v\rangle,\ldots,\langle B^{r}v,v\rangle\right):v\in S,\ \lVert v\rVert=1\right\}\subseteq\Phi^{\mathbb{B}}(\mathcal{D}_S), m S = conv { ( ⟨ B 1 v , v ⟩ , … , ⟨ B r v , v ⟩ ) : v ∈ S , ∥ v ∥ = 1 } ⊆ Φ B ( D S ) ,
which proves the other inclusion.
(2) This is item 1 written explicitly, by the definition of Φ B \Phi^{\mathbb{B}} Φ B .
(3) The inclusion { ( ⟨ B 1 z k , z k ⟩ , … , ⟨ B r z k , z k ⟩ ) } k = 1 dim ( S ) ⊂ m S \left\{\left(\langle B^{1}z_k,z_k\rangle,\ldots,\langle B^{r}z_k,z_k\rangle\right)\right\}_{k=1}^{\dim(S)}\subset m_S { ( ⟨ B 1 z k , z k ⟩ , … , ⟨ B r z k , z k ⟩ ) } k = 1 d i m ( S ) ⊂ m S , for { z k } k = 1 dim ( S ) \{z_k\}_{k=1}^{\dim(S)} { z k } k = 1 d i m ( S ) an orthonormal basis, follows from item 1, since z k z k ∗ ∈ D S z_kz_k^*\in\mathcal{D}_S z k z k ∗ ∈ D S , and the fact that m S m_S m S is convex. Hence the union of these sets is also included in m S m_S m S .
The other inclusion follows from the description of every μ ∈ D S \mu\in\mathcal{D}_S μ ∈ D S in terms of an orthonormal basis of S S S , as in the beginning of the proof of item 1.
(4) This statement can be proved using the same arguments as in the proof of (4) ⇒ \Rightarrow ⇒ (5) in [7, \Proposition 1`]`. The hypothesis S ≠ C n S\ne\mathbb{C}^n S = C n is needed for the inclusion ⊇ \supseteq ⊇ with ε < 1 \varepsilon<1 ε < 1 , which uses density matrices supported on S ⊥ S^\perp S ⊥ . If S = C n S=\mathbb{C}^n S = C n , then P S = I P_S=I P S = I and the claim is item 1 with D C n = D \mathcal{D}_{\mathbb{C}^n}=\mathcal{D} D C n = D .
For the next result, note that 0 ∈ W ( P S B 1 P S , … , P S B r P S ) 0\in W(P_SB^1P_S,\ldots,P_SB^rP_S) 0 ∈ W ( P S B 1 P S , … , P S B r P S ) whenever S ⊊ C n S\subsetneq\mathbb{C}^n S ⊊ C n , since ρ = 1 tr ( P S ⊥ ) P S ⊥ ∈ D \rho=\frac{1}{\operatorname{tr}(P_{S^\perp})}P_{S^\perp}\in\mathcal{D} ρ = tr ( P S ⊥ ) 1 P S ⊥ ∈ D satisfies tr ( P S B i P S ρ ) = 0 \operatorname{tr}(P_SB^iP_S\rho)=0 tr ( P S B i P S ρ ) = 0 , for all i = 1 , … , r i=1,\ldots,r i = 1 , … , r .
Proposition 3.5. With the previous notation, the following statements are equivalent for { 0 } ≠ S ≠ C n \{0\}\ne S\ne\mathbb{C}^n { 0 } = S = C n .
(1) 0 ‾ ∈ m S B \overline{0}\in m_S^{\mathbb{B}} 0 ∈ m S B
(2) m S B = W ( P S B 1 P S , … , P S B r P S ) m_S^{\mathbb{B}}=W(P_SB^1P_S,\ldots,P_SB^rP_S) m S B = W ( P S B 1 P S , … , P S B r P S ) ,
(3) ∃ μ ∈ D S \exists\mu\in\mathcal{D}_S ∃ μ ∈ D S such that tr ( μ B i ) = 0 \operatorname{tr}(\mu B^i)=0 tr ( μ B i ) = 0 , for i = 1 , … , r i=1,\ldots,r i = 1 , … , r ,
(4) ∃ μ ∈ D S ∩ W ∘ \exists\mu\in\mathcal{D}_S\cap W^{\circ} ∃ μ ∈ D S ∩ W ∘ ,
(5) D S ∩ W ∘ ≠ ∅ \mathcal{D}_S\cap W^{\circ}\ne\varnothing D S ∩ W ∘ = ∅ .
Proof. All the equivalences follow by direct computation from Proposition 3.4 . □
The following result is proved in [7 ] for C ∗ \mathbb{C}^* C ∗ -subalgebras of M n ( C ) M_n(\mathbb{C}) M n ( C ) ; the same proof applies here since it only uses orthonormal Hermitian bases of W \mathcal{W} W .
Lemma 3.6. If T = { T k } k = 1 r \mathbb{T}=\{T^k\}_{k=1}^r T = { T k } k = 1 r and V = { V k } k = 1 r \mathbb{V}=\{V^k\}_{k=1}^r V = { V k } k = 1 r are two orthonormal bases of Hermitian matrices of W \mathcal{W} W , then for any ρ ∈ D S \rho\in\mathcal{D}_S ρ ∈ D S
Φ V ( ρ ) = C T , V Φ T ( ρ ) , \Phi^{\mathbb{V}}(\rho)=C_{\mathbb{T},\mathbb{V}}\Phi^{\mathbb{T}}(\rho), Φ V ( ρ ) = C T , V Φ T ( ρ ) , where C T , V C_{\mathbb{T},\mathbb{V}} C T , V is the change of basis matrix from T \mathbb{T} T to V \mathbb{V} V , that is
( C T , V ) i j = ⟨ V i , T j ⟩ = tr ( T j V i ) , with i , j = 1 , … , r . (C_{\mathbb{T},\mathbb{V}})_{ij}=\langle V^i,T^j\rangle=\operatorname{tr}(T^jV^i),\quad\text{with }i,j=1,\ldots,r. ( C T , V ) ij = ⟨ V i , T j ⟩ = tr ( T j V i ) , with i , j = 1 , … , r .
Proposition 3.7 (Change of basis in W \mathcal{W} W ). If T = { T k } k = 1 dim ( W ) \mathbb{T}=\{T^k\}_{k=1}^{\dim(\mathcal{W})} T = { T k } k = 1 d i m ( W ) and V = { V k } k = 1 dim ( W ) \mathbb{V}=\{V^k\}_{k=1}^{\dim(\mathcal{W})} V = { V k } k = 1 d i m ( W ) are two orthonormal bases of Hermitian matrices of W \mathcal{W} W , then
m S V = C T , V m S T (20) m_S^{\mathbb{V}}=C_{\mathbb{T},\mathbb{V}}m_S^{\mathbb{T}} \tag*{(20)} m S V = C T , V m S T (20)
and
W ( P S V P S ) = C T , V W ( P T T P S ) , (21) W(P_S V P_S)=C_{T,V}W(P_T T P_S), \tag*{(21)} W ( P S V P S ) = C T , V W ( P T T P S ) , (21)
where C T , V C_{T,V} C T , V is the change of basis matrix.
Proof. Combining Lemma 3.6 and item 1 in Proposition 3.4 it follows that
m S V = Φ V ( D S ) = C T , V Φ T ( D S ) = C T , V m S T , m_S^V=\Phi^V(D_S)=C_{T,V}\Phi^T(D_S)=C_{T,V}m_S^T, m S V = Φ V ( D S ) = C T , V Φ T ( D S ) = C T , V m S T ,
which proves (3.8). If S = C n S=\mathbb{C}^n S = C n , then (3.9) is (3.8), by item 4 of Proposition 3.4. Otherwise, (3.9) is a direct consequence of (3.8) and item 4 of Proposition 3.4, since
W ( P S V P S ) = { ε x ′ : x ′ ∈ m S V , ε ∈ [ 0 , 1 ] } = { ε C T , V x : x ∈ m S T , ε ∈ [ 0 , 1 ] } = C T , V W ( P T T P S ) . \begin{aligned}
W(P_S V P_S)&=\{\varepsilon x':x'\in m_S^V,\varepsilon\in[0,1]\}\\
&=\{\varepsilon C_{T,V}x:x\in m_S^T,\varepsilon\in[0,1]\}\\
&=C_{T,V}W(P_T T P_S).
\end{aligned} W ( P S V P S ) = { ε x ′ : x ′ ∈ m S V , ε ∈ [ 0 , 1 ]} = { ε C T , V x : x ∈ m S T , ε ∈ [ 0 , 1 ]} = C T , V W ( P T T P S ) .
In the following proposition we relate the field (see (3.4))
F ( { ( A 0 + W 0 ) B j } j = 1 r , S M A X ) \mathcal{F}(\{(A_0+W_0)B^j\}_{j=1}^r,S_{\mathrm{MAX}}) F ({( A 0 + W 0 ) B j } j = 1 r , S MAX )
to the fields of { B j } j = 1 r \{B^j\}_{j=1}^r { B j } j = 1 r restricted to the eigenspaces S λ max S_{\lambda_{\max}} S λ m a x and S λ min S_{\lambda_{\min}} S λ m i n .
Proposition 3.8. Let W ⊂ M n h ( C ) \mathcal{W}\subset M_n^h(\mathbb{C}) W ⊂ M n h ( C ) be a subspace with dim R ( W ) = r \dim_{\mathbb{R}}(\mathcal{W})=r dim R ( W ) = r , let { B j } j = 1 r \{B^j\}_{j=1}^r { B j } j = 1 r be the fixed orthonormal basis of Hermitian matrices of W \mathcal{W} W , let A 0 ∈ M n h ( C ) ∖ W A_0\in M_n^h(\mathbb{C})\setminus\mathcal{W} A 0 ∈ M n h ( C ) ∖ W and W 0 ∈ W W_0\in\mathcal{W} W 0 ∈ W , and let S M A X = S M A X ( A 0 + W 0 ) S_{\mathrm{MAX}}=S_{\mathrm{MAX}}(A_0+W_0) S MAX = S MAX ( A 0 + W 0 ) be as in (2.10). Let λ max \lambda_{\max} λ m a x and λ min \lambda_{\min} λ m i n be the maximum and minimum eigenvalues of A 0 + W 0 A_0+W_0 A 0 + W 0 . Then:
(1) If λ max + λ min > 0 \lambda_{\max}+\lambda_{\min}>0 λ m a x + λ m i n > 0 (equivalently, S M A X = S λ max S_{\mathrm{MAX}}=S_{\lambda_{\max}} S MAX = S λ m a x ), then
F ( { ( A 0 + W 0 ) B j } j = 1 r , S M A X ) = ∥ A 0 + W 0 ∥ F ( { B j } j = 1 r , S λ max ) ⊂ R r . \mathcal{F}(\{(A_0+W_0)B^j\}_{j=1}^r,S_{\mathrm{MAX}})=\lVert A_0+W_0\rVert\mathcal{F}(\{B^j\}_{j=1}^r,S_{\lambda_{\max}})\subset\mathbb{R}^r. F ({( A 0 + W 0 ) B j } j = 1 r , S MAX ) = ∥ A 0 + W 0 ∥ F ({ B j } j = 1 r , S λ m a x ) ⊂ R r . (2) If λ max + λ min < 0 \lambda_{\max}+\lambda_{\min}<0 λ m a x + λ m i n < 0 (equivalently, S M A X = S λ min S_{\mathrm{MAX}}=S_{\lambda_{\min}} S MAX = S λ m i n ), then
F ( { ( A 0 + W 0 ) B j } j = 1 r , S M A X ) = − ∥ A 0 + W 0 ∥ F ( { B j } j = 1 r , S λ min ) ⊂ R r . \mathcal{F}(\{(A_0+W_0)B^j\}_{j=1}^r,S_{\mathrm{MAX}})=-\lVert A_0+W_0\rVert\mathcal{F}(\{B^j\}_{j=1}^r,S_{\lambda_{\min}})\subset\mathbb{R}^r. F ({( A 0 + W 0 ) B j } j = 1 r , S MAX ) = − ∥ A 0 + W 0 ∥ F ({ B j } j = 1 r , S λ m i n ) ⊂ R r . (3) If λ max + λ min = 0 \lambda_{\max}+\lambda_{\min}=0 λ m a x + λ m i n = 0 (equivalently, S M A X = S λ max ⊕ S λ min S_{\mathrm{MAX}}=S_{\lambda_{\max}}\oplus S_{\lambda_{\min}} S MAX = S λ m a x ⊕ S λ m i n ), then
F ( { ( A 0 + W 0 ) B j } j = 1 r , S M A X ) = ∥ A 0 + W 0 ∥ { t y ( u + ) − ( 1 − t ) z ( u − ) ∥ A 0 + W 0 ∥ { + 2 i t ( 1 − t ) w ( u + , u − ) } , \begin{aligned}
\mathcal{F}(\{(A_0+W_0)B^j\}_{j=1}^r,S_{\mathrm{MAX}})
={}&\lVert A_0+W_0\rVert\{t y(u_+)-(1-t)z(u_-)\\
&\mathrel{\phantom{\lVert A_0+W_0\rVert\{}}+2\mathrm{i}\sqrt{t(1-t)}w(u_+,u_-)\},
\end{aligned} F ({( A 0 + W 0 ) B j } j = 1 r , S MAX ) = ∥ A 0 + W 0 ∥ { t y ( u + ) − ( 1 − t ) z ( u − ) ∥ A 0 + W 0 ∥ { + 2 i t ( 1 − t ) w ( u + , u − )} , where the set on the right runs over t ∈ [ 0 , 1 ] t\in[0,1] t ∈ [ 0 , 1 ] and unit vectors u + ∈ S λ max u_+\in S_{\lambda_{\max}} u + ∈ S λ m a x , u − ∈ S λ min u_-\in S_{\lambda_{\min}} u − ∈ S λ m i n , and
y ( u + ) = ( ⟨ B 1 u + , u + ⟩ , … , ⟨ B r u + , u + ⟩ ) ∈ F ( { B j } j = 1 r , S λ max ) , y(u_+)=\left(\langle B^1u_+,u_+\rangle,\ldots,\langle B^ru_+,u_+\rangle\right)\in\mathcal{F}(\{B^j\}_{j=1}^r,S_{\lambda_{\max}}), y ( u + ) = ( ⟨ B 1 u + , u + ⟩ , … , ⟨ B r u + , u + ⟩ ) ∈ F ({ B j } j = 1 r , S λ m a x ) , z ( u − ) = ( ⟨ B 1 u − , u − ⟩ , … , ⟨ B r u − , u − ⟩ ) ∈ F ( { B j } j = 1 r , S λ min ) , z(u_-)=\left(\langle B^1u_-,u_-\rangle,\ldots,\langle B^ru_-,u_-\rangle\right)\in\mathcal{F}(\{B^j\}_{j=1}^r,S_{\lambda_{\min}}), z ( u − ) = ( ⟨ B 1 u − , u − ⟩ , … , ⟨ B r u − , u − ⟩ ) ∈ F ({ B j } j = 1 r , S λ m i n ) , w ( u + , u − ) = ( Im ⟨ B 1 u − , u + ⟩ , … , Im ⟨ B r u − , u + ⟩ ) ∈ R r . w(u_+,u_-)=\left(\operatorname{Im}\langle B^1u_-,u_+\rangle,\ldots,\operatorname{Im}\langle B^ru_-,u_+\rangle\right)\in\mathbb{R}^r. w ( u + , u − ) = ( Im ⟨ B 1 u − , u + ⟩ , … , Im ⟨ B r u − , u + ⟩ ) ∈ R r . In this case the field need not be contained in R r \mathbb{R}^r R r , since the cross terms ⟨ B j u − , u + ⟩ \langle B^ju_-,u_+\rangle ⟨ B j u − , u + ⟩ are in general complex.
Proof. Since A 0 ∉ W A_0\notin\mathcal{W} A 0 ∈ / W , A 0 + W 0 ≠ 0 A_0+W_0\ne0 A 0 + W 0 = 0 , so by Proposition 2.6 the three conditions on λ max + λ min \lambda_{\max}+\lambda_{\min} λ m a x + λ m i n are equivalent to the stated descriptions of S M A X S_{\mathrm{MAX}} S MAX . Let v ∈ S M A X v\in S_{\mathrm{MAX}} v ∈ S MAX be a unit vector and consider
x = ( ⟨ ( A 0 + W 0 ) B 1 v , v ⟩ , … , ⟨ ( A 0 + W 0 ) B r v , v ⟩ ) , x=\left(\langle(A_0+W_0)B^1v,v\rangle,\ldots,\langle(A_0+W_0)B^rv,v\rangle\right), x = ( ⟨( A 0 + W 0 ) B 1 v , v ⟩ , … , ⟨( A 0 + W 0 ) B r v , v ⟩ ) ,
so that x ∈ F ( { ( A 0 + W 0 ) B j } j = 1 r , S M A X ) x\in\mathcal{F}(\{(A_0+W_0)B^j\}_{j=1}^r,S_{\mathrm{MAX}}) x ∈ F ({( A 0 + W 0 ) B j } j = 1 r , S MAX ) , and every element of this field is of this form. Since B j B^j B j and A 0 + W 0 A_0+W_0 A 0 + W 0 are Hermitian, ⟨ ( A 0 + W 0 ) B j v , v ⟩ = ⟨ B j v , ( A 0 + W 0 ) v ⟩ \langle(A_0+W_0)B^jv,v\rangle=\langle B^jv,(A_0+W_0)v\rangle ⟨( A 0 + W 0 ) B j v , v ⟩ = ⟨ B j v , ( A 0 + W 0 ) v ⟩ for every $1\le j\le r. (1) If S M A X = S λ max S_{\mathrm{MAX}}=S_{\lambda_{\max}} S MAX = S λ m a x , then ( A 0 + W 0 ) v = ∥ A 0 + W 0 ∥ v (A_{0}+W_{0})v=\lVert A_{0}+W_{0}\rVert v ( A 0 + W 0 ) v = ∥ A 0 + W 0 ∥ v for every v ∈ S M A X v\in S_{\mathrm{MAX}} v ∈ S MAX , so
x = ∥ A 0 + W 0 ∥ ( ⟨ B 1 v , v ⟩ , … , ⟨ B r v , v ⟩ ) ∈ ∥ A 0 + W 0 ∥ F ( { B j } j = 1 r , S λ max ) . x=\lVert A_{0}+W_{0}\rVert\left(\langle B^{1}v,v\rangle,\ldots,\langle B^{r}v,v\rangle\right)\in\lVert A_{0}+W_{0}\rVert\mathcal{F}\left(\{B^{j}\}_{j=1}^{r},S_{\lambda_{\max}}\right). x = ∥ A 0 + W 0 ∥ ( ⟨ B 1 v , v ⟩ , … , ⟨ B r v , v ⟩ ) ∈ ∥ A 0 + W 0 ∥ F ( { B j } j = 1 r , S λ m a x ) .
Reversing the same steps gives the opposite inclusion.
(2) If S M A X = S λ min S_{\mathrm{MAX}}=S_{\lambda_{\min}} S MAX = S λ m i n , the same argument, with ( A 0 + W 0 ) v = − ∥ A 0 + W 0 ∥ v (A_{0}+W_{0})v=-\lVert A_{0}+W_{0}\rVert v ( A 0 + W 0 ) v = − ∥ A 0 + W 0 ∥ v , gives
x ∈ − ∥ A 0 + W 0 ∥ F ( { B j } j = 1 r , S λ min ) , x\in-\lVert A_{0}+W_{0}\rVert\mathcal{F}\left(\{B^{j}\}_{j=1}^{r},S_{\lambda_{\min}}\right), x ∈ − ∥ A 0 + W 0 ∥ F ( { B j } j = 1 r , S λ m i n ) ,
and again the opposite inclusion follows by reversing the steps.
(3) If λ max + λ min = 0 \lambda_{\max}+\lambda_{\min}=0 λ m a x + λ m i n = 0 , then ∥ A 0 + W 0 ∥ = λ max \lVert A_{0}+W_{0}\rVert=\lambda_{\max} ∥ A 0 + W 0 ∥ = λ m a x . Write v = v + + v − v=v_{+}+v_{-} v = v + + v − with v + ∈ S λ max v_{+}\in S_{\lambda_{\max}} v + ∈ S λ m a x and v − ∈ S λ min v_{-}\in S_{\lambda_{\min}} v − ∈ S λ m i n . Since the eigenspaces are orthogonal, t = ∥ v + ∥ 2 ∈ [ 0 , 1 ] t=\lVert v_{+}\rVert^{2}\in[0,1] t = ∥ v + ∥ 2 ∈ [ 0 , 1 ] and ∥ v − ∥ 2 = 1 − t \lVert v_{-}\rVert^{2}=1-t ∥ v − ∥ 2 = 1 − t , so that v + = t u + v_{+}=\sqrt{t}u_{+} v + = t u + and v − = 1 − t u − v_{-}=\sqrt{1-t}u_{-} v − = 1 − t u − for some unit vectors u + ∈ S λ max u_{+}\in S_{\lambda_{\max}} u + ∈ S λ m a x and u − ∈ S λ min u_{-}\in S_{\lambda_{\min}} u − ∈ S λ m i n (if t ∈ { 0 , 1 } t\in\{0,1\} t ∈ { 0 , 1 } , the vector multiplied by 0 is an arbitrary unit vector of the corresponding eigenspace). Since ( A 0 + W 0 ) v = λ max ( v + − v − ) (A_{0}+W_{0})v=\lambda_{\max}(v_{+}-v_{-}) ( A 0 + W 0 ) v = λ m a x ( v + − v − ) , B j B^{j} B j is Hermitian and ⟨ B j v + , v − ⟩ = ⟨ B j v − , v + ⟩ \langle B^{j}v_{+},v_{-}\rangle=\langle B^{j}v_{-},v_{+}\rangle ⟨ B j v + , v − ⟩ = ⟨ B j v − , v + ⟩ , we have that
⟨ ( A 0 + W 0 ) B j v , v ⟩ = λ max ⟨ B j ( v + + v − ) , v + − v − ⟩ = λ max [ ⟨ B j v + , v + ⟩ − ⟨ B j v − , v − ⟩ + ⟨ B j v − , v + ⟩ − ⟨ B j v + , v − ⟩ ] = λ max [ t ⟨ B j u + , u + ⟩ − ( 1 − t ) ⟨ B j u − , u − ⟩ + 2 i t ( 1 − t ) Im ⟨ B j u − , u + ⟩ ] \begin{aligned}
\langle(A_{0}+W_{0})B^{j}v,v\rangle
&=\lambda_{\max}\langle B^{j}(v_{+}+v_{-}),v_{+}-v_{-}\rangle\\
&=\lambda_{\max}\left[\langle B^{j}v_{+},v_{+}\rangle-\langle B^{j}v_{-},v_{-}\rangle\right.\\
&\qquad\left.+\langle B^{j}v_{-},v_{+}\rangle-\langle B^{j}v_{+},v_{-}\rangle\right]\\
&=\lambda_{\max}\left[t\langle B^{j}u_{+},u_{+}\rangle-(1-t)\langle B^{j}u_{-},u_{-}\rangle\right.\\
&\qquad\left.+2\mathrm{i}\sqrt{t(1-t)}\operatorname{Im}\langle B^{j}u_{-},u_{+}\rangle\right]
\end{aligned} ⟨( A 0 + W 0 ) B j v , v ⟩ = λ m a x ⟨ B j ( v + + v − ) , v + − v − ⟩ = λ m a x [ ⟨ B j v + , v + ⟩ − ⟨ B j v − , v − ⟩ + ⟨ B j v − , v + ⟩ − ⟨ B j v + , v − ⟩ ] = λ m a x [ t ⟨ B j u + , u + ⟩ − ( 1 − t ) ⟨ B j u − , u − ⟩ + 2 i t ( 1 − t ) Im ⟨ B j u − , u + ⟩ ]
for every 1 ≤ j ≤ r 1\leq j\leq r 1 ≤ j ≤ r . Hence
x = λ max [ t y ( u + ) − ( 1 − t ) z ( u − ) + 2 i t ( 1 − t ) w ( u + , u − ) ] , x=\lambda_{\max}\left[t\,y(u_{+})-(1-t)\,z(u_{-})+2\mathrm{i}\sqrt{t(1-t)}\,w(u_{+},u_{-})\right], x = λ m a x [ t y ( u + ) − ( 1 − t ) z ( u − ) + 2 i t ( 1 − t ) w ( u + , u − ) ] ,
which proves the inclusion ⊆ \subseteq ⊆ . Conversely, given t ∈ [ 0 , 1 ] t\in[0,1] t ∈ [ 0 , 1 ] and unit vectors u ± ∈ S λ max / min u_{\pm}\in S_{\lambda_{\max}/\min} u ± ∈ S λ m a x / m i n , the vector v = t u + + 1 − t u − v=\sqrt{t}u_{+}+\sqrt{1-t}u_{-} v = t u + + 1 − t u − is a unit vector of S M A X S_{\mathrm{MAX}} S MAX , and the same computation shows that the corresponding element of the right-hand side belongs to the field.
