Orbit quotients and modular types

Topology can describe all orbits at once, whereas a factor is built from one measure class. These are different levels of information. We connect them by turning an arbitrary ergodic probability into a nonsingular one. For a countable group, the averaging can be chosen so that every group element has a bounded Radon–Nikodym derivative. This also gives explicit entire modular orbit generators.

The topological results in the first two sections allow an arbitrary Polish group. The measured relation and factor constructions from Section 3 onward use a countable discrete group. All spaces are nonempty unless stated otherwise.

1. The four topological orbit conditions

Let a Polish group \(G\) act continuously on a Polish space \(X\), and let \(\pi:X\to Y=X/G\). Give \(Y\) the quotient topology: \(V\subset Y\) is open when \(\pi^{-1}(V)\) is open. For every open \(U\subset X\),

is open. Thus \(\pi\) is open as well as continuous. A countable base \((U_n)\) of \(X\) gives the countable quotient base \((\pi(U_n))\): if \(\pi(x)\in V\), choose \(x\in U_n\subset\pi^{-1}(V)\), and then \(\pi(U_n)\subset V\).

For \(x\in X\), the closed stabilizer is \(H_x=\{g:gx=x\}\). The map

is a continuous bijection, with the orbit carrying its relative topology.

The complete earlier theorem is Polish orbits and their quotient topology, Theorem 4.1, abbreviated POT. It proves equivalence of all four statements:

1. Every map (OQT.2) is a homeomorphism.
2. Every orbit is nonmeagre in itself.
3. Every orbit is \(G_\delta\) in \(X\).
4. \(Y\) is \(T_0\): distinct orbits are distinguished by an open quotient set.

The exact proof inputs within POT are Lemmas 1.1–1.3, the complete-parameter intersection argument of Lemma 2.1, and the open-orbit proof of Theorem 3.1. Theorem 4.1 supplies every implication. In particular, the proof never assumes an orbit is completely metrizable before deriving that conclusion. Lemma 1.3 constructs the required \(G_\delta\) open image using locally finite metric refinements and nested sets in the complete parameter space. The Borel-separation alternative following Lemma 2.1 is a second supplied proof, not an additional assumption.

There is a useful explicit code. Put

Under the equivalent conditions, this Borel map has exactly the orbits as its fibres. Equal codes mean that two quotient points belong to exactly the same basic open sets, and \(T_0\) then identifies them. This does not alone show that the image of \(c\) is Borel or that every invariant Borel set is generated by the \(GU_n\).

For the rotation in Takesaki XIII.2, Exercise 1(a), the conclusion needs irrational angle. POT Example 5.1 proves density of each irrational rotation orbit by a small-step argument and obtains an indiscrete quotient. The same example proves the full rational contrast: for angle \(p/q\) in lowest terms, \(z\mapsto z^q\) identifies the quotient with a circle. The printed exercise omits the irrationality hypothesis. We retain the existing programme correction.

2. Closures, quotient Borel sets and all ergodic measures

The quotient Borel structure is

It always contains the sigma-algebra generated by the quotient topology, but equality needs proof.

The closure hypotheses in Takesaki XIII.2, Exercise 2, are precise:

• C: for each identity neighborhood \(N\), some identity neighborhood \(M\subset N\) satisfies \(\overline{Mx}^{\,X}\subset Nx\) for every \(x\).
• D: such an \(M\) can be chosen so that, for every \(x\) and every open neighborhood base \((Q_n)\) at \(x\),

The bars denote closure in \(X\), not closure in the group. The base need not be nested. D implies C because \(Mx\subset M Q_n\) for each \(n\).

The complete proof is Locally closed orbits and measurable representatives, Theorem 1.1 and Sections 2–6, abbreviated LCR. Under C it proves equivalence of:

1. Every orbit is locally closed, meaning open in its closure.
2. \(\mathcal B_q(Y)\) is generated by the quotient topology.
3. The quotient Borel structure has a countable family separating all its points.
4. Every nonzero ergodic sigma-finite Borel measure on \(X\) is concentrated on one orbit.
5. The quotient is a standard Borel space.
6. There is a Borel transversal, meeting each orbit exactly once.

This includes all four assertions under C and the two further assertions under D requested by the source exercise. The supplied selector proof in fact establishes the stronger six-way statement already under C.

