PERSPECTIVE Fuglede–Kadison determinant: theme and variations Pierre de la Harpe1 Section de Mathématiques, Université de Genève, CH 1211 Genève 4, Switzerland
Edited by Bent Fuglede, University of Copenhagen, Copenhagen, Denmark and accepted by the Editorial Board April 10, 2013 (received for review April 6, 2013)
We review the definition of determinants for finite von Neumann algebras, due to Fuglede and Kadison [Fuglede B, Kadison R (1952) Ann Math 55:520–530], and a generalization for appropriate groups of invertible elements in Banach algebras, from a paper by Skandalis and the author (1984). After some discussion of K-theory and Whitehead torsion, we indicate the relevance of these determinants to the study of L2-torsion in topology. Contents are as follows: 1. The classical setting. 2. On von Neumann algebras and traces. 3. Fuglede–Kadison determinant for finite von Neumann algebras. 4. Motivating question. K K K top 5. Brief reminder of 0, 1, 1 , and Bott periodicity. 6. Revisiting the Fuglede–Kadison and other determinants. 7. On Whitehead torsion. 8. A few lines on L2-torsion.
1. Classical Setting 1.1. Determinants of Matrices over Commutative Rings. Let R be a ring with unit. For an integer n ≥ 1, denote by MnðRÞ the ring of n-by-n matrices over R and by GLnðRÞ its group of units. R* stands for GL1ðRÞ. Suppose R is commutative. The determinant
det :M nðRÞ → R [1]
is defined by a well-known explicit formula, a polynomial in the matrix entries. It is alternate multilinear in the columns of the matrix and
normalized by detð1nÞ = 1; when R is a field, these properties constitute an equivalent definition, as was lectured on by Weierstrass and Kronecker (probably) in the 1860s and published much later (ref. 1, p. 291; published 12 years after Kronecker’sdeath). For x; y ∈ MnðRÞ,wehavedetðxyÞ = detðxÞdetðyÞ.Forx ∈ MnðRÞ with detðxÞ invertible, an explicit formula shows that x itself is invertible, so that detðxÞ ∈ R* if and only if x ∈ GLnðRÞ. The restriction
GLnðRÞ → R*; x ↦ det x [2] is a group homomorphism.
1.2. Three Formulas for Complex Matrices Involving Determinants, Exponentials, Traces, and Logarithms. Suppose that R is the field C of complex numbers. The basic property of determinants that we wish to point out is the relation
detðexp yÞ = expðtraceðyÞÞ for all y ∈ MnðCÞ: [3]
Some expository books present this as a very basic formula (ref. 2, section 16); it reappears below as Eq. 20. It can also be written as
detðxÞ = expðtraceðlog xÞÞ for appropriate x ∈ GLnðCÞ: [4] “Appropriate” can mean several things. If kx − 1k < 1, then log x can be defined by the convergent series
X∞ k−1 ð−1Þ log x = logð1 + ðx − 1ÞÞ = ðx −1Þk: k = 1 k If x is conjugate to a diagonal matrix, then log x can be defined component-wise (in pedantic terms, this is functional calculus, justified by the spectral theorem). In [4], note that the indeterminacy in the choice of the logarithm of a complex number is swallowed by the exponential, because exp 2πi = 1.
Author contributions: P.H. performed research, analyzed data, and wrote the paper.
The author declares no conflict of interest. This article is a PNAS Direct Submission. B.F. is a guest editor invited by the Editorial Board. 1E-mail: [email protected].
15864–15877 | PNAS | October 1, 2013 | vol. 110 | no. 40 www.pnas.org/cgi/doi/10.1073/pnas.1202059110 Downloaded by guest on October 2, 2021 Let x ∈ GLnðCÞ. Because the group is connected, we can choose a piecewise smooth path ξ : ½0; 1 → GLnðCÞ from 1 to x. Because PERSPECTIVE _ −1 log ξðαÞ is a primitive of ξðαÞξðαÞ dα, it follows from [4]that Z1 ! _ −1 detðxÞ = exp trace ξðαÞξðαÞ dα: [5] 0
This is our motivating! formula for section 6, and in particular for Eq. 18. The sign = stands for a genuine equality, but indicates that some comment is in order. A priori, the integral depends on the choice of ξ, and we have also to worry about the determination of log ξðαÞ. As there is locally no obstruction to choosing a continuous determination _ −1 of the primitive log ξðαÞ of ξðαÞξðαÞ dα, the integral is invariant under small changes of the path (with fixed ends) and therefore depends only on the homotopy class of ξ,sothatitisdefined modulo its values on (homotopy classes of) closed loops. The fundamental group π1ðGLnðCÞÞ is infinite cyclic, generated by the homotopy class of 2πiα ξ : ; → ; α ↦ e 0 ; 0 ½0 1 GLnðCÞ 01n−1 R 1 ξ_ α ξ α −1 α = π fi π and we have 0 traceð 0ð Þ 0ð Þ Þd 2 i. Consequently, the integral in the right-hand side of [5]isde ned modulo 2 iZ, so that the right-hand side itself is well defined. (This is repeated in the proof of Lemma 10.) ; ...; ∈ = Because a connected group is generated by each neighborhood of the identity, there exist x1 xk GLnðCÞ such that x x1⋯xk and kxj − 1k < 1forj = 1; ...; k, and one can choose
ξ α = α ð Þ exp ð ðlog x1ÞÞ⋯exp ðαðlog xkÞÞ: A short computation with this ξ gives Z1 ξ_ α ξ α −1 α = exp trace ð Þ ð Þ d expðtraceðlog x1ÞÞ⋯expðtraceðlog xkÞÞ 0
and it is now obvious that [4] implies [5].
