3

Time
1

0
Pos
itio 0.0
0.5
n -0.5

Gradient Flow in the Wasserstein Metric

Katy Craig
University of California, Santa Barbara

NIPS, Optimal Transport & Machine Learning

December 9th, 2017
gradient flow in finite dimensions
A curve x(t): [0,T ] → ℝd is the gradient flow of an energy E: ℝd → ℝ if
d
x(t) = rE(x(t))
dt
• “x(t) evolves in the direction of steepest descent of E”
• initial value problem: given x(0), find the gradient flow x(t)

Example:
metric energy functional gradient flow

d 1 2 d
(R , | · |) E(x) = x x(t) = x(t)
2 dt
Given x(0) ∈ ℝd, x(t) = x(0)e-t is unique solution of the gradient flow.

2
gradient flow in finite dimensions
Gradient flows often arise when solving optimization problems:
min E(x)
x2Rd

Convexity of the energy determines stability and long time behavior.

Def: An energy E is λ-convex if D 2

E Id⇥d or, equivalently, if
E((1 t)x + ty)  (1 t)E(x) + tE(y) t(1 t) |x y|2
2
for all x,y ∈ ℝ, t ∈ [0,1].
x2
f (x) = , =1 f (x) = sin(x), = 1
2

3
gradient flow in finite dimensions
If E(x) is λ-convex, then…

1) Stability: for any gradient flows x(t) and y(t),

t
|x(t) y(t)|  e |x(0) y(0)|

2) long time behavior: if λ>0, there is a unique solution x̅ of min E(x)

x2Rd
and any gradient flow x(t) converges to x̅ as t → +∞:
t
|x(t) x̄|  e |x(0) x̄|

4
gradient flow

5
gradient flow

5
gradient flow with different metrics
In general, given a complete metric space (X,d), a curve x(t): ℝ → X is the
gradient flow of an energy E: X → ℝ if
“ d
x(t) = rX E(x(t)) ’’
dt
Examples: Euclidean L2
metric (X,d) (Rd , | · |) (L2 (Rd ), k · kL2 )

E(x + hv) E(x) E(f + hg) E(f )

def of ∇X hrE(x), vi = lim
h!0 h
hrE(f ), gi = lim
h!0 h
@E
formula for r
∇XX rRd E(x) = rE(x) rL2 (Rd ) E(f ) =
@f
Z
1 2 1
energy E(x) = x E(f ) = |f |2
2 2
d d
gradient flow x(t) = x(t) f (x, t) = f (x, t)
dt dt
6
gradient flow with different metrics
In general, given a complete metric space (X,d), a curve x(t): ℝ → X is the
gradient flow of an energy E: X → ℝ if
“ d
x(t) = rX E(x(t)) ’’
dt
Examples: Euclidean L2
metric (X,d) (Rd , | · |) (L2 (Rd ), k · kL2 )

E(x + hv) E(x) E(f + hg) E(f )

def of ∇X hrE(x), vi = lim
h!0 h
hrE(f ), gi = lim
h!0 h
@E
formula for r
∇XX rRd E(x) = rE(x) rL2 (Rd ) E(f ) =
@f
Z
1 2 1
energy E(x) = x E(f ) = |rf
|f |2 |2
2 2
d d
gradient flow x(t) = x(t) f (x, t) = f (x, t)
dt dt
6
gradient flow with different metrics
In general, given a complete metric space (X,d), a curve x(t): ℝ → X is the
gradient flow of an energy E: X → ℝ if
“ d
x(t) = rX E(x(t)) ’’
dt
Examples: Euclidean L2
metric (X,d) (Rd , | · |) (L2 (Rd ), k · kL2 )

E(x + hv) E(x) E(f + hg) E(f )

def of ∇X hrE(x), vi = lim
h!0 h
hrE(f ), gi = lim
h!0 h
@E
formula for r
∇XX rRd E(x) = rE(x) rL2 (Rd ) E(f ) =
@f
Z
1 2 1
energy E(x) = x E(f ) = |rf
|f |2 |2
2 2
d d
gradient flow x(t) = x(t) f (x, t) = f (x, t)
dt dt
6
 gradient flow with different metrics
finite difference approximation
f : Rd ! R approximated by {fi }i2hZd
approximate values of function
Examples: Euclidean L2
metric (X,d) (Rd , | · |) (L2 (Rd ), k · kL2 )

