betaFeedback

WorksMathematics

Transfer of similarity to Hilbert space contractions through inner functions: a proof of Douglas's conjecture

Abstract

Let AA be a bounded operator on a Hilbert space and let uu be a non-constant Blaschke product, finite or infinite. We prove that if the spectrum of AA does not meet the closure of the set of poles of uu, then u(A)u(A) is similar to a contraction if and only if AA is. This settles a conjecture of R.G. Douglas. For infinite products uu, no additional condition on the zeros beyond the Blaschke condition is required. We also show that 'u(A)u(A) is similar to a contraction if and only if AA is' for all non-constant inner functions uu whose reflected singularity set, the closure of the set of poles together with the singular support, avoids the spectrum of AA. We also prove a weighted resolvent version of the theorem, which allows the poles to accumulate on the spectrum to a limited extent; a column criterion for partial products, which is shown to force their norm convergence; and a simultaneous version for commuting tuples of operators tested by separated inner functions sharing one Lyapunov operator. For finite Blaschke products we obtain a polynomial form of the Lyapunov operator, which gives an estimate for the similarity constants. We also characterise the tuples of finite Blaschke products that form a uniform strong test collection: they are exactly the tuples that are injective, with non-vanishing joint derivative, on the closed disc. Finally, we show by examples that the separation hypothesis cannot be replaced by the existence of a bounded H∞H^\infty calculus, by the existence of a norm radial limit, or by the strong convergence of the partial products, even for Ritt operators.

2020 Mathematics Subject Classification. Primary 47A60; Secondary 47A65, 30J10, 30J15.

Introduction and preliminaries

Operators similar to contractions

Let A∈B(H)A \in\mathcal{B}(\mathcal{H}) be a bounded linear operator on a complex Hilbert space H\mathcal{H}. We say that AA is a contraction if ∥A∥≤1\lVert A\rVert\le1, and that AA is similar to a contraction if there is an invertible operator L∈B(H)L \in\mathcal{B}(\mathcal{H}) such that ∥L−1AL∥≤1\lVert L^{-1}AL\rVert\le1. Equivalently, AA is similar to a contraction if and only if there is an equivalent Hilbertian norm on H\mathcal{H}, that is, a norm ∥x∥H=⟨Hx,x⟩1/2\lVert x\rVert_H = \langle Hx,x\rangle^{1/2} with HH a coercive positive operator, with respect to which AA is a contraction. We then call AA an HH-contraction and HH a Lyapunov operator for AA (see § 1.e and 1.f). By von Neumann’s inequality, an operator similar to a contraction is polynomially bounded, that is,

sup⁡{∥p(A)∥:p∈C[z], ∥p∥∞≤1}<∞,(1)\sup\{\lVert p(A)\rVert: p \in\mathbb{C}[z],\ \lVert p\rVert_{\infty} \le1\}<\infty, \tag*{(1)}

where the supremum is taken over all polynomials pp whose supremum norm ∥p∥∞:=sup⁡{∣p(z)∣:∣z∣≤1}\|p\|_{\infty} := \sup\{|p(z)| : |z| \le1\} is at most one. The estimate (1) then extends to the disc algebra. Whether the converse holds, that is, whether every polynomially bounded operator is similar to a contraction, was the sixth of the ten problems posed by Halmos [4]; Pisier [12] showed in 1997 that the answer is negative. On the other hand, Paulsen [11] proved that AA is similar to a contraction if and only if AA is completely polynomially bounded, which means that (1) holds for all matrix-valued (equivalently, operator-valued) polynomials PP, with ∥P∥∞\|P\|_{\infty} now denoting the supremum of the operator norms ∥P(z)∥\|P(z)\| over the closed unit disc.

Douglas's conjecture

Let uu be a Blaschke product whose poles have closure disjoint from the spectrum of AA. Then u(A)u(A) is defined by the holomorphic functional calculus. If AA is similar to a contraction, then so is u(A)u(A): it suffices to apply von Neumann’s inequality in an HH-norm in which AA is a contraction. This paper is concerned with the converse implication,

u(A) similar to a contraction⟹A similar to a contraction,(2)u(A)\ \text{similar to a contraction} \quad\Longrightarrow\quad A\ \text{similar to a contraction}, \tag*{(2)}

and with its analogue for more general inner functions uu. As recorded in several papers, among them [6], [14] and [5], Ronald G. Douglas conjectured in the 1990s that the implication (2) holds for every infinite Blaschke product. For powers, u(z)=znu(z)=z^n, the implication is elementary and has been noted by several authors; see for instance [4 p. 912]. For finite Blaschke products it was proved by Mascioni [9], who also treated the more general setting of Banach space operators related to contractions; see also [13]. Two further proofs of Mascioni’s theorem in the Hilbert space case were given by Lance and Stessin [6], using a Wold-type decomposition, and by Dritschel, Estévez and Yakubovich [1 Section 3.5]. For an infinite Blaschke product with zeros aja_j, Stessin [14] proved (2) under the additional hypothesis ∑j1−∣aj∣2<∞\sum_j\sqrt{1-|a_j|^2}<\infty, which is strictly stronger than the Blaschke condition. The hypothesis is an artefact of the method: the Wold reconstruction series in [14] is estimated by the sum of the norms of its coefficient operators, and that sum converges only under the square-root summability condition above. Kazas and Kelley [5] used Clark measures to enlarge the class of inner functions for which the transfer of similarity is known.

Our contribution

The aim of this paper is to prove Douglas’s conjecture: the implication (2) holds for every Blaschke product uu and every operator AA whose spectrum avoids the closure of the set of poles of uu, and, more generally, for every inner function uu whose reflected singularity set Sp⁡(u)\operatorname{Sp}(u), defined in (5) below, avoids the spectrum of AA. The proof produces, for every Lyapunov operator XX for u(A)u(A), an explicit coercive operator YY satisfying the exact defect identity

Y−A∗YA=X−u(A)∗Xu(A),Y-A^{*}YA=X-u(A)^{*}Xu(A),

so that YY is a Lyapunov operator for AA. Consequently, every HH-norm in which u(A)u(A) is a contraction yields an HH-norm in which AA is one. For Blaschke products the operator YY is a norm-convergent series of positive operators; for the singular inner factor it is given by an integral identity of the same kind. We also establish three variants of the main theorem: a weighted resolvent version, which permits the poles to accumulate on the spectrum to a limited extent; a column criterion for the partial products, which we show forces their norm convergence; and a simultaneous version for commuting tuples of operators tested by separated inner functions sharing one Lyapunov operator. For finite Blaschke products we derive a polynomial form of the Lyapunov operator, leading to the estimate sim⁡(A)≤CBsim⁡(B(A))max⁡k<m∥Ak∥\operatorname{sim}(A) \le C_B \operatorname{sim}(B(A)) \max_{k<m} \lVert A^k\rVert for the similarity constants, and we characterise the tuples of finite Blaschke products that form a uniform strong test collection as exactly the tuples that are injective with non-vanishing joint derivative on the closed disc. To show that the separation hypothesis cannot simply be dropped, three examples are given: a polynomially bounded operator, not similar to a contraction, that is annihilated by a Blaschke product; an operator that is not power bounded although B(rA)→0B(rA) \to0 in norm as r↑1r \uparrow1; and a Ritt operator, not polynomially bounded, whose Blaschke partial products converge strongly to zero.

Organisation of the paper

The remainder of this section fixes the notation, recalls the holomorphic functional calculus for inner functions, and collects two preliminary results. § 2 proves the transfer theorem for Blaschke products, and § 3 treats the singular inner factor, so that every inner function whose reflected singularity set avoids the spectrum is covered. § 4 contains the three variants of the transfer theorem mentioned above. § 5 contains the examples showing that several natural substitutes for the separation hypothesis are insufficient. The last two sections concern finite Blaschke products: § 6 gives the polynomial form of the Lyapunov operator, and § 7 the uniform strong test theorem and its exact characterisation.

Notation and the functional calculus for inner functions

Throughout, D\mathbb{D} is the open unit disc, T\mathbb{T} the unit circle, and D‾=D∪T\overline{\mathbb{D}} = \mathbb{D} \cup\mathbb{T}. For a∈Da \in\mathbb{D} we put

ba(z)={∣a∣aa−z1−a‾z,a≠0,z,a=0,(3)b_a(z) = \begin{cases} \frac{|a|}{a}\frac{a-z}{1-\overline{a}z}, & a \ne0,\\ z, & a = 0, \end{cases} \tag*{(3)}

so that ∣ba∣=1|b_a| = 1 on T\mathbb{T} and ba(a)=0b_a(a) = 0. A Blaschke product is a product η∏jbaj\eta\prod_j b_{a_j}, with ∣η∣=1|\eta| = 1, over a finite or infinite sequence of zeros satisfying the Blaschke condition ∑j(1−∣aj∣)<∞\sum_j (1-|a_j|) < \infty; it is an inner function on D\mathbb{D}. For a finite positive singular Borel measure μ\mu on T\mathbb{T},

Sμ(z)=exp⁡(−∫T1+ζ‾z1−ζ‾z dμ(ζ))(4)S_\mu(z) = \exp\left(-\int_{\mathbb{T}} \frac{1+\overline{\zeta}z}{1-\overline{\zeta}z}\,d\mu(\zeta)\right) \tag*{(4)}

is the associated singular inner function; the integral defines a holomorphic function on C∖supp⁡μ\mathbb{C} \setminus\operatorname{supp}\mu. Every inner function factorises as θ=ηBSμ\theta= \eta B S_\mu with ∣η∣=1|\eta| = 1, BB a Blaschke product and μ\mu singular. Following [5], we call

Sp⁡(θ)={1/aj‾:aj≠0}∪supp⁡μ(5)\operatorname{Sp}(\theta) = \{1/\overline{a_j}: a_j \ne0\} \cup\operatorname{supp}\mu \tag*{(5)}

the reflected singularity set of θ\theta; it consists of the poles 1/aj‾1/\overline{a_j} of BB, their accumulation points on T\mathbb{T}, and the support of μ\mu, and θ\theta extends holomorphically to C∖Sp⁡(θ)\mathbb{C} \setminus\operatorname{Sp}(\theta). The notation Sp⁡(θ)\operatorname{Sp}(\theta) is reserved for this set, the spectrum of an operator AA being denoted by σ(A)\sigma(A).

We use the notation B(H)\mathcal{B}(\mathcal{H}) for the algebra of bounded linear operators on the Hilbert space H\mathcal{H}, and R(λ,A)=(λI−A)−1R(\lambda,A) = (\lambda I-A)^{-1} for the resolvent. The similarity constant of AA is

sim⁡(A)=inf⁡{∥X∥∥X−1∥:X invertible, ∥XAX−1∥≤1},(6)\operatorname{sim}(A) = \inf\{\lVert X\rVert\lVert X^{-1}\rVert: X\ \text{invertible},\ \lVert XAX^{-1}\rVert\le1\}, \tag*{(6)}

with the value +∞+\infty when AA is not similar to a contraction. Note that sup⁡n∥An∥≤sim⁡(A)\sup_n \lVert A^n\rVert\le\operatorname{sim}(A). A bounded self-adjoint operator HH is coercive if H≥cIH \ge cI for some c>0c>0; a coercive HH defines the equivalent Hilbertian norm ∥x∥H=⟨Hx,x⟩1/2\|x\|_{H}=\langle Hx,x\rangle^{1/2}, called the HH-norm. An operator SS is an HH-contraction if ∥Sx∥H≤∥x∥H\|Sx\|_{H}\leq\|x\|_{H} for all xx, that is, if S∗HS≤HS^{*}HS\leq H; in this case we call HH a Lyapunov operator for SS. For a Blaschke product BB with zeros (aj)(a_j), listed with multiplicity, and an operator AA for which the factors I−aj‾AI-\overline{a_j}A are invertible, we write

P0=I,PN=∏j=1Nbaj(A),Ej=1−∣aj∣2(I−aj‾A)−1Pj−1.(7)P_0=I,\qquad P_N=\prod_{j=1}^{N}b_{a_j}(A),\qquad E_j=\sqrt{1-|a_j|^2}(I-\overline{a_j}A)^{-1}P_{j-1}. \tag*{(7)}

for the partial products of BB at AA and for the factors of the Lyapunov operators constructed below.