Corollary 3.9. Under the hypotheses of Proposition 3.8, let λ max \lambda_{\max} λ m a x and λ min \lambda_{\min} λ m i n be the maximum and minimum eigenvalues of A 0 + W 0 A_{0}+W_{0} A 0 + W 0 . Recall that for every subspace S ⊂ C n S\subset\mathbb{C}^{n} S ⊂ C n , conv F ( { B j } j = 1 r , S ) = m S \operatorname{conv}\mathcal{F}(\{B^{j}\}_{j=1}^{r},S)=m_{S} conv F ({ B j } j = 1 r , S ) = m S (Remark 3.3). Hence,
m S λ max = conv F ( { B j } j = 1 r , S λ max ) , m S λ min = conv F ( { B j } j = 1 r , S λ min ) , (22) m_{S_{\lambda_{\max}}}=\operatorname{conv}\mathcal{F}\left(\{B^{j}\}_{j=1}^{r},S_{\lambda_{\max}}\right),
\qquad
m_{S_{\lambda_{\min}}}=\operatorname{conv}\mathcal{F}\left(\{B^{j}\}_{j=1}^{r},S_{\lambda_{\min}}\right), \tag*{(22)} m S λ m a x = conv F ( { B j } j = 1 r , S λ m a x ) , m S λ m i n = conv F ( { B j } j = 1 r , S λ m i n ) , (22) and, taking real parts componentwise in C r \mathbb{C}^{r} C r ,
Re ( conv F ( { ( A 0 + W 0 ) B j } j = 1 r , S M A X ) ) = ∥ A 0 + W 0 ∥ { m S M A X if λ max + λ min > 0 , − m S M A X if λ max + λ min < 0 , conv ( m S λ max ∪ − m S λ min ) if λ max + λ min = 0. (23) \operatorname{Re}\left(\operatorname{conv}\mathcal{F}\left(\{(A_{0}+W_{0})B^{j}\}_{j=1}^{r},S_{\mathrm{MAX}}\right)\right)
=\lVert A_{0}+W_{0}\rVert\begin{cases}
m_{S_{\mathrm{MAX}}} & \text{if }\lambda_{\max}+\lambda_{\min}>0,\\
-m_{S_{\mathrm{MAX}}} & \text{if }\lambda_{\max}+\lambda_{\min}<0,\\
\operatorname{conv}\left(m_{S_{\lambda_{\max}}}\cup-m_{S_{\lambda_{\min}}}\right) & \text{if }\lambda_{\max}+\lambda_{\min}=0.
\end{cases} \tag*{(23)} Re ( conv F ( {( A 0 + W 0 ) B j } j = 1 r , S MAX ) ) = ∥ A 0 + W 0 ∥ ⎩ ⎨ ⎧ m S MAX − m S MAX conv ( m S λ m a x ∪ − m S λ m i n ) if λ m a x + λ m i n > 0 , if λ m a x + λ m i n < 0 , if λ m a x + λ m i n = 0. (23)
Proof. Write
F M A X = F ( { ( A 0 + W 0 ) B j } j = 1 r , S M A X ) , \mathcal{F}_{\mathrm{MAX}}=\mathcal{F}\left(\{(A_{0}+W_{0})B^{j}\}_{j=1}^{r},S_{\mathrm{MAX}}\right), F MAX = F ( {( A 0 + W 0 ) B j } j = 1 r , S MAX ) ,
F + = F ( { B j } j = 1 r , S λ max ) , F − = F ( { B j } j = 1 r , S λ min ) . \mathcal{F}_{+}=\mathcal{F}\left(\{B^{j}\}_{j=1}^{r},S_{\lambda_{\max}}\right),
\qquad
\mathcal{F}_{-}=\mathcal{F}\left(\{B^{j}\}_{j=1}^{r},S_{\lambda_{\min}}\right). F + = F ( { B j } j = 1 r , S λ m a x ) , F − = F ( { B j } j = 1 r , S λ m i n ) .
Since Re \operatorname{Re} Re is R \mathbb{R} R -linear, Re ( conv X ) = conv ( Re X ) \operatorname{Re}(\operatorname{conv}X)=\operatorname{conv}(\operatorname{Re}X) Re ( conv X ) = conv ( Re X ) for every X ⊂ C r X\subset\mathbb{C}^{r} X ⊂ C r . If λ max + λ min > 0 \lambda_{\max}+\lambda_{\min}>0 λ m a x + λ m i n > 0 , then S M A X = S λ max S_{\mathrm{MAX}}=S_{\lambda_{\max}} S MAX = S λ m a x , and by item 1 of Proposition 3.8 we have F M A X ⊂ R r \mathcal{F}_{\mathrm{MAX}}\subset\mathbb{R}^{r} F MAX ⊂ R r and
conv F M A X = ∥ A 0 + W 0 ∥ conv F + = ∥ A 0 + W 0 ∥ m S M A X . \operatorname{conv}\mathcal{F}_{\mathrm{MAX}}=\lVert A_{0}+W_{0}\rVert\operatorname{conv}\mathcal{F}_{+}=\lVert A_{0}+W_{0}\rVert m_{S_{\mathrm{MAX}}}. conv F MAX = ∥ A 0 + W 0 ∥ conv F + = ∥ A 0 + W 0 ∥ m S MAX .
The case λ max + λ min < 0 \lambda_{\max}+\lambda_{\min}<0 λ m a x + λ m i n < 0 follows in the same way from item 2. If λ max + λ min = 0 \lambda_{\max}+\lambda_{\min}=0 λ m a x + λ m i n = 0 , the computation in the proof of item 3 holds for every t ∈ [ 0 , 1 ] t\in[0,1] t ∈ [ 0 , 1 ] and every pair of unit vectors u + ∈ S λ max u_+\in\mathcal{S}_{\lambda_{\max}} u + ∈ S λ m a x , u − ∈ S λ min u_-\in\mathcal{S}_{\lambda_{\min}} u − ∈ S λ m i n , taking v = t u + + 1 − t u − v=\sqrt{t}u_++\sqrt{1-t}u_- v = t u + + 1 − t u − . Therefore
Re F M A X = ∥ A 0 + W 0 ∥ { t y − ( 1 − t ) z : t ∈ [ 0 , 1 ] , y ∈ F + , z ∈ F − } . \operatorname{Re}\mathcal{F}_{\mathrm{MAX}}=\lVert A_0+W_0\rVert\{ty-(1-t)z:t\in[0,1],y\in\mathcal{F}_+,z\in\mathcal{F}_-\}. Re F MAX = ∥ A 0 + W 0 ∥ { t y − ( 1 − t ) z : t ∈ [ 0 , 1 ] , y ∈ F + , z ∈ F − } .
The set between braces contains F + \mathcal{F}_+ F + (take t = 1 t=1 t = 1 ) and − F − -\mathcal{F}_- − F − (take t = 0 t=0 t = 0 ), and it is contained in conv ( F + ∪ − F − ) \operatorname{conv}(\mathcal{F}_+\cup-\mathcal{F}_-) conv ( F + ∪ − F − ) . Hence its convex hull is conv ( F + ∪ − F − ) = conv ( m S λ max ∪ − m S λ min ) \operatorname{conv}(\mathcal{F}_+\cup-\mathcal{F}_-)=\operatorname{conv}(m\mathcal{S}_{\lambda_{\max}}\cup-m\mathcal{S}_{\lambda_{\min}}) conv ( F + ∪ − F − ) = conv ( m S λ m a x ∪ − m S λ m i n ) , by (22) .
Remark 3.10. By Corollary 4.6 below, the right-hand side of (23) is ∥ A 0 + W 0 ∥ \lVert A_0+W_0\rVert ∥ A 0 + W 0 ∥ times ∂ ( ∥ A ( ⋅ ) ∥ ) ( x ) \partial(\lVert A(\cdot)\rVert)(x) ∂ (∥ A ( ⋅ )∥) ( x ) , where x ∈ R r x\in\mathbb{R}^r x ∈ R r is such that A ( x ) = A 0 + W 0 A(x)=A_0+W_0 A ( x ) = A 0 + W 0 . Since ∥ A 0 + W 0 ∥ > 0 \lVert A_0+W_0\rVert>0 ∥ A 0 + W 0 ∥ > 0 , Theorem 5.9 shows that A 0 + W 0 A_0+W_0 A 0 + W 0 is W W W -minimal if and only if
0 ∈ Re ( conv F ( { ( A 0 + W 0 ) B j } j = 1 r , S M A X ) ) . \boldsymbol{0}\in\operatorname{Re}\left(\operatorname{conv}\mathcal{F}\left(\{(A_0+W_0)B^j\}_{j=1}^r,\mathcal{S}_{\mathrm{MAX}}\right)\right). 0 ∈ Re ( conv F ( {( A 0 + W 0 ) B j } j = 1 r , S MAX ) ) .
Subdifferentials
First, we introduce the definition of the subdifferential of a convex function over a Banach space.
Definition 4.1. Let X \mathcal{X} X be a Banach space with dual X ∗ \mathcal{X}^* X ∗ , and let f : X → R f:\mathcal{X}\to\mathbb{R} f : X → R be a convex function. The subdifferential of f f f at x ∈ X x\in\mathcal{X} x ∈ X is
∂ f ( x ) = { v ∈ X ∗ : f ( y ) − f ( x ) ≥ Re v ( y − x ) , ∀ y ∈ X } . (24) \partial f(x)=\{v\in\mathcal{X}^*:f(y)-f(x)\geq\operatorname{Re}v(y-x),\ \forall y\in\mathcal{X}\}. \tag*{(24)} ∂ f ( x ) = { v ∈ X ∗ : f ( y ) − f ( x ) ≥ Re v ( y − x ) , ∀ y ∈ X } . (24)
In particular for f : M n ( C ) → R f:M_n(\mathbb{C})\to\mathbb{R} f : M n ( C ) → R , f ( A ) = ∥ A ∥ f(A)=\lVert A\rVert f ( A ) = ∥ A ∥ ,
∂ ∥ A ∥ = { G ∈ M n ( C ) : ∥ B ∥ − ∥ A ∥ ≥ Re tr ( G ( B − A ) ) , ∀ B ∈ M n ( C ) } . \partial\lVert A\rVert=\{G\in M_n(\mathbb{C}):\lVert B\rVert-\lVert A\rVert\geq\operatorname{Re}\operatorname{tr}(G(B-A)),\ \forall B\in M_n(\mathbb{C})\}. ∂ ∥ A ∥ = { G ∈ M n ( C ) : ∥ B ∥ − ∥ A ∥ ≥ Re tr ( G ( B − A )) , ∀ B ∈ M n ( C )} .
and it is well known that this definition is equivalent to
∂ ∥ A ∥ = { G ∈ M n ( C ) : ∥ A ∥ = Re tr ( G ∗ A ) , ∥ G ∥ 1 ≤ 1 } . (25) \partial\lVert A\rVert=\{G\in M_n(\mathbb{C}):\lVert A\rVert=\operatorname{Re}\operatorname{tr}(G^*A),\ \lVert G\rVert_1\leq1\}. \tag*{(25)} ∂ ∥ A ∥ = { G ∈ M n ( C ) : ∥ A ∥ = Re tr ( G ∗ A ) , ∥ G ∥ 1 ≤ 1 } . (25)
Moreover, by [20 Theorem 4] , the subdifferential of the spectral norm of any A ∈ M n ( C ) A\in M_n(\mathbb{C}) A ∈ M n ( C ) has the equivalent formulation
∂ ∥ A ∥ = conv { u v ∗ : A v = ∥ A ∥ u , ∥ v ∥ = ∥ u ∥ = 1 } . (26) \partial\lVert A\rVert=\operatorname{conv}\{uv^*:Av=\lVert A\rVert u,\ \lVert v\rVert=\lVert u\rVert=1\}. \tag*{(26)} ∂ ∥ A ∥ = conv { u v ∗ : A v = ∥ A ∥ u , ∥ v ∥ = ∥ u ∥ = 1 } . (26)
For a set A ⊆ M n ( C ) \mathcal{A}\subseteq M_n(\mathbb{C}) A ⊆ M n ( C ) , we write Re ( A ) = { Re ( X ) : X ∈ A } \operatorname{Re}(\mathcal{A})=\{\operatorname{Re}(X):X\in\mathcal{A}\} Re ( A ) = { Re ( X ) : X ∈ A } , where Re ( X ) = X + X ∗ 2 \operatorname{Re}(X)=\frac{X+X^*}{2} Re ( X ) = 2 X + X ∗ . We obtain the following result for Hermitian matrices.
Lemma 4.2. Let A ∈ M n h ( C ) A\in M_n^h(\mathbb{C}) A ∈ M n h ( C ) and let G = Re ( G ) + i Im ( G ) G=\operatorname{Re}(G)+i\operatorname{Im}(G) G = Re ( G ) + i Im ( G ) be the Cartesian decomposition of a matrix G ∈ M n ( C ) G\in M_n(\mathbb{C}) G ∈ M n ( C ) , that is, Re ( G ) , Im ( G ) ∈ M n h ( C ) \operatorname{Re}(G),\operatorname{Im}(G)\in M_n^h(\mathbb{C}) Re ( G ) , Im ( G ) ∈ M n h ( C ) . Then, with respect to the real inner product ⟨ X , Y ⟩ = Re tr ( X ∗ Y ) \langle X,Y\rangle=\operatorname{Re}\operatorname{tr}(X^*Y) ⟨ X , Y ⟩ = Re tr ( X ∗ Y ) ,
G ∈ ∂ ∥ A ∥ ⟺ tr ( Re ( G ) A ) = ∥ A ∥ and ∥ G ∥ 1 ≤ 1. (27) G\in\partial\lVert A\rVert\Longleftrightarrow\operatorname{tr}(\operatorname{Re}(G)A)=\lVert A\rVert\ \text{and}\ \lVert G\rVert_1\leq1. \tag*{(27)} G ∈ ∂ ∥ A ∥ ⟺ tr ( Re ( G ) A ) = ∥ A ∥ and ∥ G ∥ 1 ≤ 1. (27)
Moreover, if ∂ h ∥ A ∥ \partial_h\lVert A\rVert ∂ h ∥ A ∥ denotes the subdifferential of the restriction of ∥ ⋅ ∥ \lVert\cdot\rVert ∥ ⋅ ∥ to M n h ( C ) M_n^h(\mathbb{C}) M n h ( C ) , then
∂ h ∥ A ∥ = ∂ ∥ A ∥ ∩ M n h ( C ) = Re ( ∂ ∥ A ∥ ) \partial_h\lVert A\rVert=\partial\lVert A\rVert\cap M_n^h(\mathbb{C})=\operatorname{Re}(\partial\lVert A\rVert) ∂ h ∥ A ∥ = ∂ ∥ A ∥ ∩ M n h ( C ) = Re ( ∂ ∥ A ∥)
and
∂ h ∥ A ∥ = conv { s v v ∗ : s ∈ { 1 , − 1 } , ∥ v ∥ = 1 , A v = s ∥ A ∥ v } . (28) \partial_h\lVert A\rVert=\operatorname{conv}\{s vv^*:s\in\{1,-1\},\ \lVert v\rVert=1,\ Av=s\lVert A\rVert v\}. \tag*{(28)} ∂ h ∥ A ∥ = conv { s v v ∗ : s ∈ { 1 , − 1 } , ∥ v ∥ = 1 , A v = s ∥ A ∥ v } . (28)
Proof. By (25) , G ∈ ∂ ∥ A ∥ G\in\partial\lVert A\rVert G ∈ ∂ ∥ A ∥ if and only if Re tr ( G ∗ A ) = ∥ A ∥ \operatorname{Re}\operatorname{tr}(G^*A)=\lVert A\rVert Re tr ( G ∗ A ) = ∥ A ∥ and ∥ G ∥ 1 ≤ 1 \lVert G\rVert_1\leq1 ∥ G ∥ 1 ≤ 1 . Since G ∗ = Re ( G ) − i Im ( G ) G^*=\operatorname{Re}(G)-i\operatorname{Im}(G) G ∗ = Re ( G ) − i Im ( G ) and the trace of a product of two Hermitian matrices is real, Re tr ( G ∗ A ) = tr ( Re ( G ) A ) \operatorname{Re}\operatorname{tr}(G^*A)=\operatorname{tr}(\operatorname{Re}(G)A) Re tr ( G ∗ A ) = tr ( Re ( G ) A ) , which proves (27) .
Clearly ∂ ∥ A ∥ ∩ M n h ( C ) ⊆ ∂ h ∥ A ∥ \partial\lVert A\rVert\cap M_n^h(\mathbb{C})\subseteq\partial_h\lVert A\rVert ∂ ∥ A ∥ ∩ M n h ( C ) ⊆ ∂ h ∥ A ∥ . Conversely, let G ∈ ∂ h ∥ A ∥ G\in\partial_h\lVert A\rVert G ∈ ∂ h ∥ A ∥ and B ∈ M n ( C ) B\in M_n(\mathbb{C}) B ∈ M n ( C ) . Since G G G is Hermitian, ⟨ G , B ⟩ = tr ( G Re ( B ) ) \langle G,B\rangle=\operatorname{tr}(G\operatorname{Re}(B)) ⟨ G , B ⟩ = tr ( G Re ( B )) . Applying the subgradient inequality to the Hermitian matrix Re ( B ) \operatorname{Re}(B) Re ( B ) and using ∥ Re ( B ) ∥ ≤ ∥ B ∥ \lVert\operatorname{Re}(B)\rVert\leq\lVert B\rVert ∥ Re ( B )∥ ≤ ∥ B ∥ , we get
∥ A ∥ + ⟨ G , B − A ⟩ = ∥ A ∥ + ⟨ G , Re ( B ) − A ⟩ ≤ ∥ Re ( B ) ∥ ≤ ∥ B ∥ , \lVert A\rVert+\langle G,B-A\rangle=\lVert A\rVert+\langle G,\operatorname{Re}(B)-A\rangle\leq\lVert\operatorname{Re}(B)\rVert\leq\lVert B\rVert, ∥ A ∥ + ⟨ G , B − A ⟩ = ∥ A ∥ + ⟨ G , Re ( B ) − A ⟩ ≤ ∥ Re ( B )∥ ≤ ∥ B ∥ ,
so G ∈ ∂ ∥ A ∥ G \in\partial\|A\| G ∈ ∂ ∥ A ∥ . Hence ∂ h ∥ A ∥ = ∂ ∥ A ∥ ∩ M n h ( C ) \partial_h\|A\| = \partial\|A\| \cap M_n^h(\mathbb{C}) ∂ h ∥ A ∥ = ∂ ∥ A ∥ ∩ M n h ( C ) . If G ∈ ∂ ∥ A ∥ G \in\partial\|A\| G ∈ ∂ ∥ A ∥ , then G ∗ ∈ ∂ ∥ A ∥ G^* \in\partial\|A\| G ∗ ∈ ∂ ∥ A ∥ by (4.4), since ∥ G ∗ ∥ 1 = ∥ G ∥ 1 \|G^*\|_1 = \|G\|_1 ∥ G ∗ ∥ 1 = ∥ G ∥ 1 and Re ( G ∗ ) = Re ( G ) \operatorname{Re}(G^*) = \operatorname{Re}(G) Re ( G ∗ ) = Re ( G ) . Then, by convexity, Re ( G ) = G + G ∗ 2 ∈ ∂ ∥ A ∥ ∩ M n h ( C ) \operatorname{Re}(G) = \frac{G+G^*}{2} \in\partial\|A\| \cap M_n^h(\mathbb{C}) Re ( G ) = 2 G + G ∗ ∈ ∂ ∥ A ∥ ∩ M n h ( C ) . As Re ( G ) = G \operatorname{Re}(G) = G Re ( G ) = G for Hermitian G G G , it follows that ∂ h ∥ A ∥ = Re ( ∂ ∥ A ∥ ) \partial_h\|A\| = \operatorname{Re}(\partial\|A\|) ∂ h ∥ A ∥ = Re ( ∂ ∥ A ∥ ) .
If A = 0 A = 0 A = 0 , then ∂ h ∥ 0 ∥ = { G ∈ M n h ( C ) : ∥ G ∥ 1 ≤ 1 } \partial_h\|0\| = \{G \in M_n^h(\mathbb{C}) : \|G\|_1 \le1\} ∂ h ∥0∥ = { G ∈ M n h ( C ) : ∥ G ∥ 1 ≤ 1 } by (4.4), and every unit vector is admissible in (4.5) with both signs. Writing G = ∑ i μ i w i w i ∗ G = \sum_i \mu_i w_iw_i^* G = ∑ i μ i w i w i ∗ , with { w i } \{w_i\} { w i } orthonormal, we have G = ∑ i ∣ μ i ∣ ( sgn ( μ i ) w i w i ∗ ) + d 2 v v ∗ + d 2 ( − v v ∗ ) G = \sum_i |\mu_i|(\operatorname{sgn}(\mu_i)w_iw_i^*) + \frac{d}{2}vv^* + \frac{d}{2}(-vv^*) G = ∑ i ∣ μ i ∣ ( sgn ( μ i ) w i w i ∗ ) + 2 d v v ∗ + 2 d ( − v v ∗ ) , where d = 1 − ∥ G ∥ 1 ≥ 0 d = 1 - \|G\|_1 \ge0 d = 1 − ∥ G ∥ 1 ≥ 0 and v v v is any unit vector. This proves that G ∈ conv { ± v v ∗ : ∥ v ∥ = 1 } G \in\operatorname{conv}\{\pm vv^* : \|v\| = 1\} G ∈ conv { ± v v ∗ : ∥ v ∥ = 1 } , hence ∂ h ∥ 0 ∥ ⊆ conv { ± v v ∗ : ∥ v ∥ = 1 } \partial_h\|0\| \subseteq\operatorname{conv}\{\pm vv^* : \|v\| = 1\} ∂ h ∥0∥ ⊆ conv { ± v v ∗ : ∥ v ∥ = 1 } . The reverse inclusion holds because ∂ h ∥ 0 ∥ \partial_h\|0\| ∂ h ∥0∥ is convex and contains every ± v v ∗ \pm vv^* ± v v ∗ with ∥ v ∥ = 1 \|v\| = 1 ∥ v ∥ = 1 . Therefore (4.5) holds in this case.
