INEQUALITY PROOF

INEQUALITY PROOF

Maths

In Loving Memory of Mamba Day, 4.13, 2016

I. Common Inequalities

(a). Lagrangian Formula for Summation Estimation

Suppose f(x) is a continuous function on x >= 1 and F(x) is its primitive function. If f(x) is monotonically decreasing,

The inequality is reversed if f(x) is monotonically increasing.

Proof:

Suppose f(x) is monotonically increasing. F’(x) = f(x). By applying Lagrange mean value theorem to F(x) on [k, k+1], we have

Note that

By summing up these inequalities above, we obtain,

Therefore,

Similarly, we can prove the other case in which f(x) is monotonically decreasing.

(b). The r-order Inequality

The r-order mean M_r(a, q) is defined as

where a = (a₁, …, a_n)^T > 0 , q = (q₁, …, q_n)^T and q₁+…+q_n = 1.

 

Theorem:M_r(a, q) is a monotonically increasing function on R with respect tor.

Proof:

Let b_i =a_i^r,i= 1, …,n,

where B is a random variable with a generalized Bernoulli distribution,

Note that f(x) = xlnx is a convex function. Thus, the expectation of a function value is larger than the function value of the expectation (See Appendix I), i.e.,

This completes the proof.#

Corollary I:

When q_i=1/n, we have,

Arithmetic mean (r= 1),

Geometric mean (r = 0),

Harmonic mean (r = -1),

H(a)<=G(a)<=A(a).

 

Corollary II (Holder Inequality)

Suppose we have m vectors with n components,

and w₁+w₂+…+w_m= 1, w_i >0, then

A relaxation on its left hand side and some changes of variables might give us more insight on the Holder inequality, i.e.,

letting r_ij^wi =b_ij, and p_i = 1/w_i,

By defining a generalized “inner product” among several vectors,

we obtain a relaxed Holder inequality,

where p_i’s are called conjugate coefficients satisfying

 

Proof:

This completes the proof.#

II. Integral Inequalities

(a). The r-order inequality

In this section, we are going to consider various means of a positive-valued continuous function f(x) on [a, b]. The interval [a, b] can be evenly divided into n sub-intervals with a=x₀<x₁<…<x_n=b, where x_k=a+k(b-a)/n, k=0, 1, …,n.

Arithmetic mean of f(x) on [a, b],

Geometric mean of f(x) on [a,b],

Harmonic mean of f(x) on [a,b],

In generalization, the r-order mean of f(x) on [a,b] is given by

Note that A(f)=M₁(f),G(f)=M₀(f) and H(f)=M_-1(f). One can also show that M_r(f) is a monotonically increasing function wrt r. In particular, H(f) <= G(f) <= A(f).

(b) Common Methods to Prove Integral Inequalities

(b1) Holder Inequality

Given m positive-valued continuous functions on [a,b], f₁(x), f₂(x), …,f_m(x), and positive scalars w₁, …,w_m satisfying w₁+…+w_m=1, then

Similarly, we have a relaxed Holder inequality

where the generalized inner product among functions is given by

and p_i’s are called conjugate coefficients satisfying

Particularly, for a “two functions and w₁=w₂=1/2” case,

or equivalently,

We obtain the Cauchy-Schwarz Inequality (See Appendix II).

(b2) Function Convexity

Definition (Convex function): f(x) is a convex function if and only if

for

Particularly, when alpha_i=1/n, we obtain the Jensen Inequality.

One can show that

If f(x) is differentiable, f(x) is convex <=> f’(x) monotonically increases;

If f(x) is twice differentiable, f(x) is convex <=> f’’(x)< 0.

Corollary: Suppose f(x) is an integrable function on [a, b], m<=f(x)<=M, g(x) is a continuous convex function on [m, M], then

Hint: This can be proved by taking a limit on Jensen Inequality.

 

APPENDIX

I. A Property of Convex Function

Prove: for a convex twice differentiable real-valued function f(x) defined on [a, b], the expectation of f(x) is equal to or larger than the function value of the expectation.

Actually this property directly follows the definition of a convex function by regarding alpha’s as probabilities or probability density. Here is another proof when the convex function is twice differentiable.

Proof:

Since the second derivative of f is nonnegative, its first derivative must be nondecreasing. Using the fundamental theorem of calculus, we obtain

Now suppose a random variable X can take every possible value x. Choosing c = E[X] and taking expectations on both sides, we have

This completes the proof.#

 

II. Cauchy-Schwarz Inequality

(1) Generic Cauchy-Schwarz Inequality

Recall that, in functional analysis,

In an inner-product space (H, K,<.,.>), for any x and y in X, the generic Cauchy-Schwarz inequality holds as

with the equality iff x and y are linearly dependent.

(2) Variants of Cauchy-Schwarz Inequality

There are myriads variants of the Cauchy-Schwarz inequality.

(a) The inner-product space of m-by-n matrices

Suppose H = {m-by-n complex matrices}, K = C, and the inner product is defined as <A, B> = trace(A*B) (One can check the validity of this inner product in (H, K, <.,.>) by examining the definition of the general inner product). Then, the Cauchy-Schwarz Inequality holds as

(b) The inner-product space of random variables

Suppose H = {random variables}, K = R, and the inner product is defined as <X,Y> = Cov(X, Y) (One can check the validity of this inner product in (H, K, <.,.>) by examining the definition of the general inner product). Then, the Cauchy-Schwarz inequality holds as

(c) The inner-product space of random matrices

Suppose A andB are p-by-n and q-by-n random matrices such that E||A||²< infinity, E||B||² < infinity and E(AA^T) is nonsingular. Then, (1)

In particular, when n=1, A:=X-EX and B:=Y-EY, we have (2)

Proof:

Let LAMBDA = E(BA^T)E^-1(AA^T). Stylistically, LAMBDA is a linear operator sequentially performing a whitening transform (eliminating the linear independence ofA’s components) and a projection onto the space of B. Then,

This completes the proof.#

Pictorially, Figure 1 shows the transformation of (2).

Philosophically, Formula (1) and (2) are both high-dimensional Cauchy-Schwarz Inequality, regarding a transformation between linear spaces and one between vectors, respectively.