1.3. Historical Note. Determinants arise naturally with linear systems of equations, first with R = R and more recently also with R = C. They have a prehistory in Chinese mathematics from the second century B.C. (3). In modern Europe, there has been an early con- † tribution by Leibniz in 1693 , unpublished until 1850. Gabriel Cramer wrote an influential book, published in 1750. Major mathe- maticians who have written about determinants include Bézout, Vandermonde, Laplace, Lagrange, Gauss, Cauchy, Jacobi, Sylvester, Cayley, and others. The connection between determinants of matrices in M3ðRÞ and volumes of parallelepipeds is often attribued to Lagrange (1773). Let us mention an amazing book on the history of determinants, ref. 4: four volumes, altogether more than 2,000 pages, an ancestor of the Mathematical Reviews, for one subject, covering the period 1693–1900. There is an extension of [1]toaskew-field k by Dieudonné, where the range of the mapping det defined on MnðkÞ is ðk*=Dk*Þ⊔f0g,where the notation “DΓ” denotes the group of commutators of a group Γ (see refs. 5 and 6, and also ref. 7 for a discussion of when k is the skew field of Hamilton quaternions). The theory of determinants, in the case of a noncommutative ring R, has motivated a lot of work, in particular by Gelfand and coauthors since the early 1990s (8). Let us also mention a version for supermathematics due to Berezin (ref. 9 and ref. 10, chap. 3), as well as “quantum determinants”, of interest in low-dimensional topology (see, for example, ref. 11). The notion of determinants extends to matrices over a ring without unit (by “adjoining a unit to the ring”). In particular, in functional analysis, there is a standard notion of determinants that appears in the theory of Fredholm integral equations, for example for operators on a Hilbert space of the form 1 + x,wherex is the “trace class” (12, 13). TheoldestoccurrenceIknowofexp y or log x, including the notation, definedbythefamiliarpowerseriesinthematrixy or x − 1, is in ref. 14, p. 374; see also ref. 15. However, exponentials of linear differential operators appear also early in Lie theory, see for example ref. 16 (p. 75) and ref. 17 (p. 82), even if Lie never uses a notation like exp X (unlike Poincaré) (see his eαX inref.18,p.177). There is a related and rather old formula known as the “Abel–Liouville–Jacobi–Ostrogradskii identity”. Consider a homogeneous linear n differential equation of the first order y′ðtÞ = AðtÞyðtÞ, for an unknown function y : ½t0; t1 → R . The columns of a set of n linearly independent solutions constitute the Wronskian matrix WðtÞ. It is quite elementary (at least nowadays!) to show that W′ðtÞ = AðtÞWðtÞ, hence d det WðtÞ =traceðAðtÞÞdet WðtÞ, and therefore dt ! Zt
det WðtÞ = det Wðt0Þexp traceðAðsÞÞds ;
t0
a close cousin of Eq. 5. The name of this identity refers to Abel (1827, case n = 2), Liouville (19), Ostrogradskii (1838), and Jacobi (1845). This was pointed out to me by Gerhard Wanner (ref. 20, section I.11); also, Philippe Henry showed me this identity on the last five lines of ref. 21 (which does not contain references to previous work).
† There are also resultants and determinants in the work of the Japanese mathematician Seki Takakazu, a contemporary of Leibniz and Newton.
de la Harpe PNAS | October 1, 2013 | vol. 110 | no. 40 | 15865 Downloaded by guest on October 2, 2021 Finally, a few words are necessary about the authors of the 1952 paper alluded to in our title. Bent Fuglede is a Danish mathematician born in 1925. He has been working on mathematical analysis; he is also known for a book on Harmonic Maps Between Riemannian Polyhedra (coauthor Jim Eells, preface by Misha Gromov). Richard Kadison is an American mathematician, born in this same year, 1925. He is known for his many contributions to operator algebras; his “global vision of the field was certainly essential for my own development” (words of Alain Connes, when Kadison was awarded the Steele Prize in 1999 for Lifetime Achievement, ref. 22, p. 461).
1.4. Plan. Section 2 is a review of von Neumann algebras based on three types of examples, section 3 is an exposition of the original Fuglede– Kadison idea, section 4 stresses the difference between the complex-valued standard determinant and the real-valued Fuglede–Kadison determinant, and section 5 is a review of some notions of K-theory. Section 6 exposes the main variations of our title: determinants defined for connected groups of invertible elements in complex Banach algebras. We end by recalling in section 7 a few facts about Whitehead torsion, 2 with values in WhðΓÞ, which is a quotient of the group K1 of a group algebra Z½Γ , and by alluding in section 8 to L -torsion, which is defined in terms of (a variant of) the Fuglede–Kadison determinant.
