the poincar é constant of a random walk in high-dimensional convex bodies
DESCRIPTION
The Poincar é Constant of a Random Walk in High-Dimensional Convex Bodies. Ivona Bez áková Thesis Advisor: Prof. Eric Vigoda. Goal: Efficient algorithm for sampling points from high-dimensional convex bodies. Approach: Random walks. Overview of the talk. Motivation & History - PowerPoint PPT PresentationTRANSCRIPT
The Poincaré Constant ofa Random Walk in High-
Dimensional Convex Bodies
Ivona Bezáková
Thesis Advisor: Prof. Eric Vigoda
Goal:
Efficient algorithm for sampling points from high-dimensional convex bodies
Approach: Random walks
Overview of the talk
• Motivation & History
• Introduction to random walks, Markov chains
• Definition and basic properties of ball walks
• Overview of the algorithm for uniform sampling
• Analysis of the Poincaré constant
Sampling points from convex bodies efficiently
• Computation of volume
• Sampling contingency tables [Morris]
• Universal portfolios [Kalai & Vempala]
• Convex programs [Bertsimas & Vempala]
Motivation
Why is computing volume difficult?
Straightforward approach:
- find bounding box
- sample sufficiently many points from the box
- compute ratio: # points in the body vs. total # points
This algorithm correctly approximates volume.
Where is the catch?
Ratio of the volume of the body vs. volume of the box decreases exponentially with dimension!
This results in exponential running time.
Note: For small dimension (2D, 3D) this algorithms works well.
Goal: Find algorithm running in time polynomial in dimension.
How efficient sampling helps with volume?
• intersect with balls doubling in volume
• this defines sequence of convex bodies
• sample points from i-th body, compute ratio: i-th vs. (i-1)-st body
• return product of ratios
Why is this better than the bounding box?
- Volume of bodies max. doubles
Result: volume of the original body vs. volume of the smallest body (a ball)
History
• negative results for volume computation:
- cannot be approximated by deterministic polytime algorithm
• randomized approximation:
- Dyer, Frieze, Kannan ’89
- improvements: combinations of [Applegate, Dyer, Frieze, Kannan, Lovász, Simonovits]
- Kannan, Lovász, Simonovits ‘97 * 5( )O n
* 23( )O n
- #P-complete
[Elekes, Bárány & Füredi, Dyer & Frieze, Khachiyan, Lawrence, Lovász & Simonovits]
Notation
• convex body of diameter
(given by membership oracle)
nK R D
• step-size • ball ( , )B x
D
K
x
( , )B x 0
unit-b
all
Ball Walks
Kx
Speedy Walk – next point ( , )y B x K
y
Metropolis Walk – next point
if then
( , )y B x y K :y x
Problem: How to implement?
Markov Chains
State space
Transition distribution
( , )P x
(likelihood of going from to )x y
For speedy walk we have
K state space = convex body
( , ) 1 ( ( , ) )P x y vol B x K if || ||x y
0 otherwise
x y
Markov Chains 2
Stationary distribution = limiting distribution
P (fix-point) ( ) ( ) ( , )x y P y x dy
i.e.
Mixing time
For given mixing time is the expected number of steps needed to get close to the stationary distribution.
0
Want: rapid mixing, i.e. time polynomial in and 1 n
Comparison KLS vs. this work
Kannan, Lovász, Simonovits study so-called conductance for bounding mixing time.
* 3 2( )O n D
? poly-logarithmic in 1
We bound so-called Poincaré constant (generalization of conductance) and get mixing time
* 3 2( )O n D
cubic in 1
spectr
al gap
New ideas in KLS
• separated analysis of speedy walk (fast mixing in principle) and Metropolis walk (efficient implementation)
• for volume computation: introduced isotropic position to reduce diameter of the body
Why Poincaré constant?