We recall how θ(A)\theta(A) is defined for an inner function θ\theta whose reflected singularity set avoids σ(A)\sigma(A); compare [6 Section 5]. Let ff be holomorphic on an open neighbourhood UU of σ(A)\sigma(A), and let VV be an open neighbourhood of σ(A)\sigma(A) with compact closure in UU whose boundary Γ=∂V\Gamma=\partial V is a finite union of smooth Jordan curves. The Riesz–Dunford functional calculus defines

f(A)=12πi∫Γf(z)R(z,A) dz.(8)f(A)=\frac{1}{2\pi i}\int_{\Gamma}f(z)R(z,A)\,dz. \tag*{(8)}

Let θ=ηBSμ\theta=\eta BS_{\mu} be an inner function and suppose that

σ(A)∩Sp⁡(θ)=∅.(9)\sigma(A)\cap\operatorname{Sp}(\theta)=\varnothing. \tag*{(9)}

The Blaschke product converges locally uniformly on C∖Sp⁡(B)\mathbb{C}\setminus\operatorname{Sp}(B), where it is holomorphic, and (4) is holomorphic on C∖supp⁡μ\mathbb{C}\setminus\operatorname{supp}\mu, so θ\theta is holomorphic on the open set C∖Sp⁡(θ)⊃σ(A)\mathbb{C}\setminus\operatorname{Sp}(\theta)\supset\sigma(A) and (8) defines θ(A)=ηB(A)Sμ(A)\theta(A)=\eta B(A)S_{\mu}(A). It is easy to see that the product

B(A)=η∏jbaj(A),baj(A)=∣aj∣aj(ajI−A)(I−aj‾A)−1(aj≠0),(10)B(A)=\eta\prod_j b_{a_j}(A),\qquad b_{a_j}(A)=\frac{|a_j|}{a_j}(a_jI-A)(I-\overline{a_j}A)^{-1}\quad(a_j\ne0), \tag*{(10)}

converges in operator norm; this is proved in Lemma 1.2 below. With this notation, the problem raised by Douglas reads as follows: let θ\theta be an inner function satisfying (9); if θ(A)\theta(A) is similar to a contraction, is AA similar to a contraction? This is the form in which the problem is stated in [6 Section 5]; Theorem 3.2 answers it in the affirmative.

Finally, for a finite non-empty set E⊂TE\subset\mathbb{T} we call AA a RittE_E operator if σ(A)⊂D‾\sigma(A)\subset\overline{\mathbb{D}} and ∥R(λ,A)∥≤C/dist⁡(λ,E)\|R(\lambda,A)\|\leq C/\operatorname{dist}(\lambda,E) for some C<∞C<\infty and all ∣λ∣>1|\lambda|>1. A Ritt operator is an operator with σ(A)⊂D‾\sigma(A)\subset\overline{\mathbb{D}} and ∥R(λ,A)∥≤C/∣λ−1∣\|R(\lambda,A)\|\leq C/|\lambda-1| for ∣λ∣>1|\lambda|>1; equivalently [10, 8], AA is power bounded and sup⁡kk∥Ak−1(I−A)∥<∞\sup_k k\|A^{k-1}(I-A)\|<\infty.

Two preliminary results

The first result is the standard reformulation of similarity to a contraction in terms of Lyapunov operators; it is the form in which similarity will be established throughout the paper.

Proposition 1.1 (Lyapunov criterion). For S∈B(H)S\in\mathcal{B}(\mathcal{H}) the following are equivalent:

  1. SS is similar to a contraction;

  2. SS is an HH-contraction for some coercive HH, that is, SS admits a Lyapunov operator.

If H≥cIH\geq cI with c>0c>0 and S∗HS≤HS^{*}HS\leq H, then H1/2SH−1/2H^{1/2}SH^{-1/2} is a contraction and sim⁡(S)≤∥H∥1/2c−1/2\operatorname{sim}(S)\leq\|H\|^{1/2}c^{-1/2}; conversely, if XSX−1XSX^{-1} is a contraction, then H=X∗XH=X^{*}X is a Lyapunov operator for SS. Moreover, if sim⁡(S)<∞\operatorname{sim}(S)<\infty, then the infimum in (6) is attained.

Proof. The equivalence follows by conjugating the inequality S∗HS≤HS^*HS \le H by H−1/2H^{-1/2}. For the attainment of the infimum, let K=sim⁡(S)K = \operatorname{sim}(S) and choose Lyapunov operators HnH_n for SS, scaled so that I≤Hn≤(K+1/n)2II \le H_n \le(K + 1/n)^2I. Any ultraweak cluster point HH of the bounded sequence (Hn)(H_n) satisfies I≤H≤K2II \le H \le K^2I and S∗HS≤HS^*HS \le H, because these inequalities define ultraweakly closed sets, and the similarity H1/2SH−1/2H^{1/2}SH^{-1/2} has condition number at most KK.

Everything in § 2–§ 4 rests on the following consequence of the separation hypothesis. Note that the lemma does not assume σ(A)⊂D‾\sigma(A) \subset\overline{\mathbb{D}}: in the theorems below this inclusion is a consequence of the conclusion.

Lemma 1.2 (Norm convergence and uniform bounds). Let B=η∏j∈JbajB = \eta\prod_{j\in J}b_{a_j} be a non-constant Blaschke product with J={1,…,m}J = \{1,\ldots,m\} or J=NJ = \mathbb{N}, its zeros being listed with multiplicity and the finitely many zeros at the origin coming first, and suppose that

σ(A)∩Sp⁡(B)=∅.(11)\sigma(A) \cap\operatorname{Sp}(B) = \varnothing. \tag*{(11)}

Then every I−aj‾AI-\overline{a_j}A is invertible and, with PNP_N as in (7),

L:=sup⁡j∥(I−aj‾A)−1∥<∞,P:=sup⁡N∥PN∥<∞,β:=∑j(1−∣aj∣2)<∞.(12)L := \sup_j \left\|(I-\overline{a_j}A)^{-1}\right\| < \infty,\qquad P := \sup_N \left\|P_N\right\| < \infty,\qquad\beta:= \sum_j(1-|a_j|^2) < \infty. \tag*{(12)}

In the infinite case the sequence (PN)(P_N) converges in operator norm, so that B(A)=ηlim⁡NPNB(A)=\eta\lim_{N}P_N is well defined, and this limit coincides with the operator given by the functional calculus ((8)).

Proof. The set {aj}⊂D‾\{a_j\} \subset\overline{\mathbb{D}} is compact, and for every aa in it the operator I−a‾AI-\overline{a}A is invertible: for a=0a=0 this is clear, and for a≠0a\ne0 the point 1/a‾1/\overline{a} lies in Sp⁡(B)\operatorname{Sp}(B), hence outside σ(A)\sigma(A). Since inversion is continuous on the group of invertible operators, L<∞L<\infty.

For a=reiθ≠0a=re^{i\theta}\ne0 one has the identity

1−ba(z)=(1−r)(1+e−iθz)(1−a‾z)−1,(13)1-b_a(z)=(1-r)(1+e^{-i\theta}z)(1-\overline{a}z)^{-1}, \tag*{(13)}

which is obtained by writing the left-hand side over the common denominator 1−a‾z1-\overline{a}z and using a‾=re−iθ\overline{a}=re^{-i\theta}. Hence

∑j:aj≠0∥I−baj(A)∥≤(1+∥A∥)L∑j:aj≠0(1−∣aj∣)<∞.\sum_{j:a_j\ne0}\left\|I-b_{a_j}(A)\right\|\le(1+\|A\|)L\sum_{j:a_j\ne0}(1-|a_j|)<\infty.

The zeros at the origin contribute a finite product. If FjF_j denotes the remaining factors, then ∑j∥Fj−I∥<∞\sum_j\|F_j-I\|<\infty and

∥∏j=N+1MFj−I∥≤exp⁡(∑j=N+1M∥Fj−I∥)−1,\left\|\prod_{j=N+1}^{M}F_j-I\right\|\le\exp\left(\sum_{j=N+1}^{M}\|F_j-I\|\right)-1,

so the partial products form a Cauchy sequence and are bounded. Writing qq for the number of zeros at the origin, we obtain

P≤max⁡{1,∥A∥}qexp⁡((1+∥A∥)L∑j:aj≠0(1−∣aj∣)).P\le\max\{1,\|A\|\}^{q}\exp\left((1+\|A\|)L\sum_{j:a_j\ne0}(1-|a_j|)\right).

Finally, 1−∣aj∣2≤2(1−∣aj∣)1-|a_j|^2\le2(1-|a_j|), so β<∞\beta<\infty. The same estimates hold uniformly on compact subsets of the complement of Sp⁡(B)\operatorname{Sp}(B), so the scalar product converges locally uniformly to a holomorphic function there, and evaluating it at AA by means of a fixed contour around σ(A)\sigma(A) shows that the norm limit of (PN)(P_N) agrees with the functional calculus.

The transfer theorem for Blaschke products

Throughout this section B=η∏j∈JbajB=\eta\prod_{j\in J}b_{a_j} is a non-constant Blaschke product, finite or infinite, and AA satisfies the separation hypothesis (11). As in (7), we put

Ej=1−∣aj∣2(I−aj‾A)−1Pj−1,j∈J,(14)E_j=\sqrt{1-|a_j|^2}(I-\overline{a_j}A)^{-1}P_{j-1},\qquad j\in J, \tag*{(14)}

where Pj−1=∏k<jbak(A)P_{j-1}=\prod_{k<j}b_{a_k}(A) and P0=IP_0=I. All the operators PjP_j and EjE_j are rational functions of AA, and therefore commute with AA and with one another.

Theorem 2.1 (Transfer for every Blaschke product). Let BB and AA be as above, so that (11) holds. Then

B(A) is similar to a contraction⟺A is similar to a contraction.(15)B(A)\text{ is similar to a contraction}\Longleftrightarrow A\text{ is similar to a contraction}. \tag*{(15)}

More precisely, if HH is a Lyapunov operator for B(A)B(A), then the series

Q=∑j∈JEj∗HEj(16)Q=\sum_{j\in J}E_j^*HE_j \tag*{(16)}

converges in operator norm, its sum is a Lyapunov operator for AA, and

Q−A∗QA=H−B(A)∗HB(A).(17)Q-A^*QA=H-B(A)^*HB(A). \tag*{(17)}

Proof. Convergence and coercivity. By Lemma 1.2,

∑j∈J∥Ej∥2≤P2L2∑j(1−∣aj∣2)=P2L2β<∞,(18)\sum_{j\in J}\|E_j\|^2\le P^2L^2\sum_j(1-|a_j|^2)=P^2L^2\beta<\infty, \tag*{(18)}

so the series of positive operators (16) converges in norm, with the tail estimate ∥Q−QN∥≤∥H∥P2L2∑j>N(1−∣aj∣2)\|Q-Q_N\|\le\|H\|P^2L^2\sum_{j>N}(1-|a_j|^2) for the partial sums QN=∑j≤NEj∗HEjQ_N=\sum_{j\le N}E_j^*HE_j, and

0≤Q≤∥H∥P2L2βI.(19)0\le Q\le\|H\|P^2L^2\beta I. \tag*{(19)}

The first term of the series already gives coercivity, since E1=1−∣a1∣2(I−a1‾A)−1E_1=\sqrt{1-|a_1|^2}(I-\overline{a_1}A)^{-1} is invertible:

Q≥E1∗HE1≥1−∣a1∣2∥H−1∥∥I−a1‾A∥2I>0.(20)Q\ge E_1^*HE_1\ge\frac{1-|a_1|^2}{\|H^{-1}\|\|I-\overline{a_1}A\|^2}I>0. \tag*{(20)}