2. On von Neumann Algebras and Traces In a series of papers from 1936 to 1949, Francis Joseph Murray and John von Neumann founded the theory of von Neumann algebras (in their terminology “rings of operators”), which are complex *-algebras representable by unital weakly closed *-subalgebras of some LðHÞ, the algebra of all bounded operators on a complex Hilbert space H. We first give three examples of pairs ðN ; τÞ,withN a finite von Neumann algebra and τ a finite trace on it. We then recall some general facts and define a few terms, such as “finite von Neumann algebra”, “finite trace”,and“factor of type II1”. ≥ fi Example 1 (factors of type In). For any n 1, the matrix algebraP MnðCÞ is a nite von Neumann algebra known as a factor of type In. n = ; ↦ 1 ; The involution is given by ðx*Þj;k xk j. The linear form x n j=1xj j is the (unique) normalized trace on MnðCÞ. Example 2 (Abelian von Neumann Algebras). Let Z be a locally compact space and ν be a positive Radon measure on Z. The space L∞ðZ; νÞ of complex-valued functions on Z that are measurable and ν-essentially bounded (modulo equality locally ν-almost everywhere) is an abelian von Neumann algebra. The involution is given by f *ðzÞ = f ðzÞ. Any abelian von Neumann algebra is of this form. R ∞ ν τν : ↦ ν ; ν τν = If is a probability measure, the linear form f Z f ðzÞd ðzÞ is a trace on L ðZ Þ, normalized in the sense ð1Þ 1. Example 3 (Group von Neumann Algebra). Let Γ be a group. The Hilbert space ℓ2ðΓÞ has a scalar product, denoted by h·j·i, and a canonical 2 orthonormal basis ðδγÞγ∈Γ, where δγðxÞ is 1 if x = γ and 0 otherwise. The left-regular representation λ of Γ on ℓ ðΓÞ is defined by − ðλðγÞξÞðxÞ = ξðγ 1xÞ for all γ; x ∈ Γ and ξ ∈ ℓ2ðΓÞ. P 2 finite The von Neumann algebra NðΓÞ of Γ is the weak closure in Lðℓ ðΓÞÞ of the set of C-linear combinations γ∈Γ zγλðγÞ; it is a finite von λ γ * = λ γ−1 ↦ δ δ NeumannP algebra. The involution is given by ðzγ ð ÞÞ zγ ð Þ. There is a canonical trace, given by x hx 1j 1i, which extends finite γ∈Γ zγλðγÞ↦ z1. ‡ Moreover, NðΓÞ is a factor of type II1 if and only if Γ is icc (this is lemma 5.3.4 in ref.23;see also ref. 24, chap. III, section 7, no.6). Remarks:
a) In the special case of a finite group, NðΓÞ of Example 3 is a finite sum of matrix algebras as in Example 1. In the special case of an abelian group, NðΓÞ of Example 3 is isomorphic, via Fourier transform, to the algebra of Example 2, with Z the Pontrjagin dual of Γ (which is a compact abelian group) and ν its normalized Haar measure.
b) The von Neumann algebra NðΓÞ is “of type I” if and only if Γ has an abelian subgroup of finite index (25). It is “of type II1” if and § only if either ½Γ : Γf = ∞ or ½Γ : Γf < ∞ and jDΓf j = ∞ (26). There exist groups Γ such that NðΓÞ is a nontrivial direct product of two two-sided ideals, one of type I and the other of type II1; see theorem 2 of ref. 28 for the result, and ref. 29 for explicit examples.
c) Suppose, in particular, that Γ is finitely generated. If Γf is of finite index in Γ, then Γf is also finitely generated and it follows that Γ has an abelian subgroup of finite index. Thus, NðΓÞ is either of type I (if and only if Γ has a free abelian group of finite index) or of type II1 (if and only if ½Γ : Γf = ∞) (29). d) Other properties of NðΓÞ are reviewed in ref. 30. Let us now recall, as promised, some general facts and some terminology: i) A von Neumann algebra N inherits several natural topologies from its representations by operators on Hilbert spaces, including the “ultraweak topology” (with respect to which the basic examples are separable) and the “operator topology” (with respect to which N is separable if and only if it is finite dimensional).
‡ A group is icc if it is infinite and if all its conjugacy classes distinct from f1g are infinite. § We denote by Γf the union of the finite conjugacy classes of a group Γ. It is easy to check that Γf is a subgroup, and it is then obvious that it is a normal subgroup.
15866 | www.pnas.org/cgi/doi/10.1073/pnas.1202059110 de la Harpe Downloaded by guest on October 2, 2021 ii) There is available a functional calculus, justified by the spectral theorem: f ðxÞ is well defined and satisfies natural properties, for x ∈ N PERSPECTIVE normal ðx*x = xx*Þ and f an essentially bounded complex-valued measurable function on the spectrum spðxÞdfz ∈ Cjz − x is not invertibleg
of x. More precisely, at least when N acts on a separable Hilbert space, we have for x normal in N a positive regular Borel measure of full support ν on spðxÞ and a natural injective morphism L∞ðspðxÞ; νÞ ∋ f ↦f ðxÞ ∈ N of von Neumann algebras.
A von Neumann algebra N is finite if, for x; y ∈ N ,therelationxy = 1impliesyx = 1. A projection in a von Neumann algebra is a self-adjoint idempotent, in equations, e = e*= e2. A von Neumann algebra N is of type I if, for any projection 0 ≠ e ∈ N , there exists a projection f ∈ N , f ≠ 0suchthatfe = ef = f and f N f is abelian. A finite von Neumann algebra N if of type II1 if, for any projection 0 ≠ e ∈ N ,thesubset eN e is not abelian. It is known that any finite von Neumann algebra is the direct product of a finite algebra of type I and an algebra of type II1. iii)AfinitetraceonavonNeumannalgebraN is a linear functional τ : N → C that is continuous with respect to all of the standard topologies on N (i.e., which is normal, in the standard jargon) and satisfies τ = τ ∈ iiia) ðx*Þ ðxÞ for all x N , τ ≥ ∈ iiib) ðx* xÞ 0, for all x N , and τ = τ ; ∈ iiic) ðxyÞ ðyxÞ for all x y N . A trace is faithful if τðx*xÞ > 0 whenever x ≠ 0.
