betaFeedback

WorksMathematics

Joint Moments of Bilinear Resolvents

Abstract

Given A,B∈B(H)+A,B\in B\left(H\right)_{+}, we consider the existence and uniqueness of a positive operator C∈B(H)+C\in B\left(H\right)_{+} such that (I+mA+nB+mnC)−1=∫[0,1]2smtndE(s,t),m,n≥0, \left(I+mA+nB+mnC\right)^{-1}=\int_{\left[0,1\right]^{2}}s^{m}t^{n}dE\left(s,t\right),\quad m,n\geq0, for a positive operator-valued measure EE. This is a joint Hausdorff moment problem. Without CC, it holds if and only if AB=BAAB=BA. In dimension two we give a complete criterion for noncommuting pairs. In higher finite dimensions, nonexistence is generic. Under certain spectral and cyclicity assumptions, we obtain a necessary and sufficient condition on arbitrary Hilbert spaces. The completion is unique in this case. We also classify these pairs up to unitary equivalence. The class includes irreducible infinite-dimensional examples, even with continuous spectrum.

Introduction

The complete monotonicity of reciprocals of scalar polynomials has been studied extensively [15, 14, 25]. In particular (see [1 Theorem 3.1]), for a,b>0a,b>0 and c≥0c\ge0,

{(1+am+bn+cmn)−1}m,n≥0 has a joint Hausdorff representation⇕c≤ab(1)\left\{\left(1+am+bn+cmn\right)^{-1}\right\}_{m,n\ge0}\text{ has a joint Hausdorff representation}\\ \Updownarrow\\ c\le ab \tag*{(1)}

Here we fix A,B∈B(H)+A,B\in\mathcal{B}(H)_{+} and ask whether a positive operator CC can be chosen so that the inverse has joint operator-valued moments. The above scalar criterion (1) does not extend directly. Specifically, we consider the following.

2020 Mathematics Subject Classification. Primary 47A57; Secondary 44A60, 47A10, 47B65.

Key words and phrases. joint Hausdorff moments, positive operator-valued measures, bilinear resolvents, positive completions, spectral rigidity.

Question 1.1 (positive completion). Given A,B∈B(H)+A,B \in\mathcal{B}(H)_{+}, can we find C∈B(H)+C \in\mathcal{B}(H)_{+} such that

(I+mA+nB+mnC)−1=∫[0,1]2smtn dE(s,t),m,n≥0,(2)(I+mA+nB+mnC)^{-1}=\int_{[0,1]^{2}}s^{m}t^{n}\,dE(s,t),\qquad m,n\geq0, \tag*{(2)}

for a positive operator-valued measure (POVM) EE?

For A,B∈B(H)+A,B\in\mathcal{B}(H)_{+}, without the mixed term, we show in Theorem 3.5 that

{(I+mA+nB)−1}m,n≥0has a joint Hausdorff representation\left\{(I+mA+nB)^{-1}\right\}_{m,n\geq0}\quad\text{has a joint Hausdorff representation}
⇕\Updownarrow
AB=BAAB=BA

So the point of adding CC is to allow noncommuting pairs, although its existence imposes much stronger spectral restrictions.

For a single operator A∈B(H)+A\in\mathcal{B}(H)_{+}, the resolvents (I+mA)−1(I+mA)^{-1}, m≥0m\geq0, form an operator-valued Hausdorff moment sequence. The representing measure is determined by the spectral measure of AA. When (2) holds, the two one-variable resolvents, (I+mA)−1(I+mA)^{-1} and (I+nB)−1(I+nB)^{-1}, are then its marginals, and the mixed coefficient CC accounts for all the joint moments at once.

The question considered here is not about a simultaneous spectral theorem. A POVM need not be projection-valued, and its marginals need not commute. Nor does positivity of

(I+xA+yB+xyC)−1(I+xA+yB+xyC)^{-1}

at each point guarantee that its values are joint moments. The latter requires positivity of all mixed finite differences, an infinite collection of operator inequalities. The desired inverse polynomial leaves little freedom to satisfy them.

In dimension two, that freedom has a simple description (Theorem 4.1). For noncommuting positive definite matrices AA and BB, a positive completion exists if and only if

A∥A∥+B∥B∥≥I.\frac{A}{\lVert A\rVert}+\frac{B}{\lVert B\rVert}\geq I.

When it exists it is unique. The condition is an ordinary operator inequality, although it decides positivity of the entire two-variable moment sequence. The situation appears to change sharply with the dimension. For every matrix size at least three, the pairs with no completion contain an open dense set (Corollary 5.2).

The exceptional pairs in higher dimensions have a rigid spectral structure, and that same structure gives a positive completion on an arbitrary Hilbert space (§ 5).

We prove the following (Theorem 5.1): Suppose the smallest spectral values a,ba,b of A,BA,B are simple isolated eigenvalues, with unit eigenvectors q,pq,p, and pp is cyclic for AA. Write R=pq∗/(q∗p)R=pq^{*}/(q^{*}p). The condition is

(A−aI)(B−bI)=γ(I−R),0<γ≤ab.(A-aI)(B-bI)=\gamma(I-R),\qquad0<\gamma\leq ab.

It is both necessary and sufficient. The completion is determined by C=BA−γRC=BA-\gamma R.

In particular, there are irreducible infinite-dimensional pairs with a positive completion, even when the operators have continuous spectrum. Up to unitary equivalence, the pairs in this class are described by three positive numbers and one scalar measure (Corollary 5.3).

Related literature. The underlying one-variable question belongs to the classical Hausdorff moment problem, in which complete monotonicity of a sequence characterizes measures on a compact interval [17 Wid41, 26]. The treatment of positive definite functions on semigroups provides a useful general setting for moment and Laplace representations [8]. These results explain the individual resolvent measures, but they do not decide whether the inverse of a prescribed noncommutative polynomial has positive joint moments.

Multivariable moment problems require positivity on an algebra of polynomials, rather than separate positivity in each coordinate. Existence and representation on semialgebraic sets have been developed through work on positive polynomials and moment functionals [24, 22, 19, 18]. Positive moment matrices, flat extensions, and dimensional extensions give other ways of approaching the multivariable problem [11, 23, 29]. These theories generally start with an unknown measure or an arbitrary moment functional. Our moments are already fixed, apart from one coefficient inside the inverse. This is what makes their positivity an operator problem rather than an unrestricted moment problem.

Positive operator-valued measures appear in the dilation theory of Naimark and the operator extensions of Halmos [Naĭ40, Hal50]. Stinespring’s representation theorem and Arveson’s work put positive operator-valued measures in the setting of completely positive maps [27, 6]. Choi’s matrix criteria are central to that setting [12]. Andô’s theorem provides commuting dilations for pairs of commuting contractions [4]. None of these results by itself produces a bilinear inverse pencil with given resolvent marginals. In particular, compatibility of positive operator-valued marginals is a weaker requirement than the one imposed here.

Moment positivity also has a long connection with subnormal operators and commuting tuples. The classical subnormality criterion and its operator-theoretic development may be found in [9, 13]. The several-operator setting is represented, for example, by joint isometries [7]. Positive kernels give another language for the same underlying positivity, starting with the reproducing kernel theory of Aronszajn [5] and its operator-valued and several-variable developments [21, 10]. The literature on multivariable operator monotone functions and Pick functions is related through positivity of analytic operator functions [3, 2]. These connections explain why joint moment positivity matters, but they do not determine which resolvent pencils have it.

A different approach studies complete monotonicity of reciprocal scalar polynomials. Early positive-coefficient questions include Szegő’s work [28]. The theory of Riesz distributions on homogeneous cones, and its treatment in harmonic analysis, gives striking criteria for positivity of inverse powers of determinants [15, 14]. Scott and Sokal developed this direction for several classes of polynomials and related it to classical coefficient-positivity questions [25]. In the specific scalar bilinear case, a criterion for the reciprocal to have joint Hausdorff moments is available [1]. Our problem differs at the first step. The coefficients are positive operators which need not commute, and positivity of a scalar reciprocal, or of the matrix entries of the inverse, does not give positivity of an operator-valued representing measure. The spectral relation and the operator-valued Laplace construction are needed to recover it.

Preliminaries: A scalar moment representation

The scalar joint Hausdorff moment problem for reciprocals of bilinear polynomials has been studied in the literature; see [1 Theorem 3.1]. We need only the existence of a representing measure under a particular condition on the coefficients. We give a direct proof.

Lemma 2.1. Let a,b>0a,b>0 and 0≤c≤ab0\le c\le ab. There exists a probability measure μ\mu on [0,1]2[0,1]^2 such that

11+am+bn+cmn=∫[0,1]2smtn dμ(s,t),m,n≥0.\frac{1}{1+am+bn+cmn}=\int_{[0,1]^2}s^m t^n\,d\mu(s,t),\quad m,n\ge0.

Proof. We consider three cases.

If c=0c=0, then

11+am+bn=∫0∞e−r(e−ar)m(e−br)n dr.\frac{1}{1+am+bn}=\int_0^\infty e^{-r}\left(e^{-ar}\right)^m\left(e^{-br}\right)^n\,dr.

The pushforward of the probability measure e−r dre^{-r}\,dr under the map

r⟼(e−ar,e−br)r\longmapsto\left(e^{-ar},e^{-br}\right)

is a representing probability measure on [0,1]2[0,1]^2.

If c=abc=ab, then

11+am+bn+abmn=11+am11+bn.\frac{1}{1+am+bn+abmn}=\frac{1}{1+am}\frac{1}{1+bn}.

For a>0a>0,

11+am=∫01sm1as1/a−1 ds.\frac{1}{1+am}=\int_0^1s^m\frac{1}{a}s^{1/a-1}\,ds.

The analogous formula holds for bb. Taking the product of these two probability measures gives the required joint representation.

Now suppose 0<c<ab0<c<ab, and let

δ=ab−c>0.\delta=ab-c>0.

We have

1+am+bn+cmn=c(m+bc)(n+ac)−δc=c(m+bc)(n+ac)[1−δc2(m+b/c)(n+a/c)].\begin{aligned} 1+am+bn+cmn=c\left(m+\frac{b}{c}\right)\left(n+\frac{a}{c}\right)-\frac{\delta}{c} \\ &=c\left(m+\frac{b}{c}\right)\left(n+\frac{a}{c}\right)\left[1-\frac{\delta}{c^2(m+b/c)(n+a/c)}\right]. \end{aligned}

Since m,n≥0m,n\ge0,

0≤δc2(m+b/c)(n+a/c)≤δab<1.0\le\frac{\delta}{c^2(m+b/c)(n+a/c)}\le\frac{\delta}{ab}<1.

We can therefore expand the last factor as a geometric series:

11+am+bn+cmn=1c∑k=0∞(δc2)k(m+bc)−k−1(n+ac)−k−1.\frac{1}{1+am+bn+cmn}=\frac{1}{c}\sum_{k=0}^{\infty}\left(\frac{\delta}{c^2}\right)^k\left(m+\frac{b}{c}\right)^{-k-1}\left(n+\frac{a}{c}\right)^{-k-1}.