The identity. Let ∣a∣<1|a|<1 and let HH be self-adjoint. Put c=I−a‾Ac=I-\overline{a}A and b(z)=(z−a)/(1−a‾z)b(z)=(z-a)/(1-\overline{a}z), a unimodular multiple of bab_a. Since b(A)c=cb(A)=A−aIb(A)c=cb(A)=A-aI, all the factors being rational functions of AA,

c∗(H−b(A)∗Hb(A))c=c∗Hc−(b(A)c)∗H(b(A)c)=(I−aA∗)H(I−a‾A)−(A∗−a‾I)H(A−aI)=(1−∣a∣2)(H−A∗HA).(21)\begin{aligned} c^*(H-b(A)^*Hb(A))c &=c^*Hc-(b(A)c)^*H(b(A)c)\\ &=(I-aA^*)H(I-\overline{a}A)-(A^*-\overline{a}I)H(A-aI)\\ &=(1-|a|^2)(H-A^*HA). \tag*{(21)} \end{aligned}

Since a unimodular factor does not affect ba(A)∗Hba(A)b_a(A)^*Hb_a(A), this can be written as

H−ba(A)∗Hba(A)=(1−∣a∣2)(I−a‾A)−∗(H−A∗HA)(I−a‾A)−1.(22)H-b_a(A)^*Hb_a(A)=(1-|a|^2)(I-\overline{a}A)^{-*}(H-A^*HA)(I-\overline{a}A)^{-1}. \tag*{(22)}

Since Pj=baj(A)Pj−1P_j=b_{a_j}(A)P_{j-1} and all the factors commute, telescoping gives, for every NN,

H−PN∗HPN=∑j=1NPj−1∗(H−baj(A)∗Hbaj(A))Pj−1,H-P_N^*HP_N=\sum_{j=1}^{N}P_{j-1}^*\bigl(H-b_{a_j}(A)^*Hb_{a_j}(A)\bigr)P_{j-1},

and inserting (22) we obtain

H−PN∗HPN=∑j=1NEj∗(H−A∗HA)Ej=QN−A∗QNA,(23)H-P_N^*HP_N=\sum_{j=1}^{N}E_j^*(H-A^*HA)E_j=Q_N-A^*Q_NA, \tag*{(23)}

the last equality because each EjE_j commutes with AA. Both sides converge in norm as N→∞N\to\infty: PN→ηˉB(A)P_N\to\bar{\eta}B(A) by Lemma 1.2, and QN→QQ_N\to Q by (18). This proves (17).

Conclusion. The right-hand side of (17) is non-negative by hypothesis, so A∗QA≤QA^*QA\le Q, and Q1/2AQ−1/2Q^{1/2}AQ^{-1/2} is a contraction by Proposition 1.1.

Converse. If XAX−1=CXAX^{-1}=C is a contraction, then PN(C)=XPN(A)X−1P_N(C)=XP_N(A)X^{-1} is a contraction for every NN by von Neumann’s inequality, and ηˉXB(A)X−1=lim⁡NPN(C)\bar{\eta}XB(A)X^{-1}=\lim_{N}P_N(C) in norm by Lemma 1.2; hence XB(A)X−1XB(A)X^{-1} is a contraction.

Remark 2.2 (Why square-root summability is not needed). Estimating the terms of (14) individually gives ∥Ej∥≤PL1−∣aj∣2\lVert E_j\rVert\le PL\sqrt{1-|a_j|^2}, and a bound for ∑j∥Ej∥\sum_j\lVert E_j\rVert requires ∑j1−∣aj∣2<∞\sum_j\sqrt{1-|a_j|^2}<\infty. The series (16), however, is a series of positive operators, whose norm is controlled by ∑j∥Ej∥2\sum_j\lVert E_j\rVert^2, and this costs only ∑j(1−∣aj∣2)<∞\sum_j(1-|a_j|^2)<\infty, which is the Blaschke condition. The passage from this estimate to the operator identity is the finite telescoping identity (23), which involves no infinite sums at all. This is what removes the square-root summability hypothesis of [14].

Corollary 2.3 (Quantitative similarity bound). In the notation of Theorem 2.1, the similarity Q1/2Q^{1/2} satisfies

∥Q1/2∥∥Q−1/2∥≤∥H∥∥H−1∥ PL∥I−aˉ1A∥β1−∣a1∣2.(24)\lVert Q^{1/2}\rVert\lVert Q^{-1/2}\rVert\le\sqrt{\lVert H\rVert\lVert H^{-1}\rVert}\,PL\lVert I-\bar{a}_1A\rVert\sqrt{\frac{\beta}{1-|a_1|^2}}. \tag*{(24)}

If a1=0a_1=0, then E1=IE_1=I, so Q≥HQ\ge H, and the bound reads ∥H∥∥H−1∥ PLβ\sqrt{\lVert H\rVert\lVert H^{-1}\rVert}\,PL\sqrt{\beta}.

Proof. Combine (19) with (20); when a1=0a_1=0 one has E1=P0=IE_1=P_0=I.

Remark 2.4. The lower bound (20) depends on the first zero only, whereas P=sup⁡N∥PN∥P=\sup_N\lVert P_N\rVert may depend on the order in which the zeros are listed; the exponential bound for PP in Lemma 1.2, however, does not. Listing a zero at the origin first removes the factor ∥I−aˉ1A∥\lVert I-\bar{a}_1A\rVert from (24). The constants PP and LL depend on AA. For finite products, a bound in terms of sim⁡(B(A))\operatorname{sim}(B(A)) and finitely many powers of AA is given in Section 6, and the dependence on the powers cannot be removed (Remark 6.3).

Remark 2.5 (On the hypotheses). The assumption that BB is non-constant cannot be dropped: a unimodular constant BB gives B(A)=ηIB(A)=\eta I for every AA. The separation hypothesis (11) cannot be replaced by the mere invertibility of the factors I−aˉjAI-\bar{a}_jA, as Section 5 shows; it can, however, be weakened to a weighted resolvent condition, see Section 4.a.

Singular inner factors

Let μ\mu be a finite positive singular Borel measure on T\mathbb{T} and let SμS_\mu be the singular inner function (4). The separation hypothesis now reads

σ(A)∩supp⁡μ=∅.(25)\sigma(A)\cap\operatorname{supp}\mu=\varnothing. \tag*{(25)}

For ζ∈supp⁡μ\zeta\in\operatorname{supp}\mu put Dζ=(I−ζˉA)−1D_\zeta=(I-\bar{\zeta}A)^{-1}, and define

F=Fμ(A)=∫T(I+ζˉA)Dζ dμ(ζ),Es=e−sF(0≤s≤1),(26)F=F_\mu(A)=\int_{\mathbb{T}}(I+\bar{\zeta}A)D_\zeta\,d\mu(\zeta),\qquad E_s=e^{-sF}\quad(0\le s\le1), \tag*{(26)}

all integrals being Bochner integrals in the operator norm. By (25) and the compactness of supp⁡μ\operatorname{supp}\mu,

Lμ:=sup⁡ζ∈supp⁡μ∥Dζ∥<∞(27)L_{\mu}:=\sup_{\zeta\in\operatorname{supp}\mu}\|D_{\zeta}\|<\infty \tag*{(27)}

when μ≠0\mu\ne0, so FF and E1=e−FE_{1}=e^{-F} are well defined. To identify E1E_{1} with the operator given by the functional calculus, put Λ(z)=−∫T1+ζ‾z1−ζ‾z dμ(ζ)\Lambda(z)=-\int_{\mathbb{T}}\frac{1+\overline{\zeta}z}{1-\overline{\zeta}z}\,d\mu(\zeta), which is holomorphic on C∖supp⁡μ\mathbb{C}\setminus\operatorname{supp}\mu and satisfies Sμ=eΛS_{\mu}=e^{\Lambda} there. Since σ(A)\sigma(A) is a compact subset of C∖supp⁡μ\mathbb{C}\setminus\operatorname{supp}\mu and the functional calculus commutes with norm-convergent Bochner integrals, Λ(A)=−F\Lambda(A)=-F, and hence Sμ(A)=e−F=E1S_{\mu}(A)=e^{-F}=E_{1}. When μ=0\mu=0 we set F=0F=0 and Sμ(A)=IS_{\mu}(A)=I.

Lemma 3.1 (A Lyapunov operator for the singular factor). Assume (25) and let H=H∗H=H^{*} be bounded. Put

Qμ(H)=2∫01∫TDζ∗Es∗HEsDζ dμ(ζ) ds.(28)Q_{\mu}(H)=2\int_{0}^{1}\int_{\mathbb{T}}D_{\zeta}^{*}E_{s}^{*}HE_{s}D_{\zeta}\,d\mu(\zeta)\,ds. \tag*{(28)}

Then

Qμ(H)−A∗Qμ(H)A=H−Sμ(A)∗HSμ(A).(29)Q_{\mu}(H)-A^{*}Q_{\mu}(H)A=H-S_{\mu}(A)^{*}HS_{\mu}(A). \tag*{(29)}

If μ≠0\mu\ne0 and H≥hIH\ge hI with h>0h>0, then Qμ(H)Q_{\mu}(H) is coercive; more precisely, with t=μ(T)t=\mu(\mathbb{T}) and f=∥F∥f=\|F\|,

2the−2f(1+∥A∥)2I≤Qμ(H)≤2tLμ2e2f∥H∥I.(30)\frac{2t h e^{-2f}}{(1+\|A\|)^{2}}I\le Q_{\mu}(H)\le2tL_{\mu}^{2}e^{2f}\|H\|I. \tag*{(30)}

Proof. For ζ∈supp⁡μ\zeta\in\operatorname{supp}\mu put Fζ=(I+ζ‾A)DζF_{\zeta}=(I+\overline{\zeta}A)D_{\zeta}. Multiplying Fζ∗H+HFζF_{\zeta}^{*}H+HF_{\zeta} on the left by I−ζA∗I-\zeta A^{*} and on the right by I−ζ‾AI-\overline{\zeta}A gives, since ∣ζ∣=1|\zeta|=1,

(I+ζA∗)H(I−ζ‾A)+(I−ζA∗)H(I+ζ‾A)=2(H−A∗HA),(I+\zeta A^{*})H(I-\overline{\zeta}A)+(I-\zeta A^{*})H(I+\overline{\zeta}A)=2(H-A^{*}HA),

that is,

Fζ∗H+HFζ=2Dζ∗(H−A∗HA)Dζ.(31)F_{\zeta}^{*}H+HF_{\zeta}=2D_{\zeta}^{*}(H-A^{*}HA)D_{\zeta}. \tag*{(31)}

Integrating (31) against μ\mu and using the identity dds(Es∗HEs)=−Es∗(F∗H+HF)Es\frac{d}{ds}(E_{s}^{*}HE_{s})=-E_{s}^{*}(F^{*}H+HF)E_{s}, which follows from Es′=−FEsE_{s}'=-FE_{s} and the fact that FF commutes with EsE_{s}, we obtain

H−E1∗HE1=2∫01∫T(EsDζ)∗(H−A∗HA)(EsDζ) dμ(ζ) ds.H-E_{1}^{*}HE_{1}=2\int_{0}^{1}\int_{\mathbb{T}}(E_{s}D_{\zeta})^{*}(H-A^{*}HA)(E_{s}D_{\zeta})\,d\mu(\zeta)\,ds.

All the factors are functions of AA and hence commute with AA, so the right-hand side equals Qμ(H)−A∗Qμ(H)AQ_{\mu}(H)-A^{*}Q_{\mu}(H)A.

As for the bounds, the integrand in (28) is jointly norm continuous on [0,1]×supp⁡μ[0,1]\times\operatorname{supp}\mu and bounded by Lμ2e2f∥H∥L_{\mu}^{2}e^{2f}\|H\|, which gives the existence of the integral and the upper estimate; positivity is clear from the definition. Each EsDζE_{s}D_{\zeta} is invertible, with ∥(EsDζ)−1∥=∥(I−ζA)esF∥≤(1+∥A∥)ef\|(E_{s}D_{\zeta})^{-1}\|=\|(I-\zeta A)e^{sF}\|\le(1+\|A\|)e^{f}, so that ⟨Dζ∗Es∗HEsDζx,x⟩≥he−2f(1+∥A∥)−2∥x∥2\langle D_{\zeta}^{*}E_{s}^{*}HE_{s}D_{\zeta}x,x\rangle\ge he^{-2f}(1+\|A\|)^{-2}\|x\|^{2}, and integrating gives the lower bound.

