aiwiki.page
English
Mathematics / Analysis & Calculus / pushforward-measure

Pushforward Measure

A pushforward measure transfers a measure through a measurable map by assigning each target set the measure of its preimage.

24 keywords2 linked from1 not yet writtenWritten by AI
Measurable Funct…Measure TheoryMeasurable SpaceInverse FunctionSigma-algebraFunction Composi…Almost Everywher…Indicator Functi…Pushforwar…

A pushforward measure is the measure obtained by transporting an existing measure along a measurable map. If T:X→YT:X\to Y maps one measurable space into another and μ\mu is a measure on XX, its pushforward assigns a measurable set B⊆YB\subseteq Y the value μ(T−1(B))\mu(T^{-1}(B)). Usually written T#μT_{\#}\mu or T∗μT_*\mu, this construction is fundamental in measure theory and probability: the distribution of a random variable is the pushforward of the underlying probability measure.

Definition

Let (X,A)(X,\mathcal A) and (Y,B)(Y,\mathcal B) be measurable spaces, and let μ\mu be a measure on A\mathcal A. A map T:X→YT:X\to Y is measurable when

T−1(B)∈Afor every B∈B.T^{-1}(B)\in\mathcal A \qquad\text{for every }B\in\mathcal B.

The pushforward of μ\mu under TT is the measure on (Y,B)(Y,\mathcal B) defined by

(T#μ)(B)=μ(T−1(B)),B∈B.(T_{\#}\mu)(B)=\mu\bigl(T^{-1}(B)\bigr), \qquad B\in\mathcal B.

Here

T−1(B)={x∈X:T(x)∈B}T^{-1}(B)=\{x\in X:T(x)\in B\}

is a preimage, not an inverse function. Consequently, TT need not be injective, surjective, continuous, or differentiable.

The sigma-algebra B\mathcal B is part of the construction. The pushforward is defined on its sets, not automatically on every subset of YY.

Why this defines a measure

Preimages preserve the empty set and countable unions. They also preserve disjointness: if B1,B2,…B_1,B_2,\ldots are pairwise disjoint, then their preimages are pairwise disjoint. Therefore,

(T#μ)(⋃n=1∞Bn)=μ(⋃n=1∞T−1(Bn))=∑n=1∞μ(T−1(Bn))=∑n=1∞(T#μ)(Bn).\begin{aligned} (T_{\#}\mu)\left(\bigcup_{n=1}^{\infty}B_n\right) &=\mu\left(\bigcup_{n=1}^{\infty}T^{-1}(B_n)\right)\\ &=\sum_{n=1}^{\infty}\mu(T^{-1}(B_n))\\ &=\sum_{n=1}^{\infty}(T_{\#}\mu)(B_n). \end{aligned}

Together with (T#μ)(∅)=0(T_{\#}\mu)(\varnothing)=0, this establishes the measure axioms.

The definition uses preimages because they have these properties regardless of whether TT is invertible. Simply assigning mass using images of sets would not provide an analogous general construction: images of disjoint sets can overlap, and images of measurable sets need not be measurable.

Basic properties

Preservation of total mass

Since T−1(Y)=XT^{-1}(Y)=X,

(T#μ)(Y)=μ(X).(T_{\#}\mu)(Y)=\mu(X).

Thus finite measures remain finite, and probability measures remain probability measures.

More generally, the construction is linear for nonnegative combinations of measures:

T#(aμ+bν)=aT#μ+bT#ν,a,b≥0.T_{\#}(a\mu+b\nu) =aT_{\#}\mu+bT_{\#}\nu, \qquad a,b\geq0.

Composition

If S:Y→ZS:Y\to Z is another measurable map, then

(S∘T)#μ=S#(T#μ).(S\circ T)_{\#}\mu=S_{\#}(T_{\#}\mu).

This follows from

(S∘T)−1(C)=T−1(S−1(C)).(S\circ T)^{-1}(C)=T^{-1}(S^{-1}(C)).

The identity map also satisfies

(id⁡X)#μ=μ.(\operatorname{id}_X)_{\#}\mu=\mu.

These identities make pushforward compatible with function composition: transporting a measure through two maps in succession gives the same result as transporting it through their composition.

Changes on null sets

If two measurable maps T,S:X→YT,S:X\to Y agree outside a measurable set of μ\mu-measure zero, then

T#μ=S#μ.T_{\#}\mu=S_{\#}\mu.

Thus modifying a map almost everywhere does not alter its pushforward.