E(x + hv) E(x) E(f + hg) E(f )

def of ∇X hrE(x), vi = lim
h!0 h
hrE(f ), gi = lim
h!0 h
@E
formula for r
∇XX rRd E(x) = rE(x) rL2 (Rd ) E(f ) =
@f
Z
1 2 1
energy E(x) = x E(f ) = |rf
|f |2 |2
2 2
d d
gradient flow x(t) = x(t) f (x, t) = f (x, t)
dt dt
7
gradient flow with different metrics

Examples: Euclidean W2
metric (X,d) (Rd , | · |) (P2 (Rd ), W2 )

E(x + hv) E(x)

def of ∇X hrE(x), vi = lim
h!0 h
✓ ◆
@E
formula for r
∇XX rRd E(x) = rE(x) rW2 E(⇢) = r · ⇢r
@⇢
Z
1 2 1
energy E(x) = x E(⇢) = x2 ⇢(x)dx
2 2
d d
gradient flow x(t) = x(t) ⇢(x, t) = r · (x⇢(x, t))
dt dt
8
gradient flow with different metrics

E((id + h⇠)#µ) E(µ)

hrE(µ), r · (⇠µ)iTanµ P2 (Rd ) = lim
h!0 h

Examples: Euclidean W2
metric (X,d) (Rd , | · |) (P2 (Rd ), W2 )

E(x + hv) E(x)

def of ∇X hrE(x), vi = lim
h!0 h
✓ ◆
@E
formula for r
∇XX rRd E(x) = rE(x) rW2 E(⇢) = r · ⇢r
@⇢
Z
1 2 1
energy E(x) = x E(⇢) = x2 ⇢(x)dx
2 2
d d
gradient flow x(t) = x(t) ⇢(x, t) = r · (x⇢(x, t))
dt dt
8
gradient flow with different metrics

Examples: Euclidean W2
metric (X,d) (Rd , | · |) (P2 (Rd ), W2 )

E(x + hv) E(x)

def of ∇X hrE(x), vi = lim
h!0 h
✓ ◆
@E
formula for r
∇XX rRd E(x) = rE(x) rW2 E(⇢) = r · ⇢r
@⇢
Z
1 2 1
energy E(x) = x E(⇢) = x2 ⇢(x)dx
2 2
d d
gradient flow x(t) = x(t) ⇢(x, t) = r · (x⇢(x, t))
dt dt
8
gradient flow with different metrics

Examples: Euclidean W2
metric (X,d) (Rd , | · |) (P2 (Rd ), W2 )

E(x + hv) E(x)

def of ∇X hrE(x), vi = lim
h!0 h
✓ ◆
@E
formula for r
∇XX rRd E(x) = rE(x) rW2 E(⇢) = r · ⇢r
@⇢
ZZ
1 2 1
energy E(x) = x E(⇢) = ⇢(x) x2 ⇢(x)dx
log(⇢(x))dx
2 2
d d
gradient flow x(t) = x(t) ⇢(x, t) = r · (x⇢(x, t))
dt dt
8
gradient flow with different metrics

Examples: Euclidean W2
metric (X,d) (Rd , | · |) (P2 (Rd ), W2 )

E(x + hv) E(x)

def of ∇X hrE(x), vi = lim
h!0 h
✓ ◆
@E
formula for r
∇XX rRd E(x) = rE(x) rW2 E(⇢) = r · ⇢r
@⇢
ZZ
1 2 1
energy E(x) = x E(⇢) = ⇢(x) x2 ⇢(x)dx
log(⇢(x))dx
2 2
d d
gradient flow x(t) = x(t) ⇢(x, t) = r⇢(x,
· (x⇢(x,
t) t))
dt dt
8
gradient flow with different metrics
particle approximation N
X
f : Rd ! R approximated by xi mi
i=1
approximate mass of function
Examples: Euclidean W2
metric (X,d) (Rd , | · |) (P2 (Rd ), W2 )