It is known that, on a finite von Neumann algebra that can be represented on a separable Hilbert space, there exists a faithful finite normal fi trace. Also, any linear form on N that is ultraweakly continuous and satis es iiic can be written canonically as a linear combination of four linear forms satisfying the three conditions of iii; this is a Jordan decomposition result of ref. 31. As we do not consider other kinds of traces, we use trace for finite trace below. A factor is a von Neumann algebra of which the center coincides with the scalar multiples of the identity. A factor of type II1 is an infinite dimensional finite factor; the discovery of such factors is one of the main results of Murray and von Neumann.
iv) Let N be a factor of type II1; • N is a simple ring{, ║ • there is a unique normalized trace τ, which is faithful.
Thus, on a factor N of type II1, it is a standard result that there exists a unique normalized normal trace τ (in the sense of iii above); but uniqueness holds in a stronger sense, because any element in the kernel of τ is a finite sum of commutators (32). For a projection e,thenumberτðeÞ is called the von Neumann dimension of e, or of the Hilbert space eðHÞ,whenN is understood to be inside some LðHÞ.
3. Fuglede–Kadison Determinant for Finite von Neumann Algebras In 1952, Fuglede and Kadison defined their determinant ( * GL1ðN Þ → R+ FK : detτ 1 ; [6] x ↦ exp τ log ðx*xÞ2
FK FK which is a partial analog of [2]. The number detτ ðxÞ is well defined by functional calculus, and most of the work in ref. 33 is to show that detτ is a homomorphism of groups. For the definition given below in section 6, it will be the opposite: some work to show that the definition makes sense, but a very short proof to show it defines a group homomorphism. In the original paper, N is a factor of type II1,andτ is its unique trace with τð1Þ = 1; but everything carries over to the case of a von Neumann FK algebra and a normalized trace (ref. 24, chap. I, section 6, no. 11). Besides being a group homomorphism, detτ has the following properties:
FK y τðyÞ ReðτðyÞÞ FK detτ ðe Þ = je j = e for all y ∈ N and in particular detτ ðλ1Þ = jλj for all λ ∈ C; FK FK 1 FK detτ ðxÞ = detτ ðx*xÞ2 for all x ∈ GL1ðN Þ and in particular detτ ðxÞ = 1 for all x ∈ U1ðN Þ. fi ∈ = = For a *-ring R with unit, U1ðRÞ denotes its unitary group, de ned to be fx Rjx*x xx* 1g. FK Instead of [6], we could equally view detτ as a family of homomorphisms GLnðN Þ → R*+,oneforeachn ≥ 1; if the traces on the MnðN Þ s FK n are normalized by τð1nÞ = n,wehavedetτ ðλ1nÞ = jλj . More generally, for any projection e ∈ MnðN Þ, we have a von Neumann algebra FK MeðN ÞdeMnðN Þe and a Fuglede–Kadison determinant detτ : GLeðN Þ → R*+ definedonitsgroupofunits.
{ See ref. 24, chap. III, section 5, no. 2. Words are often reluctant to migrate from one mathematical domain to another. Otherwise, one could define a factor of type II1 as an infinite-dimensional finite von Neumann algebra which is central simple. In the same vein, one could say that von Neumann algebras are topologically principal rings; more precisely, in a von Neumann algebra N , any ultraweakly closed left ideal is of the form N e, where e ∈ N is a projection [this is a corollary of the von Neumann density theorem (ref. 24, chap. I, section 3, no. 4)]. ║ The normalization is most often by τð1Þ = 1. It can be otherwise, for example τð1nÞ = n on a factor of the form MnðN Þ, for some factor N.