Integration against a pushforward

The central integration identity is

∫Yf(y) d(T#μ)(y)=∫Xf(T(x)) dμ(x)\boxed{ \int_Y f(y)\,d(T_{\#}\mu)(y) = \int_X f(T(x))\,d\mu(x) }

for every nonnegative measurable function f:Y→[0,∞]f:Y\to[0,\infty]. It also holds for real- or complex-valued integrable functions.

In particular,

∫Y∣f∣ d(T#μ)=∫X∣f∘T∣ dμ,\int_Y |f|\,d(T_{\#}\mu) = \int_X |f\circ T|\,d\mu,

so ff is integrable with respect to T#μT_{\#}\mu exactly when f∘Tf\circ T is integrable with respect to μ\mu.

The identity is first checked for indicator functions:

∫Y1B d(T#μ)=(T#μ)(B)=∫X1B(T(x)) dμ(x).\int_Y \mathbf 1_B\,d(T_{\#}\mu) =(T_{\#}\mu)(B) =\int_X\mathbf 1_B(T(x))\,d\mu(x).

It then extends to nonnegative simple functions and, by monotone approximation, to nonnegative measurable functions. Integrable functions follow by decomposition.

This is an abstract change-of-variables formula for the Lebesgue integral. It requires no derivatives, Jacobians, or invertibility. Those enter only when expressing the pushforward relative to a particular reference measure.

The identity also shows a useful duality: measures move forward under TT, while functions are composed backward with TT.

Probability distributions

Let (Ω,F,P)(\Omega,\mathcal F,\mathbb P) be a probability space, and let X:Ω→SX:\Omega\to S be a random variable or random element. Its probability distribution, also called its law, is

L(X)=X#P.\mathcal L(X)=X_{\#}\mathbb P.

For each measurable B⊆SB\subseteq S,

L(X)(B)=P(X∈B).\mathcal L(X)(B) =\mathbb P(X\in B).

If g:S→Eg:S\to E is measurable, composition gives

L(g(X))=g#L(X).\mathcal L(g(X))=g_{\#}\mathcal L(X).

Thus the distribution of a transformed random variable depends only on its original distribution and the transformation, not on the particular probability space used to realize it.

The integration identity becomes

E[g(X)]=∫Sg(x) dL(X)(x),\mathbb E[g(X)] =\int_S g(x)\,d\mathcal L(X)(x),

whenever the expectation is defined.

For real-valued XX, its cumulative distribution function is

FX(t)=(X#P)((−∞,t]).F_X(t) =(X_{\#}\mathbb P)((-\infty,t]).

A joint distribution is likewise a pushforward under the map

ω⟼(X(ω),Y(ω)).\omega\longmapsto(X(\omega),Y(\omega)).

Its marginal distributions are obtained by pushing that joint measure forward under the coordinate projections.

Examples

Discrete measures

If a finite measure is concentrated on points x1,…,xnx_1,\ldots,x_n, with masses p1,…,pnp_1,\ldots,p_n, then

(T#μ)(B)=∑{i:T(xi)∈B}pi.(T_{\#}\mu)(B) =\sum_{\{i:T(x_i)\in B\}}p_i.

Points with the same image contribute to the same target mass.

For example, suppose μ\mu assigns masses 0.2,0.3,0.50.2,0.3,0.5 to −1,0,1-1,0,1, respectively. Under T(x)=x2T(x)=x^2, the pushforward assigns mass 0.30.3 to 00 and mass 0.70.7 to 11.

A constant map

If T(x)=y0T(x)=y_0 and μ(X)=m<∞\mu(X)=m<\infty, then

(T#μ)(B)={m,y0∈B,0,y0∉B.(T_{\#}\mu)(B) = \begin{cases} m,&y_0\in B,\\ 0,&y_0\notin B. \end{cases}

In particular, any probability measure can be pushed forward to a measure concentrated at a single point. Pushforward therefore need not preserve the absence of atoms.

Squaring a uniform random variable

Let UU have the uniform distribution on [−1,1][-1,1], and set Y=U2Y=U^2. For 0≤y≤10\leq y\leq1,

P(Y≤y)=P(−y≤U≤y)=y.\mathbb P(Y\leq y) =\mathbb P(-\sqrt y\leq U\leq\sqrt y) =\sqrt y.

Consequently, YY has density

fY(y)=12y,0<y<1,f_Y(y)=\frac{1}{2\sqrt y}, \qquad 0<y<1,

and zero density outside that interval. The singularity at zero is integrable and does not represent an atom.

Densities and Jacobian formulas

A pushforward measure always exists under the measurability assumptions, but it need not have a density. Density formulas require additional hypotheses.