Here is the complete route to the non-elementary implications, so none is replaced by a source citation. LCR Lemmas 2.1–2.2 and Proposition 2.3 turn a countable invariant code into a locally closed orbit. Section 2 uses the precise programme Blackwell and injective-image proofs to compare the two Borel structures. Lemmas 3.1–3.2 prove the analytic Baire-property and product-category results. Proposition 3.3, Theorem 4.1, Lemma 5.1 and Proposition 5.2 construct an ergodic probability which gives every orbit zero mass whenever local closedness fails. Lemma 6.1 and Proposition 6.2 construct a Borel selector by nested closed sets while preserving every orbit-code bit. The final paragraphs of Section 6 prove both remaining reverse implications.

In statement 4, ergodicity concerns invariant Borel sets; quasi-invariance is not assumed. The counterexample measure in Proposition 5.2 is not asserted to be quasi-invariant. This distinction is essential for the passage to relation algebras below.

3. Concentration, normalization and countable actions

The implication from countable separation to ergodic concentration has a short direct proof. Pull a countable separating family back to invariant Borel sets \(A_n\subset X\). For a nonzero ergodic measure \(\mu\), choose the conull side \(C_n\) of each \(A_n\). Their intersection is conull and nonempty. All its points have the same code, so they lie in a single orbit. In fact this code fibre is that entire orbit and is a countable intersection of Borel sets, making it Borel. No quasi-invariance or sigma-finiteness is used in this direction.

A nonzero sigma-finite Borel measure can be replaced by an equivalent probability without changing this ergodicity condition. To see this directly, disjointify a countable finite-measure cover into Borel sets \(B_n\), and give each nonnull \(B_n\) the density

The resulting measure has positive finite total mass; divide by that mass. Its density is positive and finite almost everywhere, so null and conull sets are unchanged. This construction does not assert invariance of the new probability.

From now on let \(G\) be countable discrete. C holds with \(M=\{e\}\), because \(X\) is Hausdorff. D holds with the same choice: in a metric space, the closures of all members of a neighborhood base at \(x\) intersect in \(\{x\}\). Thus the full criterion of Section 2 applies to every continuous countable discrete action.

For its relation \(R\), every orbit is a countable Borel set. Given a nonsingular nonzero sigma-finite measure, 1. Whole operator fields and exact proof inputs through 4. The complete factor-type classification prove

We cannot apply this to an arbitrary probability supplied by LCR until nonsingularity has been established.

4. A controlled nonsingular probability

Let \(\rho\) be any Borel probability on a standard Borel \(G\)-space. No topology on \(X\) is needed in this section. Enumerate generators \(s_1,s_2,\ldots\) covering \(G\), permitting repetitions. Allow letters \(s_j\) and \(s_j^{-1}\), each of integer cost \(j+2\). Define \(\ell(g)\) to be the smallest total cost of a word representing \(g\), with the empty word of cost zero. Every group element has such a word. The minimum exists because the costs are nonnegative integers. Reversal and concatenation give

Only finitely many words have cost at most a fixed number, so \(\ell\) has finite balls.

The sum of \(2^{-\text{cost}}\) over all letters is

Choose one minimum-cost word for each element, for example the first in a fixed enumeration. These words are distinct for distinct elements. Summing over all words therefore gives

Put \(w_g=2^{-\ell(g)}/Z\), and define the probability

Countable additivity follows by exchanging two nonnegative sums. Every \(w_g\) is positive, so \(\nu(A)=0\) exactly when \(\rho(g^{-1}A)=0\) for every \(g\). Reindexing proves that this null ideal is invariant under every group element. Thus the action is nonsingular for \(\nu\).

There is also a quantitative estimate. For fixed \(h\in G\),

Termwise comparison gives the inequalities of measures

The positive-measure Radon–Nikodym theorem, with its full programme proof identified in 1. Counting arrows and fixing the derivative, supplies

For example, the upper bound follows by integrating over a set where the density exceeds it; a positive-measure violating set would contradict (OQT.12). The lower bound follows in the same way.

For every invariant Borel \(A\), all terms in (OQT.10) evaluate to \(\rho(A)\), so

Consequently \(\rho\)-ergodicity implies \(\nu\)-ergodicity. Each orbit is invariant; if every orbit has \(\rho\)-mass zero, every orbit has \(\nu\)-mass zero.

