Ergodic flow explained

In mathematics, ergodic flows occur in geometry, through the geodesic and horocycle flows of closed hyperbolic surfaces. Both of these examples have been understood in terms of the theory of unitary representations of locally compact groups: if Γ is the fundamental group of a closed surface, regarded as a discrete subgroup of the Möbius group G = PSL(2,R), then the geodesic and horocycle flow can be identified with the natural actions of the subgroups A of real positive diagonal matrices and N of lower unitriangular matrices on the unit tangent bundle G / Γ. The Ambrose-Kakutani theorem expresses every ergodic flow as the flow built from an invertible ergodic transformation on a measure space using a ceiling function. In the case of geodesic flow, the ergodic transformation can be understood in terms of symbolic dynamics; and in terms of the ergodic actions of Γ on the boundary S¹ = G / AN and G / A = S¹ × S¹ \ diag S¹. Ergodic flows also arise naturally as invariants in the classification of von Neumann algebras: the flow of weights for a factor of type III₀ is an ergodic flow on a measure space.

Hedlund's theorem: ergodicity of geodesic and horocycle flows

The method using representation theory relies on the following two results:

• If = acts unitarily on a Hilbert space and is a unit vector fixed by the subgroup of upper unitriangular matrices, then is fixed by .
• If = acts unitarily on a Hilbert space and is a unit vector fixed by the subgroup of diagonal matrices of determinant, then is fixed by .

(1) As a topological space, the homogeneous space = can be identified with with the standard action of as matrices. The subgroup of has two kinds of orbits: orbits parallel to the -axis with ; and points on the -axis. A continuous function on that is constant on -orbits must therefore be constant on the real axis with the origin removed. Thus the matrix coefficient = satisfies = for in . By unitarity, |||| = =, so that = for all in = = . Now let be the matrix

\begin{pmatrix} 0 & 1 \ -1 & 0 \end{pmatrix} . Then, as is easily verified, the double coset is dense in ; this is a special case of the Bruhat decomposition. Since is fixed by, the matrix coefficient is constant on . By density, = for all in . The same argument as above shows that = for all in .

(2) Suppose that is fixed by . For the unitary 1-parameter group ≅, let be the spectral subspace corresponding to the interval . Let be the diagonal matrix with entries and for || . Then . As || tends to infinity the latter projections tend to 0 in the strong operator topology if or . Since =, it follows = in either case. By the spectral theorem, it follows that is in the spectral subspace ; in other words is fixed by . But then, by the first result, must be fixed by .

The classical theorems of Gustav Hedlund from the early 1930s assert the ergodicity of the geodesic and horocycle flows corresponding to compact Riemann surfaces of constant negative curvature. Hedlund's theorem can be re-interpreted in terms of unitary representations of and its subgroups. Let be a cocompact subgroup of = for which all non-scalar elements are hyperbolic. Let = where is the subgroup of rotations

\begin{pmatrix} \cos \theta & \sin \theta \ - \sin \theta & \cos \theta \end{pmatrix} . The unit tangent bundle is =, with the geodesic flow given by the right action of and the horocycle flow by the right action of . This action if ergodic if =, i.e. the functions fixed by are just the constant functions. Since is compact, this will be the case if = . Let = . Thus acts unitarily on on the right. Any non-zero in fixed by must be fixed by, by the second result above. But in this case, if is a continuous function on of compact support with =, then = . The right hand side equals, a continuous function on . Since is right-invariant under, it follows that is constant, as required. Hence the geodesic flow is ergodic. Replacing by and using the first result above, the same argument shows that the horocycle flow is ergodic.

Ambrose−Kakutani–Krengel–Kubo theorem

Induced flows

Examples of flows induced from non-singular invertible transformations of measure spaces were defined by in his operator-theoretic approach to classical mechanics and ergodic theory. Let T be a non-singular invertible transformation of (X,μ) giving rise to an automorphism τ of A = L^∞(X). This gives rise to an invertible transformation T ⊗ id of the measure space (X × R,μ × m), where m is Lebesgue measure, and hence an automorphism τ ⊗ id of A L^∞(R). Translation L_t defines a flow on R preserving m and hence a flow λ_t on L^∞(R). Let S = L₁ with corresponding automorphism σ of L^∞(R). Thus τ ⊗ σ gives an automorphism of A L^∞(R) which commutes with the flow id ⊗ λ_t. The induced measure space Y is defined by B = L^∞(Y) = L^∞(X × R)^{τ ⊗ σ}, the functions fixed by the automorphism τ ⊗ σ. It admits the induced flow given by the restriction of id ⊗ λ_t to B. Since λ_t acts ergodically on L^∞(R), it follows that the functions fixed by the flow can be identified with L^∞(X)^τ. In particular if the original transformation is ergodic, the flow that it induces is also ergodic.