E(x + hv) E(x)

def of ∇X hrE(x), vi = lim
h!0 h
✓ ◆
@E
formula for r
∇XX rRd E(x) = rE(x) rW2 E(⇢) = r · ⇢r
@⇢
ZZ
1 2 1
energy E(x) = x E(⇢) = ⇢(x) x2 ⇢(x)dx
log(⇢(x))dx
2 2
d d
gradient flow x(t) = x(t) ⇢(x, t) = r⇢(x,
· (x⇢(x,
t) t))
dt dt
9
interpolating with different metrics
The same dichotomy between values of a function and mass of a
function is also present in the geodesics.
Def: A constant speed geodesic between two points ρ0 and ρ1 in a
metric space (X,d) is any curve ρ:[0,1]→X s.t.
⇢(0) = ⇢0 , ⇢(1) = ⇢1 , d(⇢(t), ⇢(s)) = |t s|d(⇢0 , ⇢1 )

L2 geodesic W2 geodesic
⇢(t) = (1 t)⇢0 + t⇢1 ⇢(t) = ((1 t)id + tT⇢⇢01 )#⇢0

10
interpolating with different metrics
The same dichotomy between values of a function and mass of a
function is also present in the geodesics.
Def: A constant speed geodesic between two points ρ0 and ρ1 in a
metric space (X,d) is any curve ρ:[0,1]→X s.t.
⇢(0) = ⇢0 , ⇢(1) = ⇢1 , d(⇢(t), ⇢(s)) = |t s|d(⇢0 , ⇢1 )

L2 geodesic W2 geodesic
⇢(t) = (1 t)⇢0 + t⇢1 ⇢(t) = ((1 t)id + tT⇢⇢01 )#⇢0

10
interpolating with different metrics
The same dichotomy between values of a function and mass of a
function is also present in the geodesics.
Def: A constant speed geodesic between two points ρ0 and ρ1 in a
metric space (X,d) is any curve ρ:[0,1]→X s.t.
⇢(0) = ⇢0 , ⇢(1) = ⇢1 , d(⇢(t), ⇢(s)) = |t s|d(⇢0 , ⇢1 )