Finally, \(\rho\ll\nu\), since \(w_e>0\). If the original action is already nonsingular for \(\rho\), every \(g_*\rho\) is equivalent to \(\rho\), and then \(\nu\sim\rho\). Without that hypothesis the averaging can enlarge the measure class. Only the ergodicity and orbit-mass assertions, not equivalence to \(\rho\), are used for the general witness construction.

5. Bounded derivatives give entire modular generators

Use the probability \(\nu\) from Section 4, and its principal relation algebra. Its diagonal state is \(\varphi_\nu(T)=\int E(T)\,d\nu\). Let \(V_g\) be the unitary acting on the range coordinate by \(g\). With the source-counting convention \(\delta=d\nu_r/d\nu_s\), the partial-map derivative formula of RWT-01 says

These are identities on the relation. Countability and nonsingularity allow all fixed-element identities and bounds to hold on one invariant conull set.

The modular formula of 2. The weight model at infinite mass now gives, for real \(t\),

The inverse power is forced by (OQT.15).

Since \(\|\log h_g\|_\infty\leq\ell(g)\log2\), the norm-convergent power series for the bounded multiplier defines an entire extension

Indeed the exponential series converges uniformly in operator norm on every bounded subset of \(\mathbb C\), by comparison with the scalar exponential series. Its termwise derivative converges there as well, proving complex differentiability. The norm estimate follows from \(|h_g^{-iz}|=h_g^{\operatorname{Im}z}\) and (OQT.13).

Every diagonal multiplier is fixed by the modular group. Hence for bounded \(f\),

The finite linear span of the \(M_fV_g\) is a unital star algebra: multiplication uses \(V_gM_fV_g^*=M_{f\circ g^{-1}}\), and adjoints use \(V_g^*=V_{g^{-1}}\). It generates \(M\) as a von Neumann algebra by definition, and every element is entire by the finite sum of the displayed extensions. No uniform bound in \(g\) is asserted. The cost \(\ell(g)\) records exactly how the estimate depends on the chosen generator.

6. When all ergodic relation factors are type I

Theorem. For a continuous action of a countable discrete group on a nonempty Polish space, the following are equivalent:

1. Every orbit is locally closed.
2. The quotient Borel structure is countably separated.
3. There is a Borel transversal.
4. For every nonsingular ergodic Borel probability \(\nu\), the principal relation factor \(\mathcal M(R,\nu)\) is type I.

The equivalent conditions of Sections 1–2 then hold as well.

Proof. Conditions 1–3 are equivalent by LCR and the automatic C/D observation in Section 3. Under these conditions, every nonzero ergodic Borel probability is concentrated on one orbit. For a nonsingular one, (OQT.7) gives type I, proving \(1\Rightarrow4\).

For the converse, suppose some orbit is not locally closed. The complete construction in LCR Proposition 3.3, Theorem 4.1, Lemma 5.1 and Proposition 5.2 gives a Borel probability \(\rho\) which is ergodic and gives every orbit zero mass. Its construction is explicit: continuously embed eventual equality of binary sequences into the relation on the closure of a bad orbit, and push forward the fair binary probability. The two directions of that relation embedding make each orbit preimage one countable eventual-equality class, so it has probability zero. Invariant Borel sets pull back to tail events, and the supplied zero–one proof gives ergodicity.

Apply Section 4 to this \(\rho\). The resulting \(\nu\) is a nonsingular ergodic probability and still gives every orbit zero mass. Its relation algebra is a factor, but (OQT.7) says it is not type I. This contradicts condition 4. The averaging moreover gives the bounded derivatives and entire modular generators of Section 5. \(\square\)

The theorem does not determine whether this non-type-I factor is \(II_1\), \(II_\infty\), or III. That depends on the existence and total mass of equivalent sigma-finite invariant measures, through RFT-04. No type classification by topology alone is being substituted for that measure-class criterion.

The universal quantifier is essential: one particular type I factor cannot certify the topology of all orbits, as the next section shows.

7. Two tests of the topology–measure distinction

Rational translations give type \(II_\infty\). Let the discrete countable group \(\mathbb Q\) translate \(\mathbb R\). Every orbit is countable and dense, so none is locally closed. LCR Section 7 proves directly that Lebesgue measure is ergodic: rational invariance forces equal relative masses on all dyadic cells; finite-measure uniqueness gives \(\lambda|_A=c\lambda\), and evaluation on \(A\) forces \(c=0\) or \(1\). That full proof uses no assumption that an arbitrary invariant set is open.