Flows built under a ceiling function

The induced action can also be described in terms of unitary operators and it is this approach which clarifies the generalisation to special flows, i.e. flows built under ceiling functions. Let R be the Fourier transform on L²(R,m), a unitary operator such that Rλ(t)R^∗ = V_t where λ(t) is translation by t and V_t is multiplication by e^itx. Thus V_t lies in L^∞(R). In particular V₁ = R S R^∗. A ceiling function h is a function in A with h ≥ ε1 with ε > 0. Then e^ihx gives a unitary representation of R in A, continuous in the strong operator topology and hence a unitary element W of A L^∞(R), acting on L²(X,μ) ⊗ L²(R). In particular W commutes with I ⊗V_t. So commutes with I ⊗ λ(t). The action T on L^∞(X) induces a unitary U on L²(X) using the square root of the Radon−Nikodym derivative of μ ∘ T with respect to μ. The induced algebra B is defined as the subalgebra of commuting with . The induced flow σ_t is given by .

The special flow corresponding to the ceiling function with base transformation is defined on the algebra B(H) given by the elements in commuting with . The induced flow corresponds to the ceiling function h ≡ 1, the constant function. Again W₁, and hence commutes with I ⊗ λ(t). The special flow on B(H) is again given by . The same reasoning as for induced actions shows that the functions fixed by the flow correspond to the functions in A fixed by σ, so that the special flow is ergodic if the original non-singular transformation T is ergodic.

Relation to Hopf decomposition

See main article: Hopf decomposition. If S_t is an ergodic flow on the measure space (X,μ) corresponding to a 1-parameter group of automorphisms σ_t of A = L^∞(X,μ), then by the Hopf decomposition either every S_t with t ≠ 0 is dissipative or every S_t with t ≠ 0 is conservative. In the dissipative case, the ergodic flow must be transitive, so that A can be identified with L^∞(R) under Lebesgue measure and R acting by translation.

To prove the result on the dissipative case, note that A = L^∞(X,μ) is a maximal Abelian von Neumann algebra acting on the Hilbert space L²(X,μ). The probability measure μ can be replaced by an equivalent invariant measure λ and there is a projection p in A such that σ_t(p) < p for t > 0 and λ(p – σ_t(p)) = t. In this case σ_t(p) =E([''t'',∞)) where ''E'' is a projection-valued measure on '''R'''. These projections generate a von Neumann subalgebra ''B'' of ''A''. By ergodicity σ_''t''(''p'') <math>\uparrow</math> 1 as ''t'' tends to −∞. The Hilbert space L²(''X'',λ) can be identified with the completion of the subspace of ''f'' in ''A'' with λ(|''f''|²) < ∞. The subspace corresponding to ''B'' can be identified with L²('''R''') and ''B'' with L^∞('''R'''). Since λ is invariant under ''S''_''t'', it is implemented by a unitary representation ''U''_''t''. By the [[Stone–von Neumann theorem]] for the covariant system B, U_t, the Hilbert space H = L²(X,λ) admits a decomposition L²(R) ⊗

\ell ^{2} where B and U_t act only on the first tensor factor. If there is an element a of A not in B, then it lies in the commutant of B ⊗ C, i.e. in B B(\ell ^{2} ). If can thus be realised as a matrix with entries in B. Multiplying by χ_{[''r'',''s'']} in B, the entries of a can be taken to be in L^∞(R) ∩ L¹(R). For such functions f, as an elementary case of the ergodic theorem the average of σ_t(f) over [−''R'',''R''] tends in the weak operator topology to ∫ f(t) dt. Hence for appropriate χ_{[''r'',''s'']} this will produce an element in A which lies in C ⊗ B(\ell ^{2} ) and is not a multiple of 1 ⊗ I. But such an element commutes with U_t so is fixed by σ_t, contradicting ergodicity. Hence A = B = L^∞(R).

When all the σ_t with t ≠ 0 are conservative, the flow is said to be properly ergodic. In this case it follows that for every non-zero p in A and t ≠ 0, p ≤ σ_t (p) ∨ σ_2t (p) ∨ σ_3t (p) ∨ ⋅⋅⋅ In particular ∨_±t>0 σ_t (p) = 1 for p ≠ 0.