L2 geodesic W2 geodesic
⇢(t) = (1 t)⇢0 + t⇢1 ⇢(t) = ((1 t)id + tT⇢⇢01 )#⇢0

10
gradient flow in the Wasserstein metric
Examples:
energy functional gradient flow
Z
d
E(⇢) = ⇢ log ⇢ ⇢= ⇢
dt
Z
1 d
E(⇢) = ⇢m ⇢ = ⇢m
m 1 dt
Z
d
E(⇢) = V ⇢ ⇢ = r · (rV ⇢)
dt
Z
d
E(⇢) = (K ⇤ ⇢)⇢ ⇢ = r · (r(K ⇤ ⇢)⇢)
dt
All Wasserstein gradient flows are of the form
d
⇢ + r · (v⇢) = 0
dt
continuity equation 11
gradient flow in the Wasserstein metric
Examples:
energy functional gradient flow
Z r⇢
d
E(⇢) = ⇢ log ⇢ ⇢= ⇢ v=
dt ⇢
Z
1 d
E(⇢) = ⇢m ⇢ = ⇢m
m 1 dt
Z
d
E(⇢) = V ⇢ ⇢ = r · (rV ⇢)
dt
Z
d
E(⇢) = (K ⇤ ⇢)⇢ ⇢ = r · (r(K ⇤ ⇢)⇢)
dt
All Wasserstein gradient flows are of the form
d
⇢ + r · (v⇢) = 0
dt
continuity equation 11
gradient flow in the Wasserstein metric
Examples:
energy functional gradient flow
Z r⇢
d
E(⇢) = ⇢ log ⇢ ⇢= ⇢ v=
dt ⇢
Z
1 d
E(⇢) = ⇢m ⇢ = ⇢m v= m⇢m 2
r⇢
m 1 dt
Z
d
E(⇢) = V ⇢ ⇢ = r · (rV ⇢)
dt
Z
d
E(⇢) = (K ⇤ ⇢)⇢ ⇢ = r · (r(K ⇤ ⇢)⇢)
dt
All Wasserstein gradient flows are of the form
d
⇢ + r · (v⇢) = 0
dt
continuity equation 11
gradient flow in the Wasserstein metric
Examples:
energy functional gradient flow
Z r⇢
d
E(⇢) = ⇢ log ⇢ ⇢= ⇢ v=
dt ⇢
Z
1 d
E(⇢) = ⇢m ⇢ = ⇢m v= m⇢m 2
r⇢
m 1 dt
Z
d
E(⇢) = V ⇢ ⇢ = r · (rV ⇢) v= rV
dt
Z
d
E(⇢) = (K ⇤ ⇢)⇢ ⇢ = r · (r(K ⇤ ⇢)⇢)
dt
All Wasserstein gradient flows are of the form
d
⇢ + r · (v⇢) = 0
dt
continuity equation 11
gradient flow in the Wasserstein metric
Examples:
energy functional gradient flow
Z r⇢
d
E(⇢) = ⇢ log ⇢ ⇢= ⇢ v=
dt ⇢
Z
1 d
E(⇢) = ⇢m ⇢ = ⇢m v= m⇢m 2
r⇢
m 1 dt
Z
d
E(⇢) = V ⇢ ⇢ = r · (rV ⇢) v= rV
dt
Z
d
E(⇢) = (K ⇤ ⇢)⇢ ⇢ = r · (r(K ⇤ ⇢)⇢) v = r(K ⇤ ⇢)
dt
All Wasserstein gradient flows are of the form
d
⇢ + r · (v⇢) = 0
dt
continuity equation 11
gradient flow in the Wasserstein metric
Examples:
energy functional gradient flow
Z r⇢
d
E(⇢) = ⇢ log ⇢ ⇢= ⇢ v=
dt ⇢
Z
1 d
E(⇢) = ⇢m ⇢ = ⇢m v= m⇢m 2
r⇢
m 1 dt
Z
d
E(⇢) = V ⇢ ⇢ = r · (rV ⇢) v= rV
dt
Z
d
E(⇢) = (K ⇤ ⇢)⇢ ⇢ = r · (r(K ⇤ ⇢)⇢) v = r(K ⇤ ⇢)
dt
All Wasserstein gradient flows are of the form
d @E
⇢ + r · (v⇢) = 0 v= r
dt @⇢
continuity equation 11
gradient flow in the Wasserstein metric
aggregation, drift, and degenerate diffusion:
d
⇢ = r · ((rK ⇤ ⇢)⇢) + r · (rV ⇢) + ⇢m K, V : Rd ! R, and m 1
dt
| {z } |{z} |{z}
self|interaction
{z } drift |{z}
diffusion
|{z}
Z Z Z
1 1
E(⇢) = K ⇤ ⇢d⇢ + V d⇢ + ⇢m
2 m 1

interaction kernels: degenerate diffusion:

m m 1
• granular media: K(x) = |x|3 • ⇢ = r · (m⇢ r⇢)
| {z }
• swarming: K(x) = |x|ᵃ/a
( - |x|ᵇ/b, -d<b<a
D
1
log |x| if d = 2,
• chemotaxis: K(x) = 2⇡
Cd |x|2 d otherwise.

12
biological chemotaxis
a colony of slime mold [Gregor, et. al]

13
biological chemotaxis
a colony of slime mold [Gregor, et. al]

13
gradient flow in the Wasserstein metric
aggregation, drift, and degenerate diffusion:
d
⇢ = r · ((rK ⇤ ⇢)⇢) + r · (rV ⇢) + ⇢m K, V : Rd ! R, and m 1
dt

K(x) = |x| K(x) = |x|3 /3

14
gradient flow in the Wasserstein metric
aggregation, drift, and degenerate diffusion:
d
⇢ = r · ((rK ⇤ ⇢)⇢) + r · (rV ⇢) + ⇢m K, V : Rd ! R, and m 1
dt

K(x) = |x| K(x) = |x|3 /3

14
gradient flow in the Wasserstein metric
aggregation, drift, and degenerate diffusion:
d
⇢ = r · ((rK ⇤ ⇢)⇢) + r · (rV ⇢) + ⇢m K, V : Rd ! R, and m 1
dt

K(x) = |x| K(x) = |x|3 /3

14
gradient flow in the Wasserstein metric
aggregation, drift, and degenerate diffusion:
d
⇢ = r · ((rK ⇤ ⇢)⇢) + r · (rV ⇢) + ⇢m K, V : Rd ! R, and m 1
dt