Assume now A ≠ 0 A \ne0 A = 0 . By Watson’s formula (4.3), ∂ ∥ A ∥ \partial\|A\| ∂ ∥ A ∥ is the convex hull of the matrices u v ∗ uv^* u v ∗ with ∥ u ∥ = ∥ v ∥ = 1 \|u\| = \|v\| = 1 ∥ u ∥ = ∥ v ∥ = 1 and A v = ∥ A ∥ u Av = \|A\|u A v = ∥ A ∥ u . For such v v v we have v ∈ M A v \in M_A v ∈ M A , so by Proposition 2.6 we can write v = v + + v − v = v_+ + v_- v = v + + v − with v ± ∈ S ± ∥ A ∥ v_\pm\in S_{\pm\|A\|} v ± ∈ S ± ∥ A ∥ , and then u = A v ∥ A ∥ = v + − v − u = \frac{Av}{\|A\|} = v_+ - v_- u = ∥ A ∥ A v = v + − v − . Since Re \operatorname{Re} Re is linear, ∂ h ∥ A ∥ = Re ( ∂ ∥ A ∥ ) \partial_h\|A\| = \operatorname{Re}(\partial\|A\|) ∂ h ∥ A ∥ = Re ( ∂ ∥ A ∥ ) is the convex hull of the matrices
Re ( ( v + − v − ) ( v + + v − ) ∗ ) = v + v + ∗ − v − v − ∗ = ∥ v + ∥ 2 w + w + ∗ + ∥ v − ∥ 2 ( − w − w − ∗ ) , \operatorname{Re}\bigl((v_+ - v_-)(v_+ + v_-)^*\bigr) = v_+v_+^* - v_-v_-^* = \|v_+\|^2w_+w_+^* + \|v_-\|^2(-w_-w_-^*), Re ( ( v + − v − ) ( v + + v − ) ∗ ) = v + v + ∗ − v − v − ∗ = ∥ v + ∥ 2 w + w + ∗ + ∥ v − ∥ 2 ( − w − w − ∗ ) ,
where w ± = v ± / ∥ v ± ∥ w_\pm= v_\pm/\|v_\pm\| w ± = v ± /∥ v ± ∥ (omit a term if v ± = 0 v_\pm= 0 v ± = 0 ) and ∥ v + ∥ 2 + ∥ v − ∥ 2 = ∥ v ∥ 2 = 1 \|v_+\|^2 + \|v_-\|^2 = \|v\|^2 = 1 ∥ v + ∥ 2 + ∥ v − ∥ 2 = ∥ v ∥ 2 = 1 . Each of these is a convex combination of matrices s w w ∗ sww^* s w w ∗ with ∥ w ∥ = 1 \|w\| = 1 ∥ w ∥ = 1 and A w = s ∥ A ∥ w Aw = s\|A\|w A w = s ∥ A ∥ w . Conversely, each such s w w ∗ sww^* s w w ∗ is of this form (take v = w v = w v = w , so that v − = 0 v_- = 0 v − = 0 if s = 1 s = 1 s = 1 and v + = 0 v_+ = 0 v + = 0 if s = − 1 s = -1 s = − 1 ). This proves (4.5). □ \square □
Lemma 4.3. Let A ∈ M n h ( C ) ∖ { 0 } A \in M_n^h(\mathbb{C}) \setminus\{0\} A ∈ M n h ( C ) ∖ { 0 } , and let the sets T \mathcal{T} T , T + \mathcal{T}_+ T + and T − \mathcal{T}_- T − be defined as
T = { u v ∗ : A v = ∥ A ∥ u , ∥ v ∥ = ∥ u ∥ = 1 } , T + = { u u ∗ : A u = ∥ A ∥ u , ∥ u ∥ = 1 } , T − = { − u u ∗ : A u = − ∥ A ∥ u , ∥ u ∥ = 1 } . (29) \begin{aligned}
\mathcal{T} &= \{uv^* : Av = \|A\|u,\ \|v\| = \|u\| = 1\},\\
\mathcal{T}_+ &= \{uu^* : Au = \|A\|u,\ \|u\| = 1\},\\
\mathcal{T}_- &= \{-uu^* : Au = -\|A\|u,\ \|u\| = 1\}. \tag*{(29)}
\end{aligned} T T + T − = { u v ∗ : A v = ∥ A ∥ u , ∥ v ∥ = ∥ u ∥ = 1 } , = { u u ∗ : A u = ∥ A ∥ u , ∥ u ∥ = 1 } , = { − u u ∗ : A u = − ∥ A ∥ u , ∥ u ∥ = 1 } . (29) Then, T + ∩ T − = ∅ \mathcal{T}_+ \cap\mathcal{T}_- = \varnothing T + ∩ T − = ∅ and
∂ ∥ A ∥ = conv ( T ) = conv { 1 ∥ A ∥ A v v ∗ : v ∈ S M A X ( A ) , ∥ v ∥ = 1 } . (30) \partial\|A\| = \operatorname{conv}(\mathcal{T}) = \operatorname{conv}\left\{\frac{1}{\|A\|}Avv^* : v \in S_{\mathrm{MAX}}(A),\ \|v\| = 1\right\}. \tag*{(30)} ∂ ∥ A ∥ = conv ( T ) = conv { ∥ A ∥ 1 A v v ∗ : v ∈ S MAX ( A ) , ∥ v ∥ = 1 } . (30) Additionally,
∂ ∥ A ∥ = { conv ( T + ) if λ max = ∥ A ∥ > − λ min , conv ( T − ) if − λ min = ∥ A ∥ > λ max , (31) \partial\|A\| =
\begin{cases}
\operatorname{conv}(\mathcal{T}_+) & \text{if } \lambda_{\max} = \|A\| > -\lambda_{\min},\\
\operatorname{conv}(\mathcal{T}_-) & \text{if } -\lambda_{\min} = \|A\| > \lambda_{\max},
\end{cases} \tag*{(31)} ∂ ∥ A ∥ = { conv ( T + ) conv ( T − ) if λ m a x = ∥ A ∥ > − λ m i n , if − λ m i n = ∥ A ∥ > λ m a x , (31) and if λ max = − λ min = ∥ A ∥ \lambda_{\max} = -\lambda_{\min} = \|A\| λ m a x = − λ m i n = ∥ A ∥
∂ ∥ A ∥ = conv { ( v + − v − ) ( v + + v − ) ∗ : v + ∈ S ∥ A ∥ , v − ∈ S − ∥ A ∥ , ∥ v + ∥ 2 + ∥ v − ∥ 2 = 1 } . (32) \partial\|A\| = \operatorname{conv}\{(v_+ - v_-)(v_+ + v_-)^* : v_+ \in S_{\|A\|}, v_- \in S_{-\|A\|}, \|v_+\|^2 + \|v_-\|^2 = 1\}. \tag*{(32)} ∂ ∥ A ∥ = conv {( v + − v − ) ( v + + v − ) ∗ : v + ∈ S ∥ A ∥ , v − ∈ S − ∥ A ∥ , ∥ v + ∥ 2 + ∥ v − ∥ 2 = 1 } . (32) Finally,
∂ h ∥ A ∥ = Re ( ∂ ∥ A ∥ ) = conv ( T + ∪ T − ) . (33) \partial_h\|A\| = \operatorname{Re}(\partial\|A\|) = \operatorname{conv}(\mathcal{T}_+ \cup\mathcal{T}_-). \tag*{(33)} ∂ h ∥ A ∥ = Re ( ∂ ∥ A ∥ ) = conv ( T + ∪ T − ) . (33)
Proof. First observe that if v v v is a nonzero vector in S ∥ A ∥ S_{\|A\|} S ∥ A ∥ (resp. S − ∥ A ∥ S_{-\|A\|} S − ∥ A ∥ ), then v ∉ S − ∥ A ∥ v \notin S_{-\|A\|} v ∈ / S − ∥ A ∥ (v ∉ S ∥ A ∥ v \notin S_{\|A\|} v ∈ / S ∥ A ∥ ), so v v ∗ ∈ T + vv^* \in\mathcal{T}_+ v v ∗ ∈ T + (v v ∗ ∈ T − vv^* \in\mathcal{T}_- v v ∗ ∈ T − ) but v v ∗ ∉ T − vv^* \notin\mathcal{T}_- v v ∗ ∈ / T − (v v ∗ ∉ T + vv^* \notin\mathcal{T}_+ v v ∗ ∈ / T + ). Therefore, T + ∩ T − = ∅ \mathcal{T}_+ \cap\mathcal{T}_- = \varnothing T + ∩ T − = ∅ .
By (4.3), ∂ ∥ A ∥ = conv ( T ) \partial\|A\| = \operatorname{conv}(\mathcal{T}) ∂ ∥ A ∥ = conv ( T ) . Moreover, if u v ∗ ∈ T uv^* \in\mathcal{T} u v ∗ ∈ T , then A v = ∥ A ∥ u Av = \|A\|u A v = ∥ A ∥ u with ∥ v ∥ = ∥ u ∥ = 1 \|v\| = \|u\| = 1 ∥ v ∥ = ∥ u ∥ = 1 , so ∥ A v ∥ = ∥ A ∥ \|Av\| = \|A\| ∥ A v ∥ = ∥ A ∥ , v ∈ M A ⊆ S M A X ( A ) v \in M_A \subseteq S_{\mathrm{MAX}}(A) v ∈ M A ⊆ S MAX ( A ) by item 4 of Proposition 2.6, and u = 1 ∥ A ∥ A v u = \frac{1}{\|A\|}Av u = ∥ A ∥ 1 A v . Conversely, if v ∈ S M A X ( A ) v \in S_{\mathrm{MAX}}(A) v ∈ S MAX ( A ) and ∥ v ∥ = 1 \|v\| = 1 ∥ v ∥ = 1 , then ∥ A v ∥ = ∥ A ∥ \|Av\| = \|A\| ∥ A v ∥ = ∥ A ∥ by the same item, so u = 1 ∥ A ∥ A v u = \frac{1}{\|A\|}Av u = ∥ A ∥ 1 A v is a unit vector with A v = ∥ A ∥ u Av = \|A\|u A v = ∥ A ∥ u , and u v ∗ ∈ T uv^* \in\mathcal{T} u v ∗ ∈ T . This proves (4.7). Now fix u v ∗ ∈ T uv^* \in\mathcal{T} u v ∗ ∈ T . We consider the following cases:
If λ max = ∥ A ∥ > − λ min \lambda_{\max} = \|A\| > -\lambda_{\min} λ m a x = ∥ A ∥ > − λ m i n : here S M A X ( A ) = S ∥ A ∥ S_{\mathrm{MAX}}(A) = S_{\|A\|} S MAX ( A ) = S ∥ A ∥ and it follows that u = 1 ∥ A ∥ A v = v u = \frac{1}{\|A\|}Av = v u = ∥ A ∥ 1 A v = v . Moreover, u v ∗ = v v ∗ ∈ T + uv^* = vv^* \in\mathcal{T}_+ u v ∗ = v v ∗ ∈ T + , then T = T + \mathcal{T} = \mathcal{T}_+ T = T + and ∂ ∥ A ∥ = conv ( T + ) \partial\|A\| = \operatorname{conv}(\mathcal{T}_+) ∂ ∥ A ∥ = conv ( T + ) .
If − λ min = ∥ A ∥ > λ max -\lambda_{\min} = \|A\| > \lambda_{\max} − λ m i n = ∥ A ∥ > λ m a x : in this case S M A X ( A ) = S − ∥ A ∥ S_{\mathrm{MAX}}(A) = S_{-\|A\|} S MAX ( A ) = S − ∥ A ∥ and it follows that u = 1 ∥ A ∥ A v = − v u = \frac{1}{\|A\|}Av = -v u = ∥ A ∥ 1 A v = − v . Moreover, u v ∗ = − v v ∗ ∈ T − uv^* = -vv^* \in\mathcal{T}_- u v ∗ = − v v ∗ ∈ T − , then T = T − \mathcal{T} = \mathcal{T}_- T = T − and ∂ ∥ A ∥ = conv ( T − ) \partial\|A\| = \operatorname{conv}(\mathcal{T}_-) ∂ ∥ A ∥ = conv ( T − ) .
If λ max = ∥ A ∥ = − λ min \lambda_{\max}=\lVert A\rVert=-\lambda_{\min} λ m a x = ∥ A ∥ = − λ m i n : v ∈ S M A X ( A ) = S ∥ A ∥ ⊕ S − ∥ A ∥ v\in S_{\mathrm{MAX}}(A)=S_{\lVert A\rVert}\oplus S_{-\lVert A\rVert} v ∈ S MAX ( A ) = S ∥ A ∥ ⊕ S − ∥ A ∥ , then there exist v + ∈ S ∥ A ∥ v_{+}\in S_{\lVert A\rVert} v + ∈ S ∥ A ∥ and v − ∈ S − ∥ A ∥ v_{-}\in S_{-\lVert A\rVert} v − ∈ S − ∥ A ∥ such that v = v + + v − v=v_{+}+v_{-} v = v + + v − and ∥ v ∥ 2 = ∥ v + ∥ 2 + ∥ v − ∥ 2 = 1 \lVert v\rVert^{2}=\lVert v_{+}\rVert^{2}+\lVert v_{-}\rVert^{2}=1 ∥ v ∥ 2 = ∥ v + ∥ 2 + ∥ v − ∥ 2 = 1 . By item 5 in Proposition 2.6, u = 1 ∥ A ∥ A v = v + − v − u=\frac{1}{\lVert A\rVert}Av=v_{+}-v_{-} u = ∥ A ∥ 1 A v = v + − v − . Moreover,
u v ∗ = ( v + − v − ) ( v + + v − ) ∗ , uv^{*}=(v_{+}-v_{-})(v_{+}+v_{-})^{*}, u v ∗ = ( v + − v − ) ( v + + v − ) ∗ ,
which proves (32) .
Finally, by (29) , T + ∪ T − = { s v v ∗ : s ∈ { 1 , − 1 } , ∥ v ∥ = 1 , A v = s ∥ A ∥ v } \mathcal{T}_{+}\cup\mathcal{T}_{-}=\{svv^{*}:s\in\{1,-1\},\lVert v\rVert=1,Av=s\lVert A\rVert v\} T + ∪ T − = { s v v ∗ : s ∈ { 1 , − 1 } , ∥ v ∥ = 1 , A v = s ∥ A ∥ v } , so (33) is a restatement of Lemma 4.2. □
We recall the following chain rule, which we use to obtain a closed form for the subdifferential of the norm of any Hermitian A ( x ) A(x) A ( x ) .
Lemma 4.4 (Chain rule [9 , 11 ] ). Let X \mathcal{X} X and Y \mathcal{Y} Y be Banach spaces over the field R \mathbb{R} R (or C \mathbb{C} C ). Let S : X → Y S:\mathcal{X}\to\mathcal{Y} S : X → Y be a bounded linear map, and let L : X → Y L:\mathcal{X}\to\mathcal{Y} L : X → Y be the affine map defined as L ( x ) = S ( x ) + y 0 L(x)=S(x)+y_{0} L ( x ) = S ( x ) + y 0 , where y 0 ∈ Y y_{0}\in\mathcal{Y} y 0 ∈ Y . If g : Y → R g:\mathcal{Y}\to\mathbb{R} g : Y → R is a continuous convex function, then
∂ ( g ∘ L ) ( a ) = S ∗ ∂ g ( L ( a ) ) , for all a ∈ X , \partial(g\circ L)(a)=S^{*}\partial g(L(a)),\quad\text{for all }a\in\mathcal{X}, ∂ ( g ∘ L ) ( a ) = S ∗ ∂ g ( L ( a )) , for all a ∈ X , where S ∗ S^{*} S ∗ denotes the adjoint of the operator S S S , either real or complex, depending on whether X \mathcal{X} X and Y \mathcal{Y} Y are real or complex Banach spaces.
Theorem 4.5. For A ( x ) = A 0 + ∑ k = 1 r x k B k A(x)=A_{0}+\sum_{k=1}^{r}x_{k}B^{k} A ( x ) = A 0 + ∑ k = 1 r x k B k , with A 0 ∈ M n h ( C ) A_{0}\in M_{n}^{h}(\mathbb{C}) A 0 ∈ M n h ( C ) and { B k } k = 1 r \{B^{k}\}_{k=1}^{r} { B k } k = 1 r the fixed orthonormal basis of W W W , x ∈ R r x\in\mathbb{R}^{r} x ∈ R r , let λ min ( A ( x ) ) \lambda_{\min}(A(x)) λ m i n ( A ( x )) and λ max ( A ( x ) ) \lambda_{\max}(A(x)) λ m a x ( A ( x )) be the minimum and maximum eigenvalues of A ( x ) A(x) A ( x ) , respectively. Then, the subdifferential of the spectral norm of A ( x ) A(x) A ( x ) is
∂ ( ∥ A ( ⋅ ) ∥ ) ( x ) = Φ ( Re ( ∂ ∥ A ( x ) ∥ ) ) = Φ ( conv ( T + ∪ T − ) ) \partial(\lVert A(\cdot)\rVert)(x)=\Phi\bigl(\operatorname{Re}(\partial\lVert A(x)\rVert)\bigr)=\Phi\bigl(\operatorname{conv}(\mathcal{T}_{+}\cup\mathcal{T}_{-})\bigr) ∂ (∥ A ( ⋅ )∥) ( x ) = Φ ( Re ( ∂ ∥ A ( x )∥) ) = Φ ( conv ( T + ∪ T − ) ) = { Φ ( conv ( T + ) ) if λ max + λ min > 0 , Φ ( conv ( T − ) ) if λ max + λ min < 0 , Φ ( conv ( T + ∪ T − ) ) if λ max + λ min = 0 , (34) =\begin{cases}
\Phi\bigl(\operatorname{conv}(\mathcal{T}_{+})\bigr) & \text{if }\lambda_{\max}+\lambda_{\min}>0,\\
\Phi\bigl(\operatorname{conv}(\mathcal{T}_{-})\bigr) & \text{if }\lambda_{\max}+\lambda_{\min}<0,\\
\Phi\bigl(\operatorname{conv}(\mathcal{T}_{+}\cup\mathcal{T}_{-})\bigr) & \text{if }\lambda_{\max}+\lambda_{\min}=0,
\end{cases} \tag*{(34)} = ⎩ ⎨ ⎧ Φ ( conv ( T + ) ) Φ ( conv ( T − ) ) Φ ( conv ( T + ∪ T − ) ) if λ m a x + λ m i n > 0 , if λ m a x + λ m i n < 0 , if λ m a x + λ m i n = 0 , (34) where λ max = λ max ( A ( x ) ) \lambda_{\max}=\lambda_{\max}(A(x)) λ m a x = λ m a x ( A ( x )) , λ min = λ min ( A ( x ) ) \lambda_{\min}=\lambda_{\min}(A(x)) λ m i n = λ m i n ( A ( x )) , Φ = Φ B \Phi=\Phi^{B} Φ = Φ B is the map defined in (4) , and T + , T − \mathcal{T}_{+},\mathcal{T}_{-} T + , T − are the sets defined in (29) for A = A ( x ) A=A(x) A = A ( x ) .
Proof. Apply Lemma 4.4 with X = R r \mathcal{X}=\mathbb{R}^{r} X = R r , Y = M n h ( C ) \mathcal{Y}=M_{n}^{h}(\mathbb{C}) Y = M n h ( C ) endowed with the real inner product ⟨ X , Y ⟩ = tr ( X Y ) \langle X,Y\rangle=\operatorname{tr}(XY) ⟨ X , Y ⟩ = tr ( X Y ) , S = W S=W S = W , where W ( x ) = ∑ k = 1 r x k B k W(x)=\sum_{k=1}^{r}x_{k}B^{k} W ( x ) = ∑ k = 1 r x k B k , L ( x ) = A ( x ) = A 0 + W ( x ) L(x)=A(x)=A_{0}+W(x) L ( x ) = A ( x ) = A 0 + W ( x ) , and g : M n h ( C ) → R g:M_{n}^{h}(\mathbb{C})\to\mathbb{R} g : M n h ( C ) → R , g ( R ) = ∥ R ∥ g(R)=\lVert R\rVert g ( R ) = ∥ R ∥ . Then g ∘ L = ∥ A ( ⋅ ) ∥ g\circ L=\lVert A(\cdot)\rVert g ∘ L = ∥ A ( ⋅ )∥ and ∂ g ( A ( x ) ) = ∂ h ∥ A ( x ) ∥ \partial g(A(x))=\partial_{h}\lVert A(x)\rVert ∂ g ( A ( x )) = ∂ h ∥ A ( x )∥ is the subdifferential of the restriction of the norm to M n h ( C ) M_{n}^{h}(\mathbb{C}) M n h ( C ) . For x ∈ R r x\in\mathbb{R}^{r} x ∈ R r and R ∈ M n h ( C ) R\in M_{n}^{h}(\mathbb{C}) R ∈ M n h ( C ) ,
⟨ W ( x ) , R ⟩ = ∑ k = 1 r x k tr ( B k R ) = ⟨ x , Φ ( R ) ⟩ R r , \langle W(x),R\rangle=\sum_{k=1}^{r}x_{k}\operatorname{tr}(B^{k}R)=\langle x,\Phi(R)\rangle_{\mathbb{R}^{r}}, ⟨ W ( x ) , R ⟩ = k = 1 ∑ r x k tr ( B k R ) = ⟨ x , Φ ( R ) ⟩ R r , so W ∗ = Φ W^{*}=\Phi W ∗ = Φ and
∂ ( ∥ A ( ⋅ ) ∥ ) ( x ) = Φ ( ∂ h ∥ A ( x ) ∥ ) . (35) \partial(\lVert A(\cdot)\rVert)(x)=\Phi\bigl(\partial_{h}\lVert A(x)\rVert\bigr). \tag*{(35)} ∂ (∥ A ( ⋅ )∥) ( x ) = Φ ( ∂ h ∥ A ( x )∥ ) . (35) By Lemma 4.2 we have ∂ h ∥ A ( x ) ∥ = Re ( ∂ ∥ A ( x ) ∥ ) \partial_{h}\lVert A(x)\rVert=\operatorname{Re}(\partial\lVert A(x)\rVert) ∂ h ∥ A ( x )∥ = Re ( ∂ ∥ A ( x )∥) , and by (33) in Lemma 4.3 we have Re ( ∂ ∥ A ( x ) ∥ ) = conv ( T + ∪ T − ) \operatorname{Re}(\partial\lVert A(x)\rVert)=\operatorname{conv}(\mathcal{T}_{+}\cup\mathcal{T}_{-}) Re ( ∂ ∥ A ( x )∥) = conv ( T + ∪ T − ) . Finally, if λ max + λ min > 0 \lambda_{\max}+\lambda_{\min}>0 λ m a x + λ m i n > 0 (resp. < 0 <0 < 0 ) for A ( x ) A(x) A ( x ) , then T − = ∅ \mathcal{T}_{-}=\varnothing T − = ∅ (resp. T + = ∅ \mathcal{T}_{+}=\varnothing T + = ∅ ), which gives the three cases in (34) . □
Formulas for the subdifferentials of λ max \lambda_{\max} λ m a x and λ min \lambda_{\min} λ m i n at A A A were given in [16 , 6 , 7 ] . These are
∂ λ max ( A ) = conv { q q ∗ : A q = λ max ( A ) q and ∥ q ∥ = 1 } \partial\lambda_{\max}(A)=\operatorname{conv}\{qq^{*}:Aq=\lambda_{\max}(A)q\text{ and }\lVert q\rVert=1\} ∂ λ m a x ( A ) = conv { q q ∗ : A q = λ m a x ( A ) q and ∥ q ∥ = 1 }
∂ λ min ( A ) = − ∂ λ max ( − A ) = − conv { q q ∗ : A q = λ min ( A ) q and ∥ q ∥ = 1 } . (36) \partial\lambda_{\min}(A)=-\partial\lambda_{\max}(-A)=-\operatorname{conv}\{qq^{*}:Aq=\lambda_{\min}(A)q\text{ and }\lVert q\rVert=1\}. \tag*{(36)} ∂ λ m i n ( A ) = − ∂ λ m a x ( − A ) = − conv { q q ∗ : A q = λ m i n ( A ) q and ∥ q ∥ = 1 } . (36)
If S λ max S_{\lambda_{\max}} S λ m a x and S λ min S_{\lambda_{\min}} S λ m i n are the eigenspaces of λ max ( A ( x ) ) \lambda_{\max}(A(x)) λ m a x ( A ( x )) and λ min ( A ( x ) ) \lambda_{\min}(A(x)) λ m i n ( A ( x )) , respectively, in [7 Theorem 3 and Proposition 3] we showed that
∂ ( λ max ∘ A ) ( x ) = m S λ max and ∂ ( λ min ∘ A ) ( x ) = − m S λ min . (37) \partial(\lambda_{\max}\circ A)(x)=m_{S_{\lambda_{\max}}}\quad\text{and}\quad\partial(\lambda_{\min}\circ A)(x)=-m_{S_{\lambda_{\min}}}. \tag*{(37)} ∂ ( λ m a x ∘ A ) ( x ) = m S λ m a x and ∂ ( λ m i n ∘ A ) ( x ) = − m S λ m i n . (37)
The next result follows directly from the formulas (4.11) and (4.14).
Corollary 4.6. Under the same hypotheses as in Theorem 4.5, the following statements hold for A ( x ) A(x) A ( x ) :
(1) If ∥ A ( x ) ∥ = λ max ( A ( x ) ) > − λ min ( A ( x ) ) \lVert A(x)\rVert=\lambda_{\max}(A(x))>-\lambda_{\min}(A(x)) ∥ A ( x )∥ = λ m a x ( A ( x )) > − λ m i n ( A ( x )) , then
∂ ( ∥ A ( ⋅ ) ∥ ) ( x ) = ∂ ( λ max ∘ A ) ( x ) = m S λ max = m S MAX ( A ( x ) ) . \partial(\lVert A(\cdot)\rVert)(x)=\partial(\lambda_{\max}\circ A)(x)=mS_{\lambda_{\max}}=mS_{\operatorname{MAX}(A(x))}. ∂ (∥ A ( ⋅ )∥) ( x ) = ∂ ( λ m a x ∘ A ) ( x ) = m S λ m a x = m S MAX ( A ( x )) . (2) If ∥ A ( x ) ∥ = − λ min ( A ( x ) ) > λ max ( A ( x ) ) \lVert A(x)\rVert=-\lambda_{\min}(A(x))>\lambda_{\max}(A(x)) ∥ A ( x )∥ = − λ m i n ( A ( x )) > λ m a x ( A ( x )) , then
∂ ( ∥ A ( ⋅ ) ∥ ) ( x ) = ∂ ( λ min ∘ A ) ( x ) = − m S λ min = − m S MAX ( A ( x ) ) . \partial(\lVert A(\cdot)\rVert)(x)=\partial(\lambda_{\min}\circ A)(x)=-mS_{\lambda_{\min}}=-mS_{\operatorname{MAX}(A(x))}. ∂ (∥ A ( ⋅ )∥) ( x ) = ∂ ( λ m i n ∘ A ) ( x ) = − m S λ m i n = − m S MAX ( A ( x )) . (3) If ∥ A ( x ) ∥ = − λ min ( A ( x ) ) = λ max ( A ( x ) ) \lVert A(x)\rVert=-\lambda_{\min}(A(x))=\lambda_{\max}(A(x)) ∥ A ( x )∥ = − λ m i n ( A ( x )) = λ m a x ( A ( x )) , then
∂ ( ∥ A ( ⋅ ) ∥ ) ( x ) = conv ( ∂ ( λ max ∘ A ) ( x ) ∪ ∂ ( λ min ∘ A ) ( x ) ) = conv ( m S λ max ∪ − m S λ min ) . \partial(\lVert A(\cdot)\rVert)(x)=\operatorname{conv}\left(\partial(\lambda_{\max}\circ A)(x)\cup\partial(\lambda_{\min}\circ A)(x)\right)
=\operatorname{conv}\left(mS_{\lambda_{\max}}\cup-mS_{\lambda_{\min}}\right). ∂ (∥ A ( ⋅ )∥) ( x ) = conv ( ∂ ( λ m a x ∘ A ) ( x ) ∪ ∂ ( λ m i n ∘ A ) ( x ) ) = conv ( m S λ m a x ∪ − m S λ m i n ) . In all three cases, with the convention m { 0 } = ∅ m\{0\}=\varnothing m { 0 } = ∅ and S ± ∥ A ( x ) ∥ = { v ∈ C n : A ( x ) v = ± ∥ A ( x ) ∥ v } S_{\pm\lVert A(x)\rVert}=\{v\in\mathbb{C}^{n}:A(x)v=\pm\lVert A(x)\rVert v\} S ± ∥ A ( x )∥ = { v ∈ C n : A ( x ) v = ± ∥ A ( x )∥ v } ,
∂ ( ∥ A ( ⋅ ) ∥ ) ( x ) = conv ( m S ∥ A ( x ) ∥ ∪ − m S − ∥ A ( x ) ∥ ) . \partial(\lVert A(\cdot)\rVert)(x)=\operatorname{conv}\left(mS_{\lVert A(x)\rVert}\cup-mS_{-\lVert A(x)\rVert}\right). ∂ (∥ A ( ⋅ )∥) ( x ) = conv ( m S ∥ A ( x )∥ ∪ − m S − ∥ A ( x )∥ ) .