For every α>0\alpha>0 and integer k≥0k\ge0, substitution r=−log⁡sr=-\log s gives

(m+α)−k−1=1k!∫0∞rke−(m+α)r dr=1k!∫01sm+α−1(−log⁡s)k ds.(m+\alpha)^{-k-1}=\frac{1}{k!}\int_0^\infty r^k e^{-(m+\alpha)r}\,dr=\frac{1}{k!}\int_0^1s^{m+\alpha-1}(-\log s)^k\,ds.

Thus each factor in the series has a positive one-variable Hausdorff representation. More explicitly, for each k≥0k \ge0, define a positive measure μk\mu_k on [0,1]2[0,1]^2 by

dμk(s,t)=1c(δc2)ksb/c−1(−log⁡s)kk!ta/c−1(−log⁡t)kk! dsdt.d\mu_k(s,t)=\frac{1}{c}\left(\frac{\delta}{c^2}\right)^k\frac{s^{b/c-1}(-\log s)^k}{k!}\frac{t^{a/c-1}(-\log t)^k}{k!}\,dsdt.

Its moments are

∫[0,1]2smtn dμk(s,t)=1c(δc2)k(m+bc)−k−1(n+ac)−k−1.\int_{[0,1]^2}s^m t^n\,d\mu_k(s,t)=\frac{1}{c}\left(\frac{\delta}{c^2}\right)^k\left(m+\frac{b}{c}\right)^{-k-1}\left(n+\frac{a}{c}\right)^{-k-1}.

In particular, taking m=n=0m=n=0 gives

μk([0,1]2)=1c(δc2)k(bc)−k−1(ac)−k−1=cab(δab)k.\mu_k\left([0,1]^2\right)=\frac{1}{c}\left(\frac{\delta}{c^2}\right)^k\left(\frac{b}{c}\right)^{-k-1}\left(\frac{a}{c}\right)^{-k-1}=\frac{c}{ab}\left(\frac{\delta}{ab}\right)^k.

Since δ<ab\delta< ab,

∑k=0∞μk([0,1]2)=cab∑k=0∞(δab)k=cab−δ=1.\sum_{k=0}^{\infty}\mu_k\left([0,1]^2\right)=\frac{c}{ab}\sum_{k=0}^{\infty}\left(\frac{\delta}{ab}\right)^k=\frac{c}{ab-\delta}=1.

Consequently,

μ=∑k=0∞μk\mu=\sum_{k=0}^{\infty}\mu_k

defines a probability measure on [0,1]2[0,1]^2. All the measures are positive, so we may integrate term by term. Using the geometric series above, we get

∫[0,1]2smtn dμ(s,t)=∑k=0∞1c(δc2)k(m+bc)−k−1(n+ac)−k−1=11+am+bn+cmn.\begin{aligned} \int_{[0,1]^2}s^m t^n\,d\mu(s,t)=\sum_{k=0}^{\infty}\frac{1}{c}\left(\frac{\delta}{c^2}\right)^k\left(m+\frac{b}{c}\right)^{-k-1}\left(n+\frac{a}{c}\right)^{-k-1} \\ &=\frac{1}{1+am+bn+cmn}. \end{aligned}

Resolvent measures

Let HH be a complex Hilbert space. For positive operators A,B,C∈B(H)+A,B,C\in B(H)_+, let

F(x,y)=(I+xA+yB+xyC)−1,x,y≥0.F(x,y)=(I+xA+yB+xyC)^{-1},\qquad x,y\ge0.

Definition 3.1. We call CC a positive completion of A,BA,B if there is a positive B(H)B(H)-valued measure EE on [0,1]2[0,1]^2 such that

F(m,n)=∫[0,1]2smtn dE(s,t),m,n≥0.(3)F(m,n)=\int_{[0,1]^2}s^m t^n\,dE(s,t),\qquad m,n\ge0. \tag*{(3)}

The measure has total mass II. It is unique, since polynomials are dense in C([0,1]2)C([0,1]^2) and the scalar measures obtained by taking matrix coefficients are determined by their polynomial integrals. The integrals below can all be understood in this weak sense.

The one-variable resolvent has a simple representing measure, as follows.

Lemma 3.2. For λ≥0\lambda\ge0, define the probability measure ρλ\rho_\lambda on [0,1][0,1] by

dρλ(s)={1λs1/λ−1 ds,λ>0,dδ1(s),λ=0.(4)d\rho_\lambda(s)= \begin{cases} \frac{1}{\lambda}s^{1/\lambda-1}\,ds, & \lambda>0,\\ d\delta_1(s), & \lambda=0. \end{cases} \tag*{(4)}

Then, for any T∈B(H)+T \in B(H)_{+} with spectral measure ETE_T, we have

(I+xT)−1=∫[0,1]sx dνT(s),x≥0(5)(I+xT)^{-1}=\int_{[0,1]}s^x\,d\nu_T(s),\qquad x\geq0 \tag*{(5)}

where

νT(V)=∫σ(T)ρλ(V) dET(λ)(6)\nu_T(V)=\int_{\sigma(T)}\rho_\lambda(V)\,dE_T(\lambda) \tag*{(6)}

for all Borel V⊂[0,1]V\subset[0,1]. In particular, νT\nu_T is a normalized POVM. Moreover, for u>0u>0, set

gu(λ)={e−u/λ,λ>0,0,λ=0.g_u(\lambda)= \begin{cases} e^{-u/\lambda}, & \lambda>0,\\ 0, & \lambda=0. \end{cases}

Then

νT([0,e−u])=gu(T).\nu_T([0,e^{-u}])=g_u(T).

Proof. Using (4), direct integration gives

∫[0,1]sx dρλ(s)=11+xλ,x≥0.\int_{[0,1]}s^x\,d\rho_\lambda(s)=\frac{1}{1+x\lambda},\qquad x\geq0.

By the spectral theorem,

(I+xT)−1=∫σ(T)11+xλ dET(λ)=∫σ(T)(∫[0,1]sx dρλ(s))dET(λ)=∫[0,1]sx dνT(s),\begin{aligned} (I+xT)^{-1}=\int_{\sigma(T)}\frac{1}{1+x\lambda}\,dE_T(\lambda) \\ &=\int_{\sigma(T)}\left(\int_{[0,1]}s^x\,d\rho_\lambda(s)\right)dE_T(\lambda)=\int_{[0,1]}s^x\,d\nu_T(s), \end{aligned}

which gives (5). Since each ρλ\rho_\lambda is a probability measure, νT\nu_T as in (6) is positive and has total mass II. Countable additivity follows by monotone convergence. Uniqueness follows from the one-variable Hausdorff moment theorem applied to matrix coefficients.

For u>0u>0 and λ>0\lambda>0, we have

ρλ([0,e−u])=1λ∫0e−us1/λ−1 ds=e−u/λ.\rho_\lambda([0,e^{-u}])=\frac{1}{\lambda}\int_0^{e^{-u}}s^{1/\lambda-1}\,ds=e^{-u/\lambda}.

For λ=0\lambda=0, the same interval has ρ0\rho_0-measure zero, since ρ0=δ1\rho_0=\delta_1. Thus

νT([0,e−u])=∫σ(T)ρλ([0,e−u]) dET(λ)=∫σ(T)gu(λ) dET(λ)=gu(T).\begin{aligned} \nu_T([0,e^{-u}])=\int_{\sigma(T)}\rho_\lambda([0,e^{-u}])\,dE_T(\lambda) \\ &=\int_{\sigma(T)}g_u(\lambda)\,dE_T(\lambda)=g_u(T). \end{aligned}

The function gug_u is continuous at zero since e−u/λ→0e^{-u/\lambda}\to0 as λ↓0\lambda\downarrow0. □

Next, we extend (3) to real parameters.

Proposition 3.3. Suppose AA, BB, C≥0C\geq0 and CC is a positive completion of AA, BB. Let EE be the joint representing measure. Then

F(x,y)=∫[0,1]2sxty dE(s,t),x,y≥0.(7)F(x,y)=\int_{[0,1]^2}s^xt^y\,dE(s,t),\qquad x,y\geq0. \tag*{(7)}

Proof. Fix n∈Z≥0n \in\mathbb{Z}_{\geq0}. For every Borel V⊆[0,1]V \subseteq[0,1], the positive measure

V⟼∫V×[0,1]tn dE(s,t)V \longmapsto\int_{V \times[0,1]} t^n\,dE(s,t)

has moments

∫[0,1]2smtn dE(s,t)=F(m,n)=(I+nB+m(A+nC))−1=(I+nB)−1/2(I+mKn)−1(I+nB)−1/2,\int_{[0,1]^2} s^m t^n\,dE(s,t) = F(m,n) = (I+nB+m(A+nC))^{-1} = (I+nB)^{-1/2}(I+mK_n)^{-1}(I+nB)^{-1/2},

where

Kn=(I+nB)−1/2(A+nC)(I+nB)−1/2≥0.K_n=(I+nB)^{-1/2}(A+nC)(I+nB)^{-1/2}\geq0.

By Lemma 3.2, the positive measure

V⟼(I+nB)−1/2νKn(V)(I+nB)−1/2V \longmapsto(I+nB)^{-1/2}\nu_{K_n}(V)(I+nB)^{-1/2}

has these same moments. Uniqueness of the one-variable Hausdorff representation therefore gives

∫V×[0,1]tn dE(s,t)=(I+nB)−1/2νKn(V)(I+nB)−1/2.\int_{V \times[0,1]} t^n\,dE(s,t) =(I+nB)^{-1/2}\nu_{K_n}(V)(I+nB)^{-1/2}.

We can now integrate sxs^x for any real x≥0x\geq0. Again using Lemma 3.2, we obtain

∫[0,1]2sxtn dE(s,t)=(I+nB)−1/2(I+xKn)−1(I+nB)−1/2=(I+xA+nB+xnC)−1=F(x,n).\int_{[0,1]^2} s^x t^n\,dE(s,t) =(I+nB)^{-1/2}(I+xK_n)^{-1}(I+nB)^{-1/2} =(I+xA+nB+xnC)^{-1}=F(x,n).

Now fix x≥0x\geq0. Applying the same argument in the second coordinate, with I+xAI+xA and B+xCB+xC in place of I+nBI+nB and A+nCA+nC, extends the representation from nonnegative integers nn to every real y≥0y\geq0. Thus (7) holds. □

The existence of a positive completion imposes a rigidity condition on the spectral subspaces of AA and BB.

Theorem 3.4. Suppose A,B,C∈B(H)+A,B,C\in\mathcal{B}(\mathcal{H})_+ and CC is a positive completion of A,BA,B. Let r>0r>0, r∉σ(B)r\notin\sigma(B), and write P=EB([0,r])P=E_B([0,r]). Then, for all sufficiently small x>0x>0,

(B+xC)(I+xA)−1PH⊆PH.(8)(B+xC)(I+xA)^{-1}P\mathcal{H}\subseteq P\mathcal{H}. \tag*{(8)}

The same assertion holds with AA and BB interchanged.