Theorem of Ambrose–Kakutani–Krengel–Kubo

The theorem states that every ergodic flow is isomorphic to a special flow corresponding to a ceiling function with ergodic base transformation. If the flow leaves a probability measure invariant, the same is true of the base transformation.

For simplicity only the original result of is considered, the case of an ergodic flow preserving a probability measure . Let = and let be the ergodic flow. Since the flow is conservative, for any projection p ≠ 0, 1 in A there is a T > 0 without σ_T(p) ≤ p, so that . On the other hand, as r > 0 decreases to zero

a _{r} = { 1 \over r}\int ^{r} _{0} \sigma _{t(p)} dt → p

in the strong operator topology or equivalently the weak operator topology (these topologies coincide on unitaries, hence involutions, hence projections). Indeed, it suffices to show that if ν is any finite measure on A, then ν(a_r) tends to ν(p). This follows because f(t) = ν(σ_t(p)) is a continuous function of t so that the average of f over [0,''r''] tends to f(0) as r tends to 0.

Note that . Now for fixed r > 0, following, set

q _{0(r)=\chi} _{[0,1/4]} (a _{r),} q _{1(r)=\chi} _{[3/4,1]} (a _{r).}

Set r = N^–1 for N large and f_N = a_r. Thus 0 ≤ f_N ≤ 1 in L^∞(X,μ) and f_N tends to a characteristic function p in L¹(X,μ). But then, if ε = 1/4, it follows that χ_[0,ε](f_N) tends to χ_[0,ε](p) = 1 – p in L¹(X).^[1] Using the splitting A = pA ⊕ (1 − p)A, this reduces to proving that if 0 ≤ h_N ≤ 1 in L^∞(Y,ν) and h_N tends to 0 in L¹(Y,ν), then χ_[1−ε,1](h_N) tends to 0 in L¹(Y,ν). But this follows easily by Chebyshev's inequality: indeed, so that, which tends to 0 by assumption.

Thus by definition q₀(r) ∧ q₁(r) = 0. Moreover, for r = N⁻¹ sufficiently small, q₀(r) ∧ σ_T(q₁(r)) > 0. The above reasoning shows that q₀(r) and q₁(r) tend to 1 − p and p as r = N⁻¹ tends to 0. This implies that q₀(r)σ_T(q₁(r)) tends to (1 − p)σ_T(p) ≠ 0, so is non-zero for N sufficiently large. Fixing one such N and, with r = N⁻¹, setting q₀= q₀(r) and q₁= q₁(r), it can therefore be assumed that

q _{0} \wedge q _{1} =0 , q _{0} \wedge \sigma _{T(q} _{1)>0.}

The definition of q₀ and q₁ also implies that if δ < r/4 = (4N)⁻¹, then

\sigma _{t(q} _{0)\wedge} q _{1} = 0 for |t|\le \delta.

In fact if s < t

\|\sigma _{t(a} _{r)} - \sigma _{s(a} _{r)\|} _{infty} = r ^{-1} \left\|\int ^{r + t} _{r + s} \sigma _{x(p)} dx - \int ^{t} _{s} \sigma _{x(p)} dx\right\| _{infty} \le { 2 |t - s|\over r}.

Take s = 0, so that t > 0 and suppose that e = σ_t(q₀) ∧ q₁ > 0. So e = σ_t(f) with f ≤ q₀. Then σ_t(a_r)e = σ_t(a_rf) ≤ 1/4 e and a_re ≥ 3/4 e, so that

a _{r} - \sigma _{t(a} _{r)} \ge a _{r} e - \sigma _{t(a} _{r)e} \ge { 3 \over 4 } e - { 1 \over 4 }e = { 1 \over 2 } e.

Hence ||a_r − σ_t(a_r)||_∞ ≥ 1/2. On the other hand ||a_r − σ_t(a_r)||_∞ is bounded above by 2t/r, so that t ≥ r/4. Hence σ_t(q₀) ∧ q₁ = 0 if |t| ≤ δ.

The elements a_r depend continuously in operator norm on r on (0,1]; from the above σ_t(a_r) is norm continuous in t. Let B₀ the closure in the operator norm of the unital *-algebra generated by the σ_t(a_r)'s. It is commutative and separable so, by the Gelfand–Naimark theorem, can be identified with C(Z) where Z is its spectrum, a compact metric space. By definition B₀ is a subalgebra of A and its closure B in the weak or strong operator topology can be identified with L^∞(Z,μ) where μ is also used for the restriction of μ to B. The subalgebra B is invariant under the flow σ_t, which is therefore ergodic. The analysis of this action on B₀ and B yields all the tools necessary for constructing the ergodic transformation T and ceiling function h. This will first be carried out for B (so that A will temporarily be assumed to coincide with B) and then later extended to A.

