Harmonic analysis of operators on Hilbert space (Q2637956): Difference between revisions

\textit{F. Riesz} and \textit{B. Sz.-Nagy} [Léçons d'analyse fonctionnelle. 2.~éd. (French). Budapest: Akadémiai Kiadó (Éditions de l'Académie) (1953; Zbl 0051.08403)] dealt with linear operators which are selfadjoint, unitary or normal, but not much was known about non-normal operators. The first edition got to parts hard to reach (and is hard to read), like [\textit{G. A. Heineken}, Amsterdam, 1863]. \textit{B. Sz.-Nagy} [Unitary dilations of Hilbert space operators and related topics. Conference Board of the Mathematical Sciences. Regional Conference Series in Mathematics 19. Providence, R. I.: American Mathematical Society (1974; Zbl 0281.47004)] gives a digested read of the earlier editions. The first eight chapters of the new edition are essentially the same as in previous editions (originally French [(1967; Zbl 0157.43201)] with subsequent English [(1970; Zbl 0201.45003)] and Russian [(1970; Zbl 0202.13102)] translations). The original Chapter IX on contractions \textit{similar} to a unitary operator has been dispersed amongst the earlier chapters. Several chapters conclude with notes on other developments related to the earlier material. It is inconvenient and annoying that the pagination in the indices is almost always inaccurate (though never far off). The Sz.-Nagy dilation theorem [``Sur les contractions de l'espace de Hilbert'', Acta Sci. Math. 15, 87--92 (1953; Zbl 0052.12203)] showed that every contraction \(T\) on a separable complex Hilbert space \(\mathfrak{h}\) (possibly mapping to a different Hilbert space \(\mathfrak{h}'\)) has a unitary, or unital, dilation \(U\) on a Hilbert space \(\mathfrak{k} \supset \mathfrak{h}\), such that \(T^{n} = P_{\mathfrak{h}} U^{n}P_{\mathfrak{h}}\), where \(n \in \mathbb{N}\), and that there is a \textit{minimal unital dilation} unique up to isometric isomorphism. The same applies to the adjoint \(T^{*}\). This theorem, based on a result of \textit{M. A. Naĭmark} [C. R. (Dokl.) Acad. Sci. URSS, n. Ser. 41, 359--361 (1943; Zbl 0061.25410)], led to a significant new emphasis for Hilbert space operator theory. In what follows, \(U\) will denote the minimal dilation of \(T\). In their Chapter V, the authors deal with analytic operator-valued functions denoted \(\Theta(\lambda) = \sum \lambda^{n} \theta_n\), where the \(\theta_{n}\) are contractions. For the one-dimensional case, see \textit{I. Schur} [J. Reine Angew. Math. 147, 205--232 (1917); ibid. 148, 122--145 (1918; JFM 46.0475.01)] associating a scalar analytic function on the unit disk with a sequence of complex numbers in the unit disk. \textit{B. Sz.-Nagy} and \textit{C. Foiaş} [C. R. Acad. Sci., Paris 256, 3236--3238 (1963; Zbl 0151.19504)] became concerned with operator-valued \textit{analytic characteristic functions} \(\Theta_{T}(\lambda)\), \( \lambda\) in the spectrum \(\Lambda_{T} \subset\mathbb{C}\). These are functional models and play the role of a \textit{resolvent} relative to \(\Lambda_{T}\) in the sense that it is an analytic function sneaking up on the spectrum, the problem for non-normal operators being existence of essential singularities. The characteristic functions can be easier to manage than \(T\) itself. Usually, \(T\) is assumed to be completely non-unitary (i.e., projections to subspaces are also non-unitary; see \textit{H. Langer} [Acta Math. Acad. Sci. Hung. 12, 441--445 (1961; Zbl 0102.10302)]). A general contraction can be expressed as an orthogonal sum of such a \(T\) with some unitary unitary operator [\textit{J. von Neumann}, Math. Ann. 102, 49--131 (1929; JFM 55.0824.02)]. One calls \(D_T = (I-T^*T)^{1/2}\) the \textit{defect