08.10.2016 Views

Foundations of Data Science

2dLYwbK

2dLYwbK

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

The Gaussian distribution on the interval [-1,1]<br />

Consider the interval [−1, 1]. Let δ be a “grid size” specified later and let G be the<br />

graph consisting <strong>of</strong> a path on the 2 + 1 vertices {−1, −1 + δ, −1 + 2δ, . . . , 1 − δ, 1} having<br />

δ<br />

self loops at the two ends. Let π x = ce −αx2 for x ∈ {−1, −1 + δ, −1 + 2δ, . . . , 1 − δ, 1}<br />

where α > 1 and c has been adjusted so that ∑ x π x = 1.<br />

We now describe a simple Markov chain with the π x as its stationary probability and<br />

argue its fast convergence. With the Metropolis-Hastings’ construction, the transition<br />

probabilities are<br />

( )<br />

( )<br />

p x,x+δ = 1 2 min 1, e−α(x+δ)2<br />

and p x,x−δ = 1 e −αx2 2 min 1, e−α(x−δ)2<br />

.<br />

e −αx2<br />

Let S be any subset <strong>of</strong> states with π(S) ≤ 1 . First consider the case when S is an interval<br />

2<br />

[kδ, 1] for k ≥ 2. It is easy to see that<br />

π(S) ≤<br />

≤<br />

∫ ∞<br />

x=(k−1)δ<br />

∫ ∞<br />

= O<br />

(k−1)δ<br />

(<br />

ce −αx2 dx<br />

x<br />

(k − 1)δ ce−αx2 dx<br />

)<br />

ce −α((k−1)δ)2<br />

α(k − 1)δ<br />

Now there is only one edge from S to ¯S and total conductance <strong>of</strong> edges out <strong>of</strong> S is<br />

∑ ∑<br />

π i p ij = π kδ p kδ,(k−1)δ = min(ce −αk2 δ 2 , ce −α(k−1)2 δ 2 ) = ce −αk2 δ 2 .<br />

i∈S<br />

j /∈S<br />

Using 2 ≤ k ≤ 1/δ, α ≥ 1, and π( ¯S) ≤ 1, we have<br />

Φ(S) =<br />

flow(S, ¯S)<br />

π(S)π( ¯S) ≥ δ2 α(k − 1)δ<br />

ce−αk2<br />

ce −α((k−1)δ)2<br />

≥ Ω(α(k − 1)δe −αδ2 (2k−1) ) ≥ Ω(αδe −O(αδ) ).<br />

.<br />

For δ < 1 α , we have αδ < 1, so e−O(αδ) = Ω(1), thus, Φ(S) ≥ Ω(αδ). Now, π min ≥ ce −α ≥<br />

e −1/δ , so ln(1/π min ) ≤ 1/δ.<br />

If S is not an interval <strong>of</strong> the form [k, 1] or [−1, k], then the situation is only better<br />

since there is more than one “boundary” point which contributes to flow(S, ¯S). We do<br />

not present this argument here. By Theorem 5.5 in Ω(1/α 2 δ 3 ε 3 ) steps, a walk gets within<br />

159

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!