K(x) = |x| K(x) = |x|3 /3

14
gradient flow in the Wasserstein metric
aggregation, drift, and degenerate diffusion:
d
⇢ = r · ((rK ⇤ ⇢)⇢) + r · (rV ⇢) + ⇢m K, V : Rd ! R, and m 1
dt

K(x) = |x|3 /3

14
gradient flow in the Wasserstein metric
aggregation, drift, and degenerate diffusion:
d
⇢ = r · ((rK ⇤ ⇢)⇢) + r · (rV ⇢) + ⇢m K, V : Rd ! R, and m 1
dt

K(x) = |x|3 /3

14
gradient flow in the Wasserstein metric
aggregation, drift, and degenerate diffusion:
d
⇢ = r · ((rK ⇤ ⇢)⇢) + r · (rV ⇢) + ⇢m K, V : Rd ! R, and m 1
dt

3
K(x) = |x| /3 |x| K(x) = |x|3 /3

14
gradient flow in the Wasserstein metric
aggregation, drift, and degenerate diffusion:
d
⇢ = r · ((rK ⇤ ⇢)⇢) + r · (rV ⇢) + ⇢m K, V : Rd ! R, and m 1
dt

3
K(x) = |x| /3 |x| K(x) = |x|3 /3

14
gradient flow in the Wasserstein metric
aggregation, drift, and degenerate diffusion:
d
⇢ = r · ((rK ⇤ ⇢)⇢) + r · (rV ⇢) + ⇢m K, V : Rd ! R, and m 1
dt

K(x) = |x|3 /3 |x| K(x) = |x|3 /3 |x|, V (x) = x

15
gradient flow in the Wasserstein metric
aggregation, drift, and degenerate diffusion:
d
⇢ = r · ((rK ⇤ ⇢)⇢) + r · (rV ⇢) + ⇢m K, V : Rd ! R, and m 1
dt

K(x) = |x|3 /3 |x| K(x) = |x|3 /3 |x|, V (x) = x

15
gradient flow in the Wasserstein metric
aggregation, drift, and degenerate diffusion:
d
⇢ = r · ((rK ⇤ ⇢)⇢) + r · (rV ⇢) + ⇢m K, V : Rd ! R, and m 1
dt

K(x) = |x|3 /3 |x| K(x) = |x|3 /3 |x|, V (x) = x

15
gradient flow in the Wasserstein metric
aggregation, drift, and degenerate diffusion:
d
⇢ = r · ((rK ⇤ ⇢)⇢) + r · (rV ⇢) + ⇢m K, V : Rd ! R, and m 1
dt

K(x) = |x|3 /3 |x| K(x) = |x|3 /3 |x| , m = 1

16
gradient flow in the Wasserstein metric
aggregation, drift, and degenerate diffusion:
d
⇢ = r · ((rK ⇤ ⇢)⇢) + r · (rV ⇢) + ⇢m K, V : Rd ! R, and m 1
dt

K(x) = |x|3 /3 |x| K(x) = |x|3 /3 |x| , m = 1

16
gradient flow in the Wasserstein metric
aggregation, drift, and degenerate diffusion:
d
⇢ = r · ((rK ⇤ ⇢)⇢) + r · (rV ⇢) + ⇢m K, V : Rd ! R, and m 1
dt

K(x) = |x|3 /3 |x| K(x) = |x|3 /3 |x| , m = 1

16
numerical simulation of W2 grad flow
JKO scheme
• Pros: reduces simulating grad flow to solving a sequence of optimization
problems involving W2 distance; leverages state of the art W2 solvers.
• Cons: current methods lose convexity/stability properties of gradient flow

Finite volume methods, finite element methods…

• Pros: adapts numerical approaches inspired by classical fluid mechanics
to gradient flow setting
• Cons: current methods lose convexity/stability properties of gradient flow

Without stability, we can’t prove general results on convergence.

