Scribd
Upload
54 views6 pages

Cherno Bounds For The Tail of The Binomial Distribution

This document presents Chernoff bounds for the tail probabilities of the binomial distribution. It defines relevant terms like moment generating functions and proves two theorems and three corollaries that provide upper bounds on the probability that a binomial random variable is less than or equal to a given value. The bounds are frequently used in probabilistic arguments and combinatorial analyses.

Uploaded by

Nicola Tonellotto
Copyright
© Attribution Non-Commercial (BY-NC)
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PS, PDF, TXT or read online on Scribd
54 views6 pages

Cherno Bounds For The Tail of The Binomial Distribution

This document presents Chernoff bounds for the tail probabilities of the binomial distribution. It defines relevant terms like moment generating functions and proves two theorems and three corollaries that provide upper bounds on the probability that a binomial random variable is less than or equal to a given value. The bounds are frequently used in probabilistic arguments and combinatorial analyses.

Uploaded by

Nicola Tonellotto
Copyright
© Attribution Non-Commercial (BY-NC)
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PS, PDF, TXT or read online on Scribd
You are on page 1/ 6

Cherno Bounds for the Tail of the Binomial Distribution

Lenwood S. Heath Virginia Polytechnic Institute and State University June 7, 1988 This note proves a result (see Corollaries 2 and 3) of Cherno C] that bounds the probability of the tails of the binomial distribution. This is a direct, self-contained proof that avoids the generality of the original. The Cherno bounds are frequently used in combinatorial arguments (e.g., Valiant V]) and in the probabilistic method (e.g., Erdos and Spencer ES]). Related results can be found in Feller F2], page 525.

De nitions and Notation.


Let X be a real random variable with cumulative distribution function FX (x) = P(X x): If g is a real function of a real variable, then the expected value of g(X) is E(g(X)) = The moment generating function of X is MX (t) = E(etX ): Formal calculation with power series shows that MX (t) =
1 X
Z
+

g(x) dFX (x);

if the integral exists; integrals are, in general, Lebesgue integrals.

E(X i )ti i! i=0

(see Feller F1], page 285.) De ne the shifted in mum of MX to be mX (a) = inf MX a (t) t = inf e at MX (t): t = inf E et(X t
a)

Tchebyche Inequality.
The following result is from Kolmogorov K], page 42.

Lemma 1. Let Y be a real random variable. If g is a non-negative real function of a real variable
such that for y a; g(y) b > 0; then P(Y if E(g(Y )) exists. a) E(g(Y )) ; b

Proof:

E(g(Y )) =

Z a Z a

g(y) dFY (y)

g(y) dFY (y) b dFY (y) a):

The Lemma follows.

= bP(Y

Cherno 's Results. Theorem 1. (Cherno C]) If E(X) > 1 and a E(X); then
P(X a) mX (a):

Proof: Let
g(y; t) = et(y Then, for t 0 and y a; g(y; t) 1: By the Tchebyche inequality, P(X a) E(g(X; t)) = E(g(X; t)); 1 whenever t 0: But, E g(X; t) = MX a (t): Thus, P(X a) tinf0 MX a (t):
a) :

For all real x; 1 + x ex : For all t; MX a (t) =


Z Z
+

1 +1 1

et(x

a)

dFX (x)

1 + t(x a) dFX (x)

= 1 + t(E(X) a): Since a E(X); MX a (t) 1 = MX a (0); whenever t 0: Thus,


t

inf0 MX a (t) = inf MX a (t) = mX (a): t P(X a) mX (a):

Finally,

Corollary 1. If E(X) < +1 and a E(X); then


P(X a) mX (a):

Proof: E( X) > 1 and a E( X): By Theorem 1,


P(X a) = P( X But, a) m X ( a):
X +a)

m X ( a) = inf E et( t = inf E e t

t(X a) a)

= inf E et(X t The Corollary follows. = mX (a):

Binomial Distribution.
Let X1 ; X2 ; : : :; Xn be identically distributed, independent random variables with distribution P(Xi = 1) = p; P(Xi = 0) = q = 1 p:
n X i=1

Then, E(Xi ) = p: Let the sum of the random variables be Sn = Then, and E(Sn ) = np: Xi : 0 j n;

P(Sn = j) = n pj qn j ; j
n

We will now calculate mS (a) and apply Theorem 1 to Sn : MX (t) = E(etX ) = pet + q:
i i

MS (t) = E(etS )
n n

n Y

= (pet + q)n: De ne h(a; t) = e at(pet + q)n : We wish to minimize h(a; t); for xed a < n: Take the derivative of h with respect to t : h0 (a; t) = ae at (pet + q)n + e at pet n(pet + q)n 1: Assume h0(a; t) = 0; and solve for t: Then, a(pet + q) = pet n et = p(nqa a) : Thus h(a; t) is minimized when t = ln qa=p(n a) : We obtain mS (a) = inf h(a; t) t = h a; ln p(nqa a) : Corollary 2. Let Sn be a random variable having a binomial distribution with parameters n and p: If k E(Sn ) = np and 0 < k < n; then
n

i=1

E(et Xi )

by independence

k X j =0

n pj qn j = P(S k) n j

n np n k

n k

np k : k

Proof: By the Theorem,


P(Sn k) mS (k):
n

By the above calculations, mS (k) = h k; ln p(nkq k) n peln = e k ln +q k kq + q n = p(nkq k) n k k nq n k k q = p n k n k k nq k nq n k = np n k n k nq n k np k = n k k n np n k np k : = n k k


n kq p(n k) kq p(n k)

Corollary 3. Let Sn be a random variable having a binomial distribution with parameters n and
p: If k E(Sn ) = np and 0 < k < n; then
n X j =k

n pj qn j = P(S k) n j

n np n k

n k

np k : k

Proof: Combine Corollary 1 with the calculation in the proof of Corollary 2.


In Erdos and Spencer ES], the bound is written in exponential form: mS (k) = exp (n k) ln n nq k + k ln np : k
n

References
C] ES] F1] F2] K] V] Cherno , H. A measure of asymptotic e ciency for tests of a hypothesis based on the sum of observations. Annals of Mathematical Statistics 23 (1952), 493-507. Erdos, P., and Spencer, J. Probabilistic Methods in Combinatorics. Academic Press, New York (1974). Feller, W. An Introduction to Probability Theory and Its Applications, Volume I. John Wiley & Sons, Inc., New York (1968). Feller, W. An Introduction to Probability Theory and Its Applications, Volume II. John Wiley & Sons, Inc., New York (1966). Kolmogorov, A. N. Foundations of the Theory of Probability. Chelsea Publishing Company, New York (1956). Valiant, L. G. A theory of the learnable. Communications of the ACM 27 (1984), 11341142.

You might also like