bernstein polynomial

For the Bernstein polynomial in D-module theory, see Bernstein-Sato polynomial.

In the mathematical subfield of numerical analysis, a Bernstein polynomial, named after Sergei Natanovich Bernstein, is a polynomial in the Bernstein form, that is a linear combination of Bernstein basis polynomials. A numerically stable way to evaluate polynomials in Bernstein form is de Casteljau's algorithm. Polynomials in Bernstein form were first used by Bernstein in a constructive proof for the Stone-Weierstrass approximation theorem. With the advent of computer graphics, Bernstein polynomials, restricted to the interval 0,1, became important in the form of Bzier curves.

Definition

The n + 1 Bernstein basis polynomials of degree n are defined as

b_{\nu,n}(x) = {n \choose \nu} x^{\nu} (1-x)^{n-\nu}, \qquad \nu=0,\ldots,n.

The Bernstein basis polynomials of degree n form a basis for the vector space

\Pi_n

of polynomials of degree n. A linear combination of Bernstein basis polynomials

B(x) = \sum_{\nu=0}^{n} \beta_{\nu} b_{\nu,n}(x)

is called a Bernstein polynomial or polynomial in Bernstein form of degree n. The coefficients β_ν are called Bernstein coefficients or Bzier coefficients.

Notes

The Bernstein basis polynomials have the following properties:

b_ν,n(x) has a root with multiplicity ν at point x = 0
b_ν,n(x) has a root with multiplicity n − ν at point x = 1
b_ν,n(x) ≥ 0 if x in 0,1
b_ν,n(x) has a global maximum at x = ν/n
b’_ν,n(x) = n - b_ν,n-1(x)
b_ν,n(x) = 0, if ν < 0 or ν > n

The Bernstein basis polynomials of degree n form a partition of unity:

\sum_{\nu=0}^n b_{\nu,n}(x) = \sum_{\nu=0}^n {n \choose \nu} x^{\nu}(1-x)^{n-\nu} = (x+(1-x))^n = 1.

Example

The first few Bernstein basis polynomials are

b_{0,0}(x) = 1\,

b_{0,1}(x) = 1-x \mbox{ , } b_{1,1}(x) = x\,

b_{0,2}(x) = (1-x)^2 \mbox{ , } b_{1,2}(x) = 2x(1-x) \mbox{ , } b_{2,2}(x) = x^2\,

Approximating continuous functions

Let f(x) be a continuous function on the interval 1. Consider the Bernstein polynomial

B_n(f, x) = \sum_{\nu=0}^{n} f\left(\frac{\nu}{n}\right) b_{\nu,n}(x).

It can be shown that

\lim_{n\rightarrow\infty} B_n(f,x)=f(x)

uniformly on the interval 1. This is a stronger statement than the proposition that the limit holds for each value of x separately; that would be pointwise convergence rather than uniform convergence. Specifically, the word uniformly signifies that

\lim_{n\rightarrow\infty}\sup\{\,\left|f(x)-B_n(f,x)\right|:0\leq x\leq 1\,\}=0.

Bernstein polynomials thus afford one way to prove the Stone-Weierstrass approximation theorem that every real-valued continuous function on a real interval a,b can be uniformly approximated by polynomial functions over R.

Proof

Suppose K is a random variable distributed as the number of successes in n independent Bernoulli trials with probability x of success on each trial; in other words, K has a binomial distribution with parameters n and x. Then we have the expected value E(K/n) = x. Then the weak law of large numbers of probability theory tells us that

\lim_{n\rightarrow\infty}P(\left|(K/n)-x\right|>\delta)=0.

Because f, being continuous on a closed bounded interval, must be uniformly continuous on that interval, we can infer a statement of the form

\lim_{n\rightarrow\infty} P(\left|f(K/n)-f(x)\right|>\varepsilon)=0.

Consequently

\lim_{n\rightarrow\infty}

P(\left|f(K/n)-E(f(K/n))\right|+\left|E(f(K/n))-f(x)\right|>\varepsilon)=0.

\lim_{n\rightarrow\infty}

P(\left|f(K/n)-E(f(K/n))\right|>\varepsilon/2)+P( \left|E(f(K/n))-f(x)\right|>\varepsilon/2)=0. And so the second probability above approaches 0 as n grows. But the second probability is either 0 or 1, since the only thing that is random is K, and that appears within the scope of the expectation operator E. Finally, observe that E(f(K/n)) is just the Bernstein polynomial B_n(f,x).