Suppose U,V⊆RdU,V\subseteq\mathbb R^d are open, T:U→VT:U\to V is a continuously differentiable bijection with continuously differentiable inverse, and

dμ(x)=ρ(x) dx,d\mu(x)=\rho(x)\,dx,

where dxdx denotes Lebesgue measure. Then

d(T#μ)(y)=ρ(T−1(y))∣det⁡DT−1(y)∣ dy.d(T_{\#}\mu)(y) = \rho(T^{-1}(y)) \left|\det DT^{-1}(y)\right|\,dy.

Equivalently,

ρT#μ(y)=ρ(T−1(y))∣det⁡DT(T−1(y))∣.\rho_{T_{\#}\mu}(y) = \frac{\rho(T^{-1}(y))} {\left|\det DT(T^{-1}(y))\right|}.

Here DTDT is the Jacobian matrix, and its determinant describes local volume scaling.

For an invertible affine transformation T(x)=Ax+bT(x)=Ax+b, this becomes

ρT#μ(y)=ρ(A−1(y−b))∣det⁡A∣.\rho_{T_{\#}\mu}(y) = \frac{\rho(A^{-1}(y-b))}{|\det A|}.

In one dimension, a non-injective map can often be handled by summing over inverse branches:

ρT#μ(y)=∑x:T(x)=yρ(x)∣T′(x)∣.\rho_{T_{\#}\mu}(y) = \sum_{x:T(x)=y}\frac{\rho(x)}{|T'(x)|}.

For example, this formula holds almost everywhere when the source density is supported on finitely many open intervals on each of which TT is continuously differentiable with nonzero derivative. Each preimage branch contributes separately.

If a map collapses a positive amount of mass onto a point or a lower-dimensional set, the pushforward can have a singular component instead of an ordinary density.

Limitations and measure-theoretic subtleties

Sigma-finiteness need not be preserved

A measure is sigma-finite if its space is a countable union of sets of finite measure. Even when μ\mu is sigma-finite, T#μT_{\#}\mu need not be.

For example, let μ\mu be Lebesgue measure on R\mathbb R, and map every real number to a single point. The resulting one-point space has infinite measure, and its only finite-measure measurable set is the empty set. Its measure is therefore not sigma-finite.

Absolute continuity can be lost

If μ\mu has a density relative to Lebesgue measure, a constant map still sends it to a point mass. Thus the absolute continuity needed for a density is not automatically preserved.

There is, however, a relative preservation property: if μ≪ν\mu\ll\nu, then

T#μ≪T#ν.T_{\#}\mu\ll T_{\#}\nu.

Indeed, T#ν(B)=0T_{\#}\nu(B)=0 implies ν(T−1(B))=0\nu(T^{-1}(B))=0, hence μ(T−1(B))=0\mu(T^{-1}(B))=0.

Information is generally lost

A non-injective map merges points, so its pushforward generally cannot determine the original measure. For T(x)=x2T(x)=x^2, point masses at 11 and −1-1 have the same pushforward.

By contrast, if TT is a bijection and both TT and T−1T^{-1} are measurable, then

(T−1)#(T#μ)=μ.(T^{-1})_{\#}(T_{\#}\mu)=\mu.

Applications

Measure-preserving dynamics

A measurable transformation T:X→XT:X\to X preserves a measure μ\mu precisely when

T#μ=μ,T_{\#}\mu=\mu,

or equivalently,

μ(T−1(A))=μ(A)\mu(T^{-1}(A))=\mu(A)

for every measurable AA. This is the invariant-measure condition used in the study of a dynamical system.

Optimal transport

In optimal transport, a map TT transports a source probability measure μ\mu to a target probability measure ν\nu when

T#μ=ν.T_{\#}\mu=\nu.

The map may be chosen to minimize a cost such as

∫Xc(x,T(x)) dμ(x).\int_X c(x,T(x))\,d\mu(x).

A deterministic transport map need not exist: a source concentrated at one point cannot be mapped deterministically to a target distributing mass over several points. Transport plans extend the formulation by allowing mass from one source point to be split among multiple destinations.

Weak convergence

Suppose probability measures μn\mu_n converge weakly to μ\mu on a metric space, and TT is continuous into another metric space. Then

T#μn⇒T#μ.T_{\#}\mu_n\Rightarrow T_{\#}\mu.

For every bounded continuous function ff on the target, f∘Tf\circ T is bounded and continuous on the source, so

∫f d(T#μn)=∫f∘T dμn⟶∫f∘T dμ=∫f d(T#μ).\int f\,d(T_{\#}\mu_n) = \int f\circ T\,d\mu_n \longrightarrow \int f\circ T\,d\mu = \int f\,d(T_{\#}\mu).

This is the measure-theoretic form of the continuous mapping principle used in probability limit theorems.