operator}, while \(\mathfrak{D}_T\), the closure of its range, is the \textit{defect space}. There is also a \textit{defect index} which characterises \(T\) to be isometry when the index is zero. The characteristic function is so constructed that it is a bounded operator from \(\mathfrak{D}_T\) to \(\mathfrak{D}_{T^*}\) and is a contraction when \(\lambda\) is restricted to be inside the unit disk. Matrix-valued characteristic functions date back to the 1940s, see \textit{M. S. Livšic} [Mat. Sb., N. Ser. 19(61), 239--262 (1946; Zbl 0061.25903)] and his subsequent articles on non-selfadjoint and quasi-unitary operators. Livšic, who originally set out to be a radio technician, was concerned with both mathematical and engineering applications. \textit{J. W. Helton} [Indiana Univ. Math. J. 22, 403--414 (1972; Zbl 0246.47019)] is a good reference for engineering applications. The entries of the matrices used here were rational polynomials, cf.\ also \textit{C. Foiaş} [Actes Congr. internat. Math. 1970, tom. 2, 433--440 (1978; Zbl 0232.47025)] which deals with scattering of plane waves by layered media, used in geophysical prospecting. Shift operators play a concrete role in the theory. The \textit{Hardy space} \(H^2(\text{D})\), where \(\text{D} \subset \mathfrak{h}\), is the set of Fourier series \(\sum_{n \geq 0} u_{n} e^{int}\), \(0 \leq t < 2\pi\), where \(u_{n} \in \text{D}\) with finite \(\sum \|u_{n}\|^2\). Unilateral shift operators \(T\), so-called because for \(u \in H^2\) on the Fourier coefficients \((Tu)_{n} = u_{n-1}\), \(n\geq 1\), are isometric operators. The adjoint is the left or backward shift. Bilateral shifts are similarly defined for \(n \in \mathbb{Z}\) and are unitary. \textit{G-C. Rota} [Commun. Pure Appl. Math. 13, 469--472 (1960; Zbl 0097.31604)] showed that that every (linear) Hilbert space operator with spectrum inside the unit disk has a category theory (for operators with intertwinings as morphisms) \textit{model}; this model is similar to the restriction of the adjoint of the infinite-dimensional unweighted unilateral shift restricted to a suitable invariant subspace (cf.\ [\textit{B. Sz.-Nagy}, Publ. Math. Inst. Hung. Acad. Sci. 4, 89--93 (1959; Zbl 0086.31802)] which shows that a compact operator with spectrum on the closed unit disk and whose iterates are uniformly bounded is similar to a contraction). See also \textit{C. Foiaş} in [Comun. Acad. Republ. Popul. Romine 13, 349--352 (1963; Zbl 0122.11601)], where he remarks in his abstract that Rota's result applies only to contractions such that \(T^{n} \to 0\), presumably relying on the spectral radius being less than \(1\). Also, the condition that \(T ^{n} \to 0\) is equivalent to \(U\) being a bilateral shift of \(\overline{U - T}\mathfrak{h}\subset\mathfrak{k}\); if the spectrum of \(T\) does not cover the unit circle, then \(U\) can be resolved into a part like bilateral shift and a part with spectrum contained in \(\Lambda_T\). In a similar vein, the authors show in VI \S 6 that there exist contractions \(T\) for which \(U\) is not a bilateral shift. The characteristic function is an analogue of the (analytic) characteristic function of a probability measure, though there is in general no natural corresponding convolution product of operators. Also, the unilateral shift operators are non-normal and hence do not correspond to measures. The authors do, however, deal with factorisations of operators and their characteristic functions. These can be related to the existence of invariant spaces for the contractions, but not all factorisations are linked to invariant spaces. They also deal with semigroups of contractions. There is generally no Fourier transform, but instead a \textit{Fourier representation} similar to a discrete Fourier expansion. For a functional calculus, one considers functions \(z \mapsto u(z)\), \(z \in\mathbb{C}\), which are holomorphic, bounded inside the unit disk and almost-everywhere defined on the outer circle. Suppose that \(u\) is defined at a point \(e^{i \theta_0}\) and let \(r \to 1\) while \(z = re^{i\theta_0}\), so that the limit of \(\Theta_{T} (re^{it})\) exists in the strong operator sense for almost every \(t \in (0,1)\) and is denoted by \(\Theta_{T}(e^{it})\). The `Fourier series' representation of \(u(T)\) follows the expansion of \(u(z)\), \(u \in H^2\), keeping in mind that \(T^{n}\) is the projection of \(U_{T^n}\) to the domain of \(U_{T}\) in \(\mathfrak{h}\). The operational calculus extends to functions \({u(z)} / {v(z)}\) and their limits, whenever \(v(T)\) is invertible. Nontrivial concrete examples of infinite-dimensional characteristic functions are scarce. Some multiplication operators like \(f \mapsto \phi(z)f\), \(f\) in \(L^2\) or \(H^2\), \(\phi\) analytic and \(z\) inside the unit disk, exist but can be difficult. The characteristic function of an infinite-dimensional (unweighted) unilateral shift operator \(T\) has characteristic function identically zero; indeed, a multiplication operator of the form \(f \mapsto zf\) has empty defect space and \(T^*\), the backward shift, has defect space the unit circle (cf.\ \textit{P. R. Halmos} [A Hilbert space problem book. 2nd ed., rev. and enl. Graduate Texts in Mathematics 19. New York--Heidelberg--Berlin: Springer-Verlag (1982; Zbl 0496.47001), Problem 82]). Shifts can be used as building blocks. However, most everybody would apply the Nagy-Foiaş definition of the characteristic function, being \(\Theta_{T}(\lambda) = \{-T + \lambda D_{T^*}(I - \lambda T^*)^{-1} D_T\}|_{\mathfrak{D}_T}\). In the book, this is introduced out of the blue. No derivation is given and at first sight the formula seems to have been drawn out of a hat. Nevertheless, the derivation can be found in Professor Livšic's hat; see [\textit{M. S. Livšic}, Dokl. Akad. Nauk SSSR, II. Ser. 58, 13--15 (1947; Zbl 0031.16701); Mat. Sb., N. Ser. 26(68), 247--264 (1950; Zbl 0040.35305)] and \textit{M. S. Livšic} and \textit{V. P. Potapov} [Dokl. Akad. Nauk SSSR, n. Ser. 72, 625--628 (1950; Zbl 0040.35401)] for a derivation of the formula for isometric operators with equal deficiency indices. Livšic started with the operator of multiplication on \(\mathbb{R}\) by a scalar \(\alpha\), its characteristic function operator is of the form \({\lambda - \alpha } \over {\lambda - \overline{\alpha}}\). He approximated a characteristic function by tensor products of matrices; he split his Hilbert space \(\mathbb{H}\) into a direct sum of a continuous part and a discrete part. He then modelled the characteristic function in a canonical triangular form which plays the role of a spectral resolution, cf.\ [Dokl. Akad. Nauk SSSR, n. Ser. 84, 873--876 (1952; Zbl 0046.33903)]. Assuming \(\mathfrak{h}\) to be separable, the authors generalise Livšic's construction of matrix and classes of operator-valued characteristic functions. They construct a completely nonunitary model \(\underline{T}\) of \(T\) so that it has characteristic function unitarily equivalent to that of \(T\) whenever \(\Theta(0)\) is bounded and \( \| \Theta f(0) \| \leq \|f\|\). Consider an analytic operator-valued function \(\Theta = \sum_{n \geq 0} \lambda^n e^{\theta_n}\) (cf.\ Schur [loc.