Quasi-Invariant Measure¶

A measure whose class of null sets, though not necessarily its numerical values, is preserved by every transformation in a specified action, so each pushforward remains equivalent to the original and changes density through a Radon–Nikodym cocycle.

Version: v2 · 2026-08-30 · History
Domain-specific #: 2600
Origin domain: measure theory
Subdomain: nonsingular transformations and group actions

Core Idea¶

Let a group \(G\) act by measurable automorphisms on a measurable space \((X,\mathcal B)\). A measure \(\mu\) is quasi-invariant under this action when every pushforward \(g_*\mu\), defined by

\[ (g_*\mu)(A)=\mu(g^{-1}A), \]

is equivalent to \(\mu\): each is absolutely continuous with respect to the other. Equivalently, transformations preserve the measure's null sets,

\[ \mu(A)=0 \quad\Longleftrightarrow\quad \mu(gA)=0, \]

even though they may change the positive numerical measure assigned to \(A\).^[1]^[n1]

Quasi-invariance is therefore invariance of a measure class, not invariance of a measure's values. If \(g_*\mu=\mu\) for every \(g\), the measure is invariant and automatically quasi-invariant. The converse can fail: a transformation can rescale density from place to place while preserving exactly which events are negligible.

Under standard sigma-finiteness conditions, the Radon–Nikodym theorem supplies a positive density

\[ J_g(x)=\frac{d(g_*\mu)}{d\mu}(x), \]

defined almost everywhere. These derivatives obey a chain or cocycle relation, with the precise ordering determined by pushforward convention. The cocycle records the controlled failure of exact invariance and is central to change-of-variables formulas, nonsingular dynamics, induced representations, and Gaussian translation theory.^[n1]^[2]

The abstraction's invariant is strict: one-sided absolute continuity is not enough when transformations are not being paired with an inverse, and mere overlap of supports is much weaker. What must survive is the entire null-set class under every action element in scope.

Structural Signature¶

The structure has six roles.

A measurable space \((X,\mathcal B)\).
A measure \(\mu\), commonly sigma-finite or a probability measure when Radon–Nikodym tools are used.
An invertible measurable transformation or group action whose inverse is also measurable.
The pushforward measure \(g_*\mu\), which evaluates sets after transport.
Mutual absolute continuity \(g_*\mu\sim\mu\), preserving every null set in both directions.
A Radon–Nikodym derivative or cocycle, when the hypotheses for its existence hold, describing how positive density changes.

The recognition test is not whether the Jacobian exists pointwise everywhere. It is whether the transformed and original measures belong to the same equivalence class. Radon–Nikodym derivatives are equivalence classes of functions defined almost everywhere; they need not have a special everywhere-positive version without extra regularity.

For a single invertible transformation \(T\), quasi-invariance means \(T_*\mu\sim\mu\). For a group action, it must hold for every \(g\in G\). Strong quasi-invariance can add continuity or smoothness of \(J_g(x)\), and relative invariance can require the derivative to depend only on \(g\); these are stricter variants.^[n1]

What It Is Not¶

It is not an invariant measure. Exact invariance requires \(g_*\mu=\mu\), hence \(J_g=1\) almost everywhere. Quasi-invariance permits nonconstant density change.

It is not merely absolute continuity in one direction. \(g_*\mu\ll\mu\) says every original null set remains transformed-null. Mutual absolute continuity also prevents the transformation from introducing new null sets relative to the original. With a genuine group of invertible automorphisms, checking all group elements can recover both directions, but the definition should retain equivalence explicitly.

It is not support preservation. Two measures can share topological support while being mutually singular, and equivalent measures can have different density profiles. Null-set equality is measure-theoretic, not simply geometric.

It is not relative invariance. A relatively invariant measure transforms by a scalar multiplier depending on the group element. General quasi-invariance allows the derivative to depend on both \(g\) and \(x\).

It is not a stationary distribution unless the action is the relevant dynamics and exact invariance holds. Nonsingular transformations preserve negligible events without preserving probabilities.

It is not a claim that a measure exists for every action. Infinite-dimensional spaces provide important obstructions; there is no analogue of translation-invariant Lebesgue measure under all translations, and Gaussian quasi-invariance survives only along a restricted Cameron–Martin subspace.^[1]^[2]

Scope of Application¶

The concept appears in measure theory, nonsingular ergodic theory, harmonic analysis, representation theory, probability on infinite-dimensional spaces, stochastic analysis, and geometry.

For group actions on homogeneous spaces \(G/H\), quasi-invariant measures allow integration and unitary representation construction even when no invariant measure exists. The Radon–Nikodym factor corrects the action so that operators preserve the \(L^2\) norm.^[n1]

In nonsingular dynamics, a transformation preserves measure class rather than measure. Orbit structure and null-set statements remain meaningful, while the derivative cocycle captures compression and expansion. This supports ratio ergodic theory and type classifications beyond probability-preserving systems.