de la Harpe PNAS | October 1, 2013 | vol. 110 | no. 40 | 15867 Downloaded by guest on October 2, 2021 FK There are extensions of detτ to noninvertible elements, but this raises some problems and technical difficulties. Two extensions are discussed in ref. 33: the “algebraic extension” for which the determinant vanishes on singular elements (this is not mentioned again) and the “analytic extension” that relies on Eq. 6, in which one should understand Z FK 1 det ðxÞ = exp τ log ðx*xÞ2 = exp ln λ dτðEλÞ; τ [7] = spððx*xÞ1 2Þ
1=2 λ = * −∞ = “ ” where ðE Þλ∈spððx*xÞ1 2Þ denotes the spectral resolution of ðx xÞ ; of course expð Þ 0. (Note that we write log for logarithms of matrices and operators and “ln” for logarithms of numbers.) For example, if x is such that there exists a projection e with x = xð1 − eÞ and τðeÞ > 0, we have detFK ðxÞ = 0. For all x; y ∈ N , we have τ FK 1=2 FK 1=2 detτ ðx*xÞ = lim detτ ðx*xÞ + e1 e→0+ FK FK FK detτ ðxÞdetτ ðyÞ = detτ ðxyÞ
FK (see ref. 33, respectively lemma 5 and p. 529). However, an element x with detτ ðxÞ ≠ 0 need not be invertible, and no extension N → R+ of the FK mapping detτ of [6] is norm continuous (ref. 33, theorem 6). FKL We discuss another extension detτ to singular elements, in section 8. FK More generally, detτ ðxÞ can be defined for x as an operator “affiliated” to N , and also for traces that are semifinite rather than finite as above. See refs. 34–39, among others. We do not comment further on this part of the theory. – = Example 4 [Fuglede KadisonP determinant for Mn(C)]. Let N MnðCÞ be the factor of type In, as in Example 1, let det be the usual 1 n determinant, and let τ : x ↦ = xj;j be the trace normalized by τð1nÞ = 1. Then n j 1 1=n FK 1=n 1=2 detτ ðxÞ = jdetðxÞj = det ðx*xÞ [8]
for all x ∈ MnðCÞ. ∞ Example 5 (Fuglede–Kadison Determinant for Abelian von Neumann Algebras). Let L ðZ; νÞ and τν be as in Example 2, with ν a probability measure. The corresponding Fuglede–Kadison determinant is given by Z FK detτ ðf Þ = exp lnjf ðzÞjdμðzÞ ∈ R+: [9] Z
In [9], observe that lnjf ðzÞj is bounded above on Z, because jf ðzÞj ≤ kf k∞ < ∞ for ν-almost all z. However, jf ðzÞj need not be bounded away from FK 0, so that lnjf ðzÞj = −∞ occurs. If the value of the integral is −∞, then detτ ðf Þ = expð−∞Þ = 0. Consider an integer d ≥ 1 and the von Neumann algebra NðZdÞ of the free abelian group of rank d. Fourier transform provides an isomorphism of von Neumann algebras ≈ ∞ N Zd ! L Td; ν ; x ↦ x^;
where ν denotes the normalized Haar measure on the d-dimensional torus Td. Moreover, the composition of this isomorphism with the trace τν of Example 2 is the canonical trace on NðZdÞ, in the sense of Example 3. P – fi finite λ ∈ d Example 6 (Fuglede Kadison Determinant and Mahler Measure). Let x be a nite linear combination ∈ d zn ðnÞ NðZ Þ, so that ∞ n Z ^x ∈ L ðTd; νÞ is a trigonometric polynomial. Then the τν-Fuglede–Kadison determinant of x is given by the exponential Mahler measure of ^x: Z FK = ^ d ^ ν : detτν ðxÞ M x exp lnjxðzÞjd ðzÞ Td In the one-dimensional case ðd = 1Þ,if s ^ = + + + s = ∏ − ξ ; ≠ ; xðzÞ a0 a1z ⋯ asz as z j with a0as 0 j = 1
a computation shows that
Z Z1 Xs n o ^ ν = ^ 2πiα α = + ; ξ ln xðzÞ d ðzÞ ln x e d lnjasj max 1 j j = 1 T 0 (ref. 40, proposition 16.1 or ref. 41, pp. 135–137). Mahler measures occur in particular as entropies of Zd-actions by automorphisms of compact groups. More precisely, for x ∈ Z½Zd ,which can be viewed as the inverse Fourier transform of a trigonometric polynomial, the group Zd acts naturally on the quotient Z½Zd =ðxÞ of the group ring by the principal ideal ðxÞ and hence on the Pontryagin dual ðZ½Zd =ðxÞÞ^ of this countable abelian group, which is a compact
15868 | www.pnas.org/cgi/doi/10.1073/pnas.1202059110 de la Harpe Downloaded by guest on October 2, 2021 2 −1 ^ 2
= + − ∈ ; ≈ = ≈ PERSPECTIVE abelian group. For example, if xðzÞ 1 z z Z½z z Z½Z ,thenðZ½Z ðxÞÞ T , and the corresponding action of the generator 01 of Z on T2 is described by the matrix (ref. 40, example 5.3). Every action of Zd by automorphisms of a compact abelian 11 group arises as above from some x ∈ Z½Zd . More on this is found in refs. 40, 42, and 43. The logarithm of the Fuglede–Kadison determinant occurs also in the definition of a “tree entropy”, namely in the asymptotics of the number of spanning trees in large graphs (44, 45).
4. Motivating Question It is natural to ask why R*+ appears on the right-hand side of [6], even though N is a complex algebra, for example a II1-factor, whereas C* appears on the right-hand side of [5]whenN = MnðCÞ. This is not due to some shortsightedness of Fuglede and Kadison. Indeed, for N afactoroftypeII1, it has been shown that the Fuglede– Kadison determinant provides an isomorphismfromtheabelianizedgroupGL1ðN Þ=D GL1ðN Þ onto R*+.Inotherwords,wehave the following:
Proposition 7 (Properties of Operators with Trivial Fuglede–Kadison Determinant in a Factor of Type II1). Let N be a factor of type II1.
i) Any element in U1ðN Þ is a product of finitely many multiplicative commutators of unitary elements.
ii) The kernel SL1ðN Þ of the homomorphism [6] coincides with the group of commutators of GL1ðN Þ.