Proof. For x≥0x\geq0, let

Tx=(I+xA)−1/2(B+xC)(I+xA)−1/2.(9)T_x=(I+xA)^{-1/2}(B+xC)(I+xA)^{-1/2}. \tag*{(9)}

Then Tx≥0T_x\geq0 and

(B+xC)(I+xA)−1=(I+xA)1/2Tx(I+xA)−1/2.(B+xC)(I+xA)^{-1}=(I+xA)^{1/2}T_x(I+xA)^{-1/2}.

It follows that (8) holds ⟺\Longleftrightarrow

Tx(I+xA)−1/2PH⊆(I+xA)−1/2PH.(10)T_x(I+xA)^{-1/2}P\mathcal{H}\subseteq(I+xA)^{-1/2}P\mathcal{H}. \tag*{(10)}

Since Tx→BT_x\to B in norm as x↓0x\downarrow0, and PH=EB([0,r])HP\mathcal{H}=E_B([0,r])\mathcal{H}, we will show that

(I+xA)−1/2PH=ETx([0,r])H(11)(I+xA)^{-1/2}P\mathcal{H}=E_{T_x}([0,r])\mathcal{H} \tag*{(11)}

for all sufficiently small x>0x > 0. Then (10) follows from this and the spectral theorem.

Let EE be the joint representing measure. We will use it to compare the spectral subspaces of TxT_x and BB.

Fix x≥0x \ge0. For every Borel V⊆[0,1]V \subseteq[0,1], consider the positive measure

V⟼∫[0,1]×Vsx dE(s,t).V \longmapsto\int_{[0,1]\times V} s^x \, dE(s,t).

By Proposition 3.3, its moments are

∫[0,1]2sxtn dE(s,t)=F(x,n)=(I+xA+n(B+xC))−1=(I+xA)−1/2(I+nTx)−1(I+xA)−1/2,\begin{aligned} \int_{[0,1]^2} s^x t^n \, dE(s,t) &= F(x,n) \\ &= (I+xA+n(B+xC))^{-1} \\ &= (I+xA)^{-1/2}(I+nT_x)^{-1}(I+xA)^{-1/2}, \end{aligned}

for every nonnegative integer nn.

Since Tx≥0T_x \ge0, Lemma 3.2 gives

(I+nTx)−1=∫[0,1]tn dνTx(t).(I+nT_x)^{-1}=\int_{[0,1]}t^n\,d\nu_{T_x}(t).

Therefore the positive measure

V⟼(I+xA)−1/2νTx(V)(I+xA)−1/2V \longmapsto(I+xA)^{-1/2}\nu_{T_x}(V)(I+xA)^{-1/2}

has these same moments. Uniqueness of the one-variable Hausdorff representation gives

∫[0,1]×Vsx dE(s,t)=(I+xA)−1/2νTx(V)(I+xA)−1/2.(12)\int_{[0,1]\times V}s^x\,dE(s,t)=(I+xA)^{-1/2}\nu_{T_x}(V)(I+xA)^{-1/2}. \tag*{(12)}

We can now use positivity of the joint measure. Since 0≤sx≤10 \le s^x \le1, for every Borel set V⊆[0,1]V \subseteq[0,1] we have

0≤∫[0,1]×V(1−sx) dE(s,t).0 \le\int_{[0,1]\times V}(1-s^x)\,dE(s,t).

Equivalently,

∫[0,1]×Vsx dE(s,t)≤∫[0,1]×VdE(s,t).\int_{[0,1]\times V}s^x\,dE(s,t)\leq\int_{[0,1]\times V}dE(s,t).

Using (12) on both sides, and recalling that T0=BT_0=B, this becomes

(I+xA)−1/2νTx(V)(I+xA)−1/2≤νB(V).(I+xA)^{-1/2}\nu_{T_x}(V)(I+xA)^{-1/2}\leq\nu_B(V).

Take V=[0,e−u]V=[0,e^{-u}], where u>0u>0. By Lemma 3.2,

νTx([0,e−u])=gu(Tx),νB([0,e−u])=gu(B).\nu_{T_x}([0,e^{-u}])=g_u(T_x),\qquad\nu_B([0,e^{-u}])=g_u(B).

We therefore have

(I+xA)−1/2gu(Tx)(I+xA)−1/2≤gu(B),x,u>0.(13)(I+xA)^{-1/2}g_u(T_x)(I+xA)^{-1/2}\leq g_u(B),\qquad x,u>0. \tag*{(13)}

We use this inequality to prove the inclusion in (11). Let v∈PHv\in PH and fix q>rq>r. Since P=EB([0,r])P=E_B([0,r]), the spectrum of BB on PHPH lies in [0,r][0,r].

For 0≤λ≤r0\leq\lambda\leq r, the definition of gug_u gives

0≤gu(λ)≤e−u/r.0\leq g_u(\lambda)\leq e^{-u/r}.

Hence

⟨v,gu(B)v⟩≤e−u/r∥v∥2.\langle v,g_u(B)v\rangle\leq e^{-u/r}\lVert v\rVert^2.

Combining this with (13), we get

⟨(I+xA)−1/2v,gu(Tx)(I+xA)−1/2v⟩≤e−u/r∥v∥2.\left\langle(I+xA)^{-1/2}v,g_u(T_x)(I+xA)^{-1/2}v\right\rangle\le e^{-u/r}\lVert v\rVert^2.

On the other hand, for every λ≥q\lambda\ge q,

gu(λ)=e−u/λ≥e−u/q.g_u(\lambda)=e^{-u/\lambda}\ge e^{-u/q}.

The spectral theorem therefore gives

e−u/q∥ETx([q,∞))(I+xA)−1/2v∥2≤⟨(I+xA)−1/2v,gu(Tx)(I+xA)−1/2v⟩.e^{-u/q}\left\lVert E_{T_x}([q,\infty))(I+xA)^{-1/2}v\right\rVert^2\le\left\langle(I+xA)^{-1/2}v,g_u(T_x)(I+xA)^{-1/2}v\right\rangle.

Combining the last two inequalities,

∥ETx([q,∞))(I+xA)−1/2v∥2≤e−u(1/r−1/q)∥v∥2.\left\lVert E_{T_x}([q,\infty))(I+xA)^{-1/2}v\right\rVert^2\le e^{-u(1/r-1/q)}\lVert v\rVert^2.

Since q>rq>r, we have 1/r−1/q>01/r-1/q>0. The left-hand side is independent of uu, whereas the right-hand side tends to zero as u→∞u\to\infty. Thus

ETx([q,∞))(I+xA)−1/2v=0.E_{T_x}([q,\infty))(I+xA)^{-1/2}v=0.

This holds for every q>rq>r. Letting q↓rq\downarrow r and using strong convergence of spectral projections, we obtain

ETx((r,∞))(I+xA)−1/2v=0.E_{T_x}((r,\infty))(I+xA)^{-1/2}v=0.

Consequently,

(I+xA)−1/2v∈ETx([0,r])H.(I+xA)^{-1/2}v\in E_{T_x}([0,r])H.

Since v∈PHv\in PH was arbitrary,

(I+xA)−1/2PH⊆ETx([0,r])H.(I+xA)^{-1/2}PH\subseteq E_{T_x}([0,r])H.

It remains to prove equality for sufficiently small xx. Since Tx→BT_x\to B in norm and r∉σ(B)r\notin\sigma(B), the spectral projections

Qx=ETx([0,r])Q_x=E_{T_x}([0,r])

converge in norm to PP.

Let Πx\Pi_x be the orthogonal projection onto (I+xA)−1/2PH(I+xA)^{-1/2}PH. We also have Πx→P\Pi_x\to P in norm. To see this, the operator

P(I+xA)−1PP(I+xA)^{-1}P

is positive and invertible on PHPH, and its restriction to PHPH converges in norm to IPHI_{PH}. The orthogonal projection onto (I+xA)−1/2PH(I+xA)^{-1/2}PH is

Πx=(I+xA)−1/2P⋅(P(I+xA)−1P∣PH)−1⋅P(I+xA)−1/2.\Pi_x=(I+xA)^{-1/2}P\cdot\left(P(I+xA)^{-1}P\big|_{PH}\right)^{-1}\cdot P(I+xA)^{-1/2}.

Each factor converges in norm to the corresponding factor at x=0x=0. Hence Πx→P\Pi_x\to P.

The inclusion established above gives

Πx≤Qx.\Pi_x\le Q_x.

Thus Qx−ΠxQ_x-\Pi_x is an orthogonal projection, and

∥Qx−Πx∥≤∥Qx−P∥+∥Πx−P∥⟶0.\lVert Q_x-\Pi_x\rVert\le\lVert Q_x-P\rVert+\lVert\Pi_x-P\rVert\longrightarrow0.

Since every nonzero orthogonal projection has norm 11, we have Qx=ΠxQ_x=\Pi_x for all sufficiently small x>0x>0. Therefore

(I+xA)−1/2PH=ETx([0,r])H.(I+xA)^{-1/2}PH=E_{T_x}([0,r])H.

This is (11), and hence (8) follows. The assertion with AA and BB interchanged follows by symmetry. □

The affine case (C=0C = 0) has a stronger rigidity property. The following result holds on an arbitrary Hilbert space.

Theorem 3.5. Let HH be a complex Hilbert space and let A,B∈B(H)+A,B \in B(H)_{+}. There is a positive B(H)B(H)-valued measure EE on [0,1]2[0,1]^2 such that

(I+mA+nB)−1=∫[0,1]2smtn dE(s,t)(m,n≥0)(I+mA+nB)^{-1}=\int_{[0,1]^2}s^{m}t^{n}\,\mathrm{d}E(s,t) \qquad(m,n\ge0)

if and only if AB=BAAB=BA.

Proof. Suppose the measure exists, and put

F(x,y)=(I+xA+yB)−1.F(x,y)=(I+xA+yB)^{-1}.

Proposition 3.3 shows that the measure represents F(x,y)F(x,y) for all real x,y≥0x,y\ge0. For x,y>0x,y>0, differentiation under the integral gives

(−1)k+1∂k+1F∂x∂yk(x,y)≥0(k≥0).(14)(-1)^{k+1}\frac{\partial^{k+1}F}{\partial x\partial y^{k}}(x,y)\ge0 \qquad(k\ge0). \tag*{(14)}

The logarithmic factors are bounded after multiplication by sxtys^{x}t^{y}, so the differentiation is justified. We may then let xx and yy tend to zero, since the derivatives of FF extend continuously to the origin.

Now

−∂F∂x(0,y)=(I+yB)−1A(I+yB)−1.-\frac{\partial F}{\partial x}(0,y)=(I+yB)^{-1}A(I+yB)^{-1}.

