bounding the mixing time via spectral gap

15
Bounding the mixing time via Spectral Gap Ilan Ben-Bassat Omri Weinstein

Upload: jalena

Post on 22-Feb-2016

33 views

Category:

Documents


0 download

DESCRIPTION

Bounding the mixing time via Spectral Gap. Ilan Ben-Bassat Omri Weinstein. Conductance . Properties. Reversible Chains. (*). x>=y. Upper Bound of the Spectral Gap. Path Congestion. - PowerPoint PPT Presentation

TRANSCRIPT

Page 1: Bounding the mixing time via Spectral Gap

Bounding the mixing time via Spectral Gap

Ilan Ben-Bassat Omri Weinstein

Page 2: Bounding the mixing time via Spectral Gap

Conductance For any ๐‘†โŠ† ฮฉ denote ๐›ท๐‘† = ๐œ‹(๐‘†)โˆ’1 ๐‘„(๐‘†,๐‘†าง)

where ๐‘„(๐‘†,๐‘†าง) = ฯ€(ฯ‰) P(ฯ‰,ฯƒ) Intuition: ๐›ท๐‘† is the steady state probability of moving from ๐‘† to ๐‘†าง, Conditioned on being in ๐‘†.

๐›ท= min๐›ท๐‘† โˆถ ๐‘†โŠ† ฮฉ, 0 < ๐œ‹(๐‘†) โ‰ค 1/2

Page 3: Bounding the mixing time via Spectral Gap

If โ„ณ is lazy , ๐›ท๐‘† โ‰ค 1/2. ๐›ท๐‘†๐œ‹แˆบ๐‘†แˆป= Qแˆบ๐‘†,๐‘†างแˆป= Qแˆบ๐‘†าง,๐‘†แˆป= ๐›ท๐‘†าง๐œ‹แˆบ๐‘†างแˆป

Qแˆบ๐‘†,๐‘†างแˆป= Qแˆบฮฉ,๐‘†างแˆปโˆ’ Qแˆบ๐‘†าง,๐‘†างแˆป= ๐œ‹แˆบ๐‘†างแˆปโˆ’ Qแˆบ๐‘†าง,๐‘†างแˆป= Qแˆบ๐‘†าง,๐‘†แˆป

Lemma: If โ„ณ is lazy, irreducible and aperiodic, then all eigenvalues of ๐‘ƒ are positive.

Reversible Chains

Properties

Page 4: Bounding the mixing time via Spectral Gap

Denote for any ๐‘ฆโˆˆโ„๐‘: ๐œ€แˆบ๐‘ฆ,๐‘ฆแˆป= ๐œ‹๐‘–๐‘ƒ๐‘–,๐‘—(๐‘ฆ๐‘– โˆ’ ๐‘ฆ๐‘—)2๐‘–<๐‘—

Lemma: If โ„ณ is reversible, then

1โˆ’ ๐œ†1 = ๐‘š๐‘–๐‘› ๐œ‹๐‘‡๐‘ฆ=0 ๐œ€แˆบ๐‘ฆ,๐‘ฆแˆปฯƒ ๐œ‹๐‘–๐‘– ๐‘ฆ๐‘–2 Proof. Recall ,๐‘†= ๐ท1/2๐‘ƒ๐ทโˆ’1/2, ๐‘’(0). Note that ๐œ†1 = ๐‘š๐‘Ž๐‘ฅ<๐‘’(0),๐‘ฅ>=0 <๐‘†๐‘ฅ,๐‘ฅ><๐‘ฅ,๐‘ฅ>

Now consider the matrix ๐ท1/2(๐ผโˆ’ ๐‘ƒ)๐ทโˆ’1/2. What are itโ€™s eigenvalues?โ€ฆ

Page 5: Bounding the mixing time via Spectral Gap

If โ„ณ is reversible then

1โˆ’ ๐œ†1 โ‰ฅ ๐›ท22

(*)

Page 6: Bounding the mixing time via Spectral Gap
Page 7: Bounding the mixing time via Spectral Gap
Page 8: Bounding the mixing time via Spectral Gap

( ) 1NS

Page 9: Bounding the mixing time via Spectral Gap

Corollary ๐œแˆบ๐œ€แˆป= 2|log (๐œ€๐œ‹๐‘š๐‘–๐‘› )|๐›ท2 เถˆ

Proof. 1log (๐œ†๐‘š๐‘Ž๐‘ฅโˆ’1 ) โ‰ค 1log ((1โˆ’๐›ท22 )โˆ’1) โ‰ค 2๐›ท2

Page 10: Bounding the mixing time via Spectral Gap

Consider a random walk on a Graph G=(V,E). By definition, we have

If G is a r-regular graph

Page 11: Bounding the mixing time via Spectral Gap

Example: A walk on the discrete cube ๐‘„๐‘›. For ๐‘†โŠ† ๐‘„๐‘› Let ๐‘–๐‘›(๐‘†) denote the number of edges which are wholly contained in ๐‘†. Lemma: if ๐‘† โ‰  โˆ…, ๐‘–๐‘›(๐‘†) โ‰ค 12 |๐‘†|๐‘™๐‘œ๐‘”2|๐‘†| Proof. By induction on n, using the inequality: ๐‘ฅ๐‘™๐‘œ๐‘”2๐‘ฅ + ๐‘ฆ๐‘™๐‘œ๐‘”2๐‘ฆ + 2y โ‰ค (๐‘ฅ+ ๐‘ฆ)๐‘™๐‘œ๐‘”2(x+ y) By summing the degrees at each vertex of ๐‘† we have ๐‘’แˆบ๐‘†,๐‘†างแˆป+ 2๐‘–๐‘›แˆบ๐‘†แˆป= ๐‘›|๐‘†| So by the Lemma we get

๐‘’แˆบ๐‘†,๐‘†างแˆปโ‰ฅ ๐‘›ศA๐‘†ศAโˆ’ 12ศA๐‘†ศA๐‘™๐‘œ๐‘”2ศA๐‘†ศA= ศA๐‘†ศAเตฌ๐‘› โˆ’ 12๐‘™๐‘œ๐‘”2ศA๐‘†ศAเตฐโ‰ฅ |๐‘†|

x>=y

Page 12: Bounding the mixing time via Spectral Gap

It follows that ฮฆ โ‰ฅ 1๐‘› . Adding self-loops will halve the conductance -

the denominator ฯƒ ๐‘‘๐‘ฃ๐‘ฃโˆˆ๐‘† doubles without changing the numerator in ๐›ท๐‘ . By (*), this gives us an estimate of 1โˆ’ 18๐‘›2 for the spectral gap (so weโ€™re off by a factor of n)

Page 13: Bounding the mixing time via Spectral Gap

Upper Bound of the Spectral Gap

Page 14: Bounding the mixing time via Spectral Gap

Path Congestion

Page 15: Bounding the mixing time via Spectral Gap