Characterizations of W \mathcal{W} W -minimal matrices
In this section we prove several characterizations of W \mathcal{W} W -minimal matrices. Throughout the section we drop the superscript and write Φ \Phi Φ for the map Φ B \Phi^{\mathcal{B}} Φ B of (2.4), defined on the whole of M n h ( C ) M_n^h(\mathbb{C}) M n h ( C ) , as well as for its restriction (3.7) to the set D S D_S D S of density matrices supported in a subspace S S S .
Theorem 5.1. Let W ⊆ M n h ( C ) \mathcal{W}\subseteq M_n^h(\mathbb{C}) W ⊆ M n h ( C ) be a subspace. Then A ∈ M n h ( C ) A\in M_n^h(\mathbb{C}) A ∈ M n h ( C ) is W \mathcal{W} W -minimal if and only if there exists X ∈ W 1 ∘ X\in\mathcal{W}_1^\circ X ∈ W 1 ∘ such that
A X = ∥ A ∥ ∣ X ∣ , AX=\lVert A\rVert\lvert X\rvert, A X = ∥ A ∥ ∣ X ∣ , where ∣ X ∣ \lvert X\rvert ∣ X ∣ denotes the modulus ( X ∗ X ) 1 / 2 (X^*X)^{1/2} ( X ∗ X ) 1/2 .
Proof. This is [21 Theorem 2.6] , and the proof given there remains valid not only for C ∗ C^* C ∗ -subalgebras of M n ( C ) M_n(\mathbb{C}) M n ( C ) , but also for arbitrary real subspaces of M n h ( C ) M_n^h(\mathbb{C}) M n h ( C ) , since the computations do not use the product of W \mathcal{W} W . □
When I ∈ W I\in\mathcal{W} I ∈ W , we derive the following characterization involving the intersection of the moments of the eigenspaces of the maximum and minimum eigenvalues. Observe that if I ∈ W I\in\mathcal{W} I ∈ W and A ≠ 0 A\ne0 A = 0 is W \mathcal{W} W -minimal, then { ∥ A ∥ , − ∥ A ∥ } ⊆ σ ( A ) \{\lVert A\rVert,-\lVert A\rVert\}\subseteq\sigma(A) {∥ A ∥ , − ∥ A ∥} ⊆ σ ( A ) (see also [1 ] ).
Theorem 5.2. Let A ∈ M n h ( C ) A\in M_n^h(\mathbb{C}) A ∈ M n h ( C ) , A ≠ 0 A\ne0 A = 0 , W ⊆ M n h ( C ) \mathcal{W}\subseteq M_n^h(\mathbb{C}) W ⊆ M n h ( C ) be a real subspace with I ∈ W I\in\mathcal{W} I ∈ W , and let E + E_+ E + (respectively E − E_- E − ) be the spectral projection of A A A corresponding to the eigenvalue λ max = λ max ( A ) \lambda_{\max}=\lambda_{\max}(A) λ m a x = λ m a x ( A ) (respectively λ min = λ min ( A ) \lambda_{\min}=\lambda_{\min}(A) λ m i n = λ m i n ( A ) ). Suppose that λ max + λ min = 0 \lambda_{\max}+\lambda_{\min}=0 λ m a x + λ m i n = 0 , and let B = { B i } i = 1 d \mathcal{B}=\{B_i\}_{i=1}^{d} B = { B i } i = 1 d be a fixed orthonormal basis of Hermitian matrices of W \mathcal{W} W , with respect to which the moments m S = m S B m_S=m_S^{\mathcal{B}} m S = m S B of Definition 3.2 are computed. By (3.8) the conditions below do not depend on this choice, since a change of basis acts on the moments through an invertible matrix. The following statements are equivalent:
(1) A A A is W \mathcal{W} W -minimal.
(2) There exists X ∈ W 1 ∘ X\in\mathcal{W}_1^\circ X ∈ W 1 ∘ such that
X = E + X E + + E − X E − , with E + X E + ≥ 0 , E − X E − ≤ 0 X=E_+XE_++E_-XE_-,\qquad\text{with }E_+XE_+\ge0,\quad E_-XE_-\le0 X = E + X E + + E − X E − , with E + X E + ≥ 0 , E − X E − ≤ 0 tr ( A X ) = ∥ A ∥ ∥ X ∥ 1 . (38) \operatorname{tr}(AX)=\lVert A\rVert\lVert X\rVert_1. \tag*{(38)} tr ( A X ) = ∥ A ∥ ∥ X ∥ 1 . (38) (3) m S λ max ∩ m S λ min ≠ ∅ mS_{\lambda_{\max}}\cap mS_{\lambda_{\min}}\ne\varnothing m S λ m a x ∩ m S λ m i n = ∅ .
(4) For every W ∈ W W\in\mathcal{W} W ∈ W ,
min y ∈ S λ max ∥ y ∥ = 1 ⟨ W y , y ⟩ ≤ max z ∈ S λ min ∥ z ∥ = 1 ⟨ W z , z ⟩ . \min_{\substack{y\in S_{\lambda_{\max}}\\ \lVert y\rVert=1}}\langle Wy,y\rangle
\le
\max_{\substack{z\in S_{\lambda_{\min}}\\ \lVert z\rVert=1}}\langle Wz,z\rangle. y ∈ S λ m a x ∥ y ∥ = 1 min ⟨ W y , y ⟩ ≤ z ∈ S λ m i n ∥ z ∥ = 1 max ⟨ W z , z ⟩ .
Proof. We prove the implications (1) ⇒ \Rightarrow ⇒ (2) ⇒ \Rightarrow ⇒ (3) ⇒ \Rightarrow ⇒ (4) and (4) ⇒ \Rightarrow ⇒ (3) ⇒ \Rightarrow ⇒ (2) ⇒ \Rightarrow ⇒ (1).
(1) ⇒ \Rightarrow ⇒ (2) Assume that A A A is W W W -minimal, and suppose without loss of generality that ∥ A ∥ = 1 \lVert A\rVert=1 ∥ A ∥ = 1 . The classical characterization of best approximation in a normed space, see [19 Theorem 1.1] , provides a norm one functional that annihilates W W W and attains ∥ A ∥ \lVert A\rVert ∥ A ∥ at A A A . Under the trace duality ( M n ( C ) , ∥ ⋅ ∥ ) ∗ = ( M n ( C ) , ∥ ⋅ ∥ 1 ) (M_n(\mathbb{C}),\lVert\cdot\rVert)^*=(M_n(\mathbb{C}),\lVert\cdot\rVert_1) ( M n ( C ) , ∥ ⋅ ∥ ) ∗ = ( M n ( C ) , ∥ ⋅ ∥ 1 ) this functional is given by a Hermitian matrix X X X with
X ∈ W 1 ∘ , tr ( A X ) = dist ( A , W ) = 1. (39) X\in W^\circ_1,\qquad\operatorname{tr}(AX)=\operatorname{dist}(A,W)=1. \tag*{(39)} X ∈ W 1 ∘ , tr ( A X ) = dist ( A , W ) = 1. (39)
It remains to prove the support conditions, and for this the argument of [1 Theorem 2] applies verbatim, so we include it here for the sake of completeness. Put E 1 = E + E_1=E_+ E 1 = E + , E 2 = E − E_2=E_- E 2 = E − and E 3 = I − E 1 − E 2 E_3=I-E_1-E_2 E 3 = I − E 1 − E 2 , so that E 3 E_3 E 3 is the spectral projection of A A A corresponding to the open interval ( − 1 , 1 ) (-1,1) ( − 1 , 1 ) ; then E 3 A = A E 3 E_3A=AE_3 E 3 A = A E 3 , ∥ A E 3 ∥ < 1 \lVert AE_3\rVert<1 ∥ A E 3 ∥ < 1 and A = E 1 − E 2 + A E 3 A=E_1-E_2+AE_3 A = E 1 − E 2 + A E 3 . Write X = ( X k , l ) k , l = 1 3 X=(X_{k,l})_{k,l=1}^3 X = ( X k , l ) k , l = 1 3 and A = ( A k , l ) k , l = 1 3 A=(A_{k,l})_{k,l=1}^3 A = ( A k , l ) k , l = 1 3 in the orthogonal decomposition C n = ran ( E 1 ) ⊕ ran ( E 2 ) ⊕ ran ( E 3 ) \mathbb{C}^n=\operatorname{ran}(E_1)\oplus\operatorname{ran}(E_2)\oplus\operatorname{ran}(E_3) C n = ran ( E 1 ) ⊕ ran ( E 2 ) ⊕ ran ( E 3 ) . The pinching inequality of Chandler Davis, see [2 IV.52] , gives, writing X ^ = ( X 1 , 1 X 1 , 2 X 2 , 1 X 2 , 2 ) \widehat{X}=\begin{pmatrix}X_{1,1}&X_{1,2}\\X_{2,1}&X_{2,2}\end{pmatrix} X = ( X 1 , 1 X 2 , 1 X 1 , 2 X 2 , 2 ) for the corner of X X X along ran ( E 1 ) ⊕ ran ( E 2 ) \operatorname{ran}(E_1)\oplus\operatorname{ran}(E_2) ran ( E 1 ) ⊕ ran ( E 2 ) ,
∥ X ^ ∥ 1 + ∥ X 3 , 3 ∥ 1 ≤ ∥ X ∥ 1 and ∥ X 1 , 1 ∥ 1 + ∥ X 2 , 2 ∥ 1 ≤ ∥ X ^ ∥ 1 . \lVert\widehat{X}\rVert_1+\lVert X_{3,3}\rVert_1\leq\lVert X\rVert_1\qquad\text{and}\qquad\lVert X_{1,1}\rVert_1+\lVert X_{2,2}\rVert_1\leq\lVert\widehat{X}\rVert_1. ∥ X ∥ 1 + ∥ X 3 , 3 ∥ 1 ≤ ∥ X ∥ 1 and ∥ X 1 , 1 ∥ 1 + ∥ X 2 , 2 ∥ 1 ≤ ∥ X ∥ 1 .
Together with ∥ A 3 , 3 ∥ < 1 \lVert A_{3,3}\rVert<1 ∥ A 3 , 3 ∥ < 1 and the identity tr ( A X ) = 1 = ∥ X ∥ 1 \operatorname{tr}(AX)=1=\lVert X\rVert_1 tr ( A X ) = 1 = ∥ X ∥ 1 these inequalities force X 3 , 3 = 0 X_{3,3}=0 X 3 , 3 = 0 as well as tr ( X 1 , 1 ) = ∥ X 1 , 1 ∥ 1 \operatorname{tr}(X_{1,1})=\lVert X_{1,1}\rVert_1 tr ( X 1 , 1 ) = ∥ X 1 , 1 ∥ 1 and tr ( − X 2 , 2 ) = ∥ X 2 , 2 ∥ 1 \operatorname{tr}(-X_{2,2})=\lVert X_{2,2}\rVert_1 tr ( − X 2 , 2 ) = ∥ X 2 , 2 ∥ 1 , whence X 1 , 1 ≥ 0 X_{1,1}\geq0 X 1 , 1 ≥ 0 and X 2 , 2 ≤ 0 X_{2,2}\leq0 X 2 , 2 ≤ 0 ; applying the same argument to A X AX A X and then to X ^ \widehat{X} X yields X k , l = 0 X_{k,l}=0 X k , l = 0 for k ≠ l k\ne l k = l . Therefore X = X 1 , 1 ⊕ X 2 , 2 ⊕ 0 X=X_{1,1}\oplus X_{2,2}\oplus0 X = X 1 , 1 ⊕ X 2 , 2 ⊕ 0 with X + = X 1 , 1 = E + X E + X_+=X_{1,1}=E_+XE_+ X + = X 1 , 1 = E + X E + and X − = − X 2 , 2 = − E − X E − X_-=-X_{2,2}=-E_-XE_- X − = − X 2 , 2 = − E − X E − , which is (38) .
(2) ⇒ \Rightarrow ⇒ (3) Let X X X be as in item (2), and put X + = E + X E + ≥ 0 X_+=E_+XE_+\geq0 X + = E + X E + ≥ 0 and X − = − E − X E − ≥ 0 X_-=-E_-XE_-\geq0 X − = − E − X E − ≥ 0 , so that X = X + − X − X=X_+-X_- X = X + − X − . Since I ∈ W I\in W I ∈ W and X ∈ W 1 ∘ X\in W^\circ_1 X ∈ W 1 ∘ we have tr ( X ) = 0 \operatorname{tr}(X)=0 tr ( X ) = 0 , hence t = tr ( X + ) = tr ( X − ) t=\operatorname{tr}(X_+)=\operatorname{tr}(X_-) t = tr ( X + ) = tr ( X − ) , and t > 0 t>0 t > 0 because X ≠ 0 X\ne0 X = 0 . Then,
ρ 1 = t − 1 X + = t − 1 E + X E + ∈ D S λ max , \rho_1=t^{-1}X_+=t^{-1}E_+XE_+\in DS_{\lambda_{\max}}, ρ 1 = t − 1 X + = t − 1 E + X E + ∈ D S λ m a x ,
ρ 2 = t − 1 X − = − t − 1 E − X E − ∈ D S λ min . \rho_2=t^{-1}X_-=-t^{-1}E_-XE_-\in DS_{\lambda_{\min}}. ρ 2 = t − 1 X − = − t − 1 E − X E − ∈ D S λ m i n .
and tr ( X W ) = 0 \operatorname{tr}(XW)=0 tr ( X W ) = 0 for every W ∈ W W\in W W ∈ W . Therefore tr ( ρ 1 W ) = tr ( ρ 2 W ) \operatorname{tr}(\rho_1W)=\operatorname{tr}(\rho_2W) tr ( ρ 1 W ) = tr ( ρ 2 W ) for every W ∈ W W\in W W ∈ W , and equivalently tr ( ρ 1 B i ) = tr ( ρ 2 B i ) \operatorname{tr}(\rho_1B^i)=\operatorname{tr}(\rho_2B^i) tr ( ρ 1 B i ) = tr ( ρ 2 B i ) , 1 ≤ i ≤ r 1\leq i\leq r 1 ≤ i ≤ r , for any Hermitian orthonormal basis { B i } i = 1 r \{B^i\}_{i=1}^r { B i } i = 1 r of W W W . Therefore, Φ ( ρ 1 ) = Φ ( ρ 2 ) ∈ m S λ max ∩ m S λ min \Phi(\rho_1)=\Phi(\rho_2)\in mS_{\lambda_{\max}}\cap mS_{\lambda_{\min}} Φ ( ρ 1 ) = Φ ( ρ 2 ) ∈ m S λ m a x ∩ m S λ m i n .
(3) ⇒ \Rightarrow ⇒ (4) Let ρ ∈ D S λ max \rho\in DS_{\lambda_{\max}} ρ ∈ D S λ m a x and σ ∈ D S λ min \sigma\in DS_{\lambda_{\min}} σ ∈ D S λ m i n be such that Φ ( ρ ) = Φ ( σ ) \Phi(\rho)=\Phi(\sigma) Φ ( ρ ) = Φ ( σ ) , and let W ∈ W W\in W W ∈ W . The matrix ρ \rho ρ is a convex combination of rank-one projections onto unit vectors of S λ max S_{\lambda_{\max}} S λ m a x , and similarly for σ \sigma σ with vectors of S λ min S_{\lambda_{\min}} S λ m i n , so that
min y ∈ S λ max ∥ y ∥ = 1 ⟨ W y , y ⟩ ≤ tr ( ρ W ) = tr ( σ W ) ≤ max z ∈ S λ min ∥ z ∥ = 1 ⟨ W z , z ⟩ . \min_{\substack{y\in S_{\lambda_{\max}}\\\lVert y\rVert=1}}\langle Wy,y\rangle\leq\operatorname{tr}(\rho W)=\operatorname{tr}(\sigma W)\leq\max_{\substack{z\in S_{\lambda_{\min}}\\\lVert z\rVert=1}}\langle Wz,z\rangle. y ∈ S λ m a x ∥ y ∥ = 1 min ⟨ W y , y ⟩ ≤ tr ( ρ W ) = tr ( σ W ) ≤ z ∈ S λ m i n ∥ z ∥ = 1 max ⟨ W z , z ⟩ .
(4) ⇒ \Rightarrow ⇒ (3) The sets D S λ max DS_{\lambda_{\max}} D S λ m a x and D S λ min DS_{\lambda_{\min}} D S λ m i n are compact and convex, and so are their images m S λ max mS_{\lambda_{\max}} m S λ m a x and m S λ min mS_{\lambda_{\min}} m S λ m i n under the linear map Φ \Phi Φ , which take values in the finite dimensional space R r \mathbb{R}^r R r . If these images were disjoint they could be strictly separated by a linear functional on R r \mathbb{R}^r R r ; replacing this functional by its opposite if necessary, this means that there exists an element x = ( x 1 , x 2 , … , x r ) ∈ R r x=(x_1,x_2,\ldots,x_r)\in\mathbb{R}^r x = ( x 1 , x 2 , … , x r ) ∈ R r such that
min Y ∈ D S λ max ⟨ Φ ( Y ) , x ⟩ > max Z ∈ D S λ min ⟨ Φ ( Z ) , x ⟩ . \min_{Y\in DS_{\lambda_{\max}}}\langle\Phi(Y),x\rangle>\max_{Z\in DS_{\lambda_{\min}}}\langle\Phi(Z),x\rangle. Y ∈ D S λ m a x min ⟨ Φ ( Y ) , x ⟩ > Z ∈ D S λ m i n max ⟨ Φ ( Z ) , x ⟩ .
Taking W = W ( x ) = ∑ i = 1 r x i B i W=W(x)=\sum_{i=1}^r x_iB^i W = W ( x ) = ∑ i = 1 r x i B i , we get
min Y ∈ D S λ max tr ( Y W ) > max Z ∈ D S λ min tr ( Z W ) . \min_{Y\in DS_{\lambda_{\max}}}\operatorname{tr}(YW)>\max_{Z\in DS_{\lambda_{\min}}}\operatorname{tr}(ZW). Y ∈ D S λ m a x min tr ( Y W ) > Z ∈ D S λ m i n max tr ( Z W ) .
By its spectral decomposition, every Y ∈ D S λ max Y\in DS_{\lambda_{\max}} Y ∈ D S λ m a x is a convex combination of rank-one projections y y ∗ yy^* y y ∗ onto unit vectors y ∈ S λ max y\in S_{\lambda_{\max}} y ∈ S λ m a x , and these projections belong to D S λ max DS_{\lambda_{\max}} D S λ m a x . Since tr ( y y ∗ W ) = ⟨ W y , y ⟩ \operatorname{tr}(yy^*W)=\langle Wy,y\rangle tr ( y y ∗ W ) = ⟨ W y , y ⟩ , the left-hand side equals min { ⟨ W y , y ⟩ : y ∈ S λ max , ∥ y ∥ = 1 } \min\{\langle Wy,y\rangle:y\in S_{\lambda_{\max}},\lVert y\rVert=1\} min {⟨ W y , y ⟩ : y ∈ S λ m a x , ∥ y ∥ = 1 } . In the same way, the right-hand side equals max { ⟨ W z , z ⟩ : z ∈ S λ min , ∥ z ∥ = 1 } \max\{\langle Wz,z\rangle:z\in S_{\lambda_{\min}},\lVert z\rVert=1\} max {⟨ W z , z ⟩ : z ∈ S λ m i n , ∥ z ∥ = 1 } , contradicting (4).
(3) ⇒ \Rightarrow ⇒ (2) Let Y ∈ D S λ max Y \in\mathcal{D}_{S_{\lambda_{\max}}} Y ∈ D S λ m a x and Z ∈ D S λ min Z \in\mathcal{D}_{S_{\lambda_{\min}}} Z ∈ D S λ m i n such that Φ ( Y ) = Φ ( Z ) \Phi(Y) = \Phi(Z) Φ ( Y ) = Φ ( Z ) and put X = Y − Z 2 X = \frac{Y-Z}{2} X = 2 Y − Z . Since S λ max ⊥ S λ min S_{\lambda_{\max}} \perp S_{\lambda_{\min}} S λ m a x ⊥ S λ m i n we have X + = 1 2 Y X_{+} = \frac{1}{2}Y X + = 2 1 Y , X − = 1 2 Z X_{-} = \frac{1}{2}Z X − = 2 1 Z and ∥ X ∥ 1 = 1 \lVert X\rVert_{1} = 1 ∥ X ∥ 1 = 1 , so X ≠ 0 X \ne0 X = 0 . Moreover,
X = E + X E + + E − X E − , with E + X E + = 1 2 Y ≥ 0 , E − X E − = − 1 2 Z ≤ 0 , X = E_{+}XE_{+} + E_{-}XE_{-}, \quad\text{with } E_{+}XE_{+} = \frac{1}{2}Y \ge0,\ E_{-}XE_{-} = -\frac{1}{2}Z \le0, X = E + X E + + E − X E − , with E + X E + = 2 1 Y ≥ 0 , E − X E − = − 2 1 Z ≤ 0 ,
and X ∈ W 1 ∘ X \in\mathcal{W}^{\circ}_{1} X ∈ W 1 ∘ , since Φ ( X ) = 0 ‾ \Phi(X) = \overline{0} Φ ( X ) = 0 . Finally A E + = λ max E + AE_{+} = \lambda_{\max}E_{+} A E + = λ m a x E + and A E − = λ min E − AE_{-} = \lambda_{\min}E_{-} A E − = λ m i n E − give tr ( A X ) = λ max − λ min 2 = ∥ A ∥ = ∥ A ∥ ∥ X ∥ 1 \operatorname{tr}(AX) = \frac{\lambda_{\max}-\lambda_{\min}}{2} = \lVert A\rVert= \lVert A\rVert\lVert X\rVert_{1} tr ( A X ) = 2 λ m a x − λ m i n = ∥ A ∥ = ∥ A ∥ ∥ X ∥ 1 .
(2) ⇒ \Rightarrow ⇒ (1) This implication is immediate, and is [17 Proposition 5.1] : if there exists X X X satisfying (38) , then the functional φ ( ⋅ ) = tr ( X ⋅ ) \varphi(\cdot) = \operatorname{tr}(X\cdot) φ ( ⋅ ) = tr ( X ⋅ ) satisfies
∥ A ∥ ∥ X ∥ 1 = φ ( A ) = tr ( A X ) = tr ( ( A + W ) X ) ≤ ∥ A + W ∥ ∥ X ∥ 1 , \lVert A\rVert\lVert X\rVert_{1} = \varphi(A) = \operatorname{tr}(AX) = \operatorname{tr}((A+W)X) \le\lVert A+W\rVert\lVert X\rVert_{1}, ∥ A ∥ ∥ X ∥ 1 = φ ( A ) = tr ( A X ) = tr (( A + W ) X ) ≤ ∥ A + W ∥ ∥ X ∥ 1 ,
for every W ∈ W W \in\mathcal{W} W ∈ W , since X ∈ W 1 ∘ X \in\mathcal{W}^{\circ}_{1} X ∈ W 1 ∘ and ∣ tr ( R X ) ∣ ≤ ∥ R ∥ ∥ X ∥ 1 |\operatorname{tr}(RX)| \le\lVert R\rVert\lVert X\rVert_{1} ∣ tr ( R X ) ∣ ≤ ∥ R ∥ ∥ X ∥ 1 . □
Remark 5.3 (Comments on Theorem 5.2). (1) The hypothesis I ∈ W I \in\mathcal{W} I ∈ W intervenes at exactly two points of the argument. One is through the balanced spectrum property, which for the matrices A A A of the statement holds anyway. The other is the implication (2) ⇒ \Rightarrow ⇒ (3), where tr ( X ) = 0 \operatorname{tr}(X) = 0 tr ( X ) = 0 is what allows the two parts of the witness to be normalized to density matrices. Without it one still obtains two positive matrices supported in S λ max S_{\lambda_{\max}} S λ m a x and in S λ min S_{\lambda_{\min}} S λ m i n with the same moments, but not of the same trace. The separation argument and the implications (3) ⇒ \Rightarrow ⇒ (2) ⇒ \Rightarrow ⇒ (1) do not use it.
(2) I ∈ W I \in\mathcal{W} I ∈ W implies that every X ∈ W 1 ∘ X \in\mathcal{W}^{\circ}_{1} X ∈ W 1 ∘ is a trace zero matrix.
(3) Note that condition X = E + X E + + E − X E − X = E_{+}XE_{+} + E_{-}XE_{-} X = E + X E + + E − X E − is equivalent to state that X X X equals its pinching operator.
(4) When W \mathcal{W} W is a C ∗ \mathrm{C}^{*} C ∗ -subalgebra of M n ( C ) M_n(\mathbb{C}) M n ( C ) containing I I I , the equivalence between conditions (1) and (3) follows from [7 Theorem 2 and Corollary 1] . The proofs given there do not use the product of W \mathcal{W} W : it intervenes only through the annihilator W 1 ∘ \mathcal{W}^{\circ}_{1} W 1 ∘ and through I ∈ W I \in\mathcal{W} I ∈ W .
We now consider the general case of a subspace W \mathcal{W} W of Hermitian matrices, where possibly I ∉ W I \notin\mathcal{W} I ∈ / W .