Expanding near zero gives

(I+yB)−1A(I+yB)−1=∑j,ℓ≥0(−y)j+ℓBjABℓ=∑k=0∞(−y)k(∑j=0kBjABk−j)⏟=:Hk.\begin{aligned} (I+yB)^{-1}A(I+yB)^{-1} &=\sum_{j,\ell\ge0}(-y)^{j+\ell}B^{j}AB^{\ell}\\ &=\sum_{k=0}^{\infty}(-y)^{k}\underbrace{\left(\sum_{j=0}^{k}B^{j}AB^{k-j}\right)}_{=:H_k}. \end{aligned}

Using (14), we get

(−1)k+1∂k+1F∂x∂yk(0,0)=k!Hk(-1)^{k+1}\frac{\partial^{k+1}F}{\partial x\partial y^{k}}(0,0)=k!H_k
⇓\Downarrow
Hk≥0,∀k≥0.H_k\ge0,\qquad\forall k\ge0.

We show that these inequalities force [A,B]=0[A,B]=0. To see this, consider

0<a<b≤c,ac<b20<a<b\le c,\qquad ac<b^2

and set

P=EB([0,a]),Q=EB([b,c]).P=E_B([0,a]),\qquad Q=E_B([b,c]).

Then we have

∥BjP∥≤aj,∥BjQ∥≤cj.\lVert B^{j}P\rVert\le a^{j},\qquad\lVert B^{j}Q\rVert\le c^{j}.

Note that

∥PHkP∥≤∑j=0k∥PBjABk−jP∥≤∑j=0k∥PBj∥∥A∥∥Bk−jP∥≤(k+1)ak∥A∥.\begin{aligned} \lVert P H_k P\rVert\le\sum_{j=0}^{k} \lVert P B^j A B^{k-j} P\rVert\\ &\le\sum_{j=0}^{k} \lVert P B^j\rVert\lVert A\rVert\lVert B^{k-j}P\rVert\\ &\le(k+1)a^k\lVert A\rVert. \end{aligned}

Similarly,

∥QHkQ∥≤(k+1)ck∥A∥.\lVert Q H_k Q\rVert\le(k+1)c^k\lVert A\rVert.

Therefore,

∥PHkQ∥2=∥(Hk1/2P)∗(Hk1/2Q)∥2≤∥(Hk1/2P)∗∥2∥Hk1/2Q∥2=∥PHkP∥∥QHkQ∥≤(k+1)2(ac)k∥A∥2.\begin{aligned} \lVert P H_k Q\rVert^2 = \left\lVert(H_k^{1/2}P)^*(H_k^{1/2}Q)\right\rVert^2 \\ &\le\left\lVert(H_k^{1/2}P)^*\right\rVert^2 \left\lVert H_k^{1/2}Q\right\rVert^2 \\ &= \lVert P H_k P\rVert\lVert QH_kQ\rVert \le(k+1)^2(ac)^k\lVert A\rVert^2. \end{aligned}

That is,

∥PHkQ∥≤(k+1)(ac)k/2∥A∥.\lVert P H_k Q\rVert\le(k+1)(ac)^{k/2}\lVert A\rVert.

Observe that, for k≥1k\ge1,

Hk=∑j=0kBjABk−j=ABk+BHk−1.H_k=\sum_{j=0}^{k}B^jAB^{k-j}=AB^k+BH_{k-1}.

Since Q=EB([b,c])Q=E_B([b,c]) and b>0b>0, the restriction B∣QHB|_{QH} is invertible, with

∥(B∣QH)−k∥≤b−k.\left\lVert(B|_{QH})^{-k}\right\rVert\le b^{-k}.

It follows that

PHkQ(B∣QH)−k=PABkQ(B∣QH)−k+PBHk−1Q(B∣QH)−k=PAQ+PBHk−1Q(B∣QH)−k.\begin{aligned} PH_kQ(B|_{QH})^{-k}=PAB^kQ(B|_{QH})^{-k}+PBH_{k-1}Q(B|_{QH})^{-k} \\ &=PAQ+PBH_{k-1}Q(B|_{QH})^{-k}. \end{aligned}

Therefore,

∥PAQ∥≤∥PHkQ(B∣QH)−k∥+∥PBHk−1Q(B∣QH)−k∥≤b−k(∥PHkQ∥+∥PBHk−1Q∥)≤b−k(∥PHkQ∥+a∥PHk−1Q∥)≤∥A∥((k+1)(acb)k+abk(acb)k−1).\begin{aligned} \lVert PAQ\rVert\le\left\lVert PH_kQ(B|_{QH})^{-k}\right\rVert+\left\lVert PBH_{k-1}Q(B|_{QH})^{-k}\right\rVert\\ &\le b^{-k}\bigl(\lVert PH_kQ\rVert+\lVert PBH_{k-1}Q\rVert\bigr) \\ &\le b^{-k}\bigl(\lVert PH_kQ\rVert+a\lVert PH_{k-1}Q\rVert\bigr) \\ &\le\lVert A\rVert\left((k+1)\left(\frac{\sqrt{ac}}{b}\right)^k+\frac{a}{b}k\left(\frac{\sqrt{ac}}{b}\right)^{k-1}\right). \end{aligned}

Since ac<b2ac<b^2, the right-hand side tends to zero as k→∞k\to\infty. Hence PAQ=0PAQ=0.

We have proved that

EB([0,a])AEB([b,c])=0E_B([0,a])AE_B([b,c])=0

whenever 0<a<b≤c0<a<b\le c and ac<b2ac<b^2.

For fixed a>0a>0 and b>ab>a, subdividing [b,∥B∥][b,\lVert B\rVert] into sufficiently small spectral intervals therefore gives

EB([0,a])AEB([b,∞))=0.E_B([0,a])AE_B([b,\infty))=0.

Letting b↓ab \downarrow a, we obtain

EB([0,a])AEB((a,∞))=0.E_B([0,a])AE_B((a,\infty))=0.

Taking adjoints gives the reverse off-diagonal identity. Thus AA commutes with EB([0,a])E_B([0,a]) for every a>0a>0. By the spectral theorem, AB=BAAB=BA.

Conversely, suppose AB=BAAB=BA. Then

(I+mA+nB)−1=∫0∞e−te−mtAe−ntB dt.(I+mA+nB)^{-1}=\int_{0}^{\infty}e^{-t}e^{-mtA}e^{-ntB}\,\mathrm{d}t.

For every t≥0t\geq0, the commuting positive contractions e−tAe^{-tA} and e−tBe^{-tB} have a joint spectral measure on [0,1]2[0,1]^2. Integrating these measures against e−t dte^{-t}\,\mathrm{d}t gives the required positive operator-valued measure. □\square

The following commutativity criterion, implicit in the proof of Theorem 3.5, may well be folklore. However, we have not found this exact formulation in the literature. Since it may be of independent interest, we state it separately.

Proposition 3.6. Let A,B∈B(H)+A,B\in\mathcal{B}(H)_+. For k≥0k\geq0, set

Hk=∑j=0kBjABk−j.H_k=\sum_{j=0}^{k}B^jAB^{k-j}.

Then the following holds:

AB=BA⟺Hk≥0(k≥0).AB=BA\Longleftrightarrow H_k\geq0\quad(k\geq0).

Completion in dimension two

In dimension two, the positive completion problem has a fairly rigid answer for noncommuting matrices. Although the problem asks for a positive matrix CC such that the two-parameter family

(I+mA+nB+mnC)−1(I+mA+nB+mnC)^{-1}

admits a joint Hausdorff moment representation, existence is determined by a single matrix inequality involving only AA and BB. Moreover, whenever a completion exists, it is unique and given by an explicit formula. The necessity follows from the spectral invariance established in § 3, and sufficiency is proved by reducing the matrix-valued moment problem to three scalar moment representations.

Theorem 4.1. Let A,B∈M2(C)A,B\in M_2(\mathbb{C}) be (strictly) positive definite and noncommuting. The pair A,BA,B has a positive completion if and only if

A∥A∥+B∥B∥≥I.(15)\frac{A}{\lVert A\rVert}+\frac{B}{\lVert B\rVert}\geq I. \tag*{(15)}

In that case the completion is unique, and it is given by

C0=∥B∥A+∥A∥B−∥A∥∥B∥I.(16)C_0=\lVert B\rVert A+\lVert A\rVert B-\lVert A\rVert\lVert B\rVert I. \tag*{(16)}

Proof. “⟹\Longrightarrow” Suppose first that CC is a positive completion.

Write

a=∥A∥,b=∥B∥.a=\lVert A\rVert,\qquad b=\lVert B\rVert.

Let PP be the spectral projection of BB for its smallest eigenvalue. Choose rr strictly between the two eigenvalues of $B. By Theorem 3.4, for all sufficiently small x>0x>0,

(B+xC)(I+xA)−1PH⊆PH.(B+xC)(I+xA)^{-1}PH\subseteq PH.

We have the identity

(B+xC)(I+xA)−1=B+x(C−BA)(I+xA)−1.(B+xC)(I+xA)^{-1}=B+x(C-BA)(I+xA)^{-1}.

Since PHPH reduces BB, it follows that

(C−BA)(I+xA)−1PH⊆PH.(C-BA)(I+xA)^{-1}PH\subseteq PH.

Expanding (I+xA)−1(I+xA)^{-1} near zero and comparing coefficients gives

(C−BA)AkPH⊆PH,∀k≥0.(C-BA)A^kPH\subseteq PH,\qquad\forall k\geq0.

The line PHPH is cyclic for AA. Indeed, if it were an eigenspace of AA, then both matrices would be diagonal with respect to PH⊕(I−P)HPH\oplus(I-P)H, contrary to AB≠BAAB\ne BA. Since H=PH+APHH=PH+APH, we obtain

(C−BA)H⊆PH.(C-BA)H\subseteq PH.

The operator BB acts as bIbI on (I−P)H(I-P)H, so

(I−P)(C−bA)=0.(I-P)(C-bA)=0.

The operator C0C_0 in (16) satisfies the same identity, since C0−bA=a(B−bI)C_0-bA=a(B-bI). Hence

(C−C0)H⊆PH.(C-C_0)H\subseteq PH.

Interchange AA and BB. If QQ is the spectral projection of AA for its smallest eigenvalue, the same argument gives

(C−C0)H⊆QH.(C-C_0)H\subseteq QH.

But PH∩QH={0}PH\cap QH=\{0\}, otherwise they would be a common eigenspace of AA and BB, and the two matrices would commute. Thus C=C0C=C_0. In particular, positivity of CC gives C0≥0C_0\geq0, which is (15).

“⟸\Longleftarrow” Conversely, assume that

C0=bA+aB−abI≥0.C_0=bA+aB-abI\geq0.

where a=∥A∥a=\lVert A\rVert and b=∥B∥b=\lVert B\rVert. We need to show that C0C_0 is a positive completion of A,BA,B.

Let d,ed,e be the smaller eigenvalues of A,BA,B, respectively. Then 0<d<a0<d<a and 0<e<b0<e<b.

