Understanding Machine Learning: From Theory to Algorithms

(Jeff_L) #1
Technical Lemmas 421

Using Stirling’s approximation we further have that



(em
d

)d(
1 +

(

d
e

)d
(m−d)
(d+ 1)


2 πd(d/e)d

)

=

(em
d

)d(
1 +
√ m−d
2 πd(d+ 1)

)

=

(em
d

)d
·

d+ 1 + (m−d)/


2 πd
d+ 1

(em
d

)d
·

d+ 1 + (m−d)/ 2
d+ 1
=

(em
d

)d
·

d/2 + 1 +m/ 2
d+ 1

(em
d

)d
· m
d+ 1

,

where in the last inequality we used the assumption thatd≤m−2. On the
other hand,
(
em
d+ 1


)d+1
=

(em
d

)d
·

em
d+ 1·

(

d
d+ 1

)d

=

(em
d

)d
·

em
d+ 1

·

1

(1 + 1/d)d


(em
d

)d
·

em
d+ 1

·

1

e
=

(em
d

)d
·

m
d+ 1,

which proves our inductive argument.


lemmaA.6 For alla∈Rwe have


ea+e−a
2 ≤e

a^2 / (^2).
Proof Observe that
ea=


∑∞

n=0

an
n!

.

Therefore,


ea+e−a
2 =

∑∞

n=0

a^2 n
(2n)!,

and


ea

(^2) / 2


∑∞

n=0

a^2 n
2 nn!

.

Observing that (2n)!≥ 2 nn! for everyn≥0 we conclude our proof.

Free download pdf