Theorem 3.2 (Transfer for separated inner functions). Let θ=ηBSμ\theta=\eta BS_{\mu} be a non-constant inner function, where BB is a finite or infinite Blaschke product or the empty product and μ\mu is a finite positive singular measure, and suppose that

σ(A)∩Sp⁡(θ)=∅.(32)\sigma(A)\cap\operatorname{Sp}(\theta)=\varnothing. \tag*{(32)}

Then θ(A)\theta(A) is well defined, and

θ(A) is similar to a contraction ⟺A is similar to a contraction.(33)\theta(A)\text{ is similar to a contraction }\Longleftrightarrow A\text{ is similar to a contraction}. \tag*{(33)}

If HH is a Lyapunov operator for θ(A)\theta(A), then

Q=QB(H)+B(A)∗Qμ(H)B(A),(34)Q = Q_B(H) + B(A)^*Q_\mu(H)B(A), \tag*{(34)}

where QB(H)Q_B(H) is the series (16) and QB(H)=0Q_B(H)=0 if BB is the empty product, is a Lyapunov operator for AA, and

Q−A∗QA=H−θ(A)∗Hθ(A).(35)Q-A^*QA=H-\theta(A)^*H\theta(A). \tag*{(35)}

Proof. Both constructions apply under (32), since Sp⁡(θ)\operatorname{Sp}(\theta) contains the closure of the set of poles of BB as well as supp⁡μ\operatorname{supp}\mu; the series and the integral converge, their terms are positive, and θ(A)=ηB(A)Sμ(A)\theta(A)=\eta B(A)S_\mu(A). If BB is non-constant, then QB(H)Q_B(H) is coercive by (20); if BB is the empty product, then μ≠0\mu\ne0 because θ\theta is non-constant, and (30) gives the coercivity of Qμ(H)Q_\mu(H). In either case QQ is coercive.

Since AA commutes with B(A)B(A),

Q−A∗QA=(H−B(A)∗HB(A))+B(A)∗(H−Sμ(A)∗HSμ(A))B(A)=H−θ(A)∗Hθ(A),\begin{aligned} Q-A^*QA &=(H-B(A)^*HB(A))+B(A)^*(H-S_\mu(A)^*HS_\mu(A))B(A)\\ &=H-\theta(A)^*H\theta(A), \end{aligned}

the unimodular constant η\eta playing no role. Hence A∗QA≤QA^*QA\le Q, and AA is similar to a contraction by Proposition 1.1.

Conversely, let RAR−1=CRAR^{-1}=C be a contraction. Then B(C)=RB(A)R−1B(C)=RB(A)R^{-1} is a contraction by the converse part of Theorem 2.1, and (31) with H=IH=I gives

Fμ(C)∗+Fμ(C)=2∫T(I−ζ‾C)−∗(I−C∗C)(I−ζC)−1 dμ(ζ)≥0,F_\mu(C)^*+F_\mu(C)=2\int_{\mathbb{T}}(I-\overline{\zeta}C)^{-*}(I-C^*C)(I-\zeta C)^{-1}\,d\mu(\zeta)\ge0,

whence dds(e−sFμ(C)∗e−sFμ(C))≤0\frac{d}{ds}(e^{-sF_\mu(C)^*}e^{-sF_\mu(C)})\le0 and ∥e−Fμ(C)∥≤1\|e^{-F_\mu(C)}\|\le1. Therefore θ(C)=Rθ(A)R−1\theta(C)=R\theta(A)R^{-1} is a contraction. ∎

Example 3.3 (An atomic singular factor). For t>0t>0 and ζ∈T\zeta\in\mathbb{T}, the inner function

θ(z)=exp⁡(−t1+ζ‾z1−ζ‾z)\theta(z)=\exp\left(-t\frac{1+\overline{\zeta}z}{1-\overline{\zeta}z}\right)

has Sp⁡(θ)={ζ}\operatorname{Sp}(\theta)=\{\zeta\}, so the only hypothesis is ζ∉σ(A)\zeta\notin\sigma(A). Thus

exp⁡(−t(I+ζ‾A)(I−ζ‾A)−1)\exp\left(-t(I+\overline{\zeta}A)(I-\overline{\zeta}A)^{-1}\right)

is similar to a contraction if and only if AA is, and an explicit Lyapunov operator for AA is

Q=2t∫01Dζ∗e−sF∗He−sFDζ ds,F=t(I+ζ‾A)Dζ.Q=2t\int_0^1D_\zeta^*e^{-sF^*}He^{-sF}D_\zeta\,ds,\qquad F=t(I+\overline{\zeta}A)D_\zeta.

Remark 3.4. Theorem 3.2 covers arbitrary Blaschke sequences and arbitrary singular measures, but only when the spectrum avoids the reflected singularity set (32). It says nothing about an H∞H^\infty functional calculus at a boundary singularity of θ\theta, and § 5 shows that no such statement can hold in general. Only the positivity and finiteness of μ\mu are used in the proof; singularity of μ\mu serves only to identify SμS_\mu as an inner function.

Three variants of the transfer theorem for Blaschke products

The proof of Theorem 2.1 uses less than separation of the spectrum from the whole closed set of poles: it needs a meaning for B(A)B(A) and a summability property of the factors EjE_j. In this section we vary these two ingredients and obtain three variants of the theorem: a weighted resolvent hypothesis under which the proof goes through verbatim; a column criterion for the partial products, which turns out to force their norm convergence; and a simultaneous version for commuting tuples of operators and arbitrary separated inner functions.

A weighted resolvent hypothesis

What the proof of Theorem 2.1 requires is the norm convergence of the partial products and the square summability of the factors EjE_j. Both follow from a single weighted condition on the resolvents.

Proposition 4.1. Let BB be a non-constant Blaschke product with zeros (aj)j∈J(a_j)_{j\in J}, and suppose that every I−aj‾AI-\overline{a_j}A is invertible and that

EA(B):=∑j∈J(1−∣aj∣2)∥(I−aj‾A)−1∥2<∞.(36)\mathcal{E}_A(B) := \sum_{j\in J}(1-|a_j|^2)\left\|(I-\overline{a_j}A)^{-1}\right\|^2 < \infty. \tag*{(36)}

Then the partial products PNP_N converge in operator norm, so that B(A)=ηlim⁡NPNB(A)=\eta\lim_N P_N is defined; the operator QQ of (16) is well defined and coercive, and (17) holds. Consequently (15) holds, and

1−∣a1∣2∥H−1∥∥I−a1‾A∥2I≤Q≤∥H∥P2EA(B)I,(37)\frac{1-|a_1|^2}{\left\|H^{-1}\right\|\left\|I-\overline{a_1}A\right\|^2}I\leq Q\leq\|H\|P^2\mathcal{E}_A(B)I, \tag*{(37)}

so that LβL\sqrt{\beta} may be replaced by EA(B)\sqrt{\mathcal{E}_A(B)} in (24).

Proof. Let d=∑j:aj≠0(1−∣aj∣)<∞d=\sum_{j:a_j\neq0}(1-|a_j|)<\infty. By the Cauchy–Schwarz inequality and (1−∣aj∣)/(1+∣aj∣)≤1−∣aj∣(1-|a_j|)/(1+|a_j|)\leq1-|a_j|,

∑j:aj≠0(1−∣aj∣)∥(I−aj‾A)−1∥≤(∑j:aj≠01−∣aj∣1+∣aj∣)1/2(∑j:aj≠0(1−∣aj∣2)∥(I−aj‾A)−1∥2)1/2≤dEA(B).\begin{aligned} \sum_{j:a_j\neq0}(1-|a_j|)\left\|(I-\overline{a_j}A)^{-1}\right\| &\leq\left(\sum_{j:a_j\neq0}\frac{1-|a_j|}{1+|a_j|}\right)^{1/2} \left(\sum_{j:a_j\neq0}(1-|a_j|^2)\left\|(I-\overline{a_j}A)^{-1}\right\|^2\right)^{1/2}\\ &\leq\sqrt{d\mathcal{E}_A(B)}. \end{aligned}

By (13), the product therefore converges absolutely in norm after its finite initial part, and

P≤max⁡{1,∥A∥}qexp⁡((1+∥A∥)dEA(B))<∞.P\leq\max\{1,\|A\|\}^{q}\exp((1+\|A\|)\sqrt{d\mathcal{E}_A(B)})<\infty.

Moreover ∑j∥Ej∥2≤P2EA(B)<∞\sum_j\|E_j\|^2\leq P^2\mathcal{E}_A(B)<\infty, so the remainder of the proof of Theorem 2.1 applies without change. In the converse direction, PN(C)=XPN(A)X−1P_N(C)=XP_N(A)X^{-1} is a contraction for every NN by von Neumann’s inequality and converges in norm, so no separation hypothesis is needed there either. □

Example 4.2 (The weighted condition is strictly weaker). Let A=IA=I and aj=(1−2−3j)ei2−ja_j=(1-2^{-3j})e^{i2^{-j}}. These zeros satisfy the Blaschke condition, and the poles 1/aj1/a_j accumulate at 1∈σ(A)1\in\sigma(A), so (11) fails. Writing rj=1−2−3jr_j=1-2^{-3j} and θj=2−j\theta_j=2^{-j}, we have

∣1−aj∣2=(1−rj)2+4rjsin⁡2(θj/2)≥c2−2j,1−∣aj∣2≤2⋅2−3j,|1-a_j|^2=(1-r_j)^2+4r_j\sin^2(\theta_j/2)\geq c2^{-2j},\qquad1-|a_j|^2\leq2\cdot2^{-3j},

with an absolute constant c>0c>0, so that EI(B)≤C∑j2−3j/2−2j=C∑j2−j<∞\mathcal{E}_I(B)\leq C\sum_j2^{-3j}/2^{-2j}=C\sum_j2^{-j}<\infty. Proposition 4.1 therefore applies to a Blaschke product whose poles accumulate on the spectrum. We do not claim that the weighted condition is necessary.

The column criterion forces norm convergence

The second variant replaces the summability of the norms ∥Ej∥2\|E_j\|^2 by a weaker, columnwise, condition. This condition turns out to be considerably more useful than a mere bound for strongly convergent products: combined with the Blaschke condition, it yields the norm convergence of the partial products, with an explicit rate.

Proposition 4.3 (Column criterion). Let (aj)j≥1⊂D(a_j)_{j\ge1}\subset\mathbb{D} satisfy the Blaschke condition, with the finitely many zeros at the origin listed first, and suppose that every I−aj‾AI-\overline{a_j}A is invertible. If

C:=sup⁡N∥∑j=1NEj∗Ej∥<∞,(38)C:=\sup_N\left\Vert\sum_{j=1}^{N}E_j^*E_j\right\Vert<\infty, \tag*{(38)}

then PNP_N converges in operator norm to an operator SS, and for every NN beyond the zeros at the origin

∥S−PN∥≤(1+∥A∥)C(∑j>N1−∣aj∣1+∣aj∣)1/2.(39)\left\Vert S-P_N\right\Vert\le(1+\left\Vert A\right\Vert)\sqrt{C}\left(\sum_{j>N}\frac{1-\lvert a_j\rvert}{1+\lvert a_j\rvert}\right)^{1/2}. \tag*{(39)}

If SS is similar to a contraction, then so is AA. More precisely, whenever HH is a Lyapunov operator for SS, the strong limit Q=s−limN∑j≤NEj∗HEjQ=\mathop{\mathrm{s-lim}}_{N}\sum_{j\le N}E_j^*HE_j exists, is a Lyapunov operator for AA, and satisfies

Q−A∗QA=H−S∗HS≥0.(40)Q-A^*QA=H-S^*HS\ge0. \tag*{(40)}

Conversely, if AA is similar to a contraction, then so is SS.

Proof. For aj=rjeiθj≠0a_j=r_je^{i\theta_j}\ne0, formula (13) gives

Pj−1−Pj=αj(I+e−iθjA)Ej,αj=1−rj1+rj.P_{j-1}-P_j=\alpha_j(I+e^{-i\theta_j}A)E_j,\qquad\alpha_j=\sqrt{\frac{1-r_j}{1+r_j}}.