Notation 5.4. Let W ⊂ M n h ( C ) \mathcal{W} \subset M_n^h(\mathbb{C}) W ⊂ M n h ( C ) be a subspace, A 0 ∈ M n h ( C ) ∖ W A_0 \in M_n^h(\mathbb{C}) \setminus\mathcal{W} A 0 ∈ M n h ( C ) ∖ W and W 0 ∈ W W_0 \in\mathcal{W} W 0 ∈ W . We abbreviate S M A X : = S M A X ( A 0 + W 0 ) S_{\mathrm{MAX}} := S_{\mathrm{MAX}}(A_0 + W_0) S MAX := S MAX ( A 0 + W 0 ) for the subspace introduced in (10) , which by item 1 of Proposition 2.6 can be written as
S M A X = span { v ∈ C n : ( A 0 + W 0 ) v = ± ∥ A 0 + W 0 ∥ v } . (40) S_{\mathrm{MAX}} = \operatorname{span}\{v \in\mathbb{C}^{n} : (A_0 + W_0)v = \pm\lVert A_0 + W_0\rVert v\}. \tag*{(40)} S MAX = span { v ∈ C n : ( A 0 + W 0 ) v = ± ∥ A 0 + W 0 ∥ v } . (40)
Following [19 ] , we define the extremal functionals on M n h ( C ) M_n^h(\mathbb{C}) M n h ( C ) .
Definition 5.5. Let S 1 ∗ : = { f ∈ ( M n h ( C ) ) ∗ : ∥ f ∥ = 1 } S_1^* := \{f \in(M_n^h(\mathbb{C}))^* : \lVert f\rVert= 1\} S 1 ∗ := { f ∈ ( M n h ( C ) ) ∗ : ∥ f ∥ = 1 } denote the unit sphere of the dual space. A functional f ∈ S 1 ∗ f \in S_1^* f ∈ S 1 ∗ is extremal if, whenever f = f 1 + f 2 2 f = \frac{f_1+f_2}{2} f = 2 f 1 + f 2 with f 1 , f 2 ∈ S 1 ∗ f_1, f_2 \in S_1^* f 1 , f 2 ∈ S 1 ∗ , then f = f 1 = f 2 f = f_1 = f_2 f = f 1 = f 2 . We denote by E h ⊆ S 1 ∗ E^h \subseteq S_1^* E h ⊆ S 1 ∗ the set of extremals.
Since every f ∈ S 1 ∗ f \in S_1^* f ∈ S 1 ∗ can be written as f ( ⋅ ) = tr ( F ⋅ ) f(\cdot) = \operatorname{tr}(F\cdot) f ( ⋅ ) = tr ( F ⋅ ) with a unique F ∈ M n h ( C ) F \in M_n^h(\mathbb{C}) F ∈ M n h ( C ) with ∥ F ∥ 1 = 1 \lVert F\rVert_1 = 1 ∥ F ∥ 1 = 1 , we can identify the set E h E^h E h with the set of F ∈ M n h ( C ) F \in M_n^h(\mathbb{C}) F ∈ M n h ( C ) , ∥ F ∥ 1 = 1 \lVert F\rVert_1 = 1 ∥ F ∥ 1 = 1 such that, if there exist Hermitian matrices F 1 , F 2 F_1, F_2 F 1 , F 2 with ∥ F 1 ∥ 1 = ∥ F 2 ∥ 1 = 1 \lVert F_1\rVert_1 = \lVert F_2\rVert_1 = 1 ∥ F 1 ∥ 1 = ∥ F 2 ∥ 1 = 1 and F = F 1 + F 2 2 F = \frac{F_1+F_2}{2} F = 2 F 1 + F 2 , then F = F 1 = F 2 F = F_1 = F_2 F = F 1 = F 2 .
Lemma 5.6. Let u , v ∈ C n u,v \in\mathbb{C}^{n} u , v ∈ C n be unit vectors and let F = u v ∗ ∈ M n ( C ) F = uv^* \in M_n(\mathbb{C}) F = u v ∗ ∈ M n ( C ) . If F F F is self-adjoint (F = F ∗ F = F^* F = F ∗ ), then u = v u = v u = v or u = − v u = -v u = − v .
Proof. Applying the identity u v ∗ = v u ∗ uv^* = vu^* u v ∗ = v u ∗ to u u u gives ⟨ u , v ⟩ u = v \langle u,v\rangle u = v ⟨ u , v ⟩ u = v , where ⟨ x , y ⟩ = y ∗ x \langle x,y\rangle= y^*x ⟨ x , y ⟩ = y ∗ x . Hence v = λ u v = \lambda u v = λ u with λ = ⟨ u , v ⟩ \lambda= \langle u,v\rangle λ = ⟨ u , v ⟩ , and ∣ λ ∣ = 1 |\lambda| = 1 ∣ λ ∣ = 1 because ∥ u ∥ = ∥ v ∥ = 1 \lVert u\rVert= \lVert v\rVert= 1 ∥ u ∥ = ∥ v ∥ = 1 . Then λ ‾ u u ∗ = v u ∗ = u v ∗ = λ u u ∗ \overline{\lambda}uu^* = vu^* = uv^* = \lambda uu^* λ u u ∗ = v u ∗ = u v ∗ = λ u u ∗ , so that λ ∈ R \lambda\in\mathbb{R} λ ∈ R , that is, λ = ± 1 \lambda= \pm1 λ = ± 1 . □
This result is probably known, but since we could not find a reference, we include a proof.
Proposition 5.7. The set of extremals in the unit sphere of ( M n h ( C ) ) ∗ (M_n^h(\mathbb{C}))^* ( M n h ( C ) ) ∗ is
E h = { ± v v ∗ : v ∈ C n , ∥ v ∥ = 1 } . E^h = \{\pm vv^* : v \in\mathbb{C}^{n}, \lVert v\rVert= 1\}. E h = { ± v v ∗ : v ∈ C n , ∥ v ∥ = 1 } .
Proof. Via F ↦ tr ( ⋅ F ) F \mapsto\operatorname{tr}(\cdot F) F ↦ tr ( ⋅ F ) , the dual ( M n h ( C ) ) ∗ (M_n^h(\mathbb{C}))^* ( M n h ( C ) ) ∗ is isometric to M n h ( C ) M_n^h(\mathbb{C}) M n h ( C ) with the trace norm ∥ ⋅ ∥ 1 \lVert\cdot\rVert_1 ∥ ⋅ ∥ 1 . If ∥ F ∥ 1 = 1 \lVert F\rVert_1=1 ∥ F ∥ 1 = 1 and F = F 1 + F 2 2 F=\frac{F_1+F_2}{2} F = 2 F 1 + F 2 with ∥ F 1 ∥ 1 , ∥ F 2 ∥ 1 ≤ 1 \lVert F_1\rVert_1,\lVert F_2\rVert_1\leq1 ∥ F 1 ∥ 1 , ∥ F 2 ∥ 1 ≤ 1 , then 1 ≤ ∥ F 1 ∥ 1 + ∥ F 2 ∥ 1 2 1\leq\frac{\lVert F_1\rVert_1+\lVert F_2\rVert_1}{2} 1 ≤ 2 ∥ F 1 ∥ 1 + ∥ F 2 ∥ 1 forces ∥ F 1 ∥ 1 = ∥ F 2 ∥ 1 = 1 \lVert F_1\rVert_1=\lVert F_2\rVert_1=1 ∥ F 1 ∥ 1 = ∥ F 2 ∥ 1 = 1 ; hence the extremals are exactly the extreme points of the unit ball of ( M n h ( C ) , ∥ ⋅ ∥ 1 ) (M_n^h(\mathbb{C}),\lVert\cdot\rVert_1) ( M n h ( C ) , ∥ ⋅ ∥ 1 ) , and we may work with convex combinations t G + ( 1 − t ) H tG+(1-t)H tG + ( 1 − t ) H , 0 < t < 1 0<t<1 0 < t < 1 , of elements of this ball. Let F ∈ E h F\in\mathcal{E}^h F ∈ E h , so ∥ F ∥ 1 = 1 \lVert F\rVert_1=1 ∥ F ∥ 1 = 1 . Writing F = ∑ λ i ≠ 0 ∣ λ i ∣ ( sgn λ i ) v i v i ∗ F=\sum_{\lambda_i\neq0}|\lambda_i|(\operatorname{sgn}\lambda_i)v_iv_i^* F = ∑ λ i = 0 ∣ λ i ∣ ( sgn λ i ) v i v i ∗ (spectral decomposition, with { v i } \{v_i\} { v i } orthonormal) exhibits F F F as a convex combination of the points ± v i v i ∗ \pm v_iv_i^* ± v i v i ∗ of the unit ball. Since F F F is extreme, it must coincide with each of them; therefore F F F has a single nonzero eigenvalue ± 1 \pm1 ± 1 , and F = ± v v ∗ F=\pm vv^* F = ± v v ∗ with ∥ v ∥ = 1 \lVert v\rVert=1 ∥ v ∥ = 1 .
Conversely, recall that tr G ≤ ∥ G ∥ 1 \operatorname{tr}G\leq\lVert G\rVert_1 tr G ≤ ∥ G ∥ 1 for every G ∈ M n h ( C ) G\in M_n^h(\mathbb{C}) G ∈ M n h ( C ) , with equality if and only if G ≥ 0 G\geq0 G ≥ 0 . Let ∥ v ∥ = 1 \lVert v\rVert=1 ∥ v ∥ = 1 and suppose v v ∗ = t G + ( 1 − t ) H vv^*=tG+(1-t)H v v ∗ = tG + ( 1 − t ) H with 0 < t < 1 0<t<1 0 < t < 1 and ∥ G ∥ 1 , ∥ H ∥ 1 ≤ 1 \lVert G\rVert_1,\lVert H\rVert_1\leq1 ∥ G ∥ 1 , ∥ H ∥ 1 ≤ 1 . Then 1 = t tr G + ( 1 − t ) tr H ≤ t ∥ G ∥ 1 + ( 1 − t ) ∥ H ∥ 1 ≤ 1 1=t\operatorname{tr}G+(1-t)\operatorname{tr}H\leq t\lVert G\rVert_1+(1-t)\lVert H\rVert_1\leq1 1 = t tr G + ( 1 − t ) tr H ≤ t ∥ G ∥ 1 + ( 1 − t ) ∥ H ∥ 1 ≤ 1 , so tr G = ∥ G ∥ 1 = 1 \operatorname{tr}G=\lVert G\rVert_1=1 tr G = ∥ G ∥ 1 = 1 , and likewise for H H H . Hence G , H ≥ 0 G,H\geq0 G , H ≥ 0 . For w ⊥ v w\perp v w ⊥ v we get t ⟨ G w , w ⟩ + ( 1 − t ) ⟨ H w , w ⟩ = ⟨ v v ∗ w , w ⟩ = 0 t\langle Gw,w\rangle+(1-t)\langle Hw,w\rangle=\langle vv^*w,w\rangle=0 t ⟨ Gw , w ⟩ + ( 1 − t ) ⟨ H w , w ⟩ = ⟨ v v ∗ w , w ⟩ = 0 , so ⟨ G w , w ⟩ = 0 \langle Gw,w\rangle=0 ⟨ Gw , w ⟩ = 0 and, since G ≥ 0 G\geq0 G ≥ 0 , G w = 0 Gw=0 Gw = 0 . Thus G = μ v v ∗ G=\mu vv^* G = μv v ∗ , and tr G = 1 \operatorname{tr}G=1 tr G = 1 gives G = v v ∗ G=vv^* G = v v ∗ ; likewise H = v v ∗ H=vv^* H = v v ∗ . Finally, if − v v ∗ = t G + ( 1 − t ) H -vv^*=tG+(1-t)H − v v ∗ = tG + ( 1 − t ) H , then v v ∗ = t ( − G ) + ( 1 − t ) ( − H ) vv^*=t(-G)+(1-t)(-H) v v ∗ = t ( − G ) + ( 1 − t ) ( − H ) , so − G = − H = v v ∗ -G=-H=vv^* − G = − H = v v ∗ . Hence ± v v ∗ ∈ E h \pm vv^*\in\mathcal{E}^h ± v v ∗ ∈ E h . □
By [14 Theorem 2] , the set of extremals in the unit sphere of ( M n ( C ) ) ∗ (M_n(\mathbb{C}))^* ( M n ( C ) ) ∗ is
E = { F ∈ M n ( C ) : F = u v ∗ , ∥ u ∥ = ∥ v ∥ = 1 } . \mathcal{E}=\{F\in M_n(\mathbb{C}):F=uv^*,\ \lVert u\rVert=\lVert v\rVert=1\}. E = { F ∈ M n ( C ) : F = u v ∗ , ∥ u ∥ = ∥ v ∥ = 1 } .
Theorem 5.9. Let W ⊆ M n h ( C ) \mathcal{W}\subseteq M_n^h(\mathbb{C}) W ⊆ M n h ( C ) be a subspace with dim R W = r \dim_{\mathbb{R}}\mathcal{W}=r dim R W = r , and fix A 0 ∈ M n h ( C ) ∖ W A_0\in M_n^h(\mathbb{C})\setminus\mathcal{W} A 0 ∈ M n h ( C ) ∖ W , W 0 ∈ W W_0\in\mathcal{W} W 0 ∈ W . Let S M A X S_{\mathrm{MAX}} S MAX be as in (5.3). Then the following are equivalent:
(1) ∥ A 0 + W 0 ∥ = dist ( A 0 , W ) \lVert A_0+W_0\rVert=\operatorname{dist}(A_0,\mathcal{W}) ∥ A 0 + W 0 ∥ = dist ( A 0 , W ) , that is, A 0 + W 0 A_0+W_0 A 0 + W 0 is W \mathcal{W} W -minimal.
(2) There exist l l l unit vectors v 1 , … , v l ∈ S M A X v_1,\ldots,v_l\in S_{\mathrm{MAX}} v 1 , … , v l ∈ S MAX , with 1 ≤ l ≤ r + 1 1\leq l\leq r+1 1 ≤ l ≤ r + 1 , which are moreover eigenvectors of A 0 + W 0 A_0+W_0 A 0 + W 0 associated with the eigenvalue ∥ A 0 + W 0 ∥ \lVert A_0+W_0\rVert ∥ A 0 + W 0 ∥ or with the eigenvalue − ∥ A 0 + W 0 ∥ -\lVert A_0+W_0\rVert − ∥ A 0 + W 0 ∥ , together with vectors u i ∈ S M A X u_i\in S_{\mathrm{MAX}} u i ∈ S MAX and weights 0 ≤ ω i ≤ 1 0\leq\omega_i\leq1 0 ≤ ω i ≤ 1 with ∑ i = 1 l ω i = 1 \sum_{i=1}^l\omega_i=1 ∑ i = 1 l ω i = 1 , such that
∑ i = 1 l ω i ⟨ W v i , u i ⟩ = 0 , for every W ∈ W , (41) \sum_{i=1}^{l}\omega_i\langle Wv_i,u_i\rangle=0,\quad\text{for every }W\in\mathcal{W}, \tag*{(41)} i = 1 ∑ l ω i ⟨ W v i , u i ⟩ = 0 , for every W ∈ W , (41) where the vectors u i u_i u i are given by
u i = v i if ( A 0 + W 0 ) v i = ∥ A 0 + W 0 ∥ v i , u i = − v i if ( A 0 + W 0 ) v i = − ∥ A 0 + W 0 ∥ v i . (42) \begin{aligned}
u_i&=v_i &&\text{if }(A_0+W_0)v_i=\lVert A_0+W_0\rVert v_i,\\
u_i&=-v_i &&\text{if }(A_0+W_0)v_i=-\lVert A_0+W_0\rVert v_i. \tag*{(42)}
\end{aligned} u i u i = v i = − v i if ( A 0 + W 0 ) v i = ∥ A 0 + W 0 ∥ v i , if ( A 0 + W 0 ) v i = − ∥ A 0 + W 0 ∥ v i . (42) (3)
{ 0 ‾ ∈ m S M A X if λ max + λ min ≠ 0 , 0 ‾ ∈ conv ( m S λ max ∪ − m S λ min ) if λ max + λ min = 0 , (43) \begin{cases}
\overline{0}\in m_{S_{\mathrm{MAX}}} & \text{if }\lambda_{\max}+\lambda_{\min}\neq0,\\
\overline{0}\in\operatorname{conv}\left(m_{S_{\lambda_{\max}}}\cup-m_{S_{\lambda_{\min}}}\right) & \text{if }\lambda_{\max}+\lambda_{\min}=0, \tag*{(43)}
\end{cases} { 0 ∈ m S MAX 0 ∈ conv ( m S λ m a x ∪ − m S λ m i n ) if λ m a x + λ m i n = 0 , if λ m a x + λ m i n = 0 , (43) where λ max = max σ ( A 0 + W 0 ) \lambda_{\max}=\max\sigma(A_0+W_0) λ m a x = max σ ( A 0 + W 0 ) and λ min = min σ ( A 0 + W 0 ) \lambda_{\min}=\min\sigma(A_0+W_0) λ m i n = min σ ( A 0 + W 0 ) .
(4) 0 ‾ ∈ ∂ ( ∥ A ( ⋅ ) ∥ ) ( x ) \overline{0}\in\partial(\lVert A(\cdot)\rVert)(x) 0 ∈ ∂ (∥ A ( ⋅ )∥) ( x ) , where x = ( x 1 , x 2 , … , x r ) ∈ R r x=(x_1,x_2,\ldots,x_r)\in\mathbb{R}^r x = ( x 1 , x 2 , … , x r ) ∈ R r is such that A ( x ) = A 0 + ∑ i = 1 r x i B i = A 0 + W 0 A(x)=A_0+\sum_{i=1}^r x_iB^i=A_0+W_0 A ( x ) = A 0 + ∑ i = 1 r x i B i = A 0 + W 0 .
(5) There exists F ∈ W ∘ F\in\mathcal{W}^{\circ} F ∈ W ∘ such that F ∈ ∂ ∥ A 0 + W 0 ∥ F\in\partial\lVert A_0+W_0\rVert F ∈ ∂ ∥ A 0 + W 0 ∥ , i.e.:
tr ( F ( A 0 + W 0 ) ) = tr ( F A 0 ) = ∥ A 0 + W 0 ∥ . \operatorname{tr}(F(A_0+W_0))=\operatorname{tr}(FA_0)=\lVert A_0+W_0\rVert. tr ( F ( A 0 + W 0 )) = tr ( F A 0 ) = ∥ A 0 + W 0 ∥ . Proof. (1) ⇒ \Rightarrow ⇒ (2) By Theorem 1.1 in [19 ] , page 170, − W 0 -W_0 − W 0 is a best approximation of A 0 A_0 A 0 in W \mathcal{W} W if and only if there exist l l l extremal functionals F i ∈ E h F_i\in\mathcal{E}^h F i ∈ E h and weights 0 ≤ ω i ≤ 1 0\leq\omega_i\leq1 0 ≤ ω i ≤ 1 with ∑ i = 1 l ω i = 1 \sum_{i=1}^l\omega_i=1 ∑ i = 1 l ω i = 1 such
that
∑ i = 1 l ω i tr ( F i W ) = 0 , ∀ W ∈ W (44) \sum_{i=1}^{l}\omega_i\operatorname{tr}(F_iW)=0,\quad\forall W\in\mathcal{W} \tag*{(44)} i = 1 ∑ l ω i tr ( F i W ) = 0 , ∀ W ∈ W (44)
tr ( F i ( A 0 + W 0 ) ) = ∥ A 0 + W 0 ∥ \operatorname{tr}(F_i(A_0+W_0))=\lVert A_0+W_0\rVert tr ( F i ( A 0 + W 0 )) = ∥ A 0 + W 0 ∥ , for every 1 ≤ i ≤ l 1\le i\le l 1 ≤ i ≤ l .
By Proposition 5.7, each F i F_i F i is of the form F i = ± v i v i ∗ F_i=\pm v_i v_i^* F i = ± v i v i ∗ with ∥ v i ∥ = 1 \lVert v_i\rVert=1 ∥ v i ∥ = 1 , and tr ( v i v i ∗ R ) = ⟨ R v i , v i ⟩ \operatorname{tr}(v_i v_i^*R)=\langle Rv_i,v_i\rangle tr ( v i v i ∗ R ) = ⟨ R v i , v i ⟩ for any matrix R R R ; then (5.7) is equivalent to
∑ i = 1 l ω i ⟨ W v i , u i ⟩ = 0 , ∀ W ∈ W \sum_{i=1}^{l}\omega_i\langle Wv_i,u_i\rangle=0,\quad\forall W\in\mathcal{W} i = 1 ∑ l ω i ⟨ W v i , u i ⟩ = 0 , ∀ W ∈ W
tr ( F i ( A 0 + W 0 ) ) = u i ∗ ( A 0 + W 0 ) v i = ∥ A 0 + W 0 ∥ \operatorname{tr}(F_i(A_0+W_0))=u_i^*(A_0+W_0)v_i=\lVert A_0+W_0\rVert tr ( F i ( A 0 + W 0 )) = u i ∗ ( A 0 + W 0 ) v i = ∥ A 0 + W 0 ∥ , for every 1 ≤ i ≤ l 1\le i\le l 1 ≤ i ≤ l , where each v i ∈ S M A X v_i\in S_{\mathrm{MAX}} v i ∈ S MAX and u i = ± v i u_i=\pm v_i u i = ± v i .
If u i = v i u_i=v_i u i = v i , then v i v_i v i is an eigenvector of A 0 + W 0 A_0+W_0 A 0 + W 0 associated with ∥ A 0 + W 0 ∥ \lVert A_0+W_0\rVert ∥ A 0 + W 0 ∥ .
If u i = − v i u_i=-v_i u i = − v i , then v i v_i v i is an eigenvector of A 0 + W 0 A_0+W_0 A 0 + W 0 associated with − ∥ A 0 + W 0 ∥ -\lVert A_0+W_0\rVert − ∥ A 0 + W 0 ∥ .
(2) ⇒ \Rightarrow ⇒ (1) If there exist l l l unit vectors v i , u i ∈ S M A X v_i,u_i\in S_{\mathrm{MAX}} v i , u i ∈ S MAX and weights 0 ≤ ω i ≤ 1 0\le\omega_i\le1 0 ≤ ω i ≤ 1 with ∑ i = 1 l ω i = 1 \sum_{i=1}^{l}\omega_i=1 ∑ i = 1 l ω i = 1 satisfying (5.4) and (5.5), then by (5.5) the matrix F = ∑ i = 1 l ω i u i v i ∗ F=\sum_{i=1}^{l}\omega_i u_i v_i^* F = ∑ i = 1 l ω i u i v i ∗ satisfies
{ F ∈ M n h ( C ) , tr ( F W ) = 0 , ∀ W ∈ W , tr ( F A 0 ) = tr ( F ( A 0 + W 0 ) ) = ∥ A 0 + W 0 ∥ . \begin{cases}
F\in M_n^h(\mathbb{C}),\\
\operatorname{tr}(FW)=0,\quad\forall W\in\mathcal{W},\\
\operatorname{tr}(FA_0)=\operatorname{tr}(F(A_0+W_0))=\lVert A_0+W_0\rVert.
\end{cases} ⎩ ⎨ ⎧ F ∈ M n h ( C ) , tr ( F W ) = 0 , ∀ W ∈ W , tr ( F A 0 ) = tr ( F ( A 0 + W 0 )) = ∥ A 0 + W 0 ∥ .
Moreover, ∥ F ∥ 1 ≤ ∑ i = 1 l ω i = 1 \lVert F\rVert_1\le\sum_{i=1}^{l}\omega_i=1 ∥ F ∥ 1 ≤ ∑ i = 1 l ω i = 1 , while from tr ( F ( A 0 + W 0 ) ) = ∥ A 0 + W 0 ∥ \operatorname{tr}(F(A_0+W_0))=\lVert A_0+W_0\rVert tr ( F ( A 0 + W 0 )) = ∥ A 0 + W 0 ∥ we get ∥ F ∥ 1 ≥ 1 \lVert F\rVert_1\ge1 ∥ F ∥ 1 ≥ 1 ; hence ∥ F ∥ 1 = 1 \lVert F\rVert_1=1 ∥ F ∥ 1 = 1 , and therefore F ∈ W 1 ∘ ∩ ∂ ∥ A 0 + W 0 ∥ F\in\mathcal{W}_1^\circ\cap\partial\lVert A_0+W_0\rVert F ∈ W 1 ∘ ∩ ∂ ∥ A 0 + W 0 ∥ . Consequently, since tr ( F W ) = 0 \operatorname{tr}(FW)=0 tr ( F W ) = 0 for every W ∈ W W\in\mathcal{W} W ∈ W ,
∥ A 0 + W 0 ∥ = tr ( F ( A 0 + W 0 ) ) = tr ( F ( A 0 + W ) ) ≤ ∥ A 0 + W ∥ ∥ F ∥ 1 = ∥ A 0 + W ∥ , \begin{aligned}
\lVert A_0+W_0\rVert=\operatorname{tr}(F(A_0+W_0))=\operatorname{tr}(F(A_0+W)) \\
&\le\lVert A_0+W\rVert\lVert F\rVert_1=\lVert A_0+W\rVert,
\end{aligned} ∥ A 0 + W 0 ∥ = tr ( F ( A 0 + W 0 )) = tr ( F ( A 0 + W )) ≤ ∥ A 0 + W ∥ ∥ F ∥ 1 = ∥ A 0 + W ∥ ,
for every W ∈ W W\in\mathcal{W} W ∈ W ; that is, A 0 + W 0 A_0+W_0 A 0 + W 0 is W \mathcal{W} W -minimal.