Make an equivalent probability \(\rho=r\lambda\) by putting

The sum of the interval masses is \((1+2\sum_{k\geq1}2^{-k})/3=1\), and \(r>0\). Apply Section 4. All translations preserve the Lebesgue class, so the resulting probability is still equivalent to Lebesgue measure. More explicitly,

Tonelli proves the density formula and its integral one. The bounds follow termwise. The action is ergodic, its countable orbits have mass zero, and Lebesgue measure is an equivalent infinite invariant measure. RFT-04 therefore gives a type \(II_\infty\) factor, although \(\varphi_\nu\) is a state.

For the translation by \(g\),

The numerator is a shifted density, not its reciprocal. Formula (OQT.16) then supplies the reciprocal imaginary power for the modular orbit unitary. The bounds (OQT.13) hold even though no positive global lower bound on \(q\) was imposed.

A null bad component does not change a type I factor. Let \(X\) be the disjoint union of a circle and one isolated point \(*\). Let \(\mathbb Z\) rotate the circle by an irrational angle and fix \(*\). This is a continuous action on a Polish space. The measure \(\delta_*\) is nonsingular and ergodic: the distinguished point is fixed, and a null set misses it. Its conull orbit has one point, so its relation algebra is \(\mathbb C\), of type \(I_1\).

The full topological quotient nevertheless fails \(T_0\). Distinct circle orbits cannot be separated by invariant open sets inside the circle, by POT Example 5.1. Adding the isolated fixed component does not separate those quotient points. Thus a measured algebra can ignore all the topologically bad orbits when they lie in a null component.

8. Problems with complete solutions

Problem 1. Why is a compact identity neighborhood enough for C and D in a locally compact Polish group action?

Solution. Choose a compact identity neighborhood \(M\subset N\). Its continuous image \(Mx\) is compact and closed, proving C. For \(y\in\bigcap_n\overline{M Q_n}\), choose base members inside balls of radius \(1/k\) about \(x\), and points \(m_kz_k\) within \(1/k\) of \(y\), with \(m_k\in M\) and \(z_k\) in those base members. Then \(z_k\to x\); compactness gives a subsequence \(m_k\to m\in M\), so \(y=mx\in Nx\). This proves D for a nonnested base. It is the full argument of POT Proposition 6.1.

Problem 2. Can averaging in Section 4 preserve every null set of a probability that is not nonsingular?

Solution. Not in general. Let \(C_2\) interchange two points and start with a point mass at the first. Any average with strictly positive weights on both group elements has positive mass at both points. The second singleton changes from null to nonnull. The average remains ergodic because the action is transitive. Section 4 asserts equivalence to the original probability only when the original action was already nonsingular.

Problem 3. Check the sign in (OQT.16) for the transposition \(s\) of two points with masses \(a,b>0\).

Solution. At the first point, \(h_s(1)=b/a\), and at the second, \(h_s(2)=a/b\). The swap is \(V_s=E_{12}+E_{21}\). The diagonal state has density \(\operatorname{diag}(a,b)\), so its modular action gives

Using \(h_s^{it}\) reverses both ratios.

Problem 4. Explain why the witness probability in LCR must be averaged before invoking the factor criterion.

Solution. Its ergodicity tests invariant Borel sets, and its construction does not assert that group elements preserve its null ideal. The principal relation weight construction requires nonsingularity. Section 4 supplies that missing property, keeps every orbit null, and preserves ergodicity by (OQT.14). The new measure class may be larger; no equivalence to the unaveraged witness is needed.

Problem 5. Does \(T_0\) mean the quotient is Hausdorff?

Solution. No. For positive scaling on \(\mathbb R\), the three orbits are \((0,\infty)\), \((-\infty,0)\), and \(\{0\}\). The positive and negative orbit points are open in the quotient, while every neighborhood of the zero orbit is the whole quotient. It is \(T_0\) and not Hausdorff. The Borel quotient is a three-point standard Borel space, with transversal \(\{-1,0,1\}\), as POT Example 5.3 and LCR Section 7 prove.

References

Masamichi Takesaki, Theory of Operator Algebras III, Chapter XIII, Exercises 2(1)–(2), printed pp. 30–31. Publisher record. The Effros and Glimm results and their full programme proofs are identified in Sections 1–2. The missing irrationality qualification is the existing correction in POT Example 5.1. The averaging argument, quantitative modular estimates and type-I consequence here are independently written course exposition. New text is released under CC0.