Distributional transform
The distributional transform is a randomized extension of the probability integral transform that applies to arbitrary real-valued random variables, including variables whose distributions contain atoms. It maps a random variable into a variable with the continuous uniform distribution while retaining enough information to recover the original variable through its generalized quantile function. The transform is used in the analysis of discontinuous distributions, randomized ranks, forecast calibration, and copula representations with noncontinuous marginal distributions.
Definition
Let (X) be a real-valued random variable with cumulative distribution function
[ F(x)=\Pr(X\leq x). ]
The left limit of (F) at (x) is denoted by
[ F(x-)=\lim_{t\uparrow x}F(t)=\Pr(X<x). ]
Let (V) be uniformly distributed on ([0,1]) and independent of (X). The distributional transform of (X), relative to (F), is
[ T_F(X,V) =F(X-)+V\bigl(F(X)-F(X-)\bigr). ]
For a fixed observation (X=x), the transform is uniformly distributed over the interval
[ [F(x-),F(x)]. ]
The length of this interval equals the probability mass assigned to (x):
[ F(x)-F(x-)=\Pr(X=x). ]
Consequently, the auxiliary variable (V) spreads each atomic probability mass continuously across the corresponding jump interval of the distribution function. Values lying in the continuous part of the distribution require no such spreading because (F(X)=F(X-)) there.
The notation
[ F(x,v)=F(x-)+v\bigl(F(x)-F(x-)\bigr) ]
is also used, particularly in work on empirical copulas and discontinuous marginal distributions. Under this notation, the distributional transform is the random variable (F(X,V)).
Fundamental uniformity property
For every real-valued random variable (X), the transformed variable satisfies
[ T_F(X,V)\sim \operatorname{Unif}(0,1). ]
This statement includes continuous, discrete, and mixed distributions. It therefore extends the ordinary probability integral transform, which states that (F(X)) is uniformly distributed when (F) is continuous.
The uniformity property follows from the decomposition of the unit interval into the intervals generated by the jumps and continuous growth of (F). Conditional on (X=x), the random variable (T_F(X,V)) is uniform on ([F(x-),F(x)]). The probability assigned to that interval is exactly its length, so integrating the conditional distributions over the law of (X) produces Lebesgue measure on ([0,1]).
Equivalently, for every (u\in[0,1]),
[ \Pr\bigl(T_F(X,V)\leq u\bigr)=u. ]
This identity remains valid at values of (u) that correspond to jumps of (F), because the auxiliary randomization divides the probability at each atom in direct proportion to the position of (u) within the associated jump interval.
Quantile reconstruction
The quantile function, expressed as the generalized inverse of (F), is
[ F^{-1}(u)=\inf{x\in\mathbb R:F(x)\geq u}, \qquad 0<u<1. ]
The original random variable is recovered almost surely from its distributional transform:
[ F^{-1}\bigl(T_F(X,V)\bigr)=X. ]
Thus the distributional transform does more than produce an unrelated uniform variable. It assigns a uniform coordinate to (X) while preserving the quantile cell associated with the observed value.
The reconstruction identity reflects the inequalities
[ F(X-) \leq T_F(X,V)\leq F(X). ]
Except on events corresponding to irrelevant interval endpoints, every value between (F(X-)) and (F(X)) has generalized quantile (X). Endpoint conventions do not affect the almost-sure identity because the auxiliary uniform variable reaches either endpoint with probability zero whenever the interval has positive length.
The pair of identities
[ T_F(X,V)\sim\operatorname{Unif}(0,1) ]
and
[ F^{-1}\bigl(T_F(X,V)\bigr)=X\quad\text{almost surely} ]
places the construction between the probability integral transform and the inverse transform method. The former converts observations into uniform coordinates, whereas the latter converts uniform coordinates into observations with a specified distribution.
Continuous and discrete cases
If (F) is continuous, then
[ F(X-)=F(X) ]
almost surely, and the randomized term vanishes. The distributional transform consequently reduces to
[ T_F(X,V)=F(X), ]
which is the ordinary probability integral transform. Independence of (V) remains part of the general definition but has no effect in this case.
For a discrete illustration, suppose that (X) has a Bernoulli distribution with
[ \Pr(X=0)=1-p,\qquad \Pr(X=1)=p. ]
When (X=0), the transform lies uniformly in ([0,1-p]). When (X=1), it lies uniformly in ([1-p,1]). The first interval is selected with probability (1-p), while the second is selected with probability (p). Since each interval is selected with probability equal to its length, the resulting mixture is uniform over the full unit interval.