The same computation proves (5) ⇒ \Rightarrow ⇒ (1), while (1) ⇒ \Rightarrow ⇒ (5) follows by combining (1) ⇒ \Rightarrow ⇒ (2) with the construction of F F F above. This gives a self-contained proof of
∥ A 0 + W 0 ∥ = min W ∈ W ∥ A 0 + W ∥ ⟺ ∃ F ∈ ∂ ∥ A 0 + W 0 ∥ ∩ W 1 ∘ , \lVert A_0+W_0\rVert=\min_{W\in\mathcal{W}}\lVert A_0+W\rVert\Longleftrightarrow\exists F\in\partial\lVert A_0+W_0\rVert\cap\mathcal{W}_1^\circ, ∥ A 0 + W 0 ∥ = W ∈ W min ∥ A 0 + W ∥ ⟺ ∃ F ∈ ∂ ∥ A 0 + W 0 ∥ ∩ W 1 ∘ ,
an equivalence which is also a particular case of [23 Corollary 1] .
(2) ⇒ \Rightarrow ⇒ (3) Since { B j } j = 1 r \{B^j\}_{j=1}^{r} { B j } j = 1 r is a basis of W \mathcal{W} W , condition (5.4) for any W ∈ W W\in\mathcal{W} W ∈ W is equivalent to
∑ i = 1 l ω i ⟨ B j v i , u i ⟩ = 0 , for all 1 ≤ j ≤ r , \sum_{i=1}^{l}\omega_i\langle B^jv_i,u_i\rangle=0,\quad\text{for all }1\le j\le r, i = 1 ∑ l ω i ⟨ B j v i , u i ⟩ = 0 , for all 1 ≤ j ≤ r ,
which implies that
∑ i = 1 l ω i ( ⟨ B 1 v i , u i ⟩ , … , ⟨ B r v i , u i ⟩ ) = ( 0 , … , 0 ) (45) \sum_{i=1}^{l}\omega_i\left(\langle B^1v_i,u_i\rangle,\ldots,\langle B^rv_i,u_i\rangle\right)=(0,\ldots,0) \tag*{(45)} i = 1 ∑ l ω i ( ⟨ B 1 v i , u i ⟩ , … , ⟨ B r v i , u i ⟩ ) = ( 0 , … , 0 ) (45)
that is, 0 ‾ \overline{0} 0 is a convex combination of the vectors ( ⟨ B 1 v i , u i ⟩ , … , ⟨ B r v i , u i ⟩ ) (\langle B^1v_i,u_i\rangle,\ldots,\langle B^rv_i,u_i\rangle) (⟨ B 1 v i , u i ⟩ , … , ⟨ B r v i , u i ⟩) , 1 ≤ i ≤ l 1\le i\le l 1 ≤ i ≤ l . We distinguish three cases:
If ∥ A 0 + W 0 ∥ ∈ σ ( A 0 + W 0 ) \lVert A_0+W_0\rVert\in\sigma(A_0+W_0) ∥ A 0 + W 0 ∥ ∈ σ ( A 0 + W 0 ) and − ∥ A 0 + W 0 ∥ ∉ σ ( A 0 + W 0 ) -\lVert A_0+W_0\rVert\notin\sigma(A_0+W_0) − ∥ A 0 + W 0 ∥ ∈ / σ ( A 0 + W 0 ) , then S M A X = S λ max S_{\mathrm{MAX}}=S_{\lambda_{\max}} S MAX = S λ m a x , so v i ∈ S λ max v_i\in S_{\lambda_{\max}} v i ∈ S λ m a x and ⟨ B j v i , u i ⟩ = ⟨ B j v i , v i ⟩ \langle B^jv_i,u_i\rangle=\langle B^jv_i,v_i\rangle ⟨ B j v i , u i ⟩ = ⟨ B j v i , v i ⟩ for every 1 ≤ j ≤ r 1\le j\le r 1 ≤ j ≤ r . Therefore by (5.8), 0 ‾ ∈ m S λ max = m S M A X \overline{0}\in mS_{\lambda_{\max}}=mS_{\mathrm{MAX}} 0 ∈ m S λ m a x = m S MAX .
If − ∥ A 0 + W 0 ∥ ∈ σ ( A 0 + W 0 ) -\lVert A_0+W_0\rVert\in\sigma(A_0+W_0) − ∥ A 0 + W 0 ∥ ∈ σ ( A 0 + W 0 ) and ∥ A 0 + W 0 ∥ ∉ σ ( A 0 + W 0 ) \lVert A_0+W_0\rVert\notin\sigma(A_0+W_0) ∥ A 0 + W 0 ∥ ∈ / σ ( A 0 + W 0 ) , then S M A X = S λ min S_{\mathrm{MAX}}=S_{\lambda_{\min}} S MAX = S λ m i n , so v i ∈ S λ min v_i\in S_{\lambda_{\min}} v i ∈ S λ m i n and ⟨ B j v i , u i ⟩ = − ⟨ B j v i , v i ⟩ \langle B^jv_i,u_i\rangle=-\langle B^jv_i,v_i\rangle ⟨ B j v i , u i ⟩ = − ⟨ B j v i , v i ⟩ for every 1 ≤ j ≤ r 1\le j\le r 1 ≤ j ≤ r . Therefore, by (5.8), 0 ‾ ∈ − m S λ min = − m S M A X \overline{0}\in-mS_{\lambda_{\min}}=-mS_{\mathrm{MAX}} 0 ∈ − m S λ m i n = − m S MAX which is equivalent to 0 ‾ ∈ m S M A X \overline{0}\in mS_{\mathrm{MAX}} 0 ∈ m S MAX .
If ± ∥ A 0 + W 0 ∥ ∈ σ ( A 0 + W 0 ) \pm\lVert A_{0}+W_{0}\rVert\in\sigma(A_{0}+W_{0}) ± ∥ A 0 + W 0 ∥ ∈ σ ( A 0 + W 0 ) , then S M A X = S λ min ⊕ S λ max S_{\mathrm{MAX}}=S_{\lambda_{\min}}\oplus S_{\lambda_{\max}} S MAX = S λ m i n ⊕ S λ m a x . By hypothesis, each v i v_i v i is an eigenvector of A 0 + W 0 A_{0}+W_{0} A 0 + W 0 corresponding to ∥ A 0 + W 0 ∥ \lVert A_{0}+W_{0}\rVert ∥ A 0 + W 0 ∥ or to − ∥ A 0 + W 0 ∥ -\lVert A_{0}+W_{0}\rVert − ∥ A 0 + W 0 ∥ ; after reindexing, we may assume that
( A 0 + W 0 ) v i = ∥ A 0 + W 0 ∥ v i , v i ∈ S λ max for every 1 ≤ i ≤ d , (A_{0}+W_{0})v_i=\lVert A_{0}+W_{0}\rVert v_i,\quad v_i\in S_{\lambda_{\max}}\qquad\text{for every }1\le i\le d, ( A 0 + W 0 ) v i = ∥ A 0 + W 0 ∥ v i , v i ∈ S λ m a x for every 1 ≤ i ≤ d ,
( A 0 + W 0 ) v i = − ∥ A 0 + W 0 ∥ v i , v i ∈ S λ min for every d + 1 ≤ i ≤ l . (A_{0}+W_{0})v_i=-\lVert A_{0}+W_{0}\rVert v_i,\quad v_i\in S_{\lambda_{\min}}\qquad\text{for every }d+1\le i\le l. ( A 0 + W 0 ) v i = − ∥ A 0 + W 0 ∥ v i , v i ∈ S λ m i n for every d + 1 ≤ i ≤ l .
and, writing x i = ( ⟨ B 1 v i , v i ⟩ , … , ⟨ B r v i , v i ⟩ ) x_i=(\langle B^{1}v_i,v_i\rangle,\ldots,\langle B^{r}v_i,v_i\rangle) x i = (⟨ B 1 v i , v i ⟩ , … , ⟨ B r v i , v i ⟩) for 1 ≤ i ≤ l 1\le i\le l 1 ≤ i ≤ l , condition (5.4) gives
∑ i = 1 d ω i x i + ∑ i = d + 1 l ω i ( ⟨ B j v i , − v i ⟩ ) j = 1 r = ∑ i = 1 d ω i x i − ∑ i = d + 1 l ω i x i = 0 ˉ . \sum_{i=1}^{d}\omega_i x_i+\sum_{i=d+1}^{l}\omega_i\left(\langle B^{j}v_i,-v_i\rangle\right)_{j=1}^{r}=\sum_{i=1}^{d}\omega_i x_i-\sum_{i=d+1}^{l}\omega_i x_i=\bar{0}. i = 1 ∑ d ω i x i + i = d + 1 ∑ l ω i ( ⟨ B j v i , − v i ⟩ ) j = 1 r = i = 1 ∑ d ω i x i − i = d + 1 ∑ l ω i x i = 0 ˉ .
Then,
0 ˉ ∈ conv ( m S λ max ∪ − m S λ min ) . \bar{0}\in\operatorname{conv}\left(m_{S_{\lambda_{\max}}}\cup-m_{S_{\lambda_{\min}}}\right). 0 ˉ ∈ conv ( m S λ m a x ∪ − m S λ m i n ) .
since x i ∈ m S λ max x_i\in m_{S_{\lambda_{\max}}} x i ∈ m S λ m a x for $1\le i\le d$ and − x i ∈ − m S λ min -x_i\in-m_{S_{\lambda_{\min}}} − x i ∈ − m S λ m i n for $d+1\le i\le l.
(3) ⇒ \Rightarrow ⇒ (2) First suppose that λ max + λ min ≠ 0 \lambda_{\max}+\lambda_{\min}\ne0 λ m a x + λ m i n = 0 , that is S M A X = S λ max S_{\mathrm{MAX}}=S_{\lambda_{\max}} S MAX = S λ m a x or S M A X = S λ min S_{\mathrm{MAX}}=S_{\lambda_{\min}} S MAX = S λ m i n . Since 0 ˉ ∈ m S M A X \bar{0}\in m_{S_{\mathrm{MAX}}} 0 ˉ ∈ m S MAX and m S M A X m_{S_{\mathrm{MAX}}} m S MAX is the convex hull of { ( ⟨ B 1 v , v ⟩ , … , ⟨ B r v , v ⟩ ) : v ∈ S M A X , ∥ v ∥ = 1 } ⊆ R r \{(\langle B^{1}v,v\rangle,\ldots,\langle B^{r}v,v\rangle):v\in S_{\mathrm{MAX}},\ \lVert v\rVert=1\}\subseteq\mathbb{R}^{r} {(⟨ B 1 v , v ⟩ , … , ⟨ B r v , v ⟩) : v ∈ S MAX , ∥ v ∥ = 1 } ⊆ R r , using Carathéodory’s theorem (see for instance [18 Theorem 17.1] ) there exist l ≤ r + 1 l\le r+1 l ≤ r + 1 unit vectors v i ∈ S M A X v_i\in S_{\mathrm{MAX}} v i ∈ S MAX and weights 0 ≤ α i ≤ 1 0\le\alpha_i\le1 0 ≤ α i ≤ 1 with ∑ i = 1 l α i = 1 \sum_{i=1}^{l}\alpha_i=1 ∑ i = 1 l α i = 1 such that y i = ( ⟨ B 1 v i , v i ⟩ , … , ⟨ B r v i , v i ⟩ ) y_i=(\langle B^{1}v_i,v_i\rangle,\ldots,\langle B^{r}v_i,v_i\rangle) y i = (⟨ B 1 v i , v i ⟩ , … , ⟨ B r v i , v i ⟩) for every i i i and
0 ˉ = ∑ i = 1 l α i y i = ∑ i = 1 l α i ( ⟨ B 1 v i , v i ⟩ , … , ⟨ B r v i , v i ⟩ ) . \bar{0}=\sum_{i=1}^{l}\alpha_i y_i=\sum_{i=1}^{l}\alpha_i(\langle B^{1}v_i,v_i\rangle,\ldots,\langle B^{r}v_i,v_i\rangle). 0 ˉ = i = 1 ∑ l α i y i = i = 1 ∑ l α i (⟨ B 1 v i , v i ⟩ , … , ⟨ B r v i , v i ⟩) .
This implies that 0 = ∑ i = 1 l α i ⟨ B j v i , v i ⟩ 0=\sum_{i=1}^{l}\alpha_i\langle B^{j}v_i,v_i\rangle 0 = ∑ i = 1 l α i ⟨ B j v i , v i ⟩ , for every 1 ≤ j ≤ r 1\le j\le r 1 ≤ j ≤ r . Therefore, every W ∈ W W\in\mathcal{W} W ∈ W satisfies
0 = ∑ i = 1 l α i ⟨ W v i , v i ⟩ . 0=\sum_{i=1}^{l}\alpha_i\langle Wv_i,v_i\rangle. 0 = i = 1 ∑ l α i ⟨ W v i , v i ⟩ .
We now distinguish two cases.
If S M A X = S λ max S_{\mathrm{MAX}}=S_{\lambda_{\max}} S MAX = S λ m a x , then ( A 0 + W 0 ) v i = ∥ A 0 + W 0 ∥ v i (A_{0}+W_{0})v_i=\lVert A_{0}+W_{0}\rVert v_i ( A 0 + W 0 ) v i = ∥ A 0 + W 0 ∥ v i for every $1\le i\le l.
If S M A X = S λ min S_{\mathrm{MAX}}=S_{\lambda_{\min}} S MAX = S λ m i n , then ( A 0 + W 0 ) v i = − ∥ A 0 + W 0 ∥ v i (A_{0}+W_{0})v_i=-\lVert A_{0}+W_{0}\rVert v_i ( A 0 + W 0 ) v i = − ∥ A 0 + W 0 ∥ v i for every 1 ≤ i ≤ l 1\le i\le l 1 ≤ i ≤ l , so that u i = − v i u_i=-v_i u i = − v i by (5.5). Hence, since 0 = ∑ i = 1 l α i ⟨ W v i , v i ⟩ 0=\sum_{i=1}^{l}\alpha_i\langle Wv_i,v_i\rangle 0 = ∑ i = 1 l α i ⟨ W v i , v i ⟩ ,
0 = ∑ i = 1 l α i ⟨ W v i , − v i ⟩ . 0=\sum_{i=1}^{l}\alpha_i\langle Wv_i,-v_i\rangle. 0 = i = 1 ∑ l α i ⟨ W v i , − v i ⟩ .
In both cases, conditions (5.4) and (5.5) are satisfied.
Now suppose that λ max + λ min = 0 \lambda_{\max}+\lambda_{\min}=0 λ m a x + λ m i n = 0 , that is S M A X = S λ max ⊕ S λ min S_{\mathrm{MAX}}=S_{\lambda_{\max}}\oplus S_{\lambda_{\min}} S MAX = S λ m a x ⊕ S λ m i n . Set
T max = { ( ⟨ B 1 v , v ⟩ , … , ⟨ B r v , v ⟩ ) : v ∈ S λ max , ∥ v ∥ = 1 } T_{\max}=\{(\langle B^{1}v,v\rangle,\ldots,\langle B^{r}v,v\rangle):v\in S_{\lambda_{\max}},\ \lVert v\rVert=1\} T m a x = {(⟨ B 1 v , v ⟩ , … , ⟨ B r v , v ⟩) : v ∈ S λ m a x , ∥ v ∥ = 1 }
and
T min = { − ( ⟨ B 1 v , v ⟩ , … , ⟨ B r v , v ⟩ ) : v ∈ S λ min , ∥ v ∥ = 1 } . T_{\min}=\{-(\langle B^{1}v,v\rangle,\ldots,\langle B^{r}v,v\rangle):v\in S_{\lambda_{\min}},\ \lVert v\rVert=1\}. T m i n = { − (⟨ B 1 v , v ⟩ , … , ⟨ B r v , v ⟩) : v ∈ S λ m i n , ∥ v ∥ = 1 } .
Since m S λ max = conv ( T max ) m_{S_{\lambda_{\max}}}=\operatorname{conv}(T_{\max}) m S λ m a x = conv ( T m a x ) , − m S λ min = conv ( T min ) -m_{S_{\lambda_{\min}}}=\operatorname{conv}(T_{\min}) − m S λ m i n = conv ( T m i n ) and conv ( C ∪ D ) = conv ( conv ( C ) ∪ conv ( D ) ) \operatorname{conv}(C\cup D)=\operatorname{conv}(\operatorname{conv}(C)\cup\operatorname{conv}(D)) conv ( C ∪ D ) = conv ( conv ( C ) ∪ conv ( D )) for any C , D ⊆ R r C,D\subseteq\mathbb{R}^{r} C , D ⊆ R r , the condition
0 ˉ ∈ conv ( m S λ max ∪ − m S λ min ) \bar{0}\in\operatorname{conv}\left(m_{S_{\lambda_{\max}}}\cup-m_{S_{\lambda_{\min}}}\right) 0 ˉ ∈ conv ( m S λ m a x ∪ − m S λ m i n )
is equivalent to 0 ˉ ∈ conv ( T max ∪ T min ) \bar{0}\in\operatorname{conv}(T_{\max}\cup T_{\min}) 0 ˉ ∈ conv ( T m a x ∪ T m i n ) .
By Carathéodory’s theorem in R r \mathbb{R}^{r} R r , 0 ˉ \bar{0} 0 ˉ is a convex combination of at most r + 1 r+1 r + 1 points of T max ∪ T min T_{\max}\cup T_{\min} T m a x ∪ T m i n , that is,
∑ i = 1 l α i y i = 0 ˉ \sum_{i=1}^{l}\alpha_i y_i=\bar{0} i = 1 ∑ l α i y i = 0 ˉ
for some l ≤ r + 1 l \le r+1 l ≤ r + 1 , weights α i ≥ 0 \alpha_i \ge0 α i ≥ 0 with ∑ i = 1 l α i = 1 \sum_{i=1}^{l}\alpha_i=1 ∑ i = 1 l α i = 1 , and y i ∈ T max ∪ T min y_i \in T_{\max}\cup T_{\min} y i ∈ T m a x ∪ T m i n . Each y i y_i y i is of the form
y i = s i ( ⟨ B 1 w i , w i ⟩ , … , ⟨ B r w i , w i ⟩ ) , y_i=s_i\left(\langle B^1w_i,w_i\rangle,\ldots,\langle B^rw_i,w_i\rangle\right), y i = s i ( ⟨ B 1 w i , w i ⟩ , … , ⟨ B r w i , w i ⟩ ) ,
where w i w_i w i is a unit vector, s i = 1 s_i=1 s i = 1 if w i ∈ S λ max w_i\in S_{\lambda_{\max}} w i ∈ S λ m a x and s i = − 1 s_i=-1 s i = − 1 if w i ∈ S λ min w_i\in S_{\lambda_{\min}} w i ∈ S λ m i n . Therefore,
∑ i = 1 l α i s i ⟨ B j w i , w i ⟩ = 0 , for every 1 ≤ j ≤ r . \sum_{i=1}^{l}\alpha_i s_i\langle B^j w_i,w_i\rangle=0,\quad\text{for every }1\le j\le r. i = 1 ∑ l α i s i ⟨ B j w i , w i ⟩ = 0 , for every 1 ≤ j ≤ r .
Then, regrouping the points of each type and normalizing the weights on each side, there exist 0 ≤ α ≤ 1 0\le\alpha\le1 0 ≤ α ≤ 1 , ρ ∈ D S λ max \rho\in\mathcal{D}_{S_{\lambda_{\max}}} ρ ∈ D S λ m a x of rank at most d d d and σ ∈ D S λ min \sigma\in\mathcal{D}_{S_{\lambda_{\min}}} σ ∈ D S λ m i n of rank at most k k k , with d + k ≤ r + 1 d+k\le r+1 d + k ≤ r + 1 , such that Φ ( ρ ) ∈ m S λ max \Phi(\rho)\in mS_{\lambda_{\max}} Φ ( ρ ) ∈ m S λ m a x , Φ ( σ ) ∈ m S λ min \Phi(\sigma)\in mS_{\lambda_{\min}} Φ ( σ ) ∈ m S λ m i n (see (19) ) and
0 ‾ = α Φ ( ρ ) + ( 1 − α ) Φ ( − σ ) = Φ ( α ρ − ( 1 − α ) σ ) . \overline{0}=\alpha\Phi(\rho)+(1-\alpha)\Phi(-\sigma)=\Phi\left(\alpha\rho-(1-\alpha)\sigma\right). 0 = α Φ ( ρ ) + ( 1 − α ) Φ ( − σ ) = Φ ( α ρ − ( 1 − α ) σ ) .
Thus, tr ( ( α ρ − ( 1 − α ) σ ) B j ) = 0 \operatorname{tr}\left((\alpha\rho-(1-\alpha)\sigma)B^j\right)=0 tr ( ( α ρ − ( 1 − α ) σ ) B j ) = 0 for every 1 ≤ j ≤ r 1\le j\le r 1 ≤ j ≤ r , and this implies that
tr ( ( α ρ − ( 1 − α ) σ ) W ) = 0 \operatorname{tr}\left((\alpha\rho-(1-\alpha)\sigma)W\right)=0 tr ( ( α ρ − ( 1 − α ) σ ) W ) = 0
for all W ∈ W W\in\mathcal{W} W ∈ W . Since ρ \rho ρ and σ \sigma σ have orthogonal supports, ∥ α ρ − ( 1 − α ) σ ∥ 1 = α + ( 1 − α ) = 1 \|\alpha\rho-(1-\alpha)\sigma\|_1=\alpha+(1-\alpha)=1 ∥ α ρ − ( 1 − α ) σ ∥ 1 = α + ( 1 − α ) = 1 , and therefore α ρ − ( 1 − α ) σ ∈ W ∘ \alpha\rho-(1-\alpha)\sigma\in\mathcal{W}^{\circ} α ρ − ( 1 − α ) σ ∈ W ∘ .
On the other hand, since ρ ∈ D S λ max \rho\in\mathcal{D}_{S_{\lambda_{\max}}} ρ ∈ D S λ m a x and σ ∈ D S λ min \sigma\in\mathcal{D}_{S_{\lambda_{\min}}} σ ∈ D S λ m i n , both can be written as
ρ = ∑ i = 1 d β i v i v i ∗ and σ = ∑ j = 1 k γ j z j z j ∗ , \rho=\sum_{i=1}^{d}\beta_i v_i v_i^* \quad\text{and}\quad\sigma=\sum_{j=1}^{k}\gamma_j z_j z_j^*, ρ = i = 1 ∑ d β i v i v i ∗ and σ = j = 1 ∑ k γ j z j z j ∗ ,
with 0 ≤ β i , γ j ≤ 1 0\le\beta_i,\gamma_j\le1 0 ≤ β i , γ j ≤ 1 , ∑ i = 1 d β i = ∑ j = 1 k γ j = 1 \sum_{i=1}^{d}\beta_i=\sum_{j=1}^{k}\gamma_j=1 ∑ i = 1 d β i = ∑ j = 1 k γ j = 1 , and orthonormal sets { v i } i = 1 d ⊂ S λ max \{v_i\}_{i=1}^{d}\subset S_{\lambda_{\max}} { v i } i = 1 d ⊂ S λ m a x , { z j } j = 1 k ⊂ S λ min \{z_j\}_{j=1}^{k}\subset S_{\lambda_{\min}} { z j } j = 1 k ⊂ S λ m i n . Then,
tr ( ( α ρ − ( 1 − α ) σ ) W ) = α tr ( ρ W ) − ( 1 − α ) tr ( σ W ) = α tr ( ∑ i = 1 d β i v i v i ∗ W ) − ( 1 − α ) tr ( ∑ j = 1 k γ j z j z j ∗ W ) = α ∑ i = 1 d β i tr ( v i v i ∗ W ) − ( 1 − α ) ∑ j = 1 k γ j tr ( z j z j ∗ W ) = α ∑ i = 1 d β i ⟨ W v i , v i ⟩ − ( 1 − α ) ∑ j = 1 k γ j ⟨ W z j , z j ⟩ = α ∑ i = 1 d β i ⟨ W v i , v i ⟩ + ( 1 − α ) ∑ j = 1 k γ j ⟨ W z j , − z j ⟩ = ∑ i = 1 d α β i ⟨ W v i , v i ⟩ + ∑ j = 1 k ( 1 − α ) γ j ⟨ W z j , − z j ⟩ . \begin{aligned}
\operatorname{tr}\left((\alpha\rho-(1-\alpha)\sigma)W\right)
&=\alpha\operatorname{tr}(\rho W)-(1-\alpha)\operatorname{tr}(\sigma W)\\
&=\alpha\operatorname{tr}\left(\sum_{i=1}^{d}\beta_i v_i v_i^*W\right)-(1-\alpha)\operatorname{tr}\left(\sum_{j=1}^{k}\gamma_j z_j z_j^*W\right)\\
&=\alpha\sum_{i=1}^{d}\beta_i\operatorname{tr}(v_i v_i^*W)-(1-\alpha)\sum_{j=1}^{k}\gamma_j\operatorname{tr}(z_j z_j^*W)\\
&=\alpha\sum_{i=1}^{d}\beta_i\langle Wv_i,v_i\rangle-(1-\alpha)\sum_{j=1}^{k}\gamma_j\langle Wz_j,z_j\rangle\\
&=\alpha\sum_{i=1}^{d}\beta_i\langle Wv_i,v_i\rangle+(1-\alpha)\sum_{j=1}^{k}\gamma_j\langle Wz_j,-z_j\rangle\\
&=\sum_{i=1}^{d}\alpha\beta_i\langle Wv_i,v_i\rangle+\sum_{j=1}^{k}(1-\alpha)\gamma_j\langle Wz_j,-z_j\rangle.