Property i is due to Broise (46). It is moreover known that any proper normal subgroup of U1ðN Þ is contained in its center, which is fλ idjλ ∈ C*; jλj = 1g ≈ R=Z (ref. 47, proposition 3 and its proof); this sharpens an earlier result on the classification of norm-closed normal subgroups of U1ðN Þ (ref. 48, theorem 2). Property ii is from ref. 32, proposition 2.5. It follows that the quotient of SL1ðN Þ byitscenter[whichisthesameasthecenterofU1ðN Þ]is simple,asanabstractgroup(ref.49,corollary6.6,p.123). As a kind of answer to our motivating question, we see below that, when the Fuglede–Kadison definition is adapted to a separable Banach algebra, the right-hand side of the homomorphism analogous to [6] is necessarily a quotient of the additive group C by a countable subgroup. expð·Þ For example, when A = MnðCÞ,thisquotientisC=2iπZ ≈ C*;seeCorollary 13. On the contrary, when A is a II1-factor (not separable as expðReð·ÞÞ a Banach algebra), this quotient is C=2iπR ≈ R*+;seeCorollary 14. The case of a separable Banach algebra can sometimes be seen as providing an interpolation between the two previous cases; see Remark 15.
top 5. Recalling K0, K1, K1 , and Bott Periodicity 5.1. On K0(R) and K0(A). Let R be a ring, say with unit to simplify several small technical points. Let us first recall one definition of the abelian group K0ðRÞ of K-theory. We have a nested sequence of rings of matrices and (nonunital) ring homomorphisms [ = ⊂ ⊂ ⊂ ⊂ R M1ðRÞ ⋯ MnðRÞ Mn+1ðRÞ ⋯ ⊂ M∞ðRÞd MnðRÞ; [10] n≥1 x 0 where the inclusions at finite stages are given by x ↦ . 00 An idempotent in M∞ðRÞ is an element e such that e2 = e.Twoidempotentse; f ∈ M∞ðRÞ are equivalent if there exist n ≥ 1and −1 u ∈ GLnðRÞ such that e; f ∈ MnðRÞ and f = u eu.Define an addition on equivalence classes of idempotents, by
ðclass of e ∈ MkðRÞÞ + ðclass of f ∈ MℓðRÞÞ = class of e ⊕ f ∈ Mk+ℓðRÞ; [11] e 0 where e ⊕ f denotes the matrix . Two idempotents e; f ∈ M∞ðRÞ are stably equivalent if there exists an idempotent g such that the 0 f classes of e ⊕ g and f ⊕ g are equivalent; we denote by ½e the stable equivalence class of an idempotent e. The set of stable equivalence classes of
idempotents, with the addition defined by ½e + ½f d½e ⊕ f , is a semigroup. The Grothendieck group K0ðRÞ of this semigroup is the set of formal differences ½e − ½e′ , up to the equivalence defined by ½e − ½e′ ∼ ½f − ½f ′ if ½e + ½f ′ = ½e′ + ½f . Note that K0 is a functor: To any (unital) ring homomorphism R → R′ corresponds a natural homomorphism K0ðRÞ → K0ðR′Þ of abelian groups. Note also the isomorphism K0ðMnðRÞÞ ≈ K0ðRÞ, which is a straightforward consequence of the definition and of the isomorphisms MkðMnðRÞÞ ≈ MknðRÞ. n n [To an idempotent e ∈ M∞ðRÞ is associated an R-linear mapping R → R for n large enough, of which the image is a projective R-module of finite rank. From this it can be checked that the definition of K0ðRÞ given above coincides with another standard definition, in terms of projective modules of finite rank. Details are in ref. 50, chap. 1.] Rather than a general ring R, consider now the case of a complex Banach algebra A with unit. For each n ≥ 1, the matrix algebra MnðAÞ is again a Banach algebra, for some appropriate norm, and we can furnish M∞ðAÞ with the inductive limit topology. The
de la Harpe PNAS | October 1, 2013 | vol. 110 | no. 40 | 15869 Downloaded by guest on October 2, 2021 following is rather easy to check (e.g., ref. 51, pp. 25–27): Two idempotents e; f ∈ M∞ðAÞ are equivalent if and only if there exists a continuous path ½0; 1 → idempotents of M∞ðAÞ ; α ↦ eα
such that e0 = e and e1 = f . This has the following consequence:
Proposition 8. If the Banach algebra A is separable, the abelian group K0ðAÞ is countable.
Proposition 9. If N is a factor of type II1, then K0ðN Þ ≈ R is uncountable. τ Indeed, if denotes the canonical trace on N , the mapping≈ that associates to the class of a self-adjoint idempotent e in N its von Neumann dimension τðeÞ ∈ ½0; 1 extends to an isomorphism K0ðN Þ !R. On the proof: This follows from the “comparison of projections” in von Neumann algebras (ref. 24, chap. III, section 2, no. 7). □ For historical indications on the early connections between K-theory and operator algebras, which go back to the mid-1960s, see ref. 52.