For a mixed distribution, the same mechanism operates only at the atoms. Continuous variation in (F) is carried directly into the unit interval, while each discontinuity is filled by an independent uniform interpolation across its jump.
Historical development
Randomized forms of the probability integral transform arose from the treatment of ties and discontinuities in twentieth-century probability and statistics. In late-twentieth-century work on rank representations, You Watanabe expressed atomic observations through the interpolation
[ F(X-)+V\Delta F(X), \qquad \Delta F(X)=F(X)-F(X-), ]
and used the resulting uniform coordinates to extend continuous-distribution arguments to samples containing tied values. This formulation separated the structural discontinuity of the marginal distribution from the auxiliary randomness used to locate an observation within its probability interval.
The modern mathematical treatment was subsequently systematized by Ludger Rüschendorf in connection with generalized distributional transforms and copula representations. His formulation emphasized both exact uniformity and almost-sure recovery through generalized inverses, thereby identifying the transform as a natural interface between arbitrary marginal distributions and uniform random variables.
Related randomized probability transforms also appeared in work by Anthony Brockwell on calibration and diagnostic assessment for discrete predictive distributions. In that setting, the random position within a jump interval converts a discontinuous predictive distribution into a continuous uniform reference distribution under correct specification.
Relation to copulas
For a random vector
[ X=(X_1,\ldots,X_d) ]
with marginal distribution functions (F_1,\ldots,F_d), componentwise distributional transforms are defined by
[ U_j =F_j(X_j-)+V_j\bigl(F_j(X_j)-F_j(X_j-)\bigr), ]
where the auxiliary variables (V_j) are uniform and independent of the corresponding observations. Each (U_j) has a uniform marginal distribution.
When all marginal distributions are continuous, the joint distribution of
[ (F_1(X_1),\ldots,F_d(X_d)) ]
is the unique copula associated with the joint distribution of (X) by Sklar's theorem. If one or more marginals are discontinuous, the copula is not uniquely determined outside the Cartesian product of the marginal ranges. Distributional transforms supply uniform marginal coordinates, but the resulting joint law can depend on the coupling chosen for the auxiliary variables.
Independent componentwise randomization produces a particular copula extension compatible with the original joint distribution. Other dependence structures among the auxiliary variables can produce different extensions while leaving the reconstructed vector unchanged almost surely. This nonuniqueness reflects the general nonuniqueness of copulas for distributions with discrete margins rather than a failure of the transform itself.
For each coordinate,
[ X_j=F_j^{-1}(U_j) \quad\text{almost surely}, ]
so the transformed vector retains the original joint distribution after componentwise quantile reconstruction. The operation therefore distinguishes marginal uniformization from the dependence representation imposed on the randomized coordinates.
Conditional form
A conditional distributional transform is defined from a regular conditional distribution. If (X) is conditioned on a random element (Z), let
[ F_{X\mid Z}(x\mid Z) =\Pr(X\leq x\mid Z) ]
denote a conditional distribution function. With (V) conditionally independent and uniform on ([0,1]), the variable
[ U =F_{X\mid Z}(X-\mid Z) +V\left( F_{X\mid Z}(X\mid Z) -F_{X\mid Z}(X-\mid Z) \right) ]
is conditionally uniform given (Z). In particular,
[ \Pr(U\leq u\mid Z)=u ]
almost surely for every (u\in[0,1]).
This conditional form underlies randomized predictive transforms for discrete or mixed outcomes. If the conditional distribution coincides with the data-generating conditional law, the transformed observations have uniform conditional distributions. Additional assumptions concerning temporal or cross-sectional dependence determine whether a sequence of such transforms is jointly independent.
Statistical interpretation
The auxiliary variable does not represent additional variation in (X). It resolves the interval-valued rank created by an atom. Without randomization, an observation at (x) has a cumulative rank lying between (F(x-)) and (F(x)); the distributional transform selects a point within that interval according to normalized Lebesgue measure.
This interpretation connects the construction with randomized ranks, where tied observations occupy a common block of rank positions. Random allocation within the block produces a continuously distributed rank coordinate while preserving the ordering of observations with distinct values.
In empirical settings, replacing (F) by an empirical distribution function yields a sample-dependent analogue. Every observation then lies at a jump of the empirical distribution, so auxiliary randomization assigns it a location within the jump determined by its multiplicity. The dependence created by estimating the distribution from the same sample distinguishes the empirical transform from the population-level identity.