The projections q₀ and q₁ correspond to characteristic functions of open sets. X₀ and X₁ The assumption of proper ergodicity implies that the union of either of these open sets under translates by σ_t as t runs over the positive or negative reals is conull (i.e. the complement has measure zero). Replacing X by their intersection, an open set, it can be assumed that these unions exhaust the whole space (which will now be locally compact instead of compact). Since the flow is recurrent any orbit of σ_t passes through both sets infinitely many times as t tends to either +∞ or −∞. Between a spell first in X₀ and then in X₁ f must assume the value 1/2 and then 3/4. The last time f equals 1/2 to the first time it equals 3/4 must involve a change in t of at least δ/4 by the Lipschitz continuity condition. Hence each orbit must intersect the set Ω of x for which f(x) = 1/2, f(σ_t(x)) > 1/2 for 0 < t ≤ δ/4 infinitely often. The definition implies that different insections with an orbit are separated by a distance of at least δ/4, so Ω intersects each orbit only countably many times and the intersections occur at indefinitely large negative and positive times. Thus each orbit is broken up into countably many half-open intervals [''r''_''n''(''x''),''r''_''n''+1(''x'')) of length at least δ/4 with ''r''_''n''(''x'') tending to ±∞ as ''n'' tends to ±∞. This partitioning can be normalised so that ''r''₀(''x'') ≤ 0 and ''r''₁(''x'') > 0. In particular if ''x'' lies in Ω, then ''t''₀ = 0. The function ''r''_''n''(''x'') is called the '''''n''th return time to Ω'''. The cross-section Ω is a Borel set because on each compact set {σ_''t''(''x'')} with ''t'' in [''N''⁻¹,δ/4] with N > 4/δ, the function g(t) = f(σ_t(x)) has an infimum greater than 1/2 + M⁻¹ for a sufficiently large integer M. Hence Ω can be written as a countable intersection of sets, each of which is a countable unions of closed sets; so Ω is therefore a Borel set. This implies in particular that the functions r_n are Borel functions on X. Given y in Ω, the invertible Borel transformation T is defined on Ω by S(y) = σ_t(y) where t = r₁(y), the first return time to Ω. The functions r_n(y) restrict to Borel functions on Ω and satisfy the cocycle relation:

r _{m+n} = r _{m} + \tau ^{m(r} _{n),}

where τ is the automorphism induced by T. The hitting number N_t(x) for the flow S_t on X is defined as the integer N such that t lies in [''r''_''N''(''x''),''r''_''N''+1(''x'')). It is an integer-valued Borel function on '''R''' × ''X'' satisfying the cocycle identity :<math>N_{s+t} = N_s + \sigma_s(N_t).</math> The function ''h'' = ''r''₁ is a strictly positive Borel function on Ω so formally the flow can be reconstructed from the transformation ''T'' using ''h'' a ceiling function. The missing ''T''-invariant measure class on Ω will be recovered using the second cocycle ''N''_''t''. Indeed, the discrete measure on '''Z''' defines a measure class on the product '''Z''' × ''X'' and the flow ''S''_''t'' on the second factor extends to a flow on the product given by :<math>\rho_t(m,x)=(m+N_t(x),S_t(x)).</math> Likewise the base transformation ''T'' induces a transformation ''R'' on '''R''' × Ω defined by :<math>R(s,y)=(s-h(y),T(y)).</math> These transformations are related by an invertible Borel isomorphism Φ from '''R''' × Ω onto '''Z''' × ''X'' defined by :<math>\Phi(t,y)=(N_t(y),S_t(y)).</math> Its inverse Ψ from '''Z''' × ''X'' onto '''R''' × Ω is defined by :<math>\Psi(m,x)=(-r_{-m}(y), S_{r_{-m}(y)}y).</math> Under these maps the flow ''R''_''t'' is carried onto translation by ''t'' on the first factor of '''R''' × Ω and, in the other direction, the invertible ''R'' is carried onto translation by -1 on '''Z''' × ''X''. It suffices to check that the measure class on '''Z''' × ''X'' carries over onto the same measure class as some produce measure ''m'' × ν on '''R''' × Ω, where ''m'' is Lebesgue measure and ν is a probability measure on Ω with measure class invariant under ''T''. The measure class on '''Z''' × ''X'' is invariant under '''R''', so defines a measure class on '''R''' × Ω, invariant under translation on the first factor. On the other hand, the only measure class on '''R''' invariant under translation is Lebesgue measure, so the measure class on '''R''' × Ω is equivalent to that of ''m'' × ν for some probability measure on Ω. By construction ν is quasi-invariant under ''T''. Unravelling this construction, it follows that the original flow is isomorphic to the flow built under the ceiling function ''h'' for the base transformation ''T'' on (Ω,ν).<ref>If ν is a probability measure on '''R''' such that the null sets are translation invariant, it suffices to show that ν is quasi-equivalent to Lebesgue measure, i.e. that a Borel set has zero measure for ν if and only if it has Lebesgue measure zero. But it is sufficient to check this for subsets of [0,1); and, passing to translates by '''Z''', which by assumption are null sets, to '''Z'''-invariant null sets. On the other hand the Poisson summation map ''F''(''x'') = Σ ''f''(''x''+''n'') takes bounded Borel functions on [0,1) to periodic bounded Borel functions on '''R''', so that ν can be used to define a probability measure ν₁ on '''T''' = '''R'''/'''Z''' with the same invariance properties. A simple averaging argument shows that ν₁ is quasi-equivalent to [[Haar measure]] on the circle. For, if α_θ denotes rotation by θ, ν₁ ∘ α_θ is quasi-equivalent to ν₁ and hence so is the average of these measures over 2. On the other hand that averaged measure is invariant under rotation, so bu uniqueness of Haar measure equals Lebesgue measure.