Particle methods
17
particle methods
d
Goal: Approximate a solution to ⇢(x, t) + r · (v(x, t)⇢(x, t)) = 0
dt
Assume: v(x,t) comes from a Wasserstein gradient flow and is “nice”

A general recipe for a particle methods:

(1) approximate ρ₀(x) as a sum of Dirac masses on a grid of spacing h
XN
⇢0 ⇡ xi mi
i=1

(2) evolve the locations of the Dirac masses by

d
xi (t) = v(xi (t), t) 8i
dt
XN
(3) ⇢N (x, t) = xi (t) mi is a gradient flow of the original energy;
i=1
it inherits all convexity/stability properties, hence ⇢N (x, t) ! ⇢(x, t)
18
particle methods
d
Goal: Approximate a solution to ⇢(x, t) + r · (v(x, t)⇢(x, t)) = 0
dt

Benefits of particle methods:

(1) positivity preserving

(2) inherently adaptive

(3) energy decreasing

(4) preserves convexity/stability properties of gradient flow at discrete level

but what about when v(x,t) is not “nice”?

19
v(x,t) is often not nice v(x,t)

aggregation, drift, and degenerate diffusion:

d
⇢ = r · ((rK ⇤ ⇢)⇢) + r · (rV ⇢) + ⇢m K, V : Rd ! R, and m 1
dt
m 2
v = rK ⇤ ⇢ + rV + m⇢ r⇢

• Diffusion term is worst: particles do not remain particles

• But even the interaction term can slow down convergence if it has a
strong singularity at the origin…

20
v(x,t) is often not nice v(x,t)

aggregation, drift, and degenerate diffusion:

d
⇢ = r · ((rK ⇤ ⇢)⇢) + r · (rV ⇢) + ⇢m K, V : Rd ! R, and m 1
dt
m 2
v = rK ⇤ ⇢ + rV + m⇢ r⇢

• Diffusion term is worst: particles do not remain particles

• But even the interaction term can slow down convergence if it has a
strong singularity at the origin…

“Blob Method”: regularize the velocity field to make it nice!

20
v(x,t) is often not nice v(x,t)

aggregation, drift, and degenerate diffusion:

d
⇢ = r · ((rK ⇤ ⇢)⇢) + r · (rV ⇢) + ⇢m K, V : Rd ! R, and m 1
dt
m 2
v = rK ⇤ ⇢ + rV + m⇢ r⇢

• Diffusion term is worst: particles do not remain particles

• But even the interaction term can slow down convergence if it has a
strong singularity at the origin…

“Blob Method”: regularize the velocity field to make it nice!

20
particle
a blob method
methods
for aggregation
dd
Goal: Approximate a solution to ⇢⇢(x,
= rt)· + r · (v(x,
((rK t)⇢(x,
⇤ ⇢)⇢) v =t))rK
= 0⇤ ⇢
dt
dt

A regularized particle method:

(0) regularize the interaction kernel via convolution with a mollifier
d
K ✏ (x) = K ⇤ '✏ (x), '✏ (x) = '(x/✏)/✏
(1) approximate ρ₀(x) as a sum of Dirac masses on a grid of spacing h
XN
⇢0 ⇡ xi mi
i=1
(2) evolve the locations of the Dirac masses by

N
X
(3) ⇢N (x, t) = xi (t) mi is a gradient
Z flow of the regularized energy
i=1
E✏ (⇢) = (K ✏ ⇤ ⇢)⇢
21
a blob method for aggregation
Theorem [C., Bertozzi 2014]: If ε = hq, 0<q<1, the blob method
two dimensions,
converges as h →0 . aggregation
K(x) = log |x|/2π K(x) = |x|2 /2 K(x) = |x|3 /3

Time
1

0
Pos
itio 0.0
0.5
n -0.5

22
aggregation + ?
aggregation, drift, and diffusion:
d m
⇢ = r · ((rK ⇤ ⇢)⇢) + r · (rV ⇢) + ⇢ K, V : Rd ! R, and m 1
dt

• Adding a drift term is straightforward: just do a particle method with

v = rK✏ ⇤ ⇢ + rV

• How can we add diffusion?

• Previous work: stochastic [Liu, Yang 2017], [Huang, Liu 2015],

deterministic [Carrillo, Huang, Patacchini, Wolansky 2016]

• Our idea: regularize by convolution with a mollifier.