\,cit.]), where each \(\theta_{n}\) is a contraction mapping \(\mathcal{E} \subset \mathfrak{h}\) to \(\mathcal{E}_{*} \in \mathfrak{h}\). When \(\mathcal{E} = \overline{( U - T) \mathfrak{h}}\), it is a \textit{wandering}, or ambulant, space for \(U\), i.e., the images of the \(U^{n}\) are mutually orthogonal, and similarly \(\mathcal{E}_{*}\) is wandering for \(T^{*}\). Denote the closure of orthogonal sum of \(\{ h, Uh, U^{2}h, \dots \}\) by \(\mathfrak k_{+}\), it being unitarily equivalent to \(H^2(D_{T^*})\). The authors denote \(\mathfrak{k} \ominus \{ \bigoplus _{n} U^{n} \mathcal{E} \}\) as \(\mathfrak{M}\) and use the defect operators and repeated projections onto mutually orthogonal subspaces of \(\mathfrak{h}\) to express a contraction \(T\) in terms of subcontractions of the decomposed portions of \(\mathfrak{h}\). There will be a canonical decomposition into a direct sum of four subcontractions. This provided a `triangulation' of \(T\); see [\textit{B. Sz.-Nagy} and \textit{C. Foiaš}, Acta Sci. Math. 25, 12--37 (1964; Zbl 0138.38903)] (cf.\ I \S 5 and VI \S 2 of the book under review). The authors define \(\overline{\Delta_{T} L^{2}(\mathcal{E})}\) canonically as the closure in \(L^2(\mathcal{E})\) of \(\{v :\Delta(t) v(e^{it})\}\), \(v \in L^2(\mathcal{E})\), where \(\Delta(t)=(1 - \Theta_t^* \Theta_t)^{1/2}_{\mathfrak{D}_T}\) is the bounded selfadjoint defect operator for \(\Theta\). These \(\Delta\) terms are rather complicated, but they fall away for \(T\) such that \(T^{n} \to 0\). They define a Hilbert space \(\mathbb{H}\) which is a direct sum of orthogonal spaces \(\{\{H^2(\mathcal{E}_{*}) \oplus \overline{L^2_E} \}\ominus \{ \Theta (e^{it})u(e^{it})\oplus \Delta(t) u(e^{it}) \}\), \(u \in H^{2}(\mathcal{E})\}\). Then \(\mathfrak k_{+}\) is mapped onto \(H^2(E_{*}) \oplus \overline{\Delta_{E}L^2(E)}\). The construction is such that, when restricted to \(H^2(\mathfrak{D}_{T^*})\), then \(U\) is equivalent to the adjoint of the operation of multiplication by \(e^{it}\) and, when restricted to \(\overline{\Delta_{T} L^2(\mathfrak{D}_T)}\), then it is equivalent to a multiplication by \(e^{-it}\). The model \(\underline{T}\) maps \(u \oplus v\) to \([e^{-t} u^{it} - u_{*}(0)] \oplus e^{-it}v(t)\), where \(u \oplus v\) is defined on \(H^{2}(E_{*}) \oplus \overline{\Delta_{\mathcal{E}}L^2(E)}\). The last two chapters elaborate on operator classes introduced in Chapter II, namely, the classes \(C_{\alpha,\beta} = C_{\alpha \cdot} \cap C_{\cdot \beta}\), where \(\alpha\) and \(\beta\) are \(0\) or \(1\); \( C_{0\cdot}\) and \(C_{1\cdot}\) are defined such that, respectively, \(T^{n}x \to 0\) and \(T^{n}x \to 0\) for all nonzero \(x \in \mathfrak{h}\). The classes with reversed order of \(\alpha\) and \(\beta\) are likewise defined using \(T^{*}\). Every contraction can be written in terms of triangular blocks of such operators (Theorem II 4.1). \textit{A. Beurling} [Acta Math., Uppsala 81, 239--255 (1949; Zbl 0033.37701)], concerned with spectra of contraction operators on Hilbert space, introduced the concept of inner and outer holomorphic functions in the unit circle. A~bounded holomorphic function \(f\) on the unit disk is called an \textit{inner function} if \(f^*(e^{it})=1\), where \(f^*\) denotes the radial limit of \(f\). A~characteristic function is inner if it is almost surely isometry-valued on the unit circle. The outer functions are more complicated, but are such that \(\log |f (e^{i\theta} )| = \kappa (e^{i\theta})\) almost everywhere, \(\kappa\) being a suitable real-valued integrable function; as it is an exponential, an outer function cannot have zeros inside the disk. Then \(T\) is in \(C_{\cdot 0}\) or \(C_{\cdot 1}\) if and only if \(\Theta_{T}\) is inner or outer, respectively. For \(C_{0 \cdot}\) and \(C_{1 \cdot}\), inner/outer correspond to the adjoint of the characteristic function. A~characteristic function is inner it if is almost surely isometry-valued on the unit circle and outer if and only if \(\Theta H^2(\mathcal {E})\) is everywhere dense in \(H^2(\mathcal{E}_*)\). Asymptotic properties of the \(T^n\) relate to whether the characteristic function is inner or outer. Chapter IX deals with the \(C_{1 \cdot}\) classes. \textit{B. Sz.