Kim–Milman Construction

• Kim–Milman Construction is a deterministic transport-map that reverses an OU diffusion to map the Gaussian measure onto a target measure using a score-based ODE.
• It leverages semiconcavity conditions and Lipschitz regularity, enabling precise control over functional inequalities and stability estimates via Fisher information.
• The approach extends to manifold settings and offers sharper stability bounds compared to classical optimal transport methods like Brenier maps.

The Kim–Milman construction is a deterministic transport-map construction on $\mathbb{R}^d$ that sends the standard Gaussian reference measure $\gamma=\mathcal{N}(0,I_d)$ to a target probability measure $\mu$ by reversing the Ornstein–Uhlenbeck (OU) diffusion. In recent machine-learning literature, the same object is often called the probability flow ODE associated with a diffusion model. In the notation used in recent work, the resulting map is denoted either $T_{\mathrm{KM}^\mu}$ or $\mathsf T^\mu$, and it satisfies $(T_{\mathrm{KM}^\mu})_\#\gamma=\mu$ or $(\mathsf T^\mu)_\sharp\gamma=\mu$ (Chewi et al., 3 Nov 2025, Chewi et al., 22 Jun 2026).

1. Dynamical definition through Ornstein–Uhlenbeck reversal

The standard setup fixes the ambient space $\mathbb{R}^d$, the Gaussian reference

$\gamma=\mathcal{N}(0,I_d),$

and a target measure $\mu$ with finite second moment, or in the regular setting a smooth Lebesgue density. The forward OU process is initialized at the target: $\gamma=\mathcal{N}(0,I_d)$0 If $\gamma=\mathcal{N}(0,I_d)$1, then $\gamma=\mathcal{N}(0,I_d)$2, where $\gamma=\mathcal{N}(0,I_d)$3 is the OU semigroup, and $\gamma=\mathcal{N}(0,I_d)$4 as $\gamma=\mathcal{N}(0,I_d)$5 (Chewi et al., 3 Nov 2025).

On a finite horizon $\gamma=\mathcal{N}(0,I_d)$6, reversing time in the Fokker–Planck equation yields a continuity equation for the backward-time density $\gamma=\mathcal{N}(0,I_d)$7: $\gamma=\mathcal{N}(0,I_d)$8 This is the continuity equation for the deterministic ODE

$\gamma=\mathcal{N}(0,I_d)$9

Assuming $\mu$0, the drift can also be written in semigroup form: $\mu$1 This ODE is the Kim–Milman reverse heat flow ODE, and its flow map $\mu$2 satisfies

$\mu$3

In particular, $\mu$4 (Chewi et al., 22 Jun 2026).

The construction is therefore not variational: unlike Brenier transport, it is defined by reversing a diffusion and integrating an ODE whose drift is the score of the OU-smoothed target. This direct dependence on $\mu$5 is central to both its regularity theory and its stability theory.

2. Infinite-horizon limit and existence of the map

Letting $\mu$6 sends the initial law $\mu$7 to $\mu$8. In the regular regime, the Kim–Milman map is defined as the infinite-horizon limit

$\mu$9

and for each $T_{\mathrm{KM}^\mu}$0 one also obtains a stopped map

$T_{\mathrm{KM}^\mu}$1

For regular targets, a key representation is the integral equation

$T_{\mathrm{KM}^\mu}$2

and differentiation in $T_{\mathrm{KM}^\mu}$3 gives

$T_{\mathrm{KM}^\mu}$4

The Kim–Milman map is then $T_{\mathrm{KM}^\mu}$5 (Chewi et al., 22 Jun 2026).

A major recent development is a general existence theorem for arbitrary $T_{\mathrm{KM}^\mu}$6. A regular approximating sequence $T_{\mathrm{KM}^\mu}$7 is defined by requiring that every $T_{\mathrm{KM}^\mu}$8 satisfies the regularity condition $T_{\mathrm{KM}^\mu}$9 and that $\mathsf T^\mu$0. The paper gives the example: one can truncate $\mathsf T^\mu$1 to a large ball and convolve with a small Gaussian. For such an approximation, the limits

$\mathsf T^\mu$2

exist in $\mathsf T^\mu$3, are independent of the approximating sequence, and satisfy

$\mathsf T^\mu$4

Thus, for every $\mathsf T^\mu$5, the Kim–Milman map exists as an $\mathsf T^\mu$6 limit, without requiring $\mathsf T^\mu$7 on $\mathsf T^\mu$8 (Chewi et al., 22 Jun 2026).

This general existence theorem separates the questions of definition and quantitative control. Existence holds on all of $\mathsf T^\mu$9, while Lipschitz regularity and sharpened stability estimates still require structural assumptions on at least one target.

3. Semiconcavity along the OU semigroup and Lipschitz regularity

The main regularity hypothesis used in the recent theory is the semiconcavity condition

$(T_{\mathrm{KM}^\mu})_\#\gamma=\mu$0

Here $(T_{\mathrm{KM}^\mu})_\#\gamma=\mu$1 is a scalar control on the Hessian of the OU-smoothed log-density. Under $(T_{\mathrm{KM}^\mu})_\#\gamma=\mu$2, for all $(T_{\mathrm{KM}^\mu})_\#\gamma=\mu$3, the stopped flow map $(T_{\mathrm{KM}^\mu})_\#\gamma=\mu$4 exists and is Lipschitz with constant

$(T_{\mathrm{KM}^\mu})_\#\gamma=\mu$5

In particular, the limiting map $(T_{\mathrm{KM}^\mu})_\#\gamma=\mu$6 is Lipschitz with constant $(T_{\mathrm{KM}^\mu})_\#\gamma=\mu$7. Moreover, the OU marginals satisfy the log-Sobolev inequality

$(T_{\mathrm{KM}^\mu})_\#\gamma=\mu$8

(Chewi et al., 22 Jun 2026).

Several classes of targets are known to satisfy $(T_{\mathrm{KM}^\mu})_\#\gamma=\mu$9. If $(\mathsf T^\mu)_\sharp\gamma=\mu$0 with $(\mathsf T^\mu)_\sharp\gamma=\mu$1 $(\mathsf T^\mu)_\sharp\gamma=\mu$2-strongly convex, then

$(\mathsf T^\mu)_\sharp\gamma=\mu$3

and $(\mathsf T^\mu)_\sharp\gamma=\mu$4. This is the original Kim–Milman setting and yields $(\mathsf T^\mu)_\sharp\gamma=\mu$5. The theory also covers strongly convex plus Lipschitz perturbation and more general asymptotically convex potentials characterized by a convexity profile $(\mathsf T^\mu)_\sharp\gamma=\mu$6; in these cases, precise formulas for $(\mathsf T^\mu)_\sharp\gamma=\mu$7 are available and satisfy the required integrability and boundedness properties (Chewi et al., 22 Jun 2026).

In the stability paper, the same structural condition is written as

$(\mathsf T^\mu)_\sharp\gamma=\mu$8

and explicit $(\mathsf T^\mu)_\sharp\gamma=\mu$9 are developed for three regimes: strong log-concavity, strong log-concavity plus log-Lipschitz perturbations, and asymptotically positive convex profiles (Chewi et al., 3 Nov 2025).

These bounds are curvature-type conditions along the OU flow. They control the Lipschitz modulus of the reverse-time drift and are the basic input for both Grönwall estimates on trajectories and functional inequalities on the OU marginals.

4. Fisher-information stability and the score-based viewpoint

A central recent result is that stability of the Kim–Milman flow map is controlled not by Wasserstein distance, but by relative Fisher information. For $\mathbb{R}^d$0, the relative Fisher information is

$\mathbb{R}^d$1

and the $\mathbb{R}^d$2-version is

$\mathbb{R}^d$3

If $\mathbb{R}^d$4 and $\mathbb{R}^d$5 denote the Kim–Milman maps associated with two targets $\mathbb{R}^d$6 and $\mathbb{R}^d$7, then under suitable regularity assumptions on $\mathbb{R}^d$8,

$\mathbb{R}^d$9

where $\gamma=\mathcal{N}(0,I_d),$0 is an explicit constant depending on curvature-type bounds associated with $\gamma=\mathcal{N}(0,I_d),$1. In important special cases such as strong log-concavity, $\gamma=\mathcal{N}(0,I_d),$2 takes the simple form $\gamma=\mathcal{N}(0,I_d),$3. The paper also proves

$\gamma=\mathcal{N}(0,I_d),$4

under additional assumptions (Chewi et al., 3 Nov 2025).

At finite horizon, the core estimate takes the form

$\gamma=\mathcal{N}(0,I_d),$5

with

$\gamma=\mathcal{N}(0,I_d),$6

The proof couples the two reverse ODEs, differentiates $\gamma=\mathcal{N}(0,I_d),$7, controls the drift mismatch by Cauchy–Schwarz, and propagates the score difference through the OU semigroup via a Fisher-information decay estimate due to Wibisono. In the notation of the paper,

$\gamma=\mathcal{N}(0,I_d),$8

This is combined with the Hessian bound $\gamma=\mathcal{N}(0,I_d),$9 to close the Grönwall argument (Chewi et al., 3 Nov 2025).

The resulting viewpoint is structurally adapted to the construction itself. The reverse-time drift is literally a score field, and the right-hand side of the stability estimate measures the $\mu$0-size of score differences. This is why the theory is formulated in terms of relative Fisher information rather than directly in terms of transport cost.

5. Entropy stability, near-Lipschitz $\mu$1 stability, and comparison with optimal transport

A second line of recent results proves stability in relative entropy. If $\mu$2 satisfies $\mu$3, then for any $\mu$4,

$\mu$5

This yields the sharpened Talagrand-type inequality

$\mu$6

(Chewi et al., 22 Jun 2026).

The same paper proves near-Lipschitz stability in $\mu$7, up to a logarithmic factor. Assume $\mu$8 satisfies $\mu$9, both $\gamma=\mathcal{N}(0,I_d)$00 have second moments bounded by $\gamma=\mathcal{N}(0,I_d)$01, and there exists $\gamma=\mathcal{N}(0,I_d)$02 such that

$\gamma=\mathcal{N}(0,I_d)$03

Then

$\gamma=\mathcal{N}(0,I_d)$04

For small $\gamma=\mathcal{N}(0,I_d)$05, the dominant behavior is essentially linear in $\gamma=\mathcal{N}(0,I_d)$06, multiplied by a slowly varying logarithmic factor. The proof combines the differential inequality

$\gamma=\mathcal{N}(0,I_d)$07

the de Bruijn identity

$\gamma=\mathcal{N}(0,I_d)$08

and the log-Harnack estimate

$\gamma=\mathcal{N}(0,I_d)$09

Because the resulting Fisher-information control behaves like $\gamma=\mathcal{N}(0,I_d)$10 near $\gamma=\mathcal{N}(0,I_d)$11, the argument uses an “early stopping” parameter $\gamma=\mathcal{N}(0,I_d)$12 and then optimizes over $\gamma=\mathcal{N}(0,I_d)$13, which produces the logarithmic correction (Chewi et al., 22 Jun 2026).

These results place the construction in a distinctive position relative to classical transport maps. Brenier maps are gradients of convex potentials and, in general, are only Hölder stable in $\gamma=\mathcal{N}(0,I_d)$14, with exponent up to $\gamma=\mathcal{N}(0,I_d)$15, under variations of the target. Entropic transport maps enjoy $\gamma=\mathcal{N}(0,I_d)$16-Lipschitz stability with constants depending on the regularization parameter. The Kim–Milman map is not defined by a variational problem; its stability theory is instead organized around score propagation under the OU semigroup and the functional inequalities generated by the semiconcavity condition (Chewi et al., 22 Jun 2026, Chewi et al., 3 Nov 2025).

6. Examples, corollaries, and manifold extensions

For strongly log-concave targets, the Fisher-information estimate becomes especially simple: $\gamma=\mathcal{N}(0,I_d)$17 Combined with the HWI-type bound

$\gamma=\mathcal{N}(0,I_d)$18

this gives

$\gamma=\mathcal{N}(0,I_d)$19

For log-Lipschitz perturbations of strongly log-concave measures,

$\gamma=\mathcal{N}(0,I_d)$20

with $\gamma=\mathcal{N}(0,I_d)$21 $\gamma=\mathcal{N}(0,I_d)$22-strongly convex and $\gamma=\mathcal{N}(0,I_d)$23 smooth and $\gamma=\mathcal{N}(0,I_d)$24-Lipschitz, the paper gives the explicit bound

$\gamma=\mathcal{N}(0,I_d)$25

An $\gamma=\mathcal{N}(0,I_d)$26-analogue leads to

$\gamma=\mathcal{N}(0,I_d)$27

with explicit $\gamma=\mathcal{N}(0,I_d)$28 in the same regimes (Chewi et al., 3 Nov 2025).

The construction also admits a manifold version coupled to geometric evolution. In the setting of a smooth flow of Riemannian metrics $\gamma=\mathcal{N}(0,I_d)$29 on a manifold $\gamma=\mathcal{N}(0,I_d)$30, preserving volume and satisfying

$\gamma=\mathcal{N}(0,I_d)$31

if there exist $\gamma=\mathcal{N}(0,I_d)$32 such that

$\gamma=\mathcal{N}(0,I_d)$33

then the Kim–Milman transport map

$\gamma=\mathcal{N}(0,I_d)$34

defined as the limit of the inverse characteristic flow is $\gamma=\mathcal{N}(0,I_d)$35-Lipschitz. Specializing to rescaled Ricci flow on a surface gives the criterion

$\gamma=\mathcal{N}(0,I_d)$36

where $\gamma=\mathcal{N}(0,I_d)$37 is the curvature potential defined by $\gamma=\mathcal{N}(0,I_d)$38. Under a $\gamma=\mathcal{N}(0,I_d)$39 near-roundness hypothesis, this yields a contraction from the unit round sphere to a nearly spherical positively curved surface, pushing forward the rescaled round volume onto the target volume (Serres, 19 Aug 2025).

This manifold extension shows that the construction is not confined to the Gaussian–OU setting on Euclidean space. A plausible implication is that the defining mechanism—transport generated by a continuity equation with controlled Hessian—can be coupled to geometric flows whenever one has a suitable Bakry–Émery-type differential inequality.