Let $\Omega$ be a compact domain of $\mathbb R^d$. Let $\mathcal P (\Omega)$ be the space of probability measures on $\Omega$. For each $\tau >0$, let $(\varrho^\tau_{(k)})_{k \in \mathbb N} \subset \mathcal P (\Omega)$ be the sequence of minimizers in the minimizing movement scheme of the time step $\tau$.

The following is taken from section 8.3 in Santambrogio's Optimal transport for applied mathematicians:

Let us define two interpolations between the measures $\varrho_{(k)}^\tau$. With this time-discretized method, we have obtained, for each $\tau>0$, a sequence $(\varrho_{(k)}^\tau)_k$. We can use it to build at least two interesting curves in the space of measures:

  • first we can define some piecewise constant curves, i.e., $\varrho_t^\tau:=\varrho_{(k+1)}^\tau$ for $t \in ] k \tau,(k+1) \tau]$; associated with this curve, we also define the velocities $\mathbf{v}_t^\tau=\mathbf{v}_{(k+1)}^\tau$ for $t \in] k \tau,(k+1) \tau]$, where $\mathbf{v}_{(k+1)}^\tau$ is defined as $\mathbf{v}_{(k+1)}^\tau=(\mathrm{id}-$ $\left.\mathrm{T}_{k+1}^\tau\right) / \tau$ where $\mathrm{T}_{k+1}^\tau$ is the optimal transport map from $\varrho_{(k+1)}^\tau$ to $\varrho_{(k)}^\tau$; we also define the momentum variable $E^\tau=\varrho^\tau \mathbf{v}^\tau$;
  • then, we can also consider the densities $\tilde{\varrho}_t^\tau$ that interpolate the discrete values $(\varrho_{(k)}^\tau)_k$ along geodesics: $$ \tilde{\varrho}_t^\tau=\left(\frac{k \tau-t}{\tau} \mathbf{v}_{(k)}^\tau+\mathrm{id}\right)_{\sharp} \varrho_{(k)}^\tau, \quad \text {for} \quad t \in](k-1) \tau, k \tau[; \quad (8.11) $$ the velocities $\tilde{\mathbf{v}}_t^\tau$ are defined so that $\left(\tilde{\varrho}^\tau, \tilde{\mathbf{v}}^\tau\right)$ satisfy the continuity equation and $\left\|\tilde{\mathbf{v}}_t^\tau\right\|_{L^2\left(\tilde{\varrho}_t^\tau\right)}=\left|\left(\tilde{\varrho}^\tau\right)^{\prime}\right|(t)$. To do so, we take $$ \tilde{\mathbf{v}}_t^\tau=\mathbf{v}_t^\tau \circ\left((k \tau-t) \mathbf{v}_{(k)}^\tau+\mathrm{id}\right)^{-1}; $$ as before, we define a momentum variable: $\tilde{E}_\tau=\tilde{\varrho}^\tau \tilde{\mathbf{v}}^\tau$.

When $t^* := (k-1) \tau$ we have by (8.11) that $$ \tilde{\varrho}_{t^*} = \left(\mathbf{v}_{(k)}^\tau+\mathrm{id}\right)_{\sharp} \varrho_{(k)}^\tau. $$

I could not see how $\tilde{\varrho}_{t^*}^\tau = \varrho_{(k-1)}^\tau$. As such, I could not see how $(\tilde{\varrho}_t^\tau)_t$ ''interpolates'' $(\varrho^\tau_{(k)})_n$.

Could you elaborate on my confusion?

Thank you so much for your help!


1 Answer 1


$\newcommand{\id}{\operatorname{id}}$This is because there is a typo in (8.11). One could have guessed that (8.11) is wrong for homogeneity reasons: if $x\in \Omega$ has units [m], then $\id(x)=x$ has units [m]. But the veolcity $v_k^\tau$ has units [m.s-1] so $\frac{k\tau-t}{\tau}\mathbf{v}_{(k)}^\tau$ also has unit [m.s-1] and therefore (8.11) adds up [m] with [m.s-1].

The right formula should read $$ \tilde{\varrho}_t^\tau=\left(\id-(k\tau-t)\mathbf{v}_{(k)}^\tau\right)_{\sharp} \varrho_{(k)}^\tau, \quad \text {for} \quad t \in](k-1) \tau, k \tau[; \hspace{3cm} (8.11'). $$ (Note that this has now the correct physical units, time [s] $\times$ velocity [m.s-1] = displacement [m] for the $(k\tau-t)\mathbf{v}_{(k)}^\tau$ term).

To recover (8.11'), think of it this way: by definition $$ \mathbf{v}_{(k)}^\tau(x)=\frac{x-T_k(x)}{\tau} $$ is the forward-in-time, constant velocity (effective terminal minus initial displacement $x-T_k(x)$ divided by travel time $\tau$) of a particle following a geodesic starting at $T_k(x)$ in the pointcloud $\varrho^\tau_{k-1}$ at time $t_{k-1}=(k-1)\tau$ and ending up at $x$ at time $t_k=k\tau$ in the pointcloud $\varrho^\tau_k$. (Here it is crucial that the trajectory is a geodesic, so that the velocity is constant.)

Now, we want $\tilde{\varrho}_t^\tau$ to be the time-$\tau$ geodesic interpolation between $\varrho_{k-1}$ at time $t_{k-1}$ and $\varrho_k$ at time $t_k$. From standard optimal transport theory it suffices to perform the linear interpolation between transport maps. By definition the tansport map from $\varrho_k$ to $\varrho_{k-1}$ is $T_k$ (beware, $T_k$ is defined "backward in time" from $\varrho_k$ to $\varrho_{k-1}$, which is sometimes confusing) so one should simply take $$ T^t_k=\frac{t-t_{k-1}}{\tau}\id +\frac{t_k-t}{\tau}T_k \quad\text{and}\quad \tilde{\varrho}_t^\tau=\left(T^t_k\right)_\sharp \varrho^\tau_{(k)} \qquad t\in(t_{k-1},t_k). $$ This indeed interpolates linearly between $T^{t_{k-1}}_k(x)=T_k(x)$ at time $t_{k-1}$ and $T^{t_k}(x)=\id(x)=x$ at time $t_k$. Substituting for $\mathbf{v}_{(k)}^\tau=\frac{\id-T_k}{\tau}\Leftrightarrow T_k=\id-\tau \mathbf{v}_{(k)}^\tau$ readily gives $$ T^t_k=\id -(t_k-t)\mathbf{v}_{(k)}^\tau, $$ which is exactly (8.11').

One could also have guessed it more directly: by construction $\mathbf{v}_{(k)}^\tau(x)$ is the forward-in-time velocity of the constant-speed geodesic starting from $T_{k-1}(x)$ at time $t_{k-1}$ and ending up at $x$ at time $t_k>t_{t_1}$. So, the intermediate position at time $t\in(t_{k-1},t_k)$ of this particle -- which by definition is the interpolating map $T^t_k(x)$ that is needed to define the geodesic interpolation -- is given by the rule-of-thumb "terminal position minus running time times forward velocity", i-e $T_k^t(x)=x-(t_k-t)\mathbf{v}_{(k)}^\tau(x)$. Here the minus sign is because we are running backward in time for a duration $t_k-t>0$ back from the terminal position $x$. It's a little mind game that one needs to play to figure out which objects are forward-in-time and which are backward-in-time, but once one gets used to this simply gymnastic things become much more clear (and easier!) so I'd strongly recommend thinking of it like that.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.