MathOverflow is a question and answer site for professional mathematicians. Join them; it only takes a minute:

Sign up
Here's how it works:
  1. Anybody can ask a question
  2. Anybody can answer
  3. The best answers are voted up and rise to the top

Villani gives the following formula to find the gradient of a function $F$ of a probability density function $\rho$ in the Wasserstein space : $$\nabla_W F(\rho) = -\nabla.(\rho \nabla \frac{\delta F}{\delta \rho})$$ where, if $F$ is given as $F(\rho)=\int U(\rho) dx$, then $\frac{\delta F}{\delta \rho} = U'(\rho)$.

I am trying to find how much the value at a given (fixed) point $x$ of a continuous pdf varies along a displacement interpolation. I am thus trying to apply the formula above in the distributional sense where $U = \delta_x$ (here, with $\delta$ the Dirac distribution), to get $F(\rho) = \rho(x) = <\delta_x, \rho>$.

By using $\nabla.(f G) = (\nabla f) G + (\nabla.G) f$, I thus get : $$\nabla_W F(\rho) = - (\nabla\rho.\nabla\delta_x + \rho\triangle\delta_x)$$ which equals to $0$ after using the definition of the derivatives in the distributional sense since both terms cancel each other.

So, what I get from it, is that the value at a fixed point $x$ (ie., not a point which gets advected with the function!) does not change when a function moves in the Wasserstein space... which I really don't believe to be true ! What went wrong ?

Thanks !

share|cite|improve this question
ouch.. it seems that many things went wrong :s – WhitAngl Mar 31 '12 at 7:35
up vote 1 down vote accepted

The dirac distribution $\delta_x$ is not smooth enough to carry out this computation, which only works for functionals $F(\rho)=\int U(\rho)$...

One thing though: a displacement interpolation is nothing but a geodesic in the metric space of probabilities (endowed with the Wasserstein distance). If you look for example at Villani's book (chapter 8), we know that such a geodesic $(\rho_t)_{t\in[-\varepsilon,\varepsilon]}$ satisfies a continuity equation $$ \partial_t\rho_t+\nabla\cdot(\rho_tv_t)=0 $$ for some particular vector-field $v_t$ determining the tangent vector $s=-\nabla\cdot(\rho_t v_t)\in T_{\rho_t}\mathcal{P}$. This PDE is what you are looking for, since it allows to compute the variation $\partial_t\rho_t(x)$ for a given point $x$ (forgetting about smoothness issues, of course).

share|cite|improve this answer
Great, thanks! :) – WhitAngl May 31 '13 at 0:00

Your Answer


By posting your answer, you agree to the privacy policy and terms of service.

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