23
a blob method for degenerate diffusion
diffusion equation:
d
⇢= ⇢m m 1
dt
Z
1
Solutions of diffusion equation are gradient flows of E(⇢) = ⇢m
Z m 1
1
Let’s consider gradient flows of E✏ (⇢) = (⇢ ⇤ '✏ )m 1 ⇢
m 1

• Previous work (m=2):[Lions, Mas-Gallic 2000], [P.E. Jabin, in progress]

• For ε > 0, particles remain particles, so can do a particle method for
v = r'✏ ⇤ ('✏ ⇤ ⇢)m 2
⇢ + ('✏ ⇤ ⇢)m 2
(r'✏ ⇤ ⇢)

=) a blob method for diffusion.

24
a blob method for degenerate diffusion
diffusion equation:
d
⇢= ⇢m m 1
dt

Solutions of diffusion equation are gradient flows of

Z
1
Let’s consider gradient flows of E✏ (⇢) = (⇢ ⇤ '✏ )m 1
⇢
m 1

• Previous work (m=2):[Lions, Mas-Gallic 2000], [P.E. Jabin, in progress]

• For ε > 0, particles remain particles, so can do a particle method for
v = r'✏ ⇤ ('✏ ⇤ ⇢)m 2
⇢ + ('✏ ⇤ ⇢)m 2
(r'✏ ⇤ ⇢)

=) a blob method for diffusion.

24
a blob method for degenerate diffusion
diffusion equation:
d
⇢= ⇢m m 1
dt

Solutions of diffusion equation are gradient flows of

Let’s consider gradient flows of

• Previous work (m=2):[Lions, Mas-Gallic 2000], [P.E. Jabin, in progress]

• For ε > 0, particles remain particles, so can do a particle method for
v = r'✏ ⇤ ('✏ ⇤ ⇢)m 2
⇢ + ('✏ ⇤ ⇢)m 2
(r'✏ ⇤ ⇢)

=) a blob method for diffusion.

24
a blob method for degenerate diffusion
diffusion equation:
d
⇢= ⇢m m 1
dt
Z
Solutions of diffusion equation are gradient flows of E(⇢) = log(⇢)⇢

Let’s consider gradient flows of

• Previous work (m=2):[Lions, Mas-Gallic 2000], [P.E. Jabin, in progress]

• For ε > 0, particles remain particles, so can do a particle method for
v = r'✏ ⇤ ('✏ ⇤ ⇢)m 2
⇢ + ('✏ ⇤ ⇢)m 2
(r'✏ ⇤ ⇢)

=) a blob method for diffusion.

24
a blob method for degenerate diffusion
diffusion equation:
d
⇢= ⇢m m 1
dt
Z
Solutions of diffusion equation are gradient flows of E(⇢) = log(⇢)⇢
Z
Let’s consider gradient flows of E✏ (⇢) = log(⇢ ⇤ '✏ )⇢

• Previous work (m=2):[Lions, Mas-Gallic 2000], [P.E. Jabin, in progress]

• For ε > 0, particles remain particles, so can do a particle method for
v = r'✏ ⇤ ('✏ ⇤ ⇢)m 2
⇢ + ('✏ ⇤ ⇢)m 2
(r'✏ ⇤ ⇢)

=) a blob method for diffusion.

24
a blob method for degenerate diffusion
Theorem [Carrillo, C., Patacchini 2017]: Consider
Z Z Z
1
E✏ (⇢) = (K ⇤ ⇢)⇢ + V ⇢ + (⇢ ⇤ '✏ )m 1
⇢
m 1
As ε→0,
• For all m ≥ 1, Eε Γ-converge to E.
• For m = 2 and initial data with bounded entropy, gradient flows of Eε
converge to gradient flows of E.
• For m ≥ 2 and particle initial data with ε = hq, 0<q<1, if a priori
estimates hold, gradient flows of Eε converge to gradient flows of E.

25
numerics: Keller-Segel (d=2)
subcritical mass critical mass supercritical mass
Evolution of Density

Evolution of Second Moment

26
numerics: Keller-Segel (d=2)
subcritical mass critical mass supercritical mass
Evolution of Density

Evolution of Second Moment

26
numerics: Keller-Segel (d=2)
subcritical mass critical mass supercritical mass
Evolution of Density

Evolution of Second Moment

26
Future work
• Convergence for 1 ≤ m < 2?

• Quantitative estimates for m ≥ 2?

• Utility in related fluids and kinetic equations?

Thank you!