• generalization might lead to better analysis through other quantities (log-Sobolev, [Frieze & Kannan, Jerrum & Son])
• the same difficulty
our focus: survey
Poincaré constant
where
:
( , )inf
( )f
f f
Var f
R
2( ) ( )( ( ) ( ))Var f x f x E f dx
( , ) ( ) ( )f f x h x dx
Well-studied Quantities
21( ) ( , )( ( ) ( ))
2h x P x y f x f y dy
and Dirichlet form
21( ) ( ) ( )( ( ) ( ))
2Var f x y f x f y dxdy
21( , ) ( ) ( , )( ( ) ( ))
2f f x P x y f x f y dxdy
(local variance) mea
sure
s de
cayi
ng o
f va
rianc
e
Well-studied Quantities 2
(Properties of Poincaré)
Thm: For (lazy reversible) Markov Chain
2 0 2|| || (1 / 2) || ||tt
For Markov chains defined on finite state spaces the Poincaré constant equals the spectral gap.
twhere is the distribution after stepst
with probability ½ stay at the same statecorresponds to symmetric chains
and is the stationary distribution
Thoughts about Poincaré constant
If then ( , ) ( )f f Var f ( , ) ( )P x y y
Thus, in this case and the chain mixes (very) rapidly.
Intuitively, this corresponds to a complete graph, where we can get from any point to any other point.
1
21( , ) ( ) ( , )( ( ) ( ))
2f f x P x y f x f y dxdy
21( ) ( ) ( )( ( ) ( ))
2Var f x y f x f y dxdy
Well-studied Quantities 3
10 ( ) 1/ 2inf Pr( | )t t
SX S X S
: {0,1}
( , )inf
( )f
f f
Var f
Conductance
equals Poincaré over indicator functions
trivially
Cheeger-type inequality by Jerrum and Sinclair, ‘89
2
2
Properties of Ball Walks
Local conductance( ( , ) )
( )( ( , ))
vol B x Kl x
vol B x
Ball walks:
• stationary distribution( )
( )l x
xL
( )K
L l x dxwhere
• reversible( ) 1
( ) ( , )( ( , ) )
l xx P x y
L vol B x K
From Speedy Walk to Uniform Sampling
(Overview)
• bound Poincaré constant for speedy walk
• mixing time for speedy walk
• running time of Metropolis walk (assuming good starting distribution)
• obtain a good starting distribution
• from a sample point from the speedy distribution obtain a sample point close to the uniform distribution
From Speedy Walk to Uniform Sampling
Poincaré inequality (for speedy walk):
/c D n If then2
2
c
D n
for some dimension-independent constant c
Mixing time of speedy walk:
For given 0 distribution after1 1( ln )t O
steps within from speedy distribution
(assuming reasonable starting distribution )0
Thm: For (lazy reversible) Markov Chain
2 0 2|| || (1 / 2) || ||tt
twhere is the distribution after stepstand is the stationary distribution
From Speedy Walk to Uniform Sampling
(Overview)
• bound Poincaré constant for speedy walk
• mixing time for speedy walk
• running time of Metropolis walk (assuming good starting distribution)
• obtain a good starting distribution
• from a sample point from the speedy distribution obtain a sample point close to the uniform distribution
From Speedy Walk to Uniform Sampling 2
From speedy to Metropolis walk
Run M. walk until speedy steps
Mixing time of Metropolis walk:
If 0 0 2: || || 1m then we expect the
1 1( ln )t O
total number of steps (speedy + Metropolis)
(with exception )20mto be at most
20
3
(1 )
t
m
where is the average local conductance:
( ) ( )K
l x dx vol K
From Speedy Walk to Uniform Sampling
(Overview)
• bound Poincaré constant for speedy walk
• mixing time for speedy walk
• running time of Metropolis walk (assuming good starting distribution)
• obtain a good starting distribution
• from a sample point from the speedy distribution obtain a sample point close to the uniform distribution
From Speedy Walk to Uniform Sampling 3
Obtaining a good starting distribution
Let (0,1) (0, )B K B D and /(0, (1 ) )i n
iB B
i iK B K for where0,...,i b
Algo: • Sample from according to
• For obtain :
Run Metropolis in starting at
1logb n D
0x 0K 0
1,...,i b
1ix iKix
From Speedy Walk to Uniform Sampling 4
Good starting distribution for Metropolis walk
Thm: For sufficiently small and0 2 n
the distribution of is within of .ix i
Expected total number of oracle calls
(with exception ) is less than3b
3 2 3( (1 ) ln ln(1 ))O n D D
From Speedy Walk to Uniform Sampling
(Overview)
• bound Poincaré constant for speedy walk
• mixing time for speedy walk
• running time of Metropolis walk (assuming good starting distribution)
• obtain a good starting distribution
• from a sample point from the speedy distribution obtain a sample point close to the uniform distribution
From Speedy Walk to Uniform Sampling 5
From speedy distribution to the uniform distribution
Algo: • Shrink ' (2 1 2 )K n n K K• Sample from until 'x K
• Return : (2 2 1)y n n x
Thm: If and sufficiently small
1 8 ln( / )n n 0 then the distribution of is y 10
away from the uniform distribution.