Hence, for M>NM>N beyond the zeros at the origin and x∈Hx\in\mathcal{H}, the Cauchy–Schwarz inequality gives

∥(PM−PN)x∥≤(∑j=N+1Mαj2)1/2(∑j=N+1M∥(I+e−iθjA)Ejx∥2)1/2≤(1+∥A∥)C∥x∥(∑j>N1−rj1+rj)1/2.\begin{aligned} \left\Vert(P_M-P_N)x\right\Vert\le\left(\sum_{j=N+1}^{M}\alpha_j^2\right)^{1/2}\left(\sum_{j=N+1}^{M}\left\Vert(I+e^{-i\theta_j}A)E_jx\right\Vert^2\right)^{1/2} \\ &\le(1+\left\Vert A\right\Vert)\sqrt{C}\left\Vert x\right\Vert\left(\sum_{j>N}\frac{1-r_j}{1+r_j}\right)^{1/2}. \end{aligned}

The last sum tends to zero by the Blaschke condition, which proves the norm convergence of (PN)(P_N) and the estimate (39).

The positive operators QN=∑j≤NEj∗HEjQ_N=\sum_{j\le N}E_j^*HE_j increase with NN and satisfy 0≤QN≤∥H∥CI0\le Q_N\le\left\Vert H\right\Vert CI, so they converge strongly to a bounded positive operator QQ, and Q≥E1∗HE1Q\ge E_1^*HE_1 is coercive by (20). The finite identity (23), QN−A∗QNA=H−PN∗HPNQ_N-A^*Q_NA=H-P_N^*HP_N, holds for every NN without any convergence hypothesis; passing to the limit, strongly on the left-hand side and in norm on the right-hand side, proves (40). Proposition 1.1 completes the proof of the forward implication. Conversely, in an HH-norm in which AA is a contraction all finite products PNP_N are contractions by von Neumann’s inequality, and so is their norm limit. □

Remark 4.4. Condition (38) says that x↦(Ejx)j≥1x\mapsto(E_jx)_{j\ge1} is a bounded map from H\mathcal{H} to ℓ2(H)\ell^2(\mathcal{H}). It is weaker than ∑j∥Ej∥2<∞\sum_j\left\Vert E_j\right\Vert^2<\infty, but strictly stronger than the norm convergence of the partial products. To see this, take A=IA=I on a one-dimensional space and aj=(1−εj2)eiεja_j=(1-\varepsilon_j^2)e^{i\varepsilon_j}, where 0<εj<1/20<\varepsilon_j<1/2, εj↓0\varepsilon_j\downarrow0 and ∑jεj<∞\sum_j\varepsilon_j<\infty. Then ∣PN∣=1\lvert P_N\rvert=1 for every NN, while (13) gives

∣1−baj(1)∣=2εj2cos⁡(εj/2)∣1−aj∣≤C0εj,\lvert1-b_{a_j}(1)\rvert=\frac{2\varepsilon_j^2\cos(\varepsilon_j/2)}{\lvert1-a_j\rvert}\le C_0\varepsilon_j,

because ∣1−aj∣≥cεj\lvert1-a_j\rvert\ge c\varepsilon_j; thus (PN)(P_N) converges. On the other hand ∣1−aj∣2≤2εj2\lvert1-a_j\rvert^2\le2\varepsilon_j^2 and 1−∣aj∣2≥εj21-\lvert a_j\rvert^2\ge\varepsilon_j^2, so ∥Ej∥2=(1−∣aj∣2)/∣1−aj∣2≥1/2\left\Vert E_j\right\Vert^2=(1-\lvert a_j\rvert^2)/\lvert1-a_j\rvert^2\ge1/2, and the column series diverges. In this example AA and the limit SS are unimodular scalars, so the failure of the column estimate does not by itself obstruct the transfer of similarity.

The case in which the norm limit has spectral radius less than one is elementary and requires neither separation nor the column estimate.

Proposition 4.5 (A strictly stable norm limit). Suppose that every I−aj‾AI-\overline{a_j}A is invertible and that PN→SP_N \to S in operator norm. If the spectral radius of SS is less than one, then so is that of AA, and AA is similar to a contraction.

Proof. Choose an integer kk with ∥Sk∥<1\lVert S^k\rVert< 1. By norm convergence, ∥PNk∥<1\lVert P_N^k\rVert< 1 for all sufficiently large NN, so that r(PN)<1r(P_N)<1 for such NN. Fix one such N≥1N \ge1. Since PN=RN(A)P_N=R_N(A) for the rational function RN=∏j≤NbajR_N=\prod_{j\le N}b_{a_j}, whose poles lie outside σ(A)\sigma(A), the spectral mapping theorem gives

∣∏j=1Nbaj(λ)∣<1(λ∈σ(A)).\left|\prod_{j=1}^{N} b_{a_j}(\lambda)\right|<1 \qquad(\lambda\in\sigma(A)).

Each Blaschke factor has modulus one on T\mathbb{T} and modulus greater than one outside D‾\overline{\mathbb{D}}, wherever it is finite, because ∣z−a∣2−∣1−a‾z∣2=(∣z∣2−1)(1−∣a∣2)|z-a|^2-|1-\overline{a}z|^2=(|z|^2-1)(1-|a|^2). It follows that σ(A)⊂D\sigma(A)\subset\mathbb{D}, and by compactness r(A)<1r(A)<1. Finally, Q=∑n≥0A∗nAnQ=\sum_{n\ge0}A^{*n}A^n converges in norm, satisfies Q≥IQ\ge I and Q−A∗QA=IQ-A^*QA=I, so AA is similar to a contraction.

:::

Simultaneous transfer for commuting tuples

Theorem 4.6 (Simultaneous transfer). Let A1,…,AdA_1,\ldots,A_d be pairwise commuting bounded operators, let θ1,…,θd\theta_1,\ldots,\theta_d be non-constant inner functions with σ(Ai)∩Sp⁡(θi)=∅\sigma(A_i)\cap\operatorname{Sp}(\theta_i)=\varnothing, and put Si=θi(Ai)S_i=\theta_i(A_i). If S1,…,SdS_1,\ldots,S_d admit a common Lyapunov operator HH, that is, a coercive HH with

Si∗HSi≤H(1≤i≤d),(41)S_i^*HS_i\le H \qquad(1\le i\le d), \tag*{(41)}

then A1,…,AdA_1,\ldots,A_d admit a common Lyapunov operator QQ, that is, a coercive QQ with Ai∗QAi≤QA_i^*QA_i\le Q for all ii. Consequently (S1,…,Sd)(S_1,\ldots,S_d) is simultaneously similar to a tuple of contractions if and only if (A1,…,Ad)(A_1,\ldots,A_d) is.

Proof. Write θi=ηiBiSμi\theta_i=\eta_iB_iS_{\mu_i} and define a bounded positive linear map on B(H)\mathcal{B}(\mathcal{H}) by

Φi(K)=QBi(K)+Bi(Ai)∗Qμi(K)Bi(Ai),(42)\Phi_i(K)=Q_{B_i}(K)+B_i(A_i)^*Q_{\mu_i}(K)B_i(A_i), \tag*{(42)}

where QBi(K)Q_{B_i}(K) is the series (16) and Qμi(K)Q_{\mu_i}(K) the integral (28), both formed with AiA_i in place of AA; for an empty Blaschke product the first term is zero and Bi(Ai)=IB_i(A_i)=I. These series and integrals converge in operator norm for every bounded KK. The computations in the proofs of Lemma 3.1 and Theorem 3.2 give, for every self-adjoint KK,

Φi(K)−Ai∗Φi(K)Ai=K−Si∗KSi,(43)\Phi_i(K)-A_i^*\Phi_i(K)A_i=K-S_i^*KS_i, \tag*{(43)}

and Φi(K)\Phi_i(K) is coercive whenever KK is, because θi\theta_i is non-constant. Every summand and every integrand in Φi(K)\Phi_i(K) has the form V∗KVV^*KV with VV a function of AiA_i, so an operator RR that commutes with AiA_i commutes with all these VV, and therefore

Φi(K)−R∗Φi(K)R=Φi(K−R∗KR).(44)\Phi_i(K)-R^*\Phi_i(K)R=\Phi_i(K-R^*KR). \tag*{(44)}

Put H0=HH_0=H and Hi=Φi(Hi−1)H_i=\Phi_i(H_{i-1}) for 1≤i≤d1\le i\le d. We claim that HiH_i is coercive, that Sk∗HiSk≤HiS_k^*H_iS_k\le H_i for every kk, and that Ak∗HiAk≤HiA_k^*H_iA_k\le H_i for k≤ik\le i. For i=0i=0 this is the hypothesis (41). At step ii, the inequality for AiA_i follows from (43) and the inductive hypothesis for SiS_i, while the inequalities for SkS_k (all kk) and for AkA_k (k<ik<i) are preserved by (44) and the positivity of Φi\Phi_i, since these operators commute with AiA_i. Hence Q=HdQ=H_d has the required properties. The converse follows from the converse part of Theorem 3.2, applied with one common similarity.

:::

Boundary obstructions

Theorem 2.1 is a statement about a norm-convergent product under the separation hypothesis. It must not be read as a statement about an arbitrary H∞H^\infty functional calculus, and the three examples of this section show why. In each of them the separation hypothesis fails at the point where the poles accumulate.

A bounded calculus does not suffice

Proposition 5.1. There exist a polynomially bounded operator AA on a Hilbert space that is not similar to a contraction, and an infinite Blaschke product BB with zeros aj∈(0,1)a_j \in(0,1), such that B(A)=0B(A)=0. The zeros, listed with multiplicity, can moreover be chosen so that

∑j1−aj2<∞.\sum_j \sqrt{1-a_j^2}<\infty.

In this example the point 11 lies both in σ(A)\sigma(A) and in the closure of the set of poles of BB.

Proof. We use the modification of Pisier’s counterexample due to Gamal’ [2 §5]; see also the introduction of [3]. By the finite-dimensional form of Pisier’s counterexample there are nilpotent operators RjR_j on finite-dimensional Hilbert spaces HjH_j whose polynomial bounds are bounded by a common constant KK while their complete polynomial bounds are unbounded along a subsequence. We write Rjdj=0R_j^{d_j}=0.

For 0<w<10<w<1 put βw(z)=(w−z)/(1−wz)\beta_w(z)=(w-z)/(1-wz). Disc automorphisms preserve both the polynomial and the complete polynomial bounds, by rational approximation on the closed disc in one direction and by βw∘βw=z\beta_w\circ\beta_w=z in the other. Choose wj↑1w_j \mathrel{\uparrow}1 so rapidly that

dj1−wj2≤2−j,∥βwj(Rj)−IHj∥≤2−j.d_j\sqrt{1-w_j^2}\leq2^{-j},\qquad\|\beta_{w_j}(R_j)-I_{H_j}\|\leq2^{-j}.

the second condition being attainable because I−RjI-R_j is invertible and βw(Rj)→I\beta_w(R_j)\to I as w↑1.Setw\mathrel{\uparrow}1. Set

Aj=βwj(Rj),A=⨁jAj,B=∏jβwjdj.A_j=\beta_{w_j}(R_j),\qquad A=\bigoplus_j A_j,\qquad B=\prod_j\beta_{w_j}^{d_j}.

Then BB is a Blaschke product with the stated square-root summability, and AA is polynomially bounded by KK while its complete polynomial bound is infinite, so AA is not similar to a contraction by Paulsen’s theorem [11]. An H∞H^\infty calculus is defined by f(A)=⨁jf(Aj)f(A)=\bigoplus_j f(A_j), where f(Aj)f(A_j) is given by the Riesz–Dunford calculus since σ(Aj)={wj}⊂D\sigma(A_j)=\{w_j\}\subset\mathbb{D}. The finite-dimensional calculi satisfy ∥f(Aj)∥≤K∥f∥∞\|f(A_j)\|\leq K\|f\|_\infty: indeed fr(z)=f(rz)f_r(z)=f(rz) is a uniform limit of polynomials on D\mathbb{D} with ∥fr∥∞≤∥f∥∞\|f_r\|_\infty\leq\|f\|_\infty, and fr(Aj)→f(Aj)f_r(A_j)\to f(A_j) in norm as r↑1r\mathrel{\uparrow}1. The calculus is weak-* continuous on bounded nets: bounded weak-* convergence in H∞H^\infty implies locally uniform convergence in D\mathbb{D}, hence norm convergence of fα(Aj)f_\alpha(A_j) for each jj, and a truncation of the direct sum then gives convergence in the weak operator topology. Since βwj(Aj)=Rj\beta_{w_j}(A_j)=R_j, the power βwjdj\beta_{w_j}^{d_j} annihilates AjA_j; moreover B=βwjdjGjB=\beta_{w_j}^{d_j}G_j with GjG_j holomorphic near wjw_j, because the remaining factors converge locally uniformly on D\mathbb{D}, so B(Aj)=RjdjGj(Aj)=0B(A_j)=R_j^{d_j}G_j(A_j)=0 for every jj, that is, B(A)=0B(A)=0. Finally, we have σ(A)={wj:j≥1}∪{1}\sigma(A)=\{w_j:j\geq1\}\cup\{1\}, and the poles wj−1w_j^{-1} also tend to 11.