5.2. On K1(R). For any ring R with unit, we have a nested sequence of group homomorphisms [ *= ⊂ ⊂ ⊂ ⊂ R GL1ðRÞ ⋯ GLnðRÞ GLn+1ðRÞ ⋯ ⊂ GL∞ðRÞd GLnðRÞ; [12] n≥1 x 0 where the inclusions at finite stages are given by x ↦ . 01 By definition,
K1ðRÞ = GL∞ðRÞ=D GL∞ðRÞ [13]
is an abelian group, usually written additively. Note that K1 is a functor from rings to abelian groups. For a commutative ring R, the classical determinant provides a homomorphism K1ðRÞ → R*; it is an isomorphism in several important cases, for example when R is a field or the ring of integers in a finite extension of Q (ref. 53, section 3). In general (R commutative or not), the association of an element in K1ðRÞ to a matrix in GL∞ðRÞ canbeviewedasakindofdeterminantorratherofalogofadeterminant because K1ðRÞ is written additively. Accordingly, the torsion defined in [24] below can be viewed as an alternating sum of logs of deter- minants; we recall this when defining the L2-torsion in Eq. 29. Let R be, again, an arbitrary ring with unit. The reduced K1-group is the quotient K1ðRÞ of K1ðRÞ by the image of the natural ho- momorphism f1; − 1g ⊂ GL1ðRÞ ⊂ GL∞ ðRÞ → K1ðRÞ. InthecasethatR = Z½Γ is the integral group ring of a group Γ, the Whitehead group WhðΓÞ is the cokernel K1ðZ½Γ Þ=h±1; Γi of the natural homomorphism Γ ⊂ GL1ðZ½Γ Þ → K1ðZ½Γ Þ → K1ðZ½Γ Þ. When Γ is finitely presented, there is a different (but equivalent) definition of WhðΓÞ, with geometric content. In short, let L be a connected finite CW complex with π1ðLÞ = Γ.Onedefines a group WhðLÞ of appropriate equivalence classes of pairs ðK; LÞ,withK a finite CW complex containing L in such a way that the inclusion L ⊂ K is a homotopy equivalence. The unit is represented by pairs L ⊂ K for which the inclusion is a simple homotopy equivalence. It can be shown that the functors L → WhðLÞ and L → Whðπ1ðLÞÞ are naturally equivalent (ref. 54, section 6 and theorem 21.1). d d Examples are WhðZ Þ = 0 for free abelian groups Z and WhðFdÞ = 0 for free groups Fd.Forfinite cyclic groups, WhðZ=qZÞ is a free abelian group of finite rank for all q ≥ 1andisthegroupf0g if and only if q ∈ f1; 2; 3; 4; 6g. From the standard references, let us quote from refs. 53–57.
top 5.3. On K1 (A) and on K0(A) viewed as a Fundamental Group. Let A be a Banach algebra with unit. For each n ≥ 1, the group GLnðAÞ is an open subset of the Banach space MnðAÞ, and the induced topology makes it a topological group. The group GL∞ðAÞ of [12]isalso 0 a topological group, for the inductive limit topology; we denote by GL∞ðAÞ its connected component. It is a simple consequence of the classical “Whitehead lemma” that, for any Banach algebra, the group D GL∞ðAÞ is perfect and coincides 0 0 with D GL∞ðAÞ;see,forexample,ref.58,appendix.Inparticular,D GL∞ðAÞ ⊂ GL∞ðAÞ, so that the quotient group
top dπ = = 0 K1 ðAÞ 0ðGL∞ðAÞÞ GL∞ðAÞ GL∞ðAÞ [14]
= 0 = 0 is commutative. Note that GL1ðAÞ GL1ðAÞ need not be commutative (59), even if its image in GL∞ðAÞ GL∞ðAÞ is always commutative. Moreover, we have a natural quotient homomorphism
= 0 = → top = = 0 ; GL∞ðAÞ D GL∞ðAÞ K1ðAÞ K1 ðAÞ GL∞ðAÞ GL∞ðAÞ [15]
0 which is surjective. It is an isomorphism if and only if the group GL∞ðAÞ is perfect; this is the case if A is an infinite simple C*-algebra, for
example if A is one of the Cuntz algebras On briefly mentioned below. top If the Banach algebra A is separable, the group K1 ðAÞ iscountable(comparewithProposition 8).
15870 | www.pnas.org/cgi/doi/10.1073/pnas.1202059110 de la Harpe Downloaded by guest on October 2, 2021 To an idempotent e ∈ MnðAÞ, we can associate the loop PERSPECTIVE ½0; 1 → GL ðAÞ ⊂ GL∞ðAÞ ξ : n e [16] α ↦ expð2πiαeÞ = expð2πiαÞe +ð1− eÞ;
ξ = ξ = ξ ξ note that eð0Þ eð1Þ 1. If two idempotents e and f have the same image in K0ðAÞ, it is easy to check that e and f are homotopic loops. It is ↦ ξ a fundamental fact, which is a form of Bott periodicity, that the assignment e e extends to a group isomorphism ≈ 0 K0ðAÞ ! π1 GL∞ðAÞ [17] top ≈ top ≥ (ref. 60, theorem III.1.11 or ref. 51, chap. 9). The terminology is due to a generalization of [17]: Ki ðAÞ Ki+2ðAÞ for any integer i 0; by fi top = π ≥ top = de nition, Ki ðAÞ i−1ðGL∞ðAÞÞ, for all i 1, and K0 ðAÞ K0ðAÞ. 0 5.4. A Few Standard Examples. Let A = CðTÞ be the Banach algebra of continuous functions on a compact space T.ThenK0ðAÞ = K ðTÞ top = 1 0 1 – – – and K1 ðAÞ K ðTÞ,whereK ðTÞ and K ðTÞ stand for the (Grothendieck) Atiyah Hirzebruch Bott K-theory groups of the topological space T,defined in terms of complex vector bundles. For example, if T is a sphere, we have