The above reasoning was made with the assumption that B = A. In general A is replaced by a norm closed separable unital *-subalgebra A₀containing B₀, invariant under σ_t and such that σ_t(f) is a norm continuous function of t for any f in A₀. To construct A₀, first take a generating set for the von Neumann algebra A formed of countably many projections invariant under σ_t with t rational. Replace each of this countable set of projections by averages over intervals [0,''N''⁻¹] with respect to σ_t. The norm closed unital *-algebra that these generate yields A₀. By definition it contains B₀ = C(Y). By the Gelfand-Naimark theorem A₀ has the form C(X). The construction with a_r above applies equally well here: indeed since B₀ is a subalgebra of A_{0, Y is a continuous quotient of X, so a function such as ar is equally well a function on X. The construction therefore carries over mutatis mutandis to A, through the quotient map.}

In summary there exists a measure space (Y,λ) and an ergodic action of Z × R on M = L^∞(Y,λ) given by commuting actions τⁿ and σ_t such that there is a τ-invariant subalgebra of M isomorphic to

\ell ^{infty} (Z) and a σ-invariant subalgebra of M isomorphic to L^∞(R). The original ergodic flow is given by the restriction of σ to M^τ and the corresponding base transformation given by the restriction of τ to M^σ.^[2]

Given a flow, it is possible to describe how two different single base transformations that can be used to construct the flow are related. be transformed back into an action of Z on Y, i.e. into an invertible transformation T_Y on Y. Set-theoretically T_Y (x) is defined to be T^m(x) where m ≥ 1 is the smallest integer such that T^m(x) lies in X. It is straightforward to see that applying the same process to the inverse of T yields the inverse of T_Y. The construction can be described measure theoretically as follows. Let e = χ_Y in B = L^∞(X,ν) with ν(e) ≠ 0. Then e is an orthogonal sum of projections e_n defined as follows:

e _{1=} e \tau ^{-1} (e), e ^{-1} _{2= e( 1- \tau} (e))\tau ^{-2} (e), e _{3=e} ( 1- \tau ^{-1} (e))( 1- \tau ^{-2} (e))\tau ^{-3} (e),...

Then if f lies in e_n B, the corresponding automorphism is τ_e(f) = τⁿ(f).

With these definitions two ergodic transformations τ₁, τ₂ of B₁ and B₂ arise from the same flow provided there are non-zero projections e₁ and e₂ in B₁ and B₂ such that the systems (τ₁)_e1, e₁B₁ and (τ₂)_e2, e₂B₂ are isomorphic.

See also

• Anosov flow
• Axiom A

Notes and References

1. Applying the same argument to 1 − f_N and 1 − p, shows that if g_N tends to 1 − p in L¹(X) with 0 ≤ g_N ≤ 1, then χ_[1–ε,1](g_N) tends to p in L¹(X).
2. This is a prototype for the relation of measure equivalence defined by Gromov. In that case Z and R are replaced by two discrete countable groups and the invariant subalgebras by the \ell ^{infty} functions on the two groups.