For smooth diffeomorphisms, a measure with positive smooth density relative to local volume is typically transformed to an equivalent measure; the ordinary Jacobian contributes to \(J_g\). Critical assumptions include invertibility and nonvanishing density. A singular map that collapses a region can destroy equivalence.

For Gaussian measures, the Cameron–Martin theorem identifies exactly which translations preserve measure class. In finite dimensions, nondegenerate Gaussian measures remain equivalent under every translation. In infinite dimensions, translation by a Cameron–Martin vector gives an equivalent measure with an exponential Radon–Nikodym density, while a shift outside that subspace produces a mutually singular measure.^[2]

Clarity¶

The clearest diagnostic uses a null-set probe. Choose any measurable \(A\). If \(\mu(A)=0\), must its image or preimage under each transformation also have zero measure—and conversely? If yes for every set and action element, the measure is quasi-invariant.

A density-change calculation then explains the positive sets. Suppose on \(\mathbb R\), \(\mu(dx)=e^{-x^2/2}dx\), and translate by \(a\). The translated Gaussian has density proportional to \(e^{-(x-a)^2/2}\), which differs from the original but is positive wherever the original density is positive. The measures have the same null sets and are quasi-invariant. They are not invariant unless \(a=0\).

Contrast a Dirac measure \(\delta_0\). Translation by nonzero \(a\) gives \(\delta_a\); the singleton \(\{0\}\) has full original measure and zero transformed measure, while \(\{a\}\) behaves oppositely. The measures are singular, so no quasi-invariance.

The phrase “changes by a numerical function” is shorthand for a Radon–Nikodym density, not arbitrary multiplication of a scalar total.

Manages Complexity¶

Exact invariance is often unavailable or unnecessarily strong. Quasi-invariance preserves the information needed for almost-everywhere analysis: a property true outside a null set remains true outside a transformed null set. This lets analysts transport function spaces, integrals, and probabilistic statements through transformations without demanding unchanged volume.

The Radon–Nikodym cocycle compresses all local distortion into one multiplicative object. Rather than rederive a transformed integral from scratch, one writes

\[ \int_X f\,d(g_*\mu)=\int_X f(x)J_g(x)\,d\mu(x). \]

In representation theory the square root of an appropriate density corrects pullback to form a unitary operator. In stochastic analysis, the density becomes a likelihood ratio for a shifted process. In dynamics, its products along orbits track cumulative expansion.

The abstraction therefore manages the exact gap between rigid invariance and uncontrolled singular change.

Abstract Reasoning¶

Measure equivalence is transitive. If \(g_*\mu\sim\mu\) and \(h_*\mu\sim\mu\), then \((gh)_*\mu\sim\mu\). The derivative satisfies a cocycle relation such as

\[ J_{gh}(x)=J_g(x)\,J_h(g^{-1}x) \]

under one common pushforward convention. This relation guarantees consistency under composition.

If \(\nu\sim\mu\), then \(\nu\) lies in the same measure class. Quasi-invariance is consequently a property of the measure class: every measure equivalent to a quasi-invariant measure is also quasi-invariant, although its derivative cocycle changes by a multiplicative coboundary.

If a set is null once, every group translate is null. Thus essential properties—those asserted almost everywhere—can be moved along the action. However, numerical probabilities need correction by \(J_g\); treating them as unchanged would silently upgrade quasi-invariance to invariance.

Knowledge Transfer¶

The definition transfers literally across homogeneous spaces, dynamical systems, Gaussian path spaces, and geometric transformations because all use the same pushforward–equivalence–density structure. What changes is the source of \(J_g\): a Jacobian determinant, a modular function, a Cameron–Martin exponential, or another cocycle.

The term should not be transferred metaphorically to approximate stability, robustness, or “almost unchanged” data. Quasi-invariance is exact at the null-set level. A small numerical change can fail it, and a large density change can satisfy it.

The parent Invariance captures the portable idea that some property survives transformation; this node specifies that the preserved property is the measure class rather than pointwise values.

Examples¶

Lebesgue measure under translation. On \(\mathbb R^n\), translations preserve Lebesgue measure exactly, so it is invariant and hence quasi-invariant.

A positive density under a diffeomorphism. A smooth probability density that is positive everywhere transforms by the inverse Jacobian and remains equivalent under a smooth diffeomorphism. The values change while null sets agree.

Finite-dimensional Gaussian translation. A nondegenerate Gaussian measure and any translate have mutually positive Lebesgue densities, hence the same null sets. The likelihood ratio is nonconstant.

