08.10.2016 Views

Foundations of Data Science

2dLYwbK

2dLYwbK

SHOW MORE
SHOW LESS

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

Now consider a very tall tree. If the p is the probability that a root has value +1,<br />

we can iterate the formula for the height one tree and observe that at low temperature<br />

the probability <strong>of</strong> the root being one converges to some value. At high temperature, the<br />

probability <strong>of</strong> the root being one is 1 / 2 independent <strong>of</strong> p. See Figure 9.10. At the phase<br />

transition, the slope <strong>of</strong> q at p=1/2 is one.<br />

Now the slope <strong>of</strong> the probability <strong>of</strong> the root being 1 with respect to the probability <strong>of</strong><br />

a leaf being 1 in this height one tree is<br />

∂q<br />

∂p = D ∂C − C ∂D<br />

∂p ∂p<br />

D 2<br />

Since the slope <strong>of</strong> the function q(p) at p=1/2 when the phase transition occurs is one, we<br />

can solve ∂q = 1 for the value <strong>of</strong> β where the phase transition occurs. First, we show that<br />

∂p<br />

∣<br />

= 0.<br />

∂D<br />

∂p<br />

∣<br />

p=<br />

1<br />

2<br />

Then<br />

∂q<br />

∣<br />

∂p<br />

∂D<br />

∂p<br />

∣<br />

p=<br />

1<br />

2<br />

∣<br />

D = [ pe 2β + 1 − p ] d<br />

+<br />

[<br />

p + (1 − p) e<br />

2β ] d<br />

∂D<br />

= d [ pe 2β + 1 − p ] d−1 (<br />

∂p e 2β − 1 ) + d [ p + (1 − p) e 2β] d−1 ( )<br />

1 − e<br />

2β<br />

∣<br />

p=<br />

1<br />

2<br />

= d<br />

2 d−1 [<br />

e 2β + 1 ] d−1 (<br />

e 2β − 1 ) + d<br />

2 d−1 [<br />

1 + e<br />

2β ] d−1 (<br />

1 − e<br />

2β ) = 0<br />

= D ∂C − C ∣ ∂D ∣∣∣∣p=<br />

∂p ∂p<br />

D 2 1<br />

2<br />

=<br />

∂C<br />

∂p<br />

D<br />

∣<br />

∣ 1 p=<br />

2<br />

= d [ ]<br />

1<br />

2 e2β + 1 d−1 (<br />

2 e 2β − 1 )<br />

[ 1<br />

]<br />

2 e2β + 1 d [<br />

2 + 1 + 1e2β] = d ( e 2β − 1 )<br />

d<br />

1 + e 2β<br />

2 2<br />

=<br />

d [ pe 2β + 1 − p ] d−1 (<br />

e 2β − 1 )<br />

[pe 2β + 1 − p] d + [p + (1 − p) e 2β ] d ∣ ∣∣∣∣p=<br />

1<br />

2<br />

Setting<br />

And solving for β yields<br />

d ( e 2β − 1 )<br />

1 + e 2β = 1<br />

d ( e 2β − 1 ) = 1 + e 2β<br />

e 2β = d+1<br />

d−1<br />

β = 1 2<br />

ln<br />

d+1<br />

d−1<br />

To complete the argument, we need to show that q is a monotonic function <strong>of</strong> p. To see<br />

this, write q = 1<br />

1+ B . A is a monotonically increasing function <strong>of</strong> p and B is monotonically<br />

A<br />

325

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

Saved successfully!

Ooh no, something went wrong!