The definition of C0C_0 gives

I+xA+yB+xyC0=(1+ax)(1+by)[I−x(aI−A)1+ax−y(bI−B)1+by⏟=M].(17)I+xA+yB+xyC_0=(1+ax)(1+by)\left[\underbrace{I-\frac{x(aI-A)}{1+ax}-\frac{y(bI-B)}{1+by}}_{=M}\right]. \tag*{(17)}

Both aI−AaI-A and bI−BbI-B have rank one, with traces a−da-d and b−eb-e. Recall that, for rank-one 2×22\times2 matrices U,VU,V,

det⁡(I−sU−tV)=(1−str⁡U)(1−ttr⁡V)−sttr⁡(UV).(18)\det(I-sU-tV)=(1-s\operatorname{tr}U)(1-t\operatorname{tr}V)-st\operatorname{tr}(UV). \tag*{(18)}

Indeed, the identities

det⁡(I−T)=1−tr⁡T+det⁡T\det(I-T)=1-\operatorname{tr}T+\det T
2det⁡T=(tr⁡T)2−tr⁡(T2)2\det T=(\operatorname{tr}T)^2-\operatorname{tr}(T^2)

give this formula (18), with T:=sU+tVT:=sU+tV, using det⁡U=det⁡V=0\det U=\det V=0.

Apply (18) to MM from (17) as follows. Write

M=I−sU−tVM=I-sU-tV

where U=aI−AU=aI-A, V=bI−BV=bI-B, s=x1+axs=\frac{x}{1+ax}, t=y1+byt=\frac{y}{1+by}. We get

det⁡M=(1−(a−d)x1+ax)(1−(b−e)y1+by)−xy(1+ax)(1+by)tr⁡((aI−A)(bI−B))=(1+dx)(1+ey)−xytr⁡((aI−A)(bI−B))(1+ax)(1+by).\begin{aligned} \det M=\left(1-\frac{(a-d)x}{1+ax}\right)\left(1-\frac{(b-e)y}{1+by}\right)-\frac{xy}{(1+ax)(1+by)}\operatorname{tr}\left((aI-A)(bI-B)\right) \\ &=\frac{(1+dx)(1+ey)-xy\operatorname{tr}\left((aI-A)(bI-B)\right)}{(1+ax)(1+by)}. \end{aligned}

Substitute this into (17), so

det⁡(I+xA+yB+xyC0)=(1+ax)2(1+by)2det⁡M=(1+ax)(1+by)(1+dx+ey+cxy)(19)\begin{aligned} \det\left(I+xA+yB+xyC_0\right)=(1+ax)^2(1+by)^2\det M \\ &=(1+ax)(1+by)(1+dx+ey+cxy) \tag*{(19)} \end{aligned}

where

c:=de−tr⁡((aI−A)(bI−B)).(20)c:=de-\operatorname{tr}\left((aI-A)(bI-B)\right). \tag*{(20)}

In (19), comparing the coefficients of x2y2x^2y^2, we get

det⁡C0=abc.\det C_0=abc.

Since C0≥0C_0\geq0, we have c≥0c\geq0. Also, aI−A≥0aI-A\geq0 and bI−B≥0bI-B\geq0, so

tr⁡((aI−A)(bI−B))=tr⁡((bI−B)1/2(aI−A)(bI−B)1/2)≥0.\operatorname{tr}\left((aI-A)(bI-B)\right)=\operatorname{tr}\left((bI-B)^{1/2}(aI-A)(bI-B)^{1/2}\right)\geq0.

Back to (20), it follows that

0≤c≤de.0\leq c\leq de.

By Lemma 2.1, there is a positive measure μ\mu on [0,1]2[0,1]^2 such that

11+dm+en+cmn=∫[0,1]2smtn dμ(s,t),m,n≥0.(21)\frac{1}{1+dm+en+cmn}=\int_{[0,1]^2}s^mt^n\,d\mu(s,t),\qquad m,n\geq0. \tag*{(21)}

We next calculate the matrix inverse

F(x,y)=(I+xA+yB+xyC0)−1=1(1+ax)(1+by)M−1,F(x,y)=\left(I+xA+yB+xyC_0\right)^{-1}=\frac{1}{(1+ax)(1+by)}M^{-1},

using (17).

Recall that, for any invertible 2×22\times2 matrix TT,

T−1=(tr⁡T)I−Tdet⁡T.T^{-1}=\frac{(\operatorname{tr}T)I-T}{\det T}.

For the matrix M=I−sU−tVM=I-sU-tV as above, we have

tr⁡M=2−(a−d)s−(b−e)t.\operatorname{tr}M=2-(a-d)s-(b-e)t.

Thus,

(tr⁡M)I−M=(2−(a−d)s−(b−e)t)I−(I−sU−tV)=I−s(A−dI)−t(B−eI).\begin{aligned} (\operatorname{tr}M)I-M=(2-(a-d)s-(b-e)t)I-(I-sU-tV) \\ &=I-s(A-dI)-t(B-eI). \end{aligned}

Also, from (19),

det⁡M=1+dx+ey+cxy(1+ax)(1+by).\det M=\frac{1+dx+ey+cxy}{(1+ax)(1+by)}.

Therefore,

F(x,y)=1(1+ax)(1+by)M−1=1(1+ax)(1+by)(tr⁡M)I−Mdet⁡M=11+dx+ey+cxy(I−x(A−dI)1+ax−y(B−eI)1+by)=11+dx+ey+cxy((bd+ae)I−C0ab+A−dIa(1+ax)+B−eIb(1+by)).(22)\begin{aligned} F(x,y) &= \frac{1}{(1+ax)(1+by)}M^{-1}\\ &= \frac{1}{(1+ax)(1+by)}\frac{(\operatorname{tr}M)I-M}{\det M}\\ &= \frac{1}{1+dx+ey+cxy}\left(I-\frac{x(A-dI)}{1+ax}-\frac{y(B-eI)}{1+by}\right)\\ &= \frac{1}{1+dx+ey+cxy}\left(\frac{(bd+ae)I-C_0}{ab}+\frac{A-dI}{a(1+ax)}+\frac{B-eI}{b(1+by)}\right). \tag*{(22)} \end{aligned}

All three constant matrices in (22) are positive semidefinite:

K0=(bd+ae)I−C0ab,K1=A−dIa,K2=B−eIb.K_0=\frac{(bd+ae)I-C_0}{ab},\qquad K_1=\frac{A-dI}{a},\qquad K_2=\frac{B-eI}{b}.

The last two are immediate because d,ed,e are the smallest eigenvalues of A,BA,B. Since C0=bA+aB−abIC_0=bA+aB-abI, we have

tr⁡C0=b(a+d)+a(b+e)−2ab=bd+ae.\operatorname{tr}C_0=b(a+d)+a(b+e)-2ab=bd+ae.

Recall that C0≥0C_0\ge0, so

(bd+ae)I−C0=(tr⁡C0)I−C0≥0.(bd+ae)I-C_0=(\operatorname{tr}C_0)I-C_0\ge0.

It remains to show that the scalar factors in (22) have positive joint moment representations. Specifically, we now have

F(m,n)=K01+dm+en+cmn+K1(1+dm+en+cmn)(1+am)+K2(1+dm+en+cmn)(1+bn).\begin{aligned} F(m,n)&=\frac{K_0}{1+dm+en+cmn}+\frac{K_1}{(1+dm+en+cmn)(1+am)}\\ &\quad+\frac{K_2}{(1+dm+en+cmn)(1+bn)}. \end{aligned}

The first is represented by (21).

The function (1+am)−1(1+am)^{-1} is a one-variable Hausdorff moment sequence, since

11+am=1a∫01um+1/a−1 du=∫01um dρ(u),\frac{1}{1+am}=\frac{1}{a}\int_0^1u^{m+1/a-1}\,du=\int_0^1u^m\,d\rho(u),

where ρ\rho is the probability measure on [0,1][0,1] with density a−1u1/a−1a^{-1}u^{1/a-1}.

Let μ\mu be the positive measure in (21). Multiplying the two moment representations gives

1(1+dm+en+cmn)(1+am)=∫[0,1]2∫01smtnum dρ(u) dμ(s,t)=∫[0,1]2∫01(su)mtn dρ(u) dμ(s,t).\begin{aligned} \frac{1}{(1+dm+en+cmn)(1+am)} =\int_{[0,1]^2}\int_0^1s^mt^nu^m\,d\rho(u)\,d\mu(s,t) \\ &=\int_{[0,1]^2}\int_0^1(su)^mt^n\,d\rho(u)\,d\mu(s,t). \end{aligned}

Let ν1\nu_1 be the pushforward of μ⊗ρ\mu\otimes\rho under the map

(s,t,u)⟼(su,t).(s,t,u)\longmapsto(su,t).

Since this map takes [0,1]3[0,1]^3 into [0,1]2[0,1]^2, the measure ν1\nu_1 is a positive measure on [0,1]2[0,1]^2, and

1(1+dm+en+cmn)(1+am)=∫[0,1]2smtn dν1(s,t).\frac{1}{(1+dm+en+cmn)(1+am)} =\int_{[0,1]^2}s^mt^n\,d\nu_1(s,t).

Similarly, we get a positive measure ν2\nu_2 on [0,1]2[0,1]^2 such that

1(1+dm+en+cmn)(1+bn)=∫[0,1]2smtn dν2(s,t).\frac{1}{(1+dm+en+cmn)(1+bn)}=\int_{[0,1]^2}s^m t^n\,d\nu_2(s,t).

Now define the matrix-valued measure

E(Ω)=K0μ(Ω)+K1ν1(Ω)+K2ν2(Ω).E(\Omega)=K_0\mu(\Omega)+K_1\nu_1(\Omega)+K_2\nu_2(\Omega).

for every Borel set Ω⊆[0,1]2\Omega\subseteq[0,1]^2. Since K0,K1,K2≥0K_0,K_1,K_2\geq0 and the three scalar measures are positive, EE is a positive matrix-valued measure. By the three representations above,

F(m,n)=∫[0,1]2smtn dE(s,t),m,n≥0.F(m,n)=\int_{[0,1]^2}s^m t^n\,dE(s,t),\qquad m,n\geq0.

Moreover,

E([0,1]2)=F(0,0)=I.E\left([0,1]^2\right)=F(0,0)=I.

Thus EE is a normalized positive operator-valued measure, and C0C_0 is a positive completion. □\square

Higher dimensions

The argument in Theorem 3.4 gives a restriction on AA and BB when their lowest spectral subspaces are one-dimensional. In higher dimensions this restriction is also sufficient. The proof works for bounded operators on an arbitrary Hilbert space.

Theorem 5.1. Let HH be a complex Hilbert space of dimension at least two, and let A,B∈B(H)A,B\in B(H) be strictly positive. Suppose that their smallest spectral values a,ba,b are simple isolated eigenvalues, with unit eigenvectors q,pq,p, respectively. Assume that pp is cyclic for AA. Let

R=1⟨q,p⟩∣p⟩⟨q∣.R=\frac{1}{\langle q,p\rangle}|p\rangle\langle q|.