\end{aligned} tr ( ( α ρ − ( 1 − α ) σ ) W ) = α tr ( ρ W ) − ( 1 − α ) tr ( σ W ) = α tr ( i = 1 ∑ d β i v i v i ∗ W ) − ( 1 − α ) tr ( j = 1 ∑ k γ j z j z j ∗ W ) = α i = 1 ∑ d β i tr ( v i v i ∗ W ) − ( 1 − α ) j = 1 ∑ k γ j tr ( z j z j ∗ W ) = α i = 1 ∑ d β i ⟨ W v i , v i ⟩ − ( 1 − α ) j = 1 ∑ k γ j ⟨ W z j , z j ⟩ = α i = 1 ∑ d β i ⟨ W v i , v i ⟩ + ( 1 − α ) j = 1 ∑ k γ j ⟨ W z j , − z j ⟩ = i = 1 ∑ d α β i ⟨ W v i , v i ⟩ + j = 1 ∑ k ( 1 − α ) γ j ⟨ W z j , − z j ⟩ .
Observe that α ∑ i = 1 d β i + ( 1 − α ) ∑ j = 1 k γ j = α + 1 − α = 1 \alpha\sum_{i=1}^{d}\beta_i+(1-\alpha)\sum_{j=1}^{k}\gamma_j=\alpha+1-\alpha=1 α ∑ i = 1 d β i + ( 1 − α ) ∑ j = 1 k γ j = α + 1 − α = 1 . Then if we define
ω i = { α β i if 1 ≤ i ≤ d , ( 1 − α ) γ i − d if d + 1 ≤ i ≤ d + k , \omega_i=
\begin{cases}
\alpha\beta_i & \text{if }1\le i\le d,\\
(1-\alpha)\gamma_{i-d} & \text{if }d+1\le i\le d+k,
\end{cases} ω i = { α β i ( 1 − α ) γ i − d if 1 ≤ i ≤ d , if d + 1 ≤ i ≤ d + k ,
we obtain that 0 ≤ ω i ≤ 1 0 \leq\omega_i \leq1 0 ≤ ω i ≤ 1 for every i i i and ∑ i = 1 d + k ω i = 1 \sum_{i=1}^{d+k}\omega_i = 1 ∑ i = 1 d + k ω i = 1 . Now writing v d + j : = z j v_{d+j}:=z_j v d + j := z j for 1 ≤ j ≤ k 1 \leq j \leq k 1 ≤ j ≤ k , and u i : = v i u_i:=v_i u i := v i (resp. u i : = − v i u_i:=-v_i u i := − v i ) for 1 ≤ i ≤ d 1 \leq i \leq d 1 ≤ i ≤ d (resp. d + 1 ≤ i ≤ d + k d+1 \leq i \leq d+k d + 1 ≤ i ≤ d + k ), we get
∑ i = 1 d + k ω i ⟨ W v i , u i ⟩ = 0 , \sum_{i=1}^{d+k}\omega_i\langle Wv_i,u_i\rangle=0, i = 1 ∑ d + k ω i ⟨ W v i , u i ⟩ = 0 ,
which are exactly conditions (41) and (42) with l = d + k ≤ r + 1 l=d+k\leq r+1 l = d + k ≤ r + 1 by the Carathéodory decomposition above. Moreover v i ∈ S λ max v_i\in S_{\lambda_{\max}} v i ∈ S λ m a x for 1 ≤ i ≤ d 1\leq i\leq d 1 ≤ i ≤ d and v i ∈ S λ min v_i\in S_{\lambda_{\min}} v i ∈ S λ m i n for d + 1 ≤ i ≤ d + k d+1\leq i\leq d+k d + 1 ≤ i ≤ d + k , so that all of them are unit vectors of S M A X S_{\mathrm{MAX}} S MAX and eigenvectors of A 0 + W 0 A_0+W_0 A 0 + W 0 associated with ∥ A 0 + W 0 ∥ \lVert A_0+W_0\rVert ∥ A 0 + W 0 ∥ or with − ∥ A 0 + W 0 ∥ -\lVert A_0+W_0\rVert − ∥ A 0 + W 0 ∥ , as required in item (2). This completes the proof of (3) ⇒ \Rightarrow ⇒ (2).
(3) ⇔ \Leftrightarrow ⇔ (4) By Corollary 4.6 , each condition in (43) is equivalent to 0 ‾ ∈ ∂ ( ∥ A ( ⋅ ) ∥ ) ( x ) \overline{0}\in\partial(\lVert A(\cdot)\rVert)(x) 0 ∈ ∂ (∥ A ( ⋅ )∥) ( x ) , since
∂ ( ∥ A ( ⋅ ) ∥ ) ( x ) = m S M A X ( A ( x ) ) = m S M A X . \partial(\lVert A(\cdot)\rVert)(x)=m_{S_{\mathrm{MAX}}(A(x))}=m_{S_{\mathrm{MAX}}}. ∂ (∥ A ( ⋅ )∥) ( x ) = m S MAX ( A ( x )) = m S MAX .
∂ ( ∥ A ( ⋅ ) ∥ ) ( x ) = − m S M A X ( A ( x ) ) = − m S M A X . \partial(\lVert A(\cdot)\rVert)(x)=-m_{S_{\mathrm{MAX}}(A(x))}=-m_{S_{\mathrm{MAX}}}. ∂ (∥ A ( ⋅ )∥) ( x ) = − m S MAX ( A ( x )) = − m S MAX .
In both cases, we deduce that 0 ‾ ∈ ∂ ( ∥ A ( ⋅ ) ∥ ) ( x ) ⇔ 0 ‾ ∈ m S M A X \overline{0}\in\partial(\lVert A(\cdot)\rVert)(x)\Leftrightarrow\overline{0}\in m_{S_{\mathrm{MAX}}} 0 ∈ ∂ (∥ A ( ⋅ )∥) ( x ) ⇔ 0 ∈ m S MAX .
∂ ( ∥ A ( ⋅ ) ∥ ) ( x ) = conv ( m S λ max ∪ − m S λ min ) . \partial(\lVert A(\cdot)\rVert)(x)=\operatorname{conv}\left(m_{S_{\lambda_{\max}}}\cup-m_{S_{\lambda_{\min}}}\right). ∂ (∥ A ( ⋅ )∥) ( x ) = conv ( m S λ m a x ∪ − m S λ m i n ) .
Therefore, $\overline{0}\in\partial(\lVert A(\cdot)\rVert)(x)\Leftrightarrow\overline{0}\in\operatorname{conv}\left(m_{S_{\lambda_{\max}}}\cup -m_{S_{\lambda_{\min}}}\right).
In the balanced case, condition (43) always follows from the intersection condition of Theorem 5.2 : if Φ ( ρ ) = Φ ( σ ) \Phi(\rho)=\Phi(\sigma) Φ ( ρ ) = Φ ( σ ) for some ρ ∈ D S λ max \rho\in D_{S_{\lambda_{\max}}} ρ ∈ D S λ m a x and σ ∈ D S λ min \sigma\in D_{S_{\lambda_{\min}}} σ ∈ D S λ m i n , then 0 ‾ = 1 2 Φ ( ρ ) − 1 2 Φ ( σ ) ∈ conv ( m S λ max ∪ − m S λ min ) \overline{0}=\frac{1}{2}\Phi(\rho)-\frac{1}{2}\Phi(\sigma)\in\operatorname{conv}\left(m_{S_{\lambda_{\max}}}\cup-m_{S_{\lambda_{\min}}}\right) 0 = 2 1 Φ ( ρ ) − 2 1 Φ ( σ ) ∈ conv ( m S λ m a x ∪ − m S λ m i n ) . Both conditions are equivalent when I ∈ W I\in\mathcal{W} I ∈ W , but the converse implication may fail when I ∉ W I\notin\mathcal{W} I ∈ / W , as the following remark shows.
Remark 5.10. When λ max + λ min = 0 \lambda_{\max}+\lambda_{\min}=0 λ m a x + λ m i n = 0 , the second condition in (43) means that there exist α ∈ [ 0 , 1 ] \alpha\in[0,1] α ∈ [ 0 , 1 ] , ρ ∈ D S λ max \rho\in D_{S_{\lambda_{\max}}} ρ ∈ D S λ m a x and σ ∈ D S λ min \sigma\in D_{S_{\lambda_{\min}}} σ ∈ D S λ m i n such that α Φ ( ρ ) − ( 1 − α ) Φ ( σ ) = 0 ‾ \alpha\Phi(\rho)-(1-\alpha)\Phi(\sigma)=\overline{0} α Φ ( ρ ) − ( 1 − α ) Φ ( σ ) = 0 . Equivalently, α Φ ( ρ ) = ( 1 − α ) Φ ( σ ) \alpha\Phi(\rho)=(1-\alpha)\Phi(\sigma) α Φ ( ρ ) = ( 1 − α ) Φ ( σ ) , that is,
α m S λ max ∩ ( 1 − α ) m S λ min ≠ ∅ . \alpha m_{S_{\lambda_{\max}}}\cap(1-\alpha)m_{S_{\lambda_{\min}}}\neq\varnothing. α m S λ m a x ∩ ( 1 − α ) m S λ m i n = ∅ .
By linearity Φ ( α ρ − ( 1 − α ) σ ) = 0 ‾ \Phi(\alpha\rho-(1-\alpha)\sigma)=\overline{0} Φ ( α ρ − ( 1 − α ) σ ) = 0 , so F = α ρ − ( 1 − α ) σ ∈ ker ( Φ ) = W ∘ F=\alpha\rho-(1-\alpha)\sigma\in\ker(\Phi)=\mathcal{W}^{\circ} F = α ρ − ( 1 − α ) σ ∈ ker ( Φ ) = W ∘ ; moreover, since ρ \rho ρ and σ \sigma σ have orthogonal supports, ∥ F ∥ 1 = α + ( 1 − α ) = 1 \lVert F\rVert_1=\alpha+(1-\alpha)=1 ∥ F ∥ 1 = α + ( 1 − α ) = 1 , so F ∈ W 1 ∘ F\in\mathcal{W}^{\circ}_1 F ∈ W 1 ∘ (it is the witness of item (5) of Theorem 5.9 ). Therefore
tr ( F W ) = 0 \operatorname{tr}(FW)=0 tr ( F W ) = 0
for every W ∈ W W\in\mathcal{W} W ∈ W . In particular, if I ∈ W I\in\mathcal{W} I ∈ W ,
tr ( F ) = α tr ( ρ ) − ( 1 − α ) tr ( σ ) = 2 α − 1 = 0. \operatorname{tr}(F)=\alpha\operatorname{tr}(\rho)-(1-\alpha)\operatorname{tr}(\sigma)=2\alpha-1=0. tr ( F ) = α tr ( ρ ) − ( 1 − α ) tr ( σ ) = 2 α − 1 = 0.
Hence α = 1 2 \alpha=\frac{1}{2} α = 2 1 and Φ ( ρ ) = Φ ( σ ) ∈ m S λ max ∩ m S λ min \Phi(\rho)=\Phi(\sigma)\in m_{S_{\lambda_{\max}}}\cap m_{S_{\lambda_{\min}}} Φ ( ρ ) = Φ ( σ ) ∈ m S λ m a x ∩ m S λ m i n , which is item (3) of Theorem 5.2 . If I ∉ W I\notin\mathcal{W} I ∈ / W , the weight α \alpha α is not determined in this way, and a W \mathcal{W} W -minimal matrix with balanced spectrum may satisfy
m S λ max ∩ m S λ min = ∅ and m S λ max ∩ − m S λ min = ∅ . m_{S_{\lambda_{\max}}}\cap m_{S_{\lambda_{\min}}}=\varnothing
\quad\text{and}\quad
m_{S_{\lambda_{\max}}}\cap-m_{S_{\lambda_{\min}}}=\varnothing. m S λ m a x ∩ m S λ m i n = ∅ and m S λ m a x ∩ − m S λ m i n = ∅ .
For instance, let n = 2 n=2 n = 2 , W = R B \mathcal{W}=\mathbb{R}B W = R B with B = 1 5 ( 1 0 0 2 ) B=\frac{1}{\sqrt{5}}\begin{pmatrix}1&0\\0&2\end{pmatrix} B = 5 1 ( 1 0 0 2 ) (so that I ∉ W I\notin\mathcal{W} I ∈ / W ), A 0 = ( 1 0 0 − 1 ) A_0=\begin{pmatrix}1&0\\0&-1\end{pmatrix} A 0 = ( 1 0 0 − 1 ) and W 0 = 0 W_0=0 W 0 = 0 . For t ∈ R t\in\mathbb{R} t ∈ R ,
∥ A 0 + t 5 B ∥ = max { ∣ 1 + t ∣ , ∣ 2 t − 1 ∣ } ≥ 1 , \left\lVert A_0+t\sqrt{5}B\right\rVert=\max\{|1+t|,|2t-1|\}\geq1, A 0 + t 5 B = max { ∣1 + t ∣ , ∣2 t − 1∣ } ≥ 1 ,
since the first term is greater than 1 if t > 0 t > 0 t > 0 and the second one is greater than 1 if t < 0 t < 0 t < 0 . Hence A 0 A_0 A 0 is W W W -minimal, with λ max = 1 = − λ min \lambda_{\max} = 1 = -\lambda_{\min} λ m a x = 1 = − λ m i n . Both eigenvalues are simple, S λ max = span { e 1 } S_{\lambda_{\max}} = \operatorname{span}\{e_1\} S λ m a x = span { e 1 } and S λ min = span { e 2 } S_{\lambda_{\min}} = \operatorname{span}\{e_2\} S λ m i n = span { e 2 } , so
m S λ max = { 1 5 } , m S λ min = { 2 5 } . m_{S_{\lambda_{\max}}} = \left\{\frac{1}{\sqrt{5}}\right\}, \qquad m_{S_{\lambda_{\min}}} = \left\{\frac{2}{\sqrt{5}}\right\}. m S λ m a x = { 5 1 } , m S λ m i n = { 5 2 } .
and both intersections above are empty. However, 0 ‾ = 2 3 ⋅ 1 5 + 1 3 ⋅ ( − 2 5 ) ∈ conv ( m S λ max ∪ − m S λ min ) \overline{0} = \frac{2}{3} \cdot\frac{1}{\sqrt{5}} + \frac{1}{3} \cdot\left(-\frac{2}{\sqrt{5}}\right) \in\operatorname{conv}\left(m_{S_{\lambda_{\max}}} \cup-m_{S_{\lambda_{\min}}}\right) 0 = 3 2 ⋅ 5 1 + 3 1 ⋅ ( − 5 2 ) ∈ conv ( m S λ m a x ∪ − m S λ m i n ) , as predicted by Theorem 5.9 . Here the witness is F = 2 3 e 1 e 1 ∗ − 1 3 e 2 e 2 ∗ F = \frac{2}{3}e_1e_1^* - \frac{1}{3}e_2e_2^* F = 3 2 e 1 e 1 ∗ − 3 1 e 2 e 2 ∗ . The same phenomenon occurs in Example 6.1, where W W W comes from a completion pattern: there α = 5 + 2 5 10 ≠ 1 2 \alpha= \frac{5+2\sqrt{5}}{10} \ne\frac{1}{2} α = 10 5 + 2 5 = 2 1 , and the moments are the single points ( a , 0 , 0 , 0 , 0 ) (a,0,0,0,0) ( a , 0 , 0 , 0 , 0 ) and ( b , 0 , 0 , 0 , 0 ) (b,0,0,0,0) ( b , 0 , 0 , 0 , 0 ) with a = 5 − 2 5 10 a = \frac{5-2\sqrt{5}}{10} a = 10 5 − 2 5 and b = 5 + 2 5 10 b = \frac{5+2\sqrt{5}}{10} b = 10 5 + 2 5 , which are distinct and positive.
Examples and applications
Example 6.1. Let A 0 = ( 0 1 0 1 2 0 0 0 1 ) A_0 = \begin{pmatrix} 0 & 1 & 0 \\ 1 & 2 & 0 \\ 0 & 0 & 1 \end{pmatrix} A 0 = 0 1 0 1 2 0 0 0 1 . Then
I 0 = { ( 1 , 1 ) , ( 1 , 3 ) , ( 2 , 3 ) , ( 3 , 1 ) , ( 3 , 2 ) } , I_0 = \{(1,1),(1,3),(2,3),(3,1),(3,2)\}, I 0 = {( 1 , 1 ) , ( 1 , 3 ) , ( 2 , 3 ) , ( 3 , 1 ) , ( 3 , 2 )} ,
W = span { B 1 , 1 , B 1 , 3 , B 2 , 3 , B 3 , 1 , B 3 , 2 } . W = \operatorname{span}\{B^{1,1}, B^{1,3}, B^{2,3}, B^{3,1}, B^{3,2}\}. W = span { B 1 , 1 , B 1 , 3 , B 2 , 3 , B 3 , 1 , B 3 , 2 } .
By Proposition 2.2, and since the trace norm of a block diagonal matrix is the sum of the trace norms of its blocks,
W 1 ∘ = { Y = ( 0 Y 12 0 Y 12 ‾ Y 22 0 0 0 Y 33 ) : 4 ∣ Y 12 ∣ 2 + Y 22 2 + ∣ Y 33 ∣ = 1 } . \mathcal{W}_1^\circ= \left\{Y = \begin{pmatrix} 0 & Y_{12} & 0 \\ \overline{Y_{12}} & Y_{22} & 0 \\ 0 & 0 & Y_{33} \end{pmatrix} : \sqrt{4|Y_{12}|^2 + Y_{22}^2 + |Y_{33}|} = 1\right\}. W 1 ∘ = ⎩ ⎨ ⎧ Y = 0 Y 12 0 Y 12 Y 22 0 0 0 Y 33 : 4∣ Y 12 ∣ 2 + Y 22 2 + ∣ Y 33 ∣ = 1 ⎭ ⎬ ⎫ .
because the eigenvalues 1 2 ( Y 22 ± Y 22 2 + 4 ∣ Y 12 ∣ 2 ) \frac{1}{2}\left(Y_{22} \pm\sqrt{Y_{22}^2 + 4|Y_{12}|^2}\right) 2 1 ( Y 22 ± Y 22 2 + 4∣ Y 12 ∣ 2 ) of the 2 × 2 2 \times2 2 × 2 block have opposite signs, so that its trace norm is their difference. By the duality formula (2.9), min W ∈ W ∥ A 0 + W ∥ \min_{W \in\mathcal{W}} \|A_0 + W\| min W ∈ W ∥ A 0 + W ∥ equals
max ( x , y , z ) ∈ C × R × R ∣ 2 Re ( x ) + 2 y + z ∣ , such that 4 ∣ x ∣ 2 + y 2 + ∣ z ∣ = 1. (46) \max_{(x,y,z)\in\mathbb{C}\times\mathbb{R}\times\mathbb{R}} |2\operatorname{Re}(x) + 2y + z|,\quad\text{such that } \sqrt{4|x|^2+y^2+|z|}=1. \tag*{(46)} ( x , y , z ) ∈ C × R × R max ∣2 Re ( x ) + 2 y + z ∣ , such that 4∣ x ∣ 2 + y 2 + ∣ z ∣ = 1. (46)
where Y Y Y has entries Y 12 = x Y_{12} = x Y 12 = x , Y 22 = y Y_{22} = y Y 22 = y , Y 33 = z Y_{33} = z Y 33 = z . The objective does not depend on Im ( x ) \operatorname{Im}(x) Im ( x ) , which only increases ∣ x ∣ |x| ∣ x ∣ in the restriction, so we may assume x ∈ R x \in\mathbb{R} x ∈ R ; replacing F F F by − F -F − F if necessary, also 2 x + 2 y + z ≥ 0 2x + 2y + z \ge0 2 x + 2 y + z ≥ 0 , and then z ≥ 0 z \ge0 z ≥ 0 at a maximizer, since the restriction only involves ∣ z ∣ |z| ∣ z ∣ . Put s = 4 x 2 + y 2 = 1 − z s = \sqrt{4x^2+y^2} = \sqrt{1-z} s = 4 x 2 + y 2 = 1 − z . By the Cauchy–Schwarz inequality 2 x + 2 y ≤ 5 s 2x+2y \le\sqrt{5}s 2 x + 2 y ≤ 5 s , hence 2 x + 2 y + z ≤ 1 + ( 5 − 1 ) s ≤ 5 2x+2y+z \le1+(\sqrt{5}-1)s \le\sqrt{5} 2 x + 2 y + z ≤ 1 + ( 5 − 1 ) s ≤ 5 , with equality if and only if z = 0 z=0 z = 0 and ( 2 x , y ) = 1 5 ( 1 , 2 ) (2x,y)=\frac{1}{\sqrt{5}}(1,2) ( 2 x , y ) = 5 1 ( 1 , 2 ) . Therefore
dist ( A 0 , W ) = 5 , attained at x = 1 2 5 , y = 2 5 , z = 0. (47) \operatorname{dist}(A_0,W) = \sqrt{5}, \qquad\text{attained at } x = \frac{1}{2\sqrt{5}},\quad y = \frac{2}{\sqrt{5}},\quad z=0. \tag*{(47)} dist ( A 0 , W ) = 5 , attained at x = 2 5 1 , y = 5 2 , z = 0. (47)
The minimum is attained at W 0 = − 2 B 1 , 1 W_0=-2B^{1,1} W 0 = − 2 B 1 , 1 , that is, 2 B 1 , 1 2B^{1,1} 2 B 1 , 1 is a best approximation of A 0 A_0 A 0 in W W W : the eigenvalues of A 0 + W 0 A_0+W_0 A 0 + W 0 are { 1 , 5 , − 5 } \{1,\sqrt{5},-\sqrt{5}\} { 1 , 5 , − 5 } .
Let us check this with item (3) of Theorem 5.9 . Since λ max + λ min = 0 \lambda_{\max}+\lambda_{\min}=0 λ m a x + λ m i n = 0 , the condition to be verified is the second one in (5.6). Both eigenvalues are simple, with S λ max = span { ( 5 − 2 , 1 , 0 ) } S_{\lambda_{\max}}=\operatorname{span}\{(\sqrt{5}-2,1,0)\} S λ m a x = span {( 5 − 2 , 1 , 0 )} and S λ min = span { ( − ( 2 + 5 ) , 1 , 0 ) } S_{\lambda_{\min}}=\operatorname{span}\{(-(2+\sqrt{5}),1,0)\} S λ m i n = span {( − ( 2 + 5 ) , 1 , 0 )} , so that D S λ max = { ρ + } D_{S_{\lambda_{\max}}}=\{\rho_+\} D S λ m a x = { ρ + } and D S λ min = { ρ − } D_{S_{\lambda_{\min}}}=\{\rho_-\} D S λ m i n = { ρ − } , where
ρ + = 1 10 ( 5 − 2 5 5 0 5 5 + 2 5 0 0 0 0 ) , ρ − = 1 10 ( 5 + 2 5 − 5 0 − 5 5 − 2 5 0 0 0 0 ) . \rho_+ = \frac{1}{10}\begin{pmatrix} 5-2\sqrt{5} & \sqrt{5} & 0 \\ \sqrt{5} & 5+2\sqrt{5} & 0 \\ 0 & 0 & 0 \end{pmatrix}, \qquad
\rho_- = \frac{1}{10}\begin{pmatrix} 5+2\sqrt{5} & -\sqrt{5} & 0 \\ -\sqrt{5} & 5-2\sqrt{5} & 0 \\ 0 & 0 & 0 \end{pmatrix}. ρ + = 10 1 5 − 2 5 5 0 5 5 + 2 5 0 0 0 0 , ρ − = 10 1 5 + 2 5 − 5 0 − 5 5 − 2 5 0 0 0 0 .