Abstract Wiener measure. Translation by a Cameron–Martin vector is quasi-invariant with an explicit exponential density; translation outside the Cameron–Martin space is singular. This boundary is a canonical use of the concept.^[2]

Homogeneous space. A locally compact group acting on \(G/H\) admits a quasi-invariant measure class under broad hypotheses even when a genuinely invariant measure may fail to exist.^[1]^[n1]

Dirac mass under translation. \(\delta_0\) and \(\delta_a\) for \(a\ne0\) are mutually singular. This is the minimal counterexample.

Structural Tensions¶

Exact invariance versus analytical reach. Requiring unchanged values gives stronger conservation; preserving only null sets makes many more actions tractable.

Equivalence class versus chosen density. The null-set structure is canonical at class level, while a particular \(\mu\) and its cocycle depend on normalization and representative.

Finite-dimensional intuition versus infinite-dimensional singularity. Smooth translations seem harmless in Euclidean space; most directions in infinite-dimensional Gaussian spaces make measures singular.

Pointwise formula versus almost-everywhere identity. Analysts often want a continuous derivative, but the Radon–Nikodym theorem supplies only an almost-everywhere class unless more structure is imposed.

Structural–Framed Character¶

This node is formal and domain-specific. Its identity requires sigma-algebras, measures, pushforwards, absolute continuity, and group actions. It transfers broadly across mathematical subfields but not outside measure-theoretic contexts without importing that apparatus.

Its structure is unusually crisp: preserve the null-set equivalence class while allowing density distortion. That precision makes it a strong domain-specific abstraction, but the mathematical vocabulary and exact theorem conditions keep it below the prime bar.

Structural Core vs. Domain Accent¶

The structural core is weakened invariance: a transformation preserves a selected equivalence class rather than every value. The domain accent declares the equivalence class to be mutual absolute continuity of measures and represents the deviation by a Radon–Nikodym cocycle.

Removing the domain accent leaves generic Invariance. Removing the preserved measure class leaves only arbitrary change. Both are necessary.

Quasi-Invariant Measure strictly instantiates Invariance at the level of null sets and measure class. It relates to Equivalence because mutual absolute continuity partitions measures into classes, and to Transformation because the property is action-relative.

The minimal proposed DAG parent is prime:invariance. It is not exact invariant measure coverage: the node's distinctive contribution is the controlled weakening from value preservation to null-set preservation.

Relationships to Other Abstractions¶

Current abstraction Quasi-Invariant Measure Domain-specific

Parents (1) — more general patterns this builds on

Quasi-Invariant Measure is a kind of Invariance Prime

Quasi-Invariant Measure strictly instantiates Invariance at the level of null sets and measure class.

Hierarchy path (1) — routes to 1 parentless root

Quasi-Invariant Measure → Invariance

Neighborhood in Abstraction Space¶

Quasi-Invariant Measure sits in a sparse region of the domain-specific corpus (78^th percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.

Family — Advanced Probability & Combinatorial Bounds (6 abstractions)

Nearest neighbors

Radon–Nikodym theorem — 0.84
Equivalence (measure theory) — 0.84
Hausdorff Space — 0.82
Koopman–von Neumann Classical Mechanics — 0.82
Quasiregular Representation — 0.82

Computed from structural-signature embeddings · 2026-09-08

Not to Be Confused With¶

Do not confuse quasi-invariance with invariant measure, stationary distribution, relative invariant measure, one-sided nonsingularity, equivalent support, approximate invariance, asymptotic invariance, ergodicity, or measure-preserving transformation. Each adds, removes, or changes a condition.

The phrase nonsingular transformation is closely related: it often means \(\mu\circ T^{-1}\ll\mu\). For invertible transformations with a nonsingular inverse, this yields equivalence. It should not be queued as an unrestricted alias without checking the author's convention.

Notes¶

[n1] Gerald B. Folland, A Course in Abstract Harmonic Analysis, 2^nd ed., CRC Press, sections on quasi-invariant and strongly quasi-invariant measures on homogeneous spaces. ↩a ↩b ↩c ↩d ↩e

[n2] David Pollard, A User's Guide to Measure Theoretic Probability, Cambridge University Press, chapters on absolute continuity and Radon–Nikodym derivatives.

References¶

[1] R. A. Minlos, “Quasi-invariant measure”, Encyclopedia of Mathematics, with references to Bourbaki and Gel'fand–Vilenkin. registry ↩a ↩b ↩c

[2] Vladimir I. Bogachev, Gaussian Measures, American Mathematical Society, 1998, Cameron–Martin theory and equivalence/singularity under translation. registry ↩a ↩b ↩c ↩d