-Nagy} [Acta Univ. Szeged., Acta Sci. Math. 11, 152--157 (1947; Zbl 0029.30501)] proved that every power bounded operator with power bounded inverse is similar to a unitary operator. L.\,Kérchy built this idea to construct what he called \textit{unitary} and \textit{isometric asymptotes} of a power bounded operator \(T\). These are operators on the Hilbert space completion of \(\mathfrak{h}\) with inner product provided by the sesquilinear form \(\lim_{n}(T^n x,T^n y)\), \(x, y \in \mathfrak{h}\). He writes the unitary asymptote as a pair \(\{ (X,V): XT = WX \}\), where \(W\) is unitary on \(\mathfrak{k}\) and \(X\) is a contraction embedding \(\mathfrak{h}\) into some \(\mathfrak{k}\). The mapping \(X\) is 1-1 if and only if \(T \in C_{1 \cdot}\). The isometric asymptote is constructed in a similar way. The unitary asymptote will be the residual part (see II \S 2) of the unital dilation of the isometric asymptote. The authors compare the spectra of \(T\) and \(W\); for instance, the spectrum of \(T\) contains the spectrum of \(W\). A quasiaffinity from Hilbert space \(\mathfrak{h}_1\) to a Hilbert space \(\mathfrak{h}_2\) is defined to be a 1-1 continuous transformation from \(\mathfrak{h}_1\) to an \textit{everywhere dense} manifold in \(\mathfrak{h}_2\). Operators \((S_{1},\mathfrak{h}_1)\) and (\(S_{2}, \mathfrak{h}_2\)) are called \textit{quasisimilar} if they are quasiaffine transformations of each other (II \S 2). Hyperinvariant spaces (III \S 7) for a contraction \(T\) are those subspaces of \(\mathfrak{h}\) that are invariant under every operator commuting with \(T\). Conditions are given for that \(T\) has a nontrivial hyperinvariant subspace. For the class \(C_{1 1}\), the authors consider as to when an operator is determined by its lattice of invariant subspaces. They also establish the existence of an infinite collection of pairwise disjoint hyperinvariant subspaces. Lattices of invariant subspaces and lattices of hyperinvariant spaces are related; they connect the hyperinvariant subspace lattice of \(T\) and of the hyperinvariant subspace lattice of of the adjoint of the isometric asymptote to get a bijection and lattice homomorphisms for the commutants of \(T\) and of \(W\). Chapter X concerns the \(C_{0 \cdot}\) classes. Characteristic functions for \(C_{0 0}\) operators are continuous analogues of scattering matrices. Part of this chapter concerns unicellularity and Jordan operators. Jordan cells, or blocks, are matrices with one-point spectrum. A unicellular operator will have only one eigenvector up to a scalar factor. In the finite-dimensional case, such operators have a matrix consisting of a single Jordan block. A~Jordan operator will be a direct sum of Jordan blocks. An operator on a Hilbert space is seen to be unicellular if and only if its lattice of invariant subspaces is linearly ordered. The subclass \(C_{0} \subset C_{0 0}\) consists of contractions \(T\) such that there exists a nonzero analytic function \(u\) on the disk with \(u(T) = 0\) and \(u(z) \neq 0\). It was already known for separable Hilbert spaces that a \(C_0\) operator with finite multiplicities is similar to a unique Jordan operator. One of the main results in this section, due to \textit{H. Bercovici} [Stud. Math. 60, 267--284 (1977; Zbl 0355.47006)], is that, for an arbitrary Hilbert space, every \(C_0\) operator is quasi-similar to a unique Jordan operator.