Then the pair A,BA,B has a positive completion if and only if there is a number 0<γ≤ab0<\gamma\leq ab such that

(A−aI)(B−bI)=γ(I−R).(23)(A-aI)(B-bI)=\gamma(I-R). \tag*{(23)}

In that case the completion is unique and equals

C=BA−γR=AB+BA−aB−bA+(ab−γ)I.(24)C=BA-\gamma R=AB+BA-aB-bA+(ab-\gamma)I. \tag*{(24)}

Equivalently, the spectral relation (23) can be written

(A−aI)1/2(B−bI)(A−aI)1/2=γ(I−∣q⟩⟨q∣).(25)(A-aI)^{1/2}(B-bI)(A-aI)^{1/2}=\gamma(I-|q\rangle\langle q|). \tag*{(25)}

Proof. Note that since pp is cyclic for AA and qq is an eigenvector of AA, we have ⟨q,p⟩≠0\langle q,p\rangle\ne0. Thus RR is well-defined. Moreover, R2=RR^2=R, with range Cp\mathbb{C}p and kernel q⊥q^\perp.

  1. Suppose first that CC is a positive completion. Choose a number strictly between bb and the rest of σ(B)\sigma(B). By Theorem 3.4, with PP the projection onto Cp\mathbb{C}p, we have

(B+xC)(I+xA)−1Cp⊆Cp(B+xC)(I+xA)^{-1}\mathbb{C}p\subseteq\mathbb{C}p

for all sufficiently small x>0x>0. Since

(B+xC)(I+xA)−1=B+x(C−BA)(I+xA)−1,(B+xC)(I+xA)^{-1}=B+x(C-BA)(I+xA)^{-1},

and Bp=bpBp=bp, we have

(C−BA)(I+xA)−1p∈Cp.(C-BA)(I+xA)^{-1}p\in\mathbb{C}p.

Expand (I+xA)−1(I+xA)^{-1} at zero. Comparing coefficients gives

(C−BA)Akp∈Cp(k≥0).(C-BA)A^{k}p\in\mathbb{C}p \qquad(k\geq0).

By cyclicity of pp,

(C−BA)H⊆Cp.(C-BA)H\subseteq\mathbb{C}p.

That is,

C−BA=∣p⟩⟨w∣C-BA=|p\rangle\langle w|

for some w∈Hw\in H; or, by taking adjoint,

C−AB=∣w⟩⟨p∣.(26)C-AB=|w\rangle\langle p|. \tag*{(26)}

We need only one more consequence of the measure argument. Apply Theorem 3.4 with AA, BB interchanged, using the isolated eigenvalue aa and its eigenvector qq. We obtain

(C−AB)q∈Cq.(C-AB)q\in\mathbb{C}q.

Since ⟨q,p⟩≠0\langle q,p\rangle\ne0, using (26),

(C−AB)q=w⟨p,q⟩∈Cq(C-AB)q=w\langle p,q\rangle\in\mathbb{C}q

and so w∈Cqw\in\mathbb{C}q.

We have therefore proved

C−BA=−γR(27)C-BA=-\gamma R \tag*{(27)}

for some scalar γ\gamma. It is real, since Rp=pRp=p and

⟨p,(C−BA)p⟩=⟨p,Cp⟩−b⟨p,Ap⟩\langle p,(C-BA)p\rangle=\langle p,Cp\rangle-b\langle p,Ap\rangle

is real. Notice that we did not assume qq cyclic for BB.

Taking adjoint in (27) gives

C−AB=−γR∗.(28)C-AB=-\gamma R^{*}. \tag*{(28)}

Combining (27)–(28), we get

[A−aI⏟S,B−bI⏟T]=[A,B]=γ(R∗−R).\left[\underbrace{A-aI}_{S},\underbrace{B-bI}_{T}\right]=[A,B]=\gamma(R^{*}-R).

Also, RS=0RS=0, because the range of SS is contained in q⊥q^{\perp}, and SR∗=0SR^{*}=0, because the range of R∗R^{*} is Cq\mathbb{C}q. Let

K=ST−γ(I−R).K=ST-\gamma(I-R).

Then

[S,K]=S[S,T]+γ[S,R]=γS(R∗−R)+γ(SR−RS)=γ(SR∗−RS)=0.\begin{aligned} [S,K]&=S[S,T]+\gamma[S,R]\\ &=\gamma S(R^{*}-R)+\gamma(SR-RS)\\ &=\gamma(SR^{*}-RS)=0. \end{aligned}

Since Tp=0Tp=0 and Rp=pRp=p, we also have Kp=0Kp=0. As KK commutes with SS,

KSkp=SkKp=0KS^{k}p=S^{k}Kp=0

for every k≥0k\geq0. The vector pp is cyclic for SS, so K=0K=0, i.e.,

ST=γ(I−R).(29)ST=\gamma(I-R). \tag*{(29)}

This is precisely (23).

We next show that γ>0\gamma>0. Multiplying (29) on the right by SS and using RS=0RS=0, we obtain

STS=γS.STS=\gamma S.

For any nonzero x∈q⊥x \in q^\perp,

⟨Sx,TSx⟩=γ⟨x,Sx⟩.\langle Sx,TSx\rangle= \gamma\langle x,Sx\rangle.

Here ⟨x,Sx⟩>0\langle x,Sx\rangle> 0. Also, SxSx is a nonzero vector in q⊥q^\perp, while ker⁡T=Cp\ker T = \mathbb{C}p and ⟨q,p⟩≠0\langle q,p\rangle\ne0. Hence ⟨Sx,TSx⟩>0\langle Sx,TSx\rangle> 0, giving γ>0\gamma> 0.

To obtain the upper bound for γ\gamma, observe that Aq=aqAq = aq implies

⟨q,A−1p⟩=1a⟨q,p⟩.\langle q,A^{-1}p\rangle= \frac{1}{a}\langle q,p\rangle.

Therefore RA−1p=a−1pRA^{-1}p = a^{-1}p, and

CA−1p=(BA−γR)A−1p=(b−γa)p.CA^{-1}p = (BA-\gamma R)A^{-1}p = \left(b-\frac{\gamma}{a}\right)p.

Since C≥0C \ge0,

0≤⟨A−1p,CA−1p⟩=ab−γa⟨A−1p,p⟩.0 \le\langle A^{-1}p,CA^{-1}p\rangle= \frac{ab-\gamma}{a}\langle A^{-1}p,p\rangle.

The last inner product is strictly positive. Thus γ≤ab\gamma\le ab.

Substituting

R=I−γ−1(A−aI)(B−bI)R = I-\gamma^{-1}(A-aI)(B-bI)

into C=BA−γRC=BA-\gamma R gives

C=AB+BA−aB−bA+(ab−γ)I.C = AB+BA-aB-bA+(ab-\gamma)I.

It remains to verify the equivalent spectral relation. Since RS1/2=0RS^{1/2}=0, multiplying (29) on the right by S1/2S^{1/2} gives

STS1/2=γS1/2.STS^{1/2}=\gamma S^{1/2}.

The operator S1/2S^{1/2} is boundedly invertible on q⊥q^\perp, because aa is an isolated eigenvalue of AA. Consequently,

S1/2TS1/2=γ(I−∣q⟩⟨q∣).(30)S^{1/2}TS^{1/2}=\gamma(I-\lvert q\rangle\langle q\rvert). \tag*{(30)}

That is, (25) holds.

Conversely, suppose that (30) holds. Since S1/2S^{1/2} is invertible on q⊥q^\perp, the identity implies

STx=γxSTx=\gamma x

for every x∈q⊥x\in q^\perp. Also, STp=0STp=0, since Tp=0Tp=0. As

H=q⊥⊕Cp,H=q^\perp\oplus\mathbb{C}p,

and RR is the idempotent with kernel q⊥q^\perp and Rp=pRp=p, we recover (29).

2. We next prove sufficiency. Suppose that (23) holds with 0<γ≤ab0<\gamma\le ab. Define

C=BA−γR.C=BA-\gamma R.

Expanding (23) gives

C=AB+BA−aB−bA+(ab−γ)I,C=AB+BA-aB-bA+(ab-\gamma)I,

so CC is self-adjoint.

Decompose H=Cp⊕p⊥H=\mathbb{C}p\oplus p^\perp and write

A=(uv∗vA1),B=(b00B1).A=\begin{pmatrix}u&v^*\\v&A_1\end{pmatrix},\qquad B=\begin{pmatrix}b&0\\0&B_1\end{pmatrix}.

Put D=A1−aID=A_1-aI. Since aa is a simple isolated bottom eigenvalue of AA, there is ε>0\varepsilon>0 such that

A−aI≥ε(I−∣q⟩⟨q∣).A-aI\ge\varepsilon(I-\lvert q\rangle\langle q\rvert).

For z∈p⊥z \in p^\perp,

⟨z,Dz⟩=⟨z,(A−aI)z⟩≥ε(∥z∥2−∣⟨q,z⟩∣2)≥ε∣⟨q,p⟩∣2∥z∥2.\begin{aligned} \langle z,Dz\rangle=\langle z,(A-aI)z\rangle\\ &\geq\varepsilon\left(\lVert z\rVert^2-\lvert\langle q,z\rangle\rvert^2\right)\geq\varepsilon\lvert\langle q,p\rangle\rvert^2\lVert z\rVert^2. \end{aligned}

Thus DD is strictly positive and boundedly invertible.

The equation (A−aI)q=0(A-aI)q=0, together with ⟨q,p⟩≠0\langle q,p\rangle\ne0, shows that qq is a nonzero scalar multiple of (1,−D−1v)(1,-D^{-1}v). Hence

u=a+⟨v,D−1v⟩,R=(1−v∗D−100).u=a+\langle v,D^{-1}v\rangle,\qquad R=\begin{pmatrix}1&-v^*D^{-1}\\0&0\end{pmatrix}.

Comparing the lower-right blocks in (23) gives

D(B1−bI)=γI,D(B_1-bI)=\gamma I,

so B1=bI+γD−1B_1=bI+\gamma D^{-1}. We have obtained

A=(a+⟨v,D−1v⟩v∗vaI+D),B=(b00bI+γD−1).(31)A=\begin{pmatrix}a+\langle v,D^{-1}v\rangle&v^*\\v&aI+D\end{pmatrix},\qquad B=\begin{pmatrix}b&0\\0&bI+\gamma D^{-1}\end{pmatrix}. \tag*{(31)}

Set

c=ab−γ,h(x,y)=1+ax+by+cxy.c=ab-\gamma,\qquad h(x,y)=1+ax+by+cxy.

We have 0≤c<ab0\leq c<ab. Define

K0=aA−1+bB−1−I,K_0=aA^{-1}+bB^{-1}-I,
K1=I−aA−1,K_1=I-aA^{-1},
K2=I−bB−1.K_2=I-bB^{-1}.

Clearly, K1,K2≥0K_1,K_2\geq0, because A≥aIA\geq aI and B≥bIB\geq bI. We show that K0≥0K_0\geq0 as well.

From (31), the operators A1A_1 and B1B_1 commute and satisfy

(A1−aI)(B1−bI)=γI.(A_1-aI)(B_1-bI)=\gamma I.

Expanding and multiplying by A1−1B1−1A_1^{-1}B_1^{-1} gives

aA1−1+bB1−1−I=cA1−1B1−1.aA_1^{-1}+bB_1^{-1}-I=cA_1^{-1}B_1^{-1}.