Thus a bounded, weak-* continuous H∞H^\infty calculus is compatible with B(A)=0B(A)=0 and with the failure of similarity. The refined construction of [2] even yields cyclic examples. In the other direction, for a Carleson–Newman Blaschke product BB an absolutely continuous polynomially bounded operator AA with B(A)=0B(A)=0 is similar to a contraction; see [3 Theorem 3.2].

A norm radial limit does not suffice

Example 5.2. Put an=(n2−1)/(n2+1)a_n=(n^2-1)/(n^2+1),

A=⨁n≥1(an10an),F(z)=∏n≥1an−z1−anz,B=F2.A=\bigoplus_{n\geq1}\begin{pmatrix}a_n&1\\0&a_n\end{pmatrix},\qquad F(z)=\prod_{n\geq1}\frac{a_n-z}{1-a_nz},\qquad B=F^2.

Then AA is bounded, σ(A)={an:n≥1}∪{1}\sigma(A)=\{a_n:n\geq1\}\cup\{1\}, and ∥Ak∥≥sup⁡nkank−1=k\lVert A^k\rVert\geq\sup_n ka_n^{k-1}=k, so AA is not power bounded. Nevertheless B(rA)→0B(rA)\to0 in operator norm as r↑1r\uparrow1, where each B(rA)B(rA) is defined by the holomorphic functional calculus because σ(rA)\sigma(rA) is a compact subset of D\mathbb{D}.

Proof. The product is a Blaschke product because ∑n(1−an)=∑n2/(n2+1)<∞\sum_n(1-a_n)=\sum_n2/(n^2+1)<\infty. Writing w=(1+z)/(1−z)w=(1+z)/(1-z), the Euler product formulae give

F(z)=∏n≥11−w/n21+w/n2=sin⁡(πw)sinh⁡(πw),F(z)=\prod_{n\geq1}\frac{1-w/n^2}{1+w/n^2}=\frac{\sin(\pi\sqrt{w})}{\sinh(\pi\sqrt{w})},

the quotient being independent of the choice of the square root. For the kk-th power,

Ak=⨁n(ankkank−10ank),A^k=\bigoplus_n\begin{pmatrix}a_n^k&ka_n^{k-1}\\0&a_n^k\end{pmatrix},

so ∥Ak∥≥sup⁡nkank−1=k\lVert A^k\rVert\geq\sup_nka_n^{k-1}=k.

For real x↑1x\uparrow1 put t=(1+x)/(1−x)t=\sqrt{(1+x)/(1-x)}. Then F(x)=O(e−πt)F(x)=O(e^{-\pi t}) and F′(x)=O(t3e−πt)F'(x)=O(t^3e^{-\pi t}), so BB and B′B' extend continuously to [0,1][0,1] with B(1)=B′(1)=0B(1)=B'(1)=0; moreover B(an)=B′(an)=0B(a_n)=B'(a_n)=0 for every nn, the zeros being double. Since

B(rA)=⨁n≥1(B(ran)rB′(ran)0B(ran)),B(rA)=\bigoplus_{n\geq1}\begin{pmatrix}B(ra_n)&rB'(ra_n)\\0&B(ra_n)\end{pmatrix},

the uniform continuity of BB and B′B' on [0,1][0,1] gives ∥B(rA)∥→0\lVert B(rA)\rVert\to0.

Thus the existence of a norm radial limit as r↑1r\uparrow1 does not imply the power boundedness of AA, even when that limit is zero. The example also shows that radial evaluation is quite different from the norm product: here 1∈σ(A)1\in\sigma(A) is an accumulation point of the poles an−1a_n^{-1}, so the separation hypothesis fails and the norm product is not available.

Remark 5.3. The partial products behave quite differently from B(rA)B(rA). For BN=∏n≤N((an−z)/(1−anz))2B_N=\prod_{n\leq N}((a_n-z)/(1-a_nz))^2 and n>Nn>N one has BN(A)=BN(an)I+BN′(an)NnB_N(A)=B_N(a_n)I+B_N'(a_n)N_n on the nn-th block, where NnN_n is the nilpotent part of that block, so ∥BN(A)∥≥∣BN′(an)∣\lVert B_N(A)\rVert\geq|B_N'(a_n)|; letting n→∞n\to\infty and using the continuity of BN′B_N' at 11 gives ∥BN(A)∥≥∣BN′(1)∣=2∑n≤Nn2→∞\lVert B_N(A)\rVert\geq|B_N'(1)|=2\sum_{n\leq N}n^2\to\infty. Norm radial evaluation and the convergence of the partial products are therefore unrelated hypotheses.

Strong products fail, even for Ritt operators

The third example shows that, even for Ritt operators, the strong convergence of the partial products to zero is compatible with AA not even being polynomially bounded. By a theorem of Le Merdy [7 Theorem 5.1], a polynomially bounded Ritt operator on a Hilbert space is similar to a contraction; within the class of Ritt operators, therefore, failure of similarity to a contraction is the same as failure of polynomial boundedness, and the conclusion of the following proposition is the strongest possible. Ritt operators that are not polynomially bounded were first constructed by Le Merdy [7 Proposition 5.2], on the basis of an example of Baillon and Clément; our operator AA is a variant of his construction, and the multiplier formula used below is taken from the same proof.

Proposition 5.4. There exist a Ritt operator AA on a Hilbert space that is not polynomially bounded, and hence not similar to a contraction, and an infinite Blaschke product BB whose zeros satisfy ∑n1−∣an∣2<∞\sum_n \sqrt{1-|a_n|^2}<\infty, such that the partial products BN(A)B_N(A) converge strongly to zero.

Proof. Let (xn)(x_n) be a conditional Schauder basis of a Hilbert space, with partial-sum projections PnP_n and K=sup⁡n∥Pn∥K=\sup_n\|P_n\|; in this proof PnP_n denotes these projections rather than partial Blaschke products. We shall use the following bounded-variation multiplier formula, which is [7 (5.3)]: if cn→c∞c_n\to c_\infty and V(c)=∑n∣cn+1−cn∣<∞V(c)=\sum_n|c_{n+1}-c_n|<\infty, then

Mc=c∞I−∑n≥1(cn+1−cn)Pn,Mcxn=cnxn,M_c=c_\infty I-\sum_{n\geq1}(c_{n+1}-c_n)P_n,\qquad M_cx_n=c_nx_n,

the series converging absolutely in norm, with ∥Mc∥≤∣c∞∣+KV(c)\|M_c\|\leq|c_\infty|+KV(c); indeed ∑n≥m(cn+1−cn)=c∞−cm\sum_{n\geq m}(c_{n+1}-c_n)=c_\infty-c_m.

Put λn=1−2−n\lambda_n=1-2^{-n} and Axn=λnxnAx_n=\lambda_nx_n. The multiplier formula shows that AA is bounded, and that σ(A)={λn}∪{1}\sigma(A)=\{\lambda_n\}\cup\{1\}: for λ\lambda outside this closed set, the formula applied to cn=(λ−λn)−1c_n=(\lambda-\lambda_n)^{-1}, whose variation is at most ∑n(λn+1−λn)/dist⁡(λ,{λn}∪{1})2\sum_n(\lambda_{n+1}-\lambda_n)/\operatorname{dist}(\lambda,\{\lambda_n\}\cup\{1\})^2, produces the inverse of λI−A\lambda I-A. Since λnk\lambda_n^k increases to 11, the formula also gives ∥Ak∥≤1+K\|A^k\|\leq1+K for all kk. For gk(t)=ktk−1(1−t)g_k(t)=kt^{k-1}(1-t), which is non-negative, unimodal and bounded by 11 on [0,1][0,1], the sequence gk(λn)g_k(\lambda_n) is unimodal in nn, tends to 00, and has variation at most 22; hence k∥Ak−1(I−A)∥=∥Mgk∥≤2Kk\|A^{k-1}(I-A)\|=\|M_{g_k}\|\leq2K for all kk. By the power-and-difference characterisation of Ritt operators [10, 8], AA is a Ritt operator.

The sequence (λn)(\lambda_n) is interpolating: for ∣n−m∣=j|n-m|=j,

∣λn−λm1−λnλm∣≥1−2−j1+2−j,soinf⁡n∏m≠n∣λn−λm1−λnλm∣≥∏j≥1(1−2−j1+2−j)2>0.\left|\frac{\lambda_n-\lambda_m}{1-\lambda_n\lambda_m}\right|\geq\frac{1-2^{-j}}{1+2^{-j}}, \qquad\text{so}\qquad \inf_n\prod_{m\ne n}\left|\frac{\lambda_n-\lambda_m}{1-\lambda_n\lambda_m}\right| \geq\prod_{j\geq1}\left(\frac{1-2^{-j}}{1+2^{-j}}\right)^2>0.

By Carleson’s interpolation theorem there is a constant CC such that for every choice of signs (εn)(\varepsilon_n) there is f∈H∞f\in H^\infty with ∥f∥∞≤C\|f\|_\infty\leq C and f(λn)=εnf(\lambda_n)=\varepsilon_n for all nn. Suppose that AA were polynomially bounded, with bound MM. For 0<r<10<r<1 the function fr(z)=f(rz)f_r(z)=f(rz) is a uniform limit of polynomials on the closed disc, so ∥fr(A)∥≤MC\|f_r(A)\|\leq MC, and fr(A)f_r(A) converges strongly, as r↑1r\uparrow1, to the sign multiplier xn↦εnxnx_n\mapsto\varepsilon_nx_n, since it does so on finite linear combinations of the basis vectors. All sign multipliers would then be bounded by MCMC, and the basis would be unconditional, a contradiction. Hence AA is not polynomially bounded.

Finally, let B=∏n≥1bλn2B=\prod_{n\geq1}b_{\lambda_n}^2 and BN=∏j≤Nbλj2B_N=\prod_{j\leq N}b_{\lambda_j}^2. For each j≤Nj\leq N the function bλj2b_{\lambda_j}^2 is non-negative and increasing on [λN,1)[\lambda_N,1), so for n>Nn>N the sequence cn=BN(λn)c_n=B_N(\lambda_n) increases from 00 to 11, while cn=0c_n=0 for n≤Nn\leq N; thus c∞=1c_\infty=1, V(c)≤1V(c)\leq1, and ∥BN(A)∥≤1+K\|B_N(A)\|\leq1+K for all NN. Each BN(A)B_N(A) annihilates x1,…,xNx_1,\ldots,x_N, and the uniform bound and the density of finite linear combinations give BN(A)→0B_N(A)\to0 strongly. If the double zeros are listed individually, the intermediate partial products also converge strongly to zero: the same formula gives ∥bλj(A)∥≤1+2K\|b_{\lambda_j}(A)\|\leq1+2K, because t↦(λj−t)/(1−λjt)t\mapsto(\lambda_j-t)/(1-\lambda_jt) decreases along (λn)(\lambda_n) from a value in (−1,1)(-1,1) to the limit −1-1.

Consequently the Ritt condition does not restore the implication (2) when an infinite Blaschke product is evaluated at AA as a strong limit of partial products. This is consistent with Theorem 2.1: here the poles λn−1\lambda_n^{-1} accumulate at 1∈σ(A)1\in\sigma(A).