2m ≈ 2; top 2m = ; K0ðCðS ÞÞ Z K1 ðCðS ÞÞ 0 2m+1 ≈ ; top 2m+1 ≈ ; K0ðCðS ÞÞ Z K1 ðCðS ÞÞ Z ≥ top = for all m 0. If T is a compact CW complex without cells of odd dimension, then K1 ðCðTÞÞ 0. * ⊂ ⊂ ⊂ ⊂ fi * Let A be anS AF algebra, namely a C -algebra that contains a nested sequence A1 ⋯ An An+1 ⋯ of nite-dimensional sub-C - top = : ’ algebras with n≥1An dense in A.ThenK0ðAÞ is rather well understood, and K1 ðAÞ 0 The group K0ðAÞ is the basic ingredient in Elliott s classification of AF algebras, from the 1970s; this was the beginning of a long and rich story, with numerous offspring (ref. 51, chap. 7, and refs. 61 and 62). A particular case is the so-called CAR algebra, or C*-algebra of the canonical anticommutation relations, or UHF algebra of type ð2iÞ in ref. 63: It is the C*-closure of the limit of the inductive system of finite matrix algebras ⊂ C ⋯ ⊂ M2n ðCÞ ⊂ M n + 1 ðCÞ ⊂ ⋯; 2 x 0 where the inclusions are given by x ↦ . For this, 0 x = = top = K0ðCARÞ Z½1 2 and K1 ðCARÞ 0
(for K1 of CAR and a few other AF algebras, see section 6.1). The Jiang-Su algebra Z is a simple infinite-dimensional C*-algebra with unit that plays an important role in Elliott’sclassification program of C*-algebras. It has the same K-theory as C (64). P finite The reduced C*-algebra of a group Γ is the norm-closure C*λðΓÞ of the algebra f γ∈Γ zγλðγÞg,seeExample 3,inthealgebraofall 2 bounded operators on ℓ ðΓÞ.ForthefreegroupsFd (nonabelian free groups if d ≥ 2), we have (65)
* ≈ top * ≈ d: K0ðCλðFdÞÞ Z and K1 ðCλðFdÞÞ Z π θ For a so-called irrational rotation C*-algebra Aθ, generated by two unitaries u; v satisfying the relation uv = e2 i vu,whereθ ∈ ½0; 1 with θ ∉ Q,wehave(66) ≈ 2 top ≈ 2: K0ðAθÞ Z and K1 ðAθÞ Z P fi ≥ ; ...; ∗ = δ n ∗ = For the in nite Cuntz algebras On, generated by n 2elementss1 sn satisfying sj sk j;k and j=1sjsj 1, we have (67) ≈ = − top = : K0ðOnÞ Z ðn 1ÞZ and K1 ðOnÞ 0
For N afactoroftypeII1,wehave K0ðN Þ ≈ R and K1ðN Þ = R*+:
For K0, see Proposition 9; for K1, see ref. 32, already cited for proposition 7.ii. More generally, for N a von Neumann algebra of type II1, with center denoted by Z, we have K0ðN Þ ≈ fz ∈ Zjz*= zg; where the right-hand side is viewed as a group for the addition, and
K1ðN Þ ≈ fz ∈ Zjz ≥ e > 0g ðe depends on zÞ; where the right-hand side is viewed as a group for the multiplication; see ref. 68 or ref. 41, section 9.2. For any von Neumann algebra N top = ; K1 ðN Þ 0
because GLnðN Þ is connected for all n ≥ 1; indeed, by polar decomposition and functional calculus, any x ∈ GLnðN Þ is of the form expðaÞ expðibÞ, with a; b self-adjoint in MnðN Þ, so that x is connected to 1 by the path α ↦ expðαaÞ expðiαbÞ.
de la Harpe PNAS | October 1, 2013 | vol. 110 | no. 40 | 15871 Downloaded by guest on October 2, 2021 5.5. Topology of the Group of Units in a Factor of Type II1 and Bott Periodicity. If N is a factor of type II1, the isomorphism [17]of Bott periodicity shows that
π1ðGL∞ðN ÞÞ ≈ K0ðN Þ ≈ R: Thus, by Bott periodicity,
π2jðGL∞ðN ÞÞ = 0 and π2j+1ðGL∞ðN ÞÞ ≈ R
for all j ≥ 0. For π1, it is known more precisely that π1ðGLnðN ÞÞ ≈ R and that the embedding of GLnðN Þ into GLn+1ðN Þ induces the identity on π1, for all n ≥ 1 (69, 70). Note that, still for the norm topology, polar decomposition shows that the unitary group U1ðN Þ is a deformation retract of GL1ðN Þ; in particular, we have also π1ðU1ðN ÞÞ ≈ R. For the strong topology, the situation is quite different; indeed, for “many” II1-factors, for example for those associated to infinite amenable strong topology icc groups or to nonabelian free groups, it is known that the group U1ðN Þ is contractible (71).
6. Revisiting the Fuglede–Kadison and Other Determinants Most of this section can be found in ref. 72. For other expositions of part of what follows, see ref. 73, around theorem 1.10, and ref. 41, section 3.2. Let A be a complex Banach algebra (with unit, again for reasons of simplicity), E be a Banach space, and τ : A → E be a continuous linear mapP that is tracial, namely such that τðyxÞ = τðxyÞ for all x; y ∈ A.Thenτ extends to a continuous linear map M∞ðAÞ → E,defined by ↦ τ τ ; ∈ τ = τ τ x j≥1 ðxj;jÞ, and again denoted by .Ife f M∞ðAÞ areequivalentidempotents,wehave ðeÞ ðf Þ;itfollowsthat induces a ho- momorphism of abelian groups
τ : K0ðAÞ → E; ½e ↦ τðeÞ:
For example, if A = C and τ : C → C is the identity, the stable equivalence class of an idempotent e ∈ MnðCÞ is precisely described by the n dimension of the image ImðeÞ ⊂ C , so that K0ðCÞ ≈ Z, and the image of τ is the subgroup Z of the additive group C. 0 For a piecewise differentiable path ξ : ½α1; α2 → GL∞ðAÞ,wedefine