Consequently,

K0=a(A−1−(A1−100A1−1))+c(000A1−1B1−1).K_0=a\left(A^{-1}-\begin{pmatrix}A_1^{-1}&0\\0&A_1^{-1}\end{pmatrix}\right)+c\begin{pmatrix}0&0\\0&A_1^{-1}B_1^{-1}\end{pmatrix}.

The second term is positive because A1A_1 and B1B_1 are commuting positive operators. For the first term, the block inverse formula gives

A−1−(A1−100A1−1)=1u−⟨v,A1−1v⟩(1−A1−1v)(1−v∗A1−1)≥0.A^{-1}-\begin{pmatrix}A_1^{-1}&0\\0&A_1^{-1}\end{pmatrix} =\frac{1}{u-\langle v,A_1^{-1}v\rangle} \begin{pmatrix}1\\-A_1^{-1}v\end{pmatrix} \left(1\quad-v^*A_1^{-1}\right)\geq0.

Here the denominator is strictly positive, being the Schur complement of A1A_1 in AA. Thus K0≥0K_0\geq0.

We have proved

K0,K1,K2≥0,K0+K1+K2=I.K_0,K_1,K_2\geq0,\qquad K_0+K_1+K_2=I.

Moreover, K1K_1 commutes with AA, and K2K_2 commutes with BB.

Now consider

G(x,y)=K0+K1(I+xA)−1+K2(I+yB)−1.G(x,y)=K_0+K_1(I+xA)^{-1}+K_2(I+yB)^{-1}.

Since the coefficients are positive and sum to II, and both resolvents are strictly positive, G(x,y)G(x,y) is strictly positive and boundedly invertible for every $x,y\geq 0. The elementary resolvent identity gives

G(x,y)=(1+ax)(I+xA)−1+(1+by)(I+yB)−1−I.G(x,y)=(1+ax)(I+xA)^{-1}+(1+by)(I+yB)^{-1}-I.

We claim that

G(x,y)(I+xA+yB+xyC)=h(x,y)I.G(x,y)(I+xA+yB+xyC)=h(x,y)I.

To see this, recall that

C=BA−γR=AB−γR∗,C=BA-\gamma R=AB-\gamma R^*,

and that AR∗=aR∗AR^*=aR^* and BR=bRBR=bR. Thus

(1+ax)(I+xA)−1(I+xA+yB+xyC)=(1+ax)(I+yB)−γxyR∗,(1+ax)(I+xA)^{-1}(I+xA+yB+xyC)=(1+ax)(I+yB)-\gamma xyR^*,

while

(1+by)(I+yB)−1(I+xA+yB+xyC)=(1+by)(I+xA)−γxyR.(1+by)(I+yB)^{-1}(I+xA+yB+xyC)=(1+by)(I+xA)-\gamma xyR.

Combining these identities gives

G(x,y)(I+xA+yB+xyC)=(1+ax)(I+yB)+(1+by)(I+xA)−(I+xA)(I+yB)−γxyR=(1+ax+by)I+xy(aB+bA−AB−γR)=h(x,y)I,\begin{aligned} G(x,y)(I+xA+yB+xyC)&=(1+ax)(I+yB)+(1+by)(I+xA)\\ &\quad-(I+xA)(I+yB)-\gamma xyR\\ &=(1+ax+by)I+xy(aB+bA-AB-\gamma R)\\ &=h(x,y)I, \end{aligned}

where the last equality follows from (23).

Since G(x,y)>0G(x,y)>0 and h(x,y)>0h(x,y)>0, we obtain

I+xA+yB+xyC=h(x,y)G(x,y)−1>0.I+xA+yB+xyC=h(x,y)G(x,y)^{-1}>0.

Taking x=y=tx=y=t, dividing by t2t^2, and letting t→∞t\to\infty gives C≥0C\geq0. We have also proved the resolvent formula

F(x,y)=(I+xA+yB+xyC)−1=1h(x,y)[K0+K1(I+xA)−1+K2(I+yB)−1].(32)\begin{aligned} F(x,y)&=(I+xA+yB+xyC)^{-1}\\ &=\frac{1}{h(x,y)}\left[K_0+K_1(I+xA)^{-1}+K_2(I+yB)^{-1}\right]. \tag*{(32)} \end{aligned}

It remains to obtain a joint moment representation. By Lemma 2.1, the scalar sequence

{h(m,n)−1}m,n≥0\{h(m,n)^{-1}\}_{m,n\geq0}

has a positive joint moment representation. By Lemma 3.2, the resolvent sequences {(I+mA)−1}\{(I+mA)^{-1}\} and {(I+nB)−1}\{(I+nB)^{-1}\} have normalized positive operator-valued representing measures.

Since K1K_1 commutes with AA and K2K_2 commutes with BB, the sequences

K1(I+mA)−1,K2(I+nB)−1K_1(I+mA)^{-1},\qquad K_2(I+nB)^{-1}

also have positive operator-valued representing measures. Combining these with the scalar representation, as in the proof of Theorem 4.1, gives a positive operator-valued measure representing F(m,n)F(m,n). It is normalized because F(0,0)=IF(0,0)=I.

Thus CC is a positive completion of A,BA,B. □

The condition in Theorem 5.1 is restrictive, even though there are examples in every dimension. First we retain the generic nonexistence result.

Corollary 5.2. For every d≥3d \ge3, the pairs of positive definite Hermitian d×dd \times d matrices with no positive completion contain an open dense subset. The same is true for real symmetric positive definite matrices.

For d=2d = 2, the noncommuting pairs having a positive completion contain a nonempty open subset.

Proof. For d≥3d \ge3, consider first the pairs whose smallest eigenvalues are simple and for which the bottom eigenvector of BB is cyclic for AA and the bottom eigenvector of AA is cyclic for BB. These conditions describe an open dense set. For density, perturb the two matrices to have simple spectra and then make a small change of their relative eigenbases. The required overlaps are nonzero except on sets with empty interior, and the same argument works for real orthogonal changes of basis.

For a pair in this set, let a,qa,q be the bottom eigenvalue and eigenvector of AA, and let b,pb,p be those of BB. If a completion exists, (25) implies that

(A−aI)1/2(B−bI)(A−aI)1/2(A-aI)^{1/2}(B-bI)(A-aI)^{1/2}

is a scalar operator on q⊥q^\perp. This condition is closed relative to the open set under consideration. The lowest eigenvalues and their spectral projections depend continuously on the matrices.

To see that the scalar condition has empty interior, keep AA, pp, and bb fixed and vary only the lower-right block of BB on p⊥p^\perp. Small self-adjoint perturbations preserve positivity, the simple bottom eigenvalue, and the cyclicity conditions. The restriction of (A−aI)1/2(A-aI)^{1/2} defines an invertible map

T:p⊥⟶q⊥.T:p^\perp\longrightarrow q^\perp.

Indeed, its kernel is p⊥∩Cq={0}p^\perp\cap\mathbb{C}q = \{0\}, and both spaces have dimension d−1d-1. If B1B_1 is the lower-right block of BB, then the restriction of the preceding operator to q⊥q^\perp is

T(B1−bI)T∗.T(B_1-bI)T^*.

As B1B_1 varies, this can undergo arbitrary small self-adjoint perturbations on q⊥q^\perp. When d≥3d \ge3, the space q⊥q^\perp has dimension at least two. We can therefore perturb the restriction away from scalar operators. The pairs violating the scalar condition form an open dense subset, and none has a completion.

For d=2d = 2, choose noncommuting rank-one projections P,QP,Q and put A=I+εPA=I+\varepsilon P, B=I+εQB=I+\varepsilon Q, with 0<ε<10<\varepsilon<1. Both matrices have norm 1+ε1+\varepsilon, and the matrix in (16) is

(1+ε)[(1−ε)I+ε(P+Q)]>0.(1+\varepsilon)\left[(1-\varepsilon)I+\varepsilon(P+Q)\right]>0.

Thus these pairs have a completion by Theorem 4.1. Strict positivity and noncommutativity persist under small perturbations, giving a nonempty open set. □\square

We can also describe the operators in Theorem 5.1 up to unitary equivalence. This uses only the spectral theorem and the cyclic vector in its hypotheses.

Corollary 5.3. The pairs in Theorem 5.1 which have a positive completion are, up to unitary equivalence, the following pairs. Take a,b>0a,b>0, 0<γ≤a0<\gamma\le a, and a nonzero finite positive measure ν\nu supported in a compact subset of (0,∞)(0,\infty). On

H=C⊕L2(ν)defineH=\mathbb{C}\oplus L^2(\nu)\quad\text{define}
A(z,f)=((a+∫1t dν(t))z+∫f(t) dν(t), z+(a+t)f(t)),B(z,f)=(bz,(b+γt)f(t)).\begin{aligned} A(z,f)&=\left(\left(a+\int\frac{1}{t}\,d\nu(t)\right)z+\int f(t)\,d\nu(t),\ z+(a+t)f(t)\right),\\ B(z,f)&=\left(bz,\left(b+\frac{\gamma}{t}\right)f(t)\right). \end{aligned}

In the second component of A(z,f)A(z,f), the term zz means the constant function with value zz. Every such pair has a positive completion. Two such pairs are unitarily equivalent if and only if their numbers a,b,γa,b,\gamma and measures ν\nu agree.

Proof. For a completed pair, use the block form (31). The vector vv is cyclic for DD. Otherwise, the orthogonal complement of its cyclic subspace would be a nonzero reducing subspace for DD orthogonal to vv. This would give a nonzero reducing subspace for AA orthogonal to pp, contradicting cyclicity of pp. Conversely, if vv is cyclic for DD, then p=(1,0)p=(1,0) is cyclic for AA. Indeed, a reducing subspace for AA orthogonal to pp would lie in {0}⊕p⊥\{0\}\oplus p^\perp, would reduce DD, and would be orthogonal to vv. There is no such nonzero subspace.

Let EDE_D be the spectral measure of DD and set

ν(V)=⟨v,ED(V)v⟩\nu(V)=\langle v,E_D(V)v\rangle

for Borel V⊂(0,∞)V\subset(0,\infty). The cyclic spectral theorem identifies p⊥p^\perp with L2(ν)L^2(\nu), takes DD to multiplication by tt, and takes vv to the constant function 11. Substitution in (31) gives the displayed formulas. Conversely, for a nonzero measure with the stated support, multiplication by tt has the cyclic vector 11, because polynomials are dense in L2(ν)L^2(\nu). The operators satisfy (23) and therefore have a positive completion.

The numbers a,ba,b are the bottom eigenvalues. Since RR is determined by their one-dimensional eigenspaces and I−R≠0I-R\ne0, (23) uniquely determines γ\gamma. The measure ν\nu is determined by the compression of A−aIA-aI to p⊥p^\perp and its coupling vector vv. These constructions respect unitary equivalence of the pair. Thus equal data give unitarily equivalent pairs, and equivalent pairs have equal data. □

Here are two examples. The first settles a three-dimensional case which is not covered by the two-dimensional theorem.

A=(621250104),B=(100020003).A=\begin{pmatrix}6&2&1\\2&5&0\\1&0&4\end{pmatrix},\qquad B=\begin{pmatrix}1&0&0\\0&2&0\\0&0&3\end{pmatrix}.