Finite Blaschke products: polynomial Lyapunov operators

For a finite Blaschke product the Lyapunov operator can be written without any resolvents, and this yields a quantitative bound that is used in the next section. Write

B(z)=ηN(z)D(z),N(z)=∏j=1m(z−aj),D(z)=∏j=1m(1−aj‾z),∣η∣=1,(45)B(z)=\eta\frac{N(z)}{D(z)},\qquad N(z)=\prod_{j=1}^{m}(z-a_j),\qquad D(z)=\prod_{j=1}^{m}(1-\overline{a_j}z),\qquad|\eta|=1, \tag*{(45)}

with m≥1m\ge1 and aj∈Da_j\in\mathbb{D}.

Proposition 6.1 (Polynomial Lyapunov operator). Suppose that D(A)D(A) is invertible and that B(A)B(A) is similar to a contraction with sim⁡(B(A))≤K\operatorname{sim}(B(A))\le K. Then there is a constant CB<∞C_B<\infty, depending only on BB, such that

sim⁡(A)≤CBKmax⁡0≤k<m∥Ak∥≤CBKmax⁡{1,∥A∥}m−1.(46)\operatorname{sim}(A)\le C_BK\max_{0\le k<m}\|A^k\|\le C_BK\max\{1,\|A\|\}^{m-1}. \tag*{(46)}

Remark 6.2. No hypothesis on σ(A)\sigma(A) beyond the invertibility of D(A)D(A) is needed in Proposition 6.1; in particular, the bound (46) is uniform over all operators AA with a common bound on ∥A∥\|A\| and on sim⁡(B(A))\operatorname{sim}(B(A)).

Proof of Proposition 6.1. By Proposition 1.1, there is a Lyapunov operator HH for B(A)B(A) with ∥H∥∥H−1∥≤K\sqrt{\|H\|\|H^{-1}\|}\le K. For j=1,…,mj=1,\ldots,m put

Fj(z)=1−∣aj∣2∏ℓ<j(z−aℓ)∏ℓ>j(1−aℓ‾z)∈Pm−1,(47)F_j(z)=\sqrt{1-|a_j|^2}\prod_{\ell<j}(z-a_\ell)\prod_{\ell>j}(1-\overline{a_\ell}z)\in\mathcal{P}_{m-1}, \tag*{(47)}

where Pm−1\mathcal{P}_{m-1} denotes the space of polynomials of degree at most m−1m-1, and write nℓ(z)=z−aℓn_\ell(z)=z-a_\ell and dℓ(z)=1−aℓ‾zd_\ell(z)=1-\overline{a_\ell}z. The one-factor computation

dℓ(w)‾dℓ(z)−nℓ(w)‾nℓ(z)=(1−aℓ‾w)(1−aℓ‾z)−(w‾−aℓ‾)(z−aℓ)=(1−w‾z)(1−∣aℓ∣2)\overline{d_\ell(w)}d_\ell(z)-\overline{n_\ell(w)}n_\ell(z)=(1-\overline{a_\ell}w)(1-\overline{a_\ell}z)-(\overline{w}-\overline{a_\ell})(z-a_\ell)=(1-\overline{w}z)(1-|a_\ell|^2)

and the telescoping decomposition

∏ℓ=1mdℓ(w)‾dℓ(z)−∏ℓ=1mnℓ(w)‾nℓ(z)=∑j=1m(∏ℓ<jnℓ(w)‾nℓ(z))(dj(w)‾dj(z)−nj(w)‾nj(z))(∏ℓ>jdℓ(w)‾dℓ(z))\prod_{\ell=1}^{m}\overline{d_\ell(w)}d_\ell(z)-\prod_{\ell=1}^{m}\overline{n_\ell(w)}n_\ell(z) =\sum_{j=1}^{m}\left(\prod_{\ell<j}\overline{n_\ell(w)}n_\ell(z)\right)\left(\overline{d_j(w)}d_j(z)-\overline{n_j(w)}n_j(z)\right)\left(\prod_{\ell>j}\overline{d_\ell(w)}d_\ell(z)\right)

together give the Takenaka–Malmquist identity

D(w)‾D(z)−N(w)‾N(z)=(1−w‾z)∑j=1mFj(w)‾Fj(z).(48)\overline{D(w)}D(z)-\overline{N(w)}N(z)=(1-\overline{w}z)\sum_{j=1}^{m}\overline{F_j(w)}F_j(z). \tag*{(48)}

an identity between polynomials in zz and w‾\overline{w}. Replacing zz by AA and w‾\overline{w} by A∗A^*, and inserting HH between the adjoint and the non-adjoint factors, turns (48) into

Q^−A∗Q^A=D(A)∗HD(A)−N(A)∗HN(A)=D(A)∗[H−B(A)∗HB(A)]D(A)≥0.(49)\widehat{Q}-A^*\widehat{Q}A=D(A)^*HD(A)-N(A)^*HN(A)=D(A)^*[H-B(A)^*HB(A)]D(A)\ge0. \tag*{(49)}

where Q^=∑j=1mFj(A)∗HFj(A)\widehat{Q}=\sum_{j=1}^{m}F_j(A)^*HF_j(A) and we have used N(A)=η‾B(A)D(A)N(A)=\overline{\eta}B(A)D(A). The polynomials F1,…,FmF_1,\ldots,F_m form a basis of Pm−1\mathcal{P}_{m-1}. Indeed, if ∑jcjFj=0\sum_j c_jF_j=0, then evaluation at a1a_1 gives c1=0c_1=0, since F1(a1)≠0F_1(a_1)\ne0 while every other summand contains the factor z−a1z-a_1; dividing by z−a1z-a_1 and evaluating at a2a_2 gives c2=0c_2=0, and so on, repeated zeros causing no difficulty. Hence there are scalars αj\alpha_j with ∑jαjFj=1\sum_j\alpha_jF_j=1, and with cα=(∑j∣αj∣2)1/2c_\alpha=\left(\sum_j|\alpha_j|^2\right)^{1/2} we have

∥x∥2≤cα2∑j∥Fj(A)x∥2\|x\|^2\le c_\alpha^2\sum_j\|F_j(A)x\|^2

and therefore Q^≥1cα2∥H−1∥I\widehat{Q}\ge\frac{1}{c_\alpha^2\|H^{-1}\|}I.

Writing Fj(z)=∑k<mcjkzkF_j(z)=\sum_{k<m}c_{jk}z^k, cF=(∑j(∑k∣cjk∣)2)1/2c_F=\left(\sum_j\left(\sum_k|c_{jk}|\right)^2\right)^{1/2} and MA=max⁡k<m∥Ak∥M_A=\max_{k<m}\|A^k\|, we also have Q^≤∥H∥cF2MA2I\widehat{Q}\le\|H\|c_F^2M_A^2I. Thus Q^1/2AQ^−1/2\widehat{Q}^{1/2}A\widehat{Q}^{-1/2} is a contraction, with condition number at most cαcFKMAc_\alpha c_FKM_A, which is (46) with CB=cαcFC_B=c_\alpha c_F.

Remark 6.3 (Sharpness of the exponent). For B(z)=zmB(z)=z^m with m≥2m\ge2, let JmJ_m be the nilpotent shift on Cm\mathbb{C}^m and At=tJmA_t=tJ_m with t≥1t\ge1. Then B(At)=0B(A_t)=0, so sim⁡(B(At))=1\operatorname{sim}(B(A_t))=1, while any similarity of AtA_t to a contraction has condition number at least ∥Atm−1∥=tm−1\|A_t^{m-1}\|=t^{m-1}. Hence the exponent m−1m-1 in a bound in terms of ∥A∥\|A\| alone cannot be lowered. For B(z)=zmB(z)=z^m the construction above reproduces the elementary Lyapunov operator ∑k<mA∗kHAk\sum_{k<m}A^{*k}HA^k.

Uniform strong test collections

This section is inspired by the work of Dritschel, Estévez and Yakubovich [1]. Fix non-constant finite Blaschke products B1,…,BnB_1,\ldots,B_n, write Ψ=(B1,…,Bn)\Psi=(B_1,\ldots,B_n), and let AΨ\mathcal{A}_\Psi be the algebra of finite sums of functions of the form

∏j=1nfj(Bj(z)),fj∈A(D),(50)\prod_{j=1}^{n}f_j(B_j(z)),\qquad f_j\in A(\mathbb{D}), \tag*{(50)}

where A(D)A(\mathbb{D}) is the disc algebra. We say that Ψ\Psi is a uniform strong test collection on D\mathbb{D} if for every K≥1K\ge1 there is a constant L(Ψ,K)L(\Psi,K) such that

σ(A)⊂D‾,sim⁡(Bj(A))≤K for all j⟹sim⁡(A)≤L(Ψ,K),\sigma(A)\subset\overline{\mathbb{D}},\qquad\operatorname{sim}(B_j(A))\le K\ \text{for all }j\quad\Longrightarrow\quad\operatorname{sim}(A)\le L(\Psi,K),

where the similarities for different jj are not required to coincide; and that Ψ\Psi is a uniform test collection if the same holds with ∥Bj(A)∥≤1\|B_j(A)\|\le1 in place of sim⁡(Bj(A))≤K\operatorname{sim}(B_j(A))\le K.

Theorem 7.1 (Uniform strong testing from a dense algebra). Suppose that AΨ\mathcal{A}_\Psi is dense in A(D)A(\mathbb{D}) in the uniform norm. Then Ψ\Psi is a uniform strong test collection on D\mathbb{D}.

Proof. Let m=deg⁡B1m=\deg B_1, let C=CB1C=C_{B_1} be the constant of Proposition 6.1, and put MA=max⁡k<m∥Ak∥M_A=\max_{k<m}\|A^k\|. That proposition shows that AA is similar to a contraction with sim⁡(A)≤CKMA\operatorname{sim}(A)\le CKM_A, so that ∥g(A)∥≤CKMA∥g∥∞\|g(A)\|\le CKM_A\|g\|_\infty for every g∈A(D)g\in A(\mathbb{D}); this bound depends on AA and will be used only to control approximation errors.

Choose ε=(2CK)−1\varepsilon=(2CK)^{-1} and, by density, gk∈AΨg_k\in\mathcal{A}_\Psi with ∥zk−gk∥∞<ε\|z^k-g_k\|_\infty<\varepsilon for 0≤k<m0\le k<m, where g0=1g_0=1. Writing gk=∑ℓ=1rk∏j=1nfkℓj(Bj(z))g_k=\sum_{\ell=1}^{r_k}\prod_{j=1}^{n}f_{k\ell j}(B_j(z)), we have

∥gk(A)∥≤∑ℓ=1rkKn∏j=1n∥fkℓj∥∞≤G,\|g_k(A)\|\le\sum_{\ell=1}^{r_k}K^n\prod_{j=1}^{n}\|f_{k\ell j}\|_\infty\le G,

where GG is the largest of these finitely many bounds over 0≤k<m0\le k<m; here we have used that each fkℓj(Bj(A))f_{k\ell j}(B_j(A)) has norm at most K∥fkℓj∥∞K\|f_{k\ell j}\|_\infty, by the similarity of Bj(A)B_j(A) to a contraction and von Neumann’s inequality. The bound GG depends on Ψ\Psi, KK and the chosen gkg_k, but not on AA. On the other hand, ∥Ak−gk(A)∥≤CKMAε=MA/2\|A^k-g_k(A)\|\le CKM_A\varepsilon=M_A/2, so MA≤G+MA/2M_A\le G+M_A/2, that is, MA≤2GM_A\le2G. Applying Proposition 6.1 once more gives sim⁡(A)≤2CKG\operatorname{sim}(A)\le2CKG, which is the required uniform bound.

Corollary 7.2 (Injective unramified tuples). Let Ψ=(B1,…,Bn)\Psi=(B_1,\ldots,B_n) be a tuple of non-constant finite Blaschke products that is injective on D‾\overline{\mathbb D} and satisfies

(B1′(z),…,Bn′(z))≠0(z∈D‾).(51)(B_1'(z),\ldots,B_n'(z))\ne0 \qquad(z\in\overline{\mathbb D}). \tag*{(51)}

Then Ψ\Psi is a uniform strong test collection on D‾\overline{\mathbb D}.