Since ρ ± \rho_{\pm} ρ ± vanish at the positions ( 1 , 3 ) (1,3) ( 1 , 3 ) and ( 2 , 3 ) (2,3) ( 2 , 3 ) , by (7) and (8) only the coordinate corresponding to B 1 , 1 B^{1,1} B 1 , 1 survives: m S λ max = { ( a , 0 , 0 , 0 , 0 ) } m_{S_{\lambda_{\max}}}=\{(a,0,0,0,0)\} m S λ m a x = {( a , 0 , 0 , 0 , 0 )} and − m S λ min = { ( − b , 0 , 0 , 0 , 0 ) } -m_{S_{\lambda_{\min}}}=\{(-b,0,0,0,0)\} − m S λ m i n = {( − b , 0 , 0 , 0 , 0 )} , with a = ( ρ + ) 11 a=(\rho_+)_{11} a = ( ρ + ) 11 and b = ( ρ − ) 11 b=(\rho_-)_{11} b = ( ρ − ) 11 . As ρ + + ρ − \rho_+ + \rho_- ρ + + ρ − is the orthogonal projection onto span { e 1 , e 2 } \operatorname{span}\{e_1,e_2\} span { e 1 , e 2 } , we have a + b = 1 a+b=1 a + b = 1 , so α a − ( 1 − α ) b = 0 \alpha a-(1-\alpha)b=0 α a − ( 1 − α ) b = 0 for
α = b = 5 + 2 5 10 ∈ ( 0 , 1 ) . (48) \alpha=b=\frac{5+2\sqrt{5}}{10}\in(0,1). \tag*{(48)} α = b = 10 5 + 2 5 ∈ ( 0 , 1 ) . (48)
Hence 0 ‾ ∈ conv ( m S λ max ∪ − m S λ min ) \overline{0}\in\operatorname{conv}\left(m_{S_{\lambda_{\max}}}\cup-m_{S_{\lambda_{\min}}}\right) 0 ∈ conv ( m S λ m a x ∪ − m S λ m i n ) , as expected. Note that, since a ≠ b a\ne b a = b , m S λ max ∩ m S λ min = ∅ m_{S_{\lambda_{\max}}}\cap m_{S_{\lambda_{\min}}}=\varnothing m S λ m a x ∩ m S λ m i n = ∅ : here I ∉ W I\notin\mathcal{W} I ∈ / W (see Remark 5.10). Moreover, the same weight recovers the maximizer of (46) : F = α ρ + − ( 1 − α ) ρ − F=\alpha\rho_+-(1-\alpha)\rho_- F = α ρ + − ( 1 − α ) ρ − has F 11 = F 33 = 0 F_{11}=F_{33}=0 F 11 = F 33 = 0 , F 12 = 1 2 5 F_{12}=\frac{1}{2\sqrt{5}} F 12 = 2 5 1 and F 22 = 2 5 F_{22}=\frac{2}{\sqrt{5}} F 22 = 5 2 , which is (47) . This illustrates the equivalence between items (3) and (5) of Theorem 5.9 .
Example 6.2. This example shows that best approximations need not be unique, and how S M A X S_{\mathrm{MAX}} S MAX and its moment change along the set of best approximations. Let A 0 = Diag ( 0 , 2 , 1 ) A_0=\operatorname{Diag}(0,2,1) A 0 = Diag ( 0 , 2 , 1 ) . Then
I 0 = { ( 1 , 1 ) , ( 1 , 2 ) , ( 1 , 3 ) , ( 2 , 1 ) , ( 2 , 3 ) , ( 3 , 1 ) , ( 3 , 2 ) } , I_0=\{(1,1),(1,2),(1,3),(2,1),(2,3),(3,1),(3,2)\}, I 0 = {( 1 , 1 ) , ( 1 , 2 ) , ( 1 , 3 ) , ( 2 , 1 ) , ( 2 , 3 ) , ( 3 , 1 ) , ( 3 , 2 )} ,
W = span { B 1 , 1 , B 1 , 2 , B 1 , 3 , B 2 , 1 , B 2 , 3 , B 3 , 1 , B 3 , 2 } , \mathcal{W}=\operatorname{span}\{B^{1,1},B^{1,2},B^{1,3},B^{2,1},B^{2,3},B^{3,1},B^{3,2}\}, W = span { B 1 , 1 , B 1 , 2 , B 1 , 3 , B 2 , 1 , B 2 , 3 , B 3 , 1 , B 3 , 2 } ,
so that the only specified entries are the diagonal entries ( 2 , 2 ) (2,2) ( 2 , 2 ) and ( 3 , 3 ) (3,3) ( 3 , 3 ) . For every W ∈ W W\in\mathcal{W} W ∈ W ,
∥ A 0 + W ∥ ≥ ∣ ( A 0 + W ) 22 ∣ = 2 = ∥ A 0 ∥ , \lVert A_0+W\rVert\geq\lvert(A_0+W)_{22}\rvert=2=\lVert A_0\rVert, ∥ A 0 + W ∥ ≥ ∣( A 0 + W ) 22 ∣ = 2 = ∥ A 0 ∥ ,
hence dist ( A 0 , W ) = 2 \operatorname{dist}(A_0,\mathcal{W})=2 dist ( A 0 , W ) = 2 , A 0 A_0 A 0 is W \mathcal{W} W -minimal and 0 0 0 is a best approximation of A 0 A_0 A 0 in W \mathcal{W} W . It is not unique: for every − 2 ≤ α ≤ 2 -2\leq\alpha\leq2 − 2 ≤ α ≤ 2 , A 0 + α B 1 , 1 = Diag ( α , 2 , 1 ) A_0+\alpha B^{1,1}=\operatorname{Diag}(\alpha,2,1) A 0 + α B 1 , 1 = Diag ( α , 2 , 1 ) has norm 2 2 2 , so − α B 1 , 1 -\alpha B^{1,1} − α B 1 , 1 is also a best approximation. Along this segment the different situations of Theorem 5.9 occur:
If ∣ α ∣ < 2 \lvert\alpha\rvert<2 ∣ α ∣ < 2 , then λ max = 2 > − λ min \lambda_{\max}=2>-\lambda_{\min} λ m a x = 2 > − λ m i n , S M A X = span { e 2 } S_{\mathrm{MAX}}=\operatorname{span}\{e_2\} S MAX = span { e 2 } and m S M A X = { 0 ‾ } m_{S_{\mathrm{MAX}}}=\{\overline{0}\} m S MAX = { 0 } , because ⟨ B k e 2 , e 2 ⟩ = ( B k ) 22 = 0 \langle B^k e_2,e_2\rangle=(B^k)_{22}=0 ⟨ B k e 2 , e 2 ⟩ = ( B k ) 22 = 0 for every k k k .
If α = 2 \alpha=2 α = 2 , then still λ max = 2 > − λ min \lambda_{\max}=2>-\lambda_{\min} λ m a x = 2 > − λ m i n , but now S M A X = span { e 1 , e 2 } S_{\mathrm{MAX}}=\operatorname{span}\{e_1,e_2\} S MAX = span { e 1 , e 2 } and every ρ ∈ D S M A X \rho\in\mathcal{D}_{S_{\mathrm{MAX}}} ρ ∈ D S MAX is of the form
ρ = ( β c c ‾ 1 − β ) ⊕ 0 , 0 ≤ β ≤ 1 , ∣ c ∣ 2 ≤ β ( 1 − β ) . \rho=\begin{pmatrix}\beta&c\\\overline{c}&1-\beta\end{pmatrix}\oplus0,\qquad0\leq\beta\leq1,\quad\lvert c\rvert^2\leq\beta(1-\beta). ρ = ( β c c 1 − β ) ⊕ 0 , 0 ≤ β ≤ 1 , ∣ c ∣ 2 ≤ β ( 1 − β ) .
Since tr ( ρ B 1 , 1 ) = β \operatorname{tr}(\rho B^{1,1})=\beta tr ( ρ B 1 , 1 ) = β , tr ( ρ B 1 , 2 ) = 2 Re c \operatorname{tr}(\rho B^{1,2})=\sqrt{2}\operatorname{Re}c tr ( ρ B 1 , 2 ) = 2 Re c , tr ( ρ B 2 , 1 ) = − 2 Im c \operatorname{tr}(\rho B^{2,1})=-\sqrt{2}\operatorname{Im}c tr ( ρ B 2 , 1 ) = − 2 Im c , and tr ( ρ B k ) = 0 \operatorname{tr}(\rho B^k)=0 tr ( ρ B k ) = 0 for the remaining elements of B \mathbb{B} B (which involve the third row or column), item 1 of Proposition 3.4 gives, in the coordinates corresponding to B 1 , 1 , B 1 , 2 , B 2 , 1 B^{1,1},B^{1,2},B^{2,1} B 1 , 1 , B 1 , 2 , B 2 , 1 (all other coordinates being zero),
m S M A X = { ( β , 2 Re c , − 2 Im c ) : 0 ≤ β ≤ 1 , ∣ c ∣ 2 ≤ β ( 1 − β ) } , m_{S_{\mathrm{MAX}}}=\{(\beta,\sqrt{2}\operatorname{Re}c,-\sqrt{2}\operatorname{Im}c):0\leq\beta\leq1,\ \lvert c\rvert^2\leq\beta(1-\beta)\}, m S MAX = {( β , 2 Re c , − 2 Im c ) : 0 ≤ β ≤ 1 , ∣ c ∣ 2 ≤ β ( 1 − β )} ,
a solid ellipsoid. It contains 0 ‾ = Φ ( e 2 e 2 ∗ ) \overline{0}=\Phi(e_2e_2^*) 0 = Φ ( e 2 e 2 ∗ ) , as required by condition (3) of Theorem 5.9 , but only as a boundary point (see Figure 1 ).
Figure 1. Example 6.2 with α = 2 \alpha= 2 α = 2 . Projection of m S max mS_{\max} m S m a x onto the coordinates corresponding to B 1 , 1 B^{1,1} B 1 , 1 , B 1 , 2 B^{1,2} B 1 , 2 , which is { ( β , 2 Re c ) : ( β − 1 2 ) 2 + ( Re c ) 2 ≤ 1 4 } \{(\beta,\sqrt{2}\operatorname{Re}c):(\beta-\frac{1}{2})^{2}+(\operatorname{Re}c)^{2}\leq\frac{1}{4}\} {( β , 2 Re c ) : ( β − 2 1 ) 2 + ( Re c ) 2 ≤ 4 1 } . The origin 0 ‾ = Φ ( e 2 e 2 ∗ ) \overline{0}=\Phi(e_{2}e_{2}^{*}) 0 = Φ ( e 2 e 2 ∗ ) belongs to the moment, but only as a boundary point.
If α = − 2 \alpha=-2 α = − 2 , the spectrum is balanced: S λ max = span { e 2 } S_{\lambda_{\max}}=\operatorname{span}\{e_2\} S λ m a x = span { e 2 } , S λ min = span { e 1 } S_{\lambda_{\min}}=\operatorname{span}\{e_1\} S λ m i n = span { e 1 } , and the second condition of (43) holds since m S λ max = { 0 ‾ } m_{S_{\lambda_{\max}}}=\{\overline{0}\} m S λ m a x = { 0 } .
Adjacency matrices of graphs
Let G G G be a connected simple graph with vertices { 1 , … , n } \{1,\ldots,n\} { 1 , … , n } , n ≥ 2 n\geq2 n ≥ 2 , and let A 0 = A ( G ) A_0=A(G) A 0 = A ( G ) be its adjacency matrix. Then I 0 I_0 I 0 consists of the non-edges of G G G together with the diagonal, and by Proposition 2.2 , W ∘ \mathcal{W}^{\circ} W ∘ consists of the Hermitian matrices of trace norm one supported on the edges of G G G ; in particular, they have zero diagonal. Completing A 0 A_0 A 0 within W \mathcal{W} W means choosing freely the weights of the non-edges and of the diagonal (that is, adding loops), and the natural question is when A ( G ) A(G) A ( G ) itself is a best completion. We shall use the following facts.
Lemma 6.3 (Perron–Frobenius for connected graphs). Let G G G be a connected graph with n ≥ 2 n\geq2 n ≥ 2 vertices, A = A ( G ) A=A(G) A = A ( G ) and ρ \rho ρ its spectral radius. Then:
(i) ρ \rho ρ is a simple eigenvalue of A A A , with an eigenvector v v v such that v i > 0 v_i>0 v i > 0 for every i i i ;
(ii) every eigenvalue λ \lambda λ of A A A satisfies ∣ λ ∣ ≤ ρ \lvert\lambda\rvert\leq\rho ∣ λ ∣ ≤ ρ , and − ρ -\rho − ρ is an eigenvalue of A A A if and only if G G G is bipartite;
(iii) If G G G is bipartite with classes V 1 ∪ V 2 V_{1}\cup V_{2} V 1 ∪ V 2 and D = Diag ( ε 1 , … , ε n ) D=\operatorname{Diag}(\varepsilon_{1},\ldots,\varepsilon_{n}) D = Diag ( ε 1 , … , ε n ) , where ε j = 1 \varepsilon_{j}=1 ε j = 1 if j ∈ V 1 j\in V_{1} j ∈ V 1 and ε j = − 1 \varepsilon_{j}=-1 ε j = − 1 if j ∈ V 2 j\in V_{2} j ∈ V 2 , then D A D = − A DAD=-A D A D = − A . Consequently − ρ -\rho − ρ is also simple, with eigenvector D v Dv D v . In particular, ∥ A ∥ = ρ \lVert A\rVert=\rho ∥ A ∥ = ρ .
Proof. The matrix A A A is nonnegative, and it is irreducible because G G G is connected [12 Theorem 6.2.24] . Then (i) and the first part of (ii) follow from the Perron–Frobenius theorem for irreducible nonnegative matrices [12 Theorem 8.4.4] , and the characterization of bipartite graphs in (ii) is [8 Proposition 3.4.1] . For (iii), ( D A D ) i j = ε i ε j a i j (DAD)_{ij}=\varepsilon_{i}\varepsilon_{j}a_{ij} ( D A D ) ij = ε i ε j a ij and every edge joins different classes; since D = D − 1 D=D^{-1} D = D − 1 , the similarity D A D = − A DAD=-A D A D = − A maps the eigenspace of ρ \rho ρ onto that of − ρ -\rho − ρ .
Proposition 6.4. Let G G G be a connected graph and A 0 = A ( G ) A_{0}=A(G) A 0 = A ( G ) . Then A 0 A_{0} A 0 is W \mathcal{W} W -minimal if and only if G G G is complete bipartite. In that case, dist ( A 0 , W ) = ∥ A 0 ∥ = m k \operatorname{dist}(A_{0},\mathcal{W})=\lVert A_{0}\rVert=\sqrt{mk} dist ( A 0 , W ) = ∥ A 0 ∥ = mk for G = K m , k G=K_{m,k} G = K m , k .
Proof. By the equivalence of items (1) and (5) of Theorem 5.9, A 0 A_{0} A 0 is W \mathcal{W} W -minimal if and only if there exists F ∈ ∂ ∥ A 0 ∥ ∩ W 1 ∘ F\in\partial\lVert A_{0}\rVert\cap\mathcal{W}_{1}^{\circ} F ∈ ∂ ∥ A 0 ∥ ∩ W 1 ∘ , and such an F F F has zero diagonal. Let ρ = ∥ A 0 ∥ \rho=\lVert A_{0}\rVert ρ = ∥ A 0 ∥ and let v v v be the unit eigenvector of Lemma 6.3 (i).
If G G G is not bipartite, by Lemma 6.3 (i) and (ii), λ max = ρ > − λ min \lambda_{\max}=\rho>-\lambda_{\min} λ m a x = ρ > − λ m i n and ρ \rho ρ is simple. By (4.8), ∂ ∥ A 0 ∥ = { v v ∗ } \partial\lVert A_{0}\rVert=\{vv^{*}\} ∂ ∥ A 0 ∥ = { v v ∗ } , whose diagonal entries v i 2 v_{i}^{2} v i 2 are positive. Hence A 0 A_{0} A 0 is not W \mathcal{W} W -minimal.
If G G G is bipartite, by Lemma 6.3 (iii), ± ρ \pm\rho ± ρ are simple eigenvalues with unit eigenvectors v v v and w = D v w=Dv w = D v , that is, w j = ε j v j w_{j}=\varepsilon_{j}v_{j} w j = ε j v j . By (4.10) in Lemma 4.3, the Hermitian elements of ∂ ∥ A 0 ∥ \partial\lVert A_{0}\rVert ∂ ∥ A 0 ∥ are
F = α v v ∗ − ( 1 − α ) w w ∗ , α ∈ [ 0 , 1 ] . F=\alpha vv^{*}-(1-\alpha)ww^{*},\qquad\alpha\in[0,1]. F = α v v ∗ − ( 1 − α ) w w ∗ , α ∈ [ 0 , 1 ] .
The diagonal of such an F F F is ( ( 2 α − 1 ) v i 2 ) i ((2\alpha-1)v_{i}^{2})_{i} (( 2 α − 1 ) v i 2 ) i , so that a zero diagonal forces α = 1 2 \alpha=\frac{1}{2} α = 2 1 . For this α \alpha α ,
F i j = 1 2 v i v j ( 1 − ε i ε j ) = { 0 if i , j belong to the same class , v i v j ≠ 0 if i , j belong to different classes . F_{ij}=\frac{1}{2}v_{i}v_{j}(1-\varepsilon_{i}\varepsilon_{j})=
\begin{cases}
0 & \text{if }i,j\text{ belong to the same class},\\
v_{i}v_{j}\ne0 & \text{if }i,j\text{ belong to different classes}.
\end{cases} F ij = 2 1 v i v j ( 1 − ε i ε j ) = { 0 v i v j = 0 if i , j belong to the same class , if i , j belong to different classes .
Therefore F ∈ W 1 ∘ F\in\mathcal{W}_{1}^{\circ} F ∈ W 1 ∘ if and only if every pair of vertices in different classes is an edge, that is, if and only if G G G is complete bipartite. For G = K m , k G=K_{m,k} G = K m , k we have ρ = m k \rho=\sqrt{mk} ρ = mk .
If the vertices of K m , k K_{m,k} K m , k are numbered so that the class of size m m m comes first, then
A 0 = ( 0 J J T 0 ) , A_{0}=\begin{pmatrix}0&J\\J^{T}&0\end{pmatrix}, A 0 = ( 0 J T J 0 ) ,
where J J J is the m × k m\times k m × k matrix of ones, and W = M m h ( C ) ⊕ M k h ( C ) \mathcal{W}=M_{m}^{h}(\mathbb{C})\oplus M_{k}^{h}(\mathbb{C}) W = M m h ( C ) ⊕ M k h ( C ) is the Hermitian part of a unital C ∗ C^{*} C ∗ -subalgebra, so this case is covered by [7 ] . Here the witness of the proof is F = 1 2 ( v v ∗ − w w ∗ ) = A 0 / ∥ A 0 ∥ 1 F=\frac{1}{2}(vv^{*}-ww^{*})=A_{0}/\lVert A_{0}\rVert_{1} F = 2 1 ( v v ∗ − w w ∗ ) = A 0 / ∥ A 0 ∥ 1 . Note also that, by Lemma 6.3 (ii), the spectrum of A ( G ) A(G) A ( G ) is balanced exactly when G G G is bipartite, so bipartite graphs fall into the second condition of (5.6) and non-bipartite graphs into the first one; both conditions are needed.
Figure 2. Two complete bipartite graphs, with the classes of the bipartition in different colors. By Proposition 6.4 their adjacency matrices are W \mathcal{W} W -minimal: non-edges always join vertices of the same class, which is exactly where the witness F = 1 2 ( v v ∗ − w w ∗ ) F=\frac{1}{2}(vv^{*}-ww^{*}) F = 2 1 ( v v ∗ − w w ∗ ) vanishes.
Example 6.5. Completeness cannot be dropped in Proposition 6.4 . Let A 0 = A ( P 4 ) A_0=A(P_4) A 0 = A ( P 4 ) , where P 4 P_4 P 4 is the path 1 − 2 − 3 − 4 1-2-3-4 1 − 2 − 3 − 4 , which is bipartite but not complete. Then ∥ A 0 ∥ = φ = 1 + 5 2 \lVert A_0\rVert=\varphi=\frac{1+\sqrt{5}}{2} ∥ A 0 ∥ = φ = 2 1 + 5 , while dist ( A 0 , W ) = 2 \operatorname{dist}(A_0,\mathcal{W})=\sqrt{2} dist ( A 0 , W ) = 2 , attained at W 0 = − 2 B 1 , 4 W_0=-\sqrt{2}B^{1,4} W 0 = − 2 B 1 , 4 , that is, with weight − 1 -1 − 1 at the non-edge { 1 , 4 } \{1,4\} { 1 , 4 } (see Figure 3 ):
A 0 + W 0 = ( 0 1 0 − 1 1 0 1 0 0 1 0 1 − 1 0 1 0 ) , ( A 0 + W 0 ) 2 = 2 I . A_0+W_0=
\begin{pmatrix}
0 & 1 & 0 & -1\\
1 & 0 & 1 & 0\\
0 & 1 & 0 & 1\\
-1 & 0 & 1 & 0
\end{pmatrix},
\qquad
(A_0+W_0)^2=2I. A 0 + W 0 = 0 1 0 − 1 1 0 1 0 0 1 0 1 − 1 0 1 0 , ( A 0 + W 0 ) 2 = 2 I . In particular σ ( A 0 + W 0 ) = { ± 2 } \sigma(A_0+W_0)=\{\pm\sqrt{2}\} σ ( A 0 + W 0 ) = { ± 2 } and S M A X = C 4 S_{\mathrm{MAX}}=\mathbb{C}^{4} S MAX = C 4 . A witness is the matrix F ∈ W 1 ∘ F\in\mathcal{W}^{\circ}_{1} F ∈ W 1 ∘ supported on the edges, with weights a = c = 1 4 2 a=c=\frac{1}{4\sqrt{2}} a = c = 4 2 1 at { 1 , 2 } , { 3 , 4 } \{1,2\},\{3,4\} { 1 , 2 } , { 3 , 4 } and b = 1 2 2 b=\frac{1}{2\sqrt{2}} b = 2 2 1 at { 2 , 3 } \{2,3\} { 2 , 3 } . Indeed, for such tridiagonal matrices with a , c ≥ 0 a,c\geq0 a , c ≥ 0 one has ∥ F ∥ 1 = 2 ( a + c ) 2 + b 2 = 1 \lVert F\rVert_{1}=2\sqrt{(a+c)^2+b^2}=1 ∥ F ∥ 1 = 2 ( a + c ) 2 + b 2 = 1 , so that for every W ∈ W W\in\mathcal{W} W ∈ W
2 = 2 ( a + b + c ) = tr ( F A 0 ) = tr ( F ( A 0 + W ) ) ≤ ∥ A 0 + W ∥ , \sqrt{2}=2(a+b+c)=\operatorname{tr}(FA_0)=\operatorname{tr}(F(A_0+W))\leq\lVert A_0+W\rVert, 2 = 2 ( a + b + c ) = tr ( F A 0 ) = tr ( F ( A 0 + W )) ≤ ∥ A 0 + W ∥ , with equality at W 0 W_0 W 0 . Every choice with a , c ≥ 0 a,c\geq0 a , c ≥ 0 and a + c = b = 1 2 2 a+c=b=\frac{1}{2\sqrt{2}} a + c = b = 2 2 1 works, so the dual witness is not unique.
The minimality depends on W \mathcal{W} W : if only the diagonal is free, that is, I 0 = { ( i , i ) } 1 ≤ i ≤ 4 I_0=\{(i,i)\}_{1\leq i\leq4} I 0 = {( i , i ) } 1 ≤ i ≤ 4 , then A ( P 4 ) A(P_4) A ( P 4 ) is minimal, since it is tridiagonal with zero diagonal [3 Lemma 6.3] , and dist ( A ( P 4 ) , Diag ( M 4 h ( C ) ) ) = φ > 2 \operatorname{dist}(A(P_4),\operatorname{Diag}(M_4^{h}(\mathbb{C})))=\varphi>\sqrt{2} dist ( A ( P 4 ) , Diag ( M 4 h ( C ))) = φ > 2 . More generally, A ( G ) A(G) A ( G ) is minimal with respect to the real diagonal matrices for every connected bipartite graph G G G , because the matrix F = 1 2 ( v v ∗ − w w ∗ ) F=\frac{1}{2}(vv^{*}-ww^{*}) F = 2 1 ( v v ∗ − w w ∗ ) of the proof of Proposition 6.4 always has zero diagonal; what fails outside complete bipartite graphs is only that F F F does not vanish at the non-edges joining different classes.
Figure 3. The path P 4 P_4 P 4 and its optimal completion. The pair { 1 , 4 } \{1,4\} { 1 , 4 } joins different classes but is not an edge, and this is exactly where the witness of Proposition 6.4 fails. The optimal completion has weight − 1 -1 − 1 at this position (dashed).
Example 6.6. For G = K n G = K_n G = K n there are no non-edges, so W \mathcal{W} W is the space of real diagonal matrices, which is the case studied in [1 ] . For n ≥ 3 n \ge3 n ≥ 3 the graph is not bipartite, and indeed ∥ A ( K n ) ∥ = n − 1 \|A(K_n)\| = n - 1 ∥ A ( K n ) ∥ = n − 1 while
dist ( A ( K n ) , W ) = n 2 , \operatorname{dist}(A(K_n),\mathcal{W}) = \frac{n}{2}, dist ( A ( K n ) , W ) = 2 n ,
attained at J − n 2 I J - \frac{n}{2}I J − 2 n I , where J J J is the n × n n \times n n × n matrix of ones; its spectrum is { n 2 , − n 2 } \left\{\frac{n}{2},-\frac{n}{2}\right\} { 2 n , − 2 n } , with − n 2 -\frac{n}{2} − 2 n of multiplicity n − 1 n - 1 n − 1 . With v = 1 n 1 v = \frac{1}{\sqrt{n}}\mathbf{1} v = n 1 1 and P 1 ⊥ = I − v v ∗ P_{1^\perp} = I - vv^* P 1 ⊥ = I − v v ∗ , a witness is
F = 1 2 ( v v ∗ − 1 n − 1 P 1 ⊥ ) , F = \frac{1}{2}\left(vv^* - \frac{1}{n-1}P_{1^\perp}\right), F = 2 1 ( v v ∗ − n − 1 1 P 1 ⊥ ) ,
which has zero diagonal, because both terms have constant diagonal equal to 1 2 n \frac{1}{2n} 2 n 1 , and satisfies ∥ F ∥ 1 = 1 \|F\|_1 = 1 ∥ F ∥ 1 = 1 and tr ( F ( J − n 2 I ) ) = n 2 \operatorname{tr}\left(F\left(J - \frac{n}{2}I\right)\right) = \frac{n}{2} tr ( F ( J − 2 n I ) ) = 2 n .
References