Expected number of samples needed .2
From Speedy Walk to Uniform Sampling
(Overview)
• bound Poincaré constant for speedy walk
• mixing time for speedy walk
• running time of Metropolis walk (assuming good starting distribution)
• obtain a good starting distribution
• from a sample point from the speedy distribution obtain a sample point close to the uniform distribution
Proof of the Poincaré Inequality
Restricted variance, Dirichlet form, expected value
' '[ ( ) ( ')] ( )K K
f x K f x dx 2
' ''( ) ( )( ( ) )K KK
Var f x f x f dx ' '( , ) ( ) ( )K Kf f x h x dx
Poincaré inequality (for speedy walk):
/c D n If then for any function2
2
c
D n
for some dimension-independent constant c
where
:f K R
( , ) ( )f f Var f
Idea of the proof:
• For a sufficiently small set such that does not vary much within
• Assuming Poincaré does not hold, we find a set contradicting the above
'K K'K
( )l x
' '( , ) ( ) 1 [2 ( 1)]K K lf f Var f c e
( , ) ( )K Kf f Var f
Find needle-like s.t.
1 1( , ) ( )K Kf f Var f
1K K
wlog ( ) 0KE f f
and10Kf
Chop to obtain desired set1K 0K
Needle-like Body
Eliminate dimensions one by one (inductively)
( , ) ( )i iK Kf f Var f 0
iKf • assume has fat dimensions and iK 1i
while
• projection of onto two fat dim.J iK
• there exists a point s.t. any lineJthrough cuts into appx. half
• take hyperplane s.t. 0i iK H K H
f f H
( , ) ( )i iK H K H
f f Var f
• at least one of these must be true
( , ) ( )i iK H K H
f f Var f
or
iK
J H
Shrinking Last Dimension1K
'D
Goal: find s.t. last dim. of is0 1K K 0K'
:c D
D n
and0 0 0( , ) ( )K Kf f Var f c
where is a constant (dependent on )0c c
How? Chop into1K 0 1,..., mS S
Ideally
1 1 0( , ) ( , ) ( ) ( )i iS K K Sf f f f Var f c Var f
?
But
1 1
2( ) ( ) ( )( )
i iK S i S KVar f Var f S f f
Assumption1 1
2( ) 2 ( )( )
iK i S KVar f S f f Idea: relate to ( )
iUVar f where 1:i i iU S S
We get
1 ,0
1( ) ( ) ( )ki j
i j mK UK Var f Var fa
where
, : 2 1j
i j i j kk i
a w w w
and : ( )k kw S
Next goal: bound ,0
i ji j m
a
Chopping of 1K
What do we need?
• does not vary much within , i.e. for any ( )l x
, ix y S let ( ) / ( ) 2l x l y
• width of is at mostiS / 2
iS
We will show that this chopping allows us to bound
,0
i ji j m
a appropriately
Properties of local conductance
• is concave over 1/( ) nl x K
From Brunn-Minkowski Thm:
• is Lipschitz over : For anyln ( )l x K ,x y K
| ln ( ) ln ( ) | || ||n
l x l y x y
Implications for the iS
• The width of increases, then it is (full width)
iS
and then it decreases
/ 2
• For sufficiently small the width of any is
at least
(2 )n
iS
From Dinghas’ Thm:
For the middle section is convex.11/ ,...,1/l rw w
Now we can split into several sums and estimate them separately
,i ja
, , , , ,0 ,
...i j i j i j i j i ji j m i j l i l j r i l r j l i j r
a a a a a
Thus 2, 1( ( ))i j
l i j r
a O M K
where is the number of slabs in the middle sectionM
What to do outside the middle section?
, 1( )i j li j l
a O lw
In the left section, the increase exponentiallyiw
This allows us to bound
We obtain similar bounds for other parts of the sum, putting them together we get
22
, 1 12 2(( ) ( )) ( ( ))i j
D na O l M m r K O K
c
We wanted 0
, 1( )i jl i j r
ca K
Thus we proved the Poincaré inequality.
THANK YOU