Proof. We verify the density hypothesis of Theorem 7.1 by means of the algebra-generation results of [1]. Finite Blaschke products extend holomorphically across T\mathbb T with non-vanishing derivative there, so the injectivity of Ψ\Psi on T\mathbb T makes Ψ\Psi weakly admissible in the sense of [1]. By [1 Lemma 17] there is an admissible family Φ\Phi whose coordinates have the form φℓ(z)=∏j=1nhℓj(Bj(z))\varphi_\ell(z)=\prod_{j=1}^n h_{\ell j}(B_j(z)) with hℓj∈A(D‾)h_{\ell j}\in A(\overline{\mathbb D}) and ∥hℓj∥∞≤1\|h_{\ell j}\|_\infty\le1. Adjoin the original coordinates to Φ\Phi, assigning empty boundary arcs to the added functions, as in the proof of [1 Theorem 18]. The enlarged family Φ0\Phi_0 is admissible, injective, and has non-vanishing joint derivative, so [1 Theorem II] gives AΦ0=A(D‾)\mathcal A_{\Phi_0}=A(\overline{\mathbb D}).

Let CC be the uniform closure of AΨ\mathcal A_\Psi. It contains every coordinate of Φ0\Phi_0; and if φ\varphi is such a coordinate and f∈A(D‾)f\in A(\overline{\mathbb D}), then uniform polynomial approximation of ff shows that f∘φ∈Cf\circ\varphi\in C. Hence AΦ0⊂C\mathcal A_{\Phi_0}\subset C, so C=A(D‾)C=A(\overline{\mathbb D}), and Theorem 7.1 applies.

Theorem 7.3 (Exact characterisation). For a tuple Ψ=(B1,…,Bn)\Psi=(B_1,\ldots,B_n) of non-constant finite Blaschke products the following are equivalent:

  1. Ψ\Psi is injective on D‾\overline{\mathbb D} and (51) holds;

  2. Ψ\Psi is a uniform strong test collection on D‾\overline{\mathbb D};

  3. Ψ\Psi is a uniform test collection on D‾\overline{\mathbb D}.

Proof. The implication (i)⇒\Rightarrow(ii) is Corollary 7.2, and (ii)⇒\Rightarrow(iii) is immediate, since ∥Bj(A)∥≤1\|B_j(A)\|\le1 is the case K=1K=1 of sim⁡(Bj(A))≤K\operatorname{sim}(B_j(A))\le K.

For (iii)⇒\Rightarrow(i), suppose first that z≠wz\ne w in D‾\overline{\mathbb D} satisfy Bj(z)=Bj(w)B_j(z)=B_j(w) for all jj, and for t>0t>0 put

At=(zt(w−z)0w).A_t=\begin{pmatrix} z & t(w-z) \\ 0 & w \end{pmatrix}.

Then AtA_t is diagonalisable with eigenvalues z≠wz\ne w, so Bj(At)=Bj(z)IB_j(A_t)=B_j(z)I for every jj, a contraction, while ∥At∥≥t∣w−z∣→∞\|A_t\|\ge t|w-z|\to\infty. A uniform test collection would give sim⁡(At)≤L(Ψ,1)\operatorname{sim}(A_t)\le L(\Psi,1), whereas sim⁡(At)≥∥At∥\operatorname{sim}(A_t)\ge\|A_t\|; this is impossible. If instead the joint derivative vanishes at some z0∈D‾z_0\in\overline{\mathbb D}, take At=z0I+tNA_t=z_0I+tN with

N=(0100);N=\begin{pmatrix}0&1\\0&0\end{pmatrix};

then Bj(At)=Bj(z0)I+tBj′(z0)N=Bj(z0)IB_j(A_t)=B_j(z_0)I+tB_j'(z_0)N=B_j(z_0)I, again a contraction, while ∥At∥≥t→∞\|A_t\|\ge t\to\infty.

Corollary 7.4 (Common peripheral collapse). Let E⊂TE\subset\mathbb T be a finite set with at least two points, and suppose that Bj(ζ)=1B_j(\zeta)=1 for every ζ∈E\zeta\in E and every jj. Then Ψ\Psi is not a uniform test collection on D‾\overline{\mathbb D}, even when attention is restricted to Ritt⁡E\operatorname{Ritt}_E operators, with no bound imposed on their Ritt⁡E\operatorname{Ritt}_E constants.

Proof. Take z=ζ1≠ζ2=wz=\zeta_1\ne\zeta_2=w in EE in the first matrix of the preceding proof. Then Bj(At)=IB_j(A_t)=I for all jj, and AtA_t is similar to diag⁡(ζ1,ζ2)\operatorname{diag}(\zeta_1,\zeta_2):

(λI−At)−1=Ptdiag⁡((λ−ζ1)−1,(λ−ζ2)−1)Pt−1,Pt=(1t01).(\lambda I-A_t)^{-1}=P_t\operatorname{diag}((\lambda-\zeta_1)^{-1},(\lambda-\zeta_2)^{-1})P_t^{-1},\qquad P_t=\begin{pmatrix}1&t\\0&1\end{pmatrix}.

Hence dist⁡(λ,E)∥(λI−At)−1∥≤∥Pt∥∥Pt−1∥\operatorname{dist}(\lambda,E)\|(\lambda I-A_t)^{-1}\|\le\|P_t\|\|P_t^{-1}\| for ∣λ∣>1|\lambda|>1, so AtA_t is a Ritt⁡E\operatorname{Ritt}_E operator, whereas sim⁡(At)≥∥At∥≥t∣ζ1−ζ2∣\operatorname{sim}(A_t)\ge\|A_t\|\ge t|\zeta_1-\zeta_2| is unbounded.

Example 7.5 (Three tests, none injective). Fix 0<a<10<a<1 and put B1(z)=z2B_1(z)=z^2, B2(z)=z3B_2(z)=z^3 and B3(z)=(z−a1−az)2B_3(z)=\left(\frac{z-a}{1-az}\right)^2. The pair (z2,z3)(z^2,z^3) is injective on D\mathbb{D}, and its joint derivative vanishes only at 00; since B3′(0)=−2a(1−a2)≠0B_3'(0)=-2a(1-a^2)\ne0, the joint derivative of the triple vanishes nowhere. Thus Corollary 7.2 applies, although no single test is injective.

In this example the constant can be computed. With u=z2u=z^2, v=z3v=z^3 and w=B3(z)w=B_3(z),

z=v+a2(u−1)2+(1−a2)2u−w(1−a2u)22a(1−a2).(52)z=v+\frac{a^2(u-1)^2+(1-a^2)^2u-w(1-a^2u)^2}{2a(1-a^2)}. \tag*{(52)}

Indeed, w(1−a2z2)2=[a(z2−1)+(1−a2)z2]2w(1-a^2z^2)^2=\left[a(z^2-1)+(1-a^2)z^2\right]^2, so the numerator equals −2a(1−a2)z3+2a(1−a2)z-2a(1-a^2)z^3+2a(1-a^2)z, and (52) is an identity. Suppose that sim⁡(Bj(A))≤K\operatorname{sim}(B_j(A))\le K for j=1,2,3j=1,2,3. Then ∥A3∥≤K\|A^3\|\le K and ∥B3(A)∥≤K\|B_3(A)\|\le K; moreover, with q(u)=a2(u−1)2+(1−a2)2uq(u)=a^2(u-1)^2+(1-a^2)^2u, the similarity of A2A^2 to a contraction and von Neumann’s inequality give ∥q(A2)∥≤K∥q∥∞≤K(1+a2)2\|q(A^2)\|\le K\|q\|_\infty\le K(1+a^2)^2 and ∥(I−a2A2)2∥≤K(1+a2)2\|(I-a^2A^2)^2\|\le K(1+a^2)^2. Evaluating (52) at AA therefore gives

∥A∥≤L(a,K):=K+K(1+K)(1+a2)22a(1−a2).\|A\|\le L(a,K):=K+\frac{K(1+K)(1+a^2)^2}{2a(1-a^2)}.

If HH is a Lyapunov operator for A2A^2 with ∥H∥∥H−1∥≤K\sqrt{\|H\|\|H^{-1}\|}\le K, which exists by Proposition 1.1, then Q=H+A∗HAQ=H+A^*HA satisfies Q−A∗QA=H−(A2)∗HA2≥0Q-A^*QA=H-(A^2)^*HA^2\ge0 and H≤Q≤(1+∥A∥2)∥H∥IH\le Q\le(1+\|A\|^2)\|H\|I, so that

sim⁡(A)≤K1+L(a,K)2.(53)\operatorname{sim}(A)\le K\sqrt{1+L(a,K)^2}. \tag*{(53)}

Remark 7.6 (What is used and what is not). In Example 7.5 the bound ∥A3∥≤K\|A^3\|\le K comes from the similarity of B2(A)=A3B_2(A)=A^3, and ∥B3(A)∥≤K\|B_3(A)\|\le K from that of B3(A)B_3(A); neither follows from the similarity of A2A^2 alone. The final Lyapunov operator in (53), however, uses only the similarity of A2A^2 together with the crude norm bound on AA, which is why the constant is so large.

Remark 7.7 (Relation to [1]). Corollary 19 of [1] gives a uniform test collection for injective unramified families of inner functions on finitely connected analytic domains, and the discussion on p. 1017 of [1] leaves the strong version open. Theorem 7.3 settles the strong version on the disc, by an argument that is special to the disc: a single non-constant finite Blaschke coordinate already makes AA similar to a contraction, with a bound involving finitely many powers of AA, and the absorption argument in the proof of Theorem 7.1 upgrades this to a uniform bound.

Declaration of AI use and acknowledgements

All mathematical content and arguments were developed solely by the author. Large Language Models assistance was limited strictly to the writing and proofreading stages of this manuscript. This work was supported by the CDP C2EMPI and its institutional partners through the project R-CDP-24-004-C2EMPI and by the EU COST network OC2025-1-28418.

References

References

  1. [1]M. A. Dritschel, D. Estévez and D. Yakubovich, Tests for complete K-spectral sets, J. Funct. Anal. 273 (2017), no. 3, 984–1019.
  2. [2]M. F. Gamal, Examples of cyclic polynomially bounded operators that are not similar to contractions, Acta Sci. Math. (Szeged) 82 (2016), 597–628.DOI
  3. [3]M. F. Gamal’, A sufficient condition for the similarity of a polynomially bounded operator to a contraction, J. Math. Sci. (N.Y.) 234 (2018), no. 3, 318–329; Russian original: Zap. Nauchn. Sem. POMI 456 (2017), 77–95.DOI
  4. [4]P. R. Halmos, Ten problems in Hilbert space, Bull. Amer. Math. Soc. 76 (1970), 887–933.
  5. [5]A. Kazas and A. Kelley, Clark measures and operators similar to a contraction, Complex Var. Elliptic Equ. 51 (2006), no. 1, 9–31.DOI
  6. [6]T. L. Lance and M. I. Stessin, Multiplication invariant subspaces of Hardy spaces, Canad. J. Math. 49 (1997), no. 1, 100–118.DOI
  7. [7]C. Le Merdy, The similarity problem for bounded analytic semigroups on Hilbert space, Semigroup Forum 56 (1998), no. 2, 205–224.DOI
  8. [8]Yu. Lyubich, Spectral localization, power boundedness and invariant subspaces under Ritt’s type condition, Studia Math. 134 (1999), no. 2, 153–167.DOI
  9. [9]V. Mascioni, Ideals of the disc algebra, operators related to Hilbert space contractions and complete boundedness, Houston J. Math. 20 (1994), no. 2, 299–311.
  10. [10]B. Nagy and J. Zemánek, A resolvent condition implying power boundedness, Studia Math. 134 (1999), no. 2, 143–151.DOI
  11. [11]V. I. Paulsen, Every completely polynomially bounded operator is similar to a contraction, J. Funct. Anal. 55 (1984), 1–17.DOI
  12. [12]G. Pisier, A polynomially bounded operator on Hilbert space which is not similar to a contraction, J. Amer. Math. Soc. 10 (1997), no. 2, 351–369.arxiv.org/abs/math/9602207
  13. [13]G. Pisier, Similarity Problems and Completely Bounded Maps, Lecture Notes in Mathematics 1618, Springer, second edition, 2001.
  14. [14]M. I. Stessin, Wold decomposition of the Hardy space and Blaschke products similar to a contraction, Colloq. Math. 81 (1999), no. 2, 271–284.DOI

Paper details

vlnc:W_5afrerva
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 5

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