In (31), a=3a=3, b=1b=1, γ=2\gamma=2, D=diag⁡(2,1)D=\operatorname{diag}(2,1), and v=(2,1)v=(2,1). Its positive completion is

C=(44341003012).C=\begin{pmatrix}4&4&3\\4&10&0\\3&0&12\end{pmatrix}.

The pair generates M3(C)M_3(\mathbb{C}).

For an infinite-dimensional example, let H=C⊕L2([1,2])H=\mathbb{C}\oplus L^2([1,2]), let DD be multiplication by tt, and let vv be the constant function 11. In (31) take a=b=1a=b=1 and γ=1/2\gamma=1/2. The theorem gives a positive completion. This pair is irreducible. To see this, an operator commuting with BB preserves its simple eigenspace Cp\mathbb{C}p. On its orthogonal complement it commutes with multiplication by tt. Commutation with AA then requires that the lower block send the constant function to a scalar multiple of itself. Since 11 is cyclic for multiplication by tt, the operator is scalar on the lower block as well. Hence the joint commutant consists of scalars. The multiplication operator has continuous spectrum.

There are also operator-algebraic restrictions on these pairs. They hold for every pair in Theorem 5.1 in dimension at least two, whether finite or infinite.

Corollary 5.4. For a completed pair under the hypotheses of Theorem 5.1,

rank⁡(AB−BA)=2.\operatorname{rank}(AB-BA)=2.

If HH is infinite dimensional, then

C∗(A,B)=C∗(A)+K(H)\mathcal{C}^{*}(A,B)=\mathcal{C}^{*}(A)+\mathcal{K}(H)

and

σess(B)={b+γt−a:t∈σess(A)}.\sigma_{\mathrm{ess}}(B)=\left\{b+\frac{\gamma}{t-a}:t\in\sigma_{\mathrm{ess}}(A)\right\}.

Proof. From (24) and self-adjointness of C\mathcal{C},

AB−BA=γ(R∗−R).AB-BA=\gamma(R^{*}-R).

The vectors pp and qq are not proportional. Otherwise pp would be an eigenvector of AA, and cyclicity of pp would force H=CpH=\mathbb{C}p. Since q∗p≠0q^{*}p\ne0, the operator R∗−RR^{*}-R has rank two. This proves the first assertion.

Let GG be zero on Cq\mathbb{C}q and equal to (A−aI)−1(A-aI)^{-1} on q⊥q^{\perp}. Since aa is isolated, GG is bounded and belongs to C∗(A)\mathcal{C}^{*}(A). Multiplying (23) on the left by GG gives

(I−qq∗)(B−bI)=γG(I−R).(I-qq^{*})(B-bI)=\gamma G(I-R).

Consequently,

B−(bI+γG)=qq∗(B−bI)−γGR.B-(bI+\gamma G)=qq^{*}(B-bI)-\gamma GR.

The right side has rank at most two. Hence BB is a finite-rank perturbation of bI+γGbI+\gamma G and the essential-spectrum identity follows.

The pair acts irreducibly. Indeed, any operator commuting with both AA and BB preserves the one-dimensional eigenspace Cp\mathbb{C}p and acts on it as a scalar. Since pp is cyclic for AA, the operator is scalar on HH. Thus C∗(A,B)\mathcal{C}^{*}(A,B) is an irreducible C∗\mathcal{C}^{*}-algebra. It contains the nonzero finite-rank commutator AB−BAAB-BA. The standard theorem on irreducible C∗\mathcal{C}^{*}-algebras containing a nonzero compact operator gives K(H)⊆C∗(A,B)\mathcal{K}(H)\subseteq\mathcal{C}^{*}(A,B). The preceding finite-rank identity also gives B∈C∗(A)+K(H)B\in\mathcal{C}^{*}(A)+\mathcal{K}(H). Combining the two inclusions proves the asserted equality. □

Final remark

We end with two questions, one concerns the spectral structure of pairs admitting positive completions, the other is about the dilation of their representing measures.

  1. Theorem 5.1 assumes that the smallest spectral values of AA and BB are simple isolated eigenvalues. It is not clear whether an irreducible pair of strictly positive operators can admit a positive completion when neither operator has eigenvalues.

  1. A positive completion also gives a simultaneous dilation. Indeed, let EE be the representing POVM, then there exist a spectral measure PP on [0,1]2[0,1]^{2} and an isometry V:H⟶KV:H\longrightarrow K such that

E(U)=V∗P(U)V.E(U)=V^{*}P(U)V.

Set

S=∫[0,1]2s dP(s,t),T=∫[0,1]2t dP(s,t).S=\int_{[0,1]^2} s\,dP(s,t),\qquad T=\int_{[0,1]^2} t\,dP(s,t).

Then S,TS,T are commuting positive contractions and

(I+mA+nB+mnC)−1=V∗SmTnV,m,n≥0.(I+mA+nB+mnC)^{-1}=V^*S^mT^nV,\qquad m,n\geq0.

We may take

K=span⁡‾{SmTnVh:m,n≥0, h∈H}.K=\overline{\operatorname{span}}\{S^mT^nVh:m,n\geq0,\ h\in H\}.

This minimal dilation is unique up to unitary equivalence. It would be interesting to determine its spectral multiplicity and the structure of its spectral measure, particularly how these depend on the completion CC.

References

Mathematical Reviews, 535 W. William St, Suite 210, Ann Arbor, MI 48103, USA

Email address: [email protected]

References

  1. [1][ACN24] Akash Anand, Sameer Chavan, and Rajkamal Nailwal, Joint complete monotonicity of rational functions in two variables and toral m-isometric pairs, Journal of Operator Theory 92 (2024), no. 1, 101–130.arxiv.org/abs/2207.13903
  2. [2][AM14] Jim Agler and John E. McCarthy, Hankel vector moment sequences and the non-tangential regularity at infinity of two variable Pick functions, Transactions of the American Mathematical Society 366 (2014), no. 3, 1379–1411. MR 3145735arxiv.org/abs/1111.2075
  3. [3][AMY12] Jim Agler, John E. McCarthy, and Nicholas J. Young, Operator monotone functions and Löwner functions of several variables, Annals of Mathematics 176 (2012), no. 3, 1783–1826. MR 2979860
  4. [4][And63] Tsuyoshi Andō, On a pair of commutative contractions, Acta Scientiarum Mathematicarum (Szeged) 24 (1963), 88–90. MR 0155193
  5. [5][Aro50] N. Aronszajn, Theory of reproducing kernels, Transactions of the American Mathematical Society 68 (1950), 337–404. MR 0051437
  6. [6][Arv69] William B. Arveson, Subalgebras of C*-algebras, Acta Mathematica 123 (1969), 141–224. MR 0253059DOI
  7. [7][Ath90] Ameer Athavale, On the intertwining of joint isometries, Journal of Operator Theory 23 (1990), no. 2, 339–350. MR 1066811
  8. [8][BCR84] Christian Berg, Jens Peter Reus Christensen, and Paul Ressel, Harmonic analysis on semigroups: Theory of positive definite and related functions, Graduate Texts in Mathematics, vol. 100, Springer-Verlag, New York, 1984. MR 0747302
  9. [9][Bra55] Joseph Bram, Subnormal operators, Duke Mathematical Journal 22 (1955), 75–94. MR 0068129DOI
  10. [10][BT98] Joseph A. Ball and Tavan T. Trent, Unitary colligations, reproducing kernel Hilbert spaces, and Nevanlinna-Pick interpolation in several variables, Journal of Functional Analysis 157 (1998), no. 1, 1–61. MR 1637941DOI
  11. [11][CF98] Raúl E. Curto and Luis A. Fialkow, Flat extensions of positive moment matrices: Recursively generated relations, Memoirs of the American Mathematical Society 136 (1998), no. 648, x+56. MR 1445490DOI
  12. [12][Cho75] Man Duen Choi, Completely positive linear maps on complex matrices, Linear Algebra and its Applications 10 (1975), 285–290. MR 0376726DOI
  13. [13][Con91] John B. Conway, The theory of subnormal operators, Mathematical Surveys and Monographs, vol. 36, American Mathematical Society, Providence, RI, 1991. MR 1112128DOI
  14. [14][FK94] Jacques Faraut and Adam Korányi, Analysis on symmetric cones, Oxford Mathematical Monographs, The Clarendon Press, Oxford University Press, New York, 1994. MR 1446489DOI
  15. [15][Gin75] S. G. Gindikin, Invariant generalized functions in homogeneous domains, Funktsional’ nyi Analiz i ego Prilozheniya 9 (1975), no. 1, 56–58, In Russian. MR 0377423DOI
  16. [17][Hau23] Felix Hausdorff, Momentprobleme für ein endliches intervall, Mathematische Zeitschrift 16 (1923), no. 1, 220–248. MR 1544592
  17. [18]Jean Bernard Lasserre, Moments, positive polynomials and their applications, Imperial College Press Optimization Series, vol. 1, Imperial College Press, London, 2010. MR 2589247
  18. [19]Murray Marshall, Positive polynomials and sums of squares, Mathematical Surveys and Monographs, vol. 146, American Mathematical Society, Providence, RI, 2008. MR 2383959
  19. [21]Vern I. Paulsen and Mrinal Raghupathi, An introduction to the theory of reproducing kernel hilbert spaces, Cambridge Studies in Advanced Mathematics, vol. 152, Cambridge University Press, Cambridge, 2016. MR 3526117DOI
  20. [22]Mihai Putinar, Positive polynomials on compact semi-algebraic sets, Indiana University Mathematics Journal 42 (1993), no. 3, 969–984. MR 1254128
  21. [23]Mihai Putinar and Florian-Horia Vasilescu, Solving moment problems by dimensional extension, Annals of Mathematics 149 (1999), no. 3, 1087–1107. MR 1709313
  22. [24]Konrad Schmüdgen, The K-moment problem for compact semi-algebraic sets, Mathematische Annalen 289 (1991), 203–206. MR 1092173
  23. [25]Alexander D. Scott and Alan D. Sokal, Complete monotonicity for inverse powers of some combinatorially defined polynomials, Acta Mathematica 213 (2014), no. 2, 323–392.DOI
  24. [26]J. A. Shohat and J. D. Tamarkin, The problem of moments, Mathematical Surveys, vol. 1, American Mathematical Society, New York, 1943. MR 0008438DOI
  25. [27]W. Forrest Stinespring, Positive functions on C*-algebras, Proceedings of the American Mathematical Society 6 (1955), 211–216. MR 0069403
  26. [28]Gábor Szegő, Über gewisse potenzreihen mit lauter positiven koeffizienten, Mathematische Zeitschrift 37 (1933), no. 1, 674–688. MR 1545428DOI
  27. [29]Florian-Horia Vasilescu, Hamburger and Stieltjes moment problems in several variables, Transactions of the American Mathematical Society 354 (2002), no. 3, 1265–1278. MR 1867381DOI

Paper details

vlnc:W_525wf3t9
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

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

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents