---
title: Kim–Milman Construction
url: https://www.emergentmind.com/topics/kim-milman-construction
type: topic
---

# Kim–Milman Construction

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 \((\mathsf T^\mu)_\sharp\gamma=\mu\) [2511.01154] [2606.23383].

## 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:
\[
\mathrm{d}X_t^\rightarrow=-X_t^\rightarrow\,\mathrm{d}t+\sqrt{2}\,\mathrm{d}B_t,
\qquad X_0^\rightarrow\sim\mu.
\]
If \(q_t^\mu:=\mathrm{Law}(X_t^\rightarrow)\), then \(q_t^\mu=Q_t\mu\), where \(Q_t\) is the OU semigroup, and \(q_t^\mu\to\gamma\) as \(t\to\infty\) [2511.01154].

On a finite horizon \([0,T]\), reversing time in the Fokker–Planck equation yields a continuity equation for the backward-time density \(q^\mu_{T-t}\):
\[
\partial_t q^\mu_{T-t}+\nabla\cdot\bigl(q^\mu_{T-t}(\mathrm{id}+\nabla\log q^\mu_{T-t})\bigr)=0.
\]
This is the continuity equation for the deterministic ODE
\[
\dot X_t^\leftarrow
= X_t^\leftarrow+\nabla\log q^\mu_{T-t}(X_t^\leftarrow)
= \nabla\log\Bigl(\frac{\mathrm{d}q^\mu_{T-t}}{\mathrm{d}\gamma}\Bigr)(X_t^\leftarrow),
\qquad X_0^\leftarrow\sim q_T^\mu.
\]
Assuming \(\mu\ll\gamma\), the drift can also be written in semigroup form:
\[
\dot X_t^\leftarrow
=\nabla\log Q_{T-t}\Bigl[\frac{\mathrm{d}\mu}{\mathrm{d}\gamma}\Bigr](X_t^\leftarrow).
\]
This ODE is the Kim–Milman reverse heat flow ODE, and its flow map \(\mathsf T^{\mu,T}=(\mathsf T_t^{\mu,T})_{t\in[0,T]}\) satisfies
\[
(\mathsf T^{\mu,T}_{T-t})_\sharp q_T^\mu=q_t^\mu.
\]
In particular, \((\mathsf T_T^{\mu,T})_\sharp q_T^\mu=\mu\) [2606.23383].

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 \(\nabla\log q_t^\mu\) is central to both its regularity theory and its stability theory.

## 2. Infinite-horizon limit and existence of the map

Letting \(T\to\infty\) sends the initial law \(q_T^\mu\) to \(\gamma\). In the regular regime, the Kim–Milman map is defined as the infinite-horizon limit
\[
\mathsf T^\mu=\mathsf T^{\mu,\infty}=\lim_{T\to\infty}\mathsf T^{\mu,T},
\]
and for each \(t>0\) one also obtains a stopped map
\[
\mathsf T^\mu_{t\gets}:=\lim_{T\to\infty}\mathsf T^{\mu,T}_{T-t},
\qquad
(\mathsf T^\mu_{t\gets})_\sharp\gamma=q_t^\mu.
\]
For regular targets, a key representation is the integral equation
\[
\mathsf T^\mu_{\delta\gets}
=\mathrm{id}+\int_\delta^\infty
\nabla\log\frac{q_t^\mu}{\gamma}\circ\mathsf T^\mu_{t\gets}\,\mathrm{d}t,
\qquad \delta>0,
\]
and differentiation in \(t\) gives
\[
\partial_t\mathsf T^\mu_{t\gets}
=-\nabla\log\frac{q_t^\mu}{\gamma}\circ\mathsf T^\mu_{t\gets},
\qquad
(\mathsf T^\mu_{t\gets})_\sharp\gamma=q_t^\mu.
\]
The Kim–Milman map is then \(\mathsf T^\mu=\lim_{t\searrow0}\mathsf T^\mu_{t\gets}\) [2606.23383].

A major recent development is a general existence theorem for arbitrary \(\nu\in\mathcal{P}_2(\mathbb{R}^d)\). A regular approximating sequence \(\{\nu_n\}\) is defined by requiring that every \(\nu_n\) satisfies the regularity condition \((\Theta)\) and that \(W_2(\nu_n,\nu)\to0\). The paper gives the example: one can truncate \(\nu\) to a large ball and convolve with a small Gaussian. For such an approximation, the limits
\[
\mathsf T^\nu_{t\gets}:=\lim_{n\to\infty}\mathsf T^{\nu_n}_{t\gets},
\qquad
\mathsf T^\nu:=\lim_{t\searrow0}\mathsf T^\nu_{t\gets}
\]
exist in \(L^2(\gamma)\), are independent of the approximating sequence, and satisfy
\[
(\mathsf T^\nu_{t\gets})_\sharp\gamma=q_t^\nu,
\qquad
(\mathsf T^\nu)_\sharp\gamma=\nu.
\]
Thus, for every \(\nu\in\mathcal{P}_2(\mathbb{R}^d)\), the Kim–Milman map exists as an \(L^2(\gamma)\) limit, without requiring \((\Theta)\) on \(\nu\) [2606.23383].

This general existence theorem separates the questions of definition and quantitative control. Existence holds on all of \(\mathcal{P}_2(\mathbb{R}^d)\), 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
\[
\tag{\(\Theta\)}
\nabla^2\log Q_s\Bigl[\frac{\mathrm{d}\mu}{\mathrm{d}\gamma}\Bigr](x)
=
I+\nabla^2\log q_s^\mu(x)
\preceq \theta_s I,
\qquad
\sup_{t\ge0}\int_t^\infty \theta_s\,\mathrm{d}s<+\infty.
\]
Here \(\theta_s\) is a scalar control on the Hessian of the OU-smoothed log-density. Under \((\Theta)\), for all \(t>0\), the stopped flow map \(\mathsf T^\mu_{t\gets}\) exists and is Lipschitz with constant
\[
L_t(\mu):=\exp\Bigl(\int_t^\infty\theta_s\,\mathrm{d}s\Bigr).
\]
In particular, the limiting map \(\mathsf T^\mu\) is Lipschitz with constant \(L_0(\mu)\). Moreover, the OU marginals satisfy the log-Sobolev inequality
\[
\mathsf H(\rho\mid q_t^\mu)\le \frac{L_t(\mu)^2}{2}\,\mathsf{FI}(\rho\mid q_t^\mu)
\qquad\text{for all }\rho\ll q_t^\mu
\]
[2606.23383].

Several classes of targets are known to satisfy \((\Theta)\). If \(\mu\propto e^{-V}\) with \(V\) \(\alpha\)-strongly convex, then
\[
I+\nabla^2\log q_s^\mu(\cdot)
\preceq
\frac{1-\alpha}{\alpha(e^{2s}-1)+1}\,I,
\]
and \(\int_0^\infty\theta_s\,\mathrm{d}s<\infty\). This is the original Kim–Milman setting and yields \(L_0(\mu)\sim \alpha^{-1/2}\). The theory also covers strongly convex plus Lipschitz perturbation and more general asymptotically convex potentials characterized by a convexity profile \(\kappa_V(r)\); in these cases, precise formulas for \(\theta_t\) are available and satisfy the required integrability and boundedness properties [2606.23383].

In the stability paper, the same structural condition is written as
\[
\nabla^2 \log Q_{s}\Bigl[\frac{\mu}{\gamma}\Bigr](x)
= I+\nabla^2\log q_s^\mu(x)\preceq \theta_s I
\quad\text{for all }s\ge0,
\]
and explicit \(\theta_s\) are developed for three regimes: strong log-concavity, strong log-concavity plus log-Lipschitz perturbations, and asymptotically positive convex profiles [2511.01154].

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 \(\nu\ll\mu\), the relative Fisher information is
\[
\mathrm{FI}(\nu\mid\mu)
:=\int \bigl\|\nabla\log(\nu/\mu)(x)\bigr\|^2\,\nu(\mathrm{d}x),
\]
and the \(L^\infty\)-version is
\[
\mathrm{FI}_\infty(\nu\mid\mu)
:=\|\nabla\log(\nu/\mu)\|_{L^\infty(\nu)}^2.
\]
If \(T_{\mathrm{KM}^\mu}\) and \(T_{\mathrm{KM}^\nu}\) denote the Kim–Milman maps associated with two targets \(\mu\) and \(\nu\), then under suitable regularity assumptions on \(\mu\),
\[
\|T_{\mathrm{KM}^\mu}-T_{\mathrm{KM}^\nu}\|_{L^2(\gamma)}
\le \Lambda_\infty \sqrt{\mathrm{FI}(\nu\mid\mu)},
\]
where \(\Lambda_\infty\) is an explicit constant depending on curvature-type bounds associated with \(\mu\). In important special cases such as strong log-concavity, \(\Lambda_\infty\) takes the simple form \(\alpha^{-1}\). The paper also proves
\[
\|T_{\mathrm{KM}^\mu}-T_{\mathrm{KM}^\nu}\|_{L^\infty(\gamma)}
\lesssim \sqrt{\mathrm{FI}_\infty(\nu\mid\mu)}
\]
under additional assumptions [2511.01154].

At finite horizon, the core estimate takes the form
\[
\bigl(\mathbb{E}\|X_T-Y_T\|^2\bigr)^{1/2}
\le
\sqrt{\mathrm{FI}(\nu\mid\mu)}\,\Lambda_T,
\]
with
\[
\Lambda_T
:=
\int_0^T
\exp\Bigl(\int_0^{T-s}(3\theta_u-1)\,\mathrm{d}u\Bigr)\,\mathrm{d}s.
\]
The proof couples the two reverse ODEs, differentiates \(\mathbb{E}\|X_t-Y_t\|^2\), 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,
\[
\frac{\mathrm{d}}{\mathrm{d}t}\mathrm{FI}(q_t^\nu\mid q_t^\mu)
\le -2(1-2\theta_t)\,\mathrm{FI}(q_t^\nu\mid q_t^\mu).
\]
This is combined with the Hessian bound \((\Theta)\) to close the Grönwall argument [2511.01154].

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 \(L^2\)-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 \(W_2\) stability, and comparison with optimal transport

A second line of recent results proves stability in relative entropy. If \(\mu\in\mathcal{P}_2(\mathbb{R}^d)\) satisfies \((\Theta)\), then for any \(\nu\in\mathcal{P}_2(\mathbb{R}^d)\),
\[
\|\mathsf T^\mu-\mathsf T^\nu\|_{L^2(\gamma)}^2
\le
2\,L_0(\mu)^2\,\mathsf H(\nu\mid\mu),
\qquad
L_0(\mu)=\exp\Bigl(\int_0^\infty\theta_t\,\mathrm{d}t\Bigr).
\]
This yields the sharpened Talagrand-type inequality
\[
\frac12 W_2^2(\mu,\nu)
\le
\frac12\|\mathsf T^\mu-\mathsf T^\nu\|_{L^2(\gamma)}^2
\le
L_0(\mu)^2\,\mathsf H(\nu\mid\mu)
\]
[2606.23383].

The same paper proves near-Lipschitz stability in \(W_2\), up to a logarithmic factor. Assume \(\mu\) satisfies \((\Theta)\), both \(\mu,\nu\) have second moments bounded by \(M\), and there exists \(\mathsf C_\mu>0\) such that
\[
\int_s^t\theta_u\,\mathrm{d}u\le \mathsf C_\mu
\qquad\text{for all }0\le s\le t.
\]
Then
\[
\|\mathsf T^\mu-\mathsf T^\nu\|_{L^2(\gamma)}
\lesssim
W_2(\mu,\nu)\,\exp(3\mathsf C_\mu)
\log\Bigl[
e^2\max\Bigl\{
1,\,
\frac{4M^2+8d}{W_2^2(\mu,\nu)\,\exp(6\mathsf C_\mu)}
\Bigr\}
\Bigr].
\]
For small \(W_2(\mu,\nu)\), the dominant behavior is essentially linear in \(W_2\), multiplied by a slowly varying logarithmic factor. The proof combines the differential inequality
\[
\frac{\mathrm{d}}{\mathrm{d}t}
\|\mathsf T^\mu_{t\gets}-\mathsf T^\nu_{t\gets}\|_{L^2(\gamma)}
\le
-\theta_t\,
\|\mathsf T^\mu_{t\gets}-\mathsf T^\nu_{t\gets}\|_{L^2(\gamma)}
+
\sqrt{\mathsf{FI}(q_t^\nu\mid q_t^\mu)},
\]
the de Bruijn identity
\[
\frac{\mathrm{d}}{\mathrm{d}t}\mathsf H(q_t^\nu\mid q_t^\mu)
=
-\mathsf{FI}(q_t^\nu\mid q_t^\mu),
\]
and the log-Harnack estimate
\[
\mathsf H(q_t^\nu\mid q_t^\mu)
\le
\frac{1}{2(e^{2t}-1)}\,W_2^2(\nu,\mu).
\]
Because the resulting Fisher-information control behaves like \(t^{-1}\) near \(t=0\), the argument uses an “early stopping” parameter \(\delta>0\) and then optimizes over \(\delta\), which produces the logarithmic correction [2606.23383].

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 \(W_2\), with exponent up to \(1/3\), under variations of the target. Entropic transport maps enjoy \(W_2\)-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 [2606.23383] [2511.01154].

## 6. Examples, corollaries, and manifold extensions

For strongly log-concave targets, the Fisher-information estimate becomes especially simple:
\[
\|T_{\mathrm{KM}^\mu}-T_{\mathrm{KM}^\nu}\|_{L^2(\gamma)}
\le
\alpha^{-1}\sqrt{\mathrm{FI}(\nu\mid\mu)}.
\]
Combined with the HWI-type bound
\[
W_2^2(\mu,\nu)\le \alpha^{-2}\,\mathrm{FI}(\nu\mid\mu),
\]
this gives
\[
W_2^2(\mu,\nu)
\le
\|T_{\mathrm{KM}^\mu}-T_{\mathrm{KM}^\nu}\|_{L^2(\gamma)}^2
\le
\alpha^{-2}\,\mathrm{FI}(\nu\mid\mu).
\]
For log-Lipschitz perturbations of strongly log-concave measures,
\[
\mu\propto \exp(-V+H),
\]
with \(V\) \(\alpha\)-strongly convex and \(H\) smooth and \(L\)-Lipschitz, the paper gives the explicit bound
\[
\|T_{\mathrm{KM}^\mu}-T_{\mathrm{KM}^\nu}\|_{L^2(\gamma)}
\le
\frac{1}{\alpha}\,
\exp\Bigl(\frac{3L^2}{2\alpha}+\frac{6L}{\sqrt{\alpha}}\Bigr)
\sqrt{\mathrm{FI}(\nu\mid\mu)}.
\]
An \(L^\infty\)-analogue leads to
\[
W_\infty(\mu,\nu)
\le
\|T_{\mathrm{KM}^\mu}-T_{\mathrm{KM}^\nu}\|_{L^\infty(\gamma)}
\le
\eta_\infty \sqrt{\mathrm{FI}_\infty(\nu\mid\mu)}
\]
with explicit \(\eta_\infty\) in the same regimes [2511.01154].

The construction also admits a manifold version coupled to geometric evolution. In the setting of a smooth flow of Riemannian metrics \(g_t\) on a manifold \(M\), preserving volume and satisfying
\[
\partial_t\nu_t=-\nabla\cdot(\nabla\xi_t\,\nu_t),
\]
if there exist \(\dot\lambda_t\in\mathbb{R}\) such that
\[
2\nabla^2\xi_t-\partial_t g_t\ge -\dot\lambda_t g_t
\qquad\text{for all }t\ge0,
\]
then the Kim–Milman transport map
\[
T:(M,g_\infty)\to(M,g_0)
\]
defined as the limit of the inverse characteristic flow is \(e^{\int_0^\infty\dot\lambda_s\,\mathrm{d}s}\)-Lipschitz. Specializing to rescaled Ricci flow on a surface gives the criterion
\[
\nabla^2\xi_t+\mathrm{Ric}_{g_t}
\ge
\frac12(r-\dot\lambda_t)g_t,
\]
where \(\xi_t\) is the curvature potential defined by \(\Delta_{g_t}\xi_t=R_t-r\). Under a \(C^{5,\alpha}\) 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 [2508.13688].

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.

Source: https://www.emergentmind.com/topics/kim-milman-construction