The Bernstein Convergence Theorem. Binomial Series

Let f ( x ) be a function defined for a x b ; suppose also that for every integer n 0 , f ( n ) ( x ) exists and f ( n ) ( x ) 0 for a x b . Let x 1 be some point of the interval and choose some fixed positive integer n . Then consider Taylor's formula for f ( x ) about x 1 , where the polynomial of the formula has more than n terms, say m terms. In particular, consider what this gives us for f ( b ) , when we use the Cauchy remainder:

\begin{aligned} f(b) = f(x_1) &+ \frac{(b-x_1)}{1!} f'(x_1) + \dots + \frac{(b-x_1)^n}{n!} f^{(n)}(x_1) + \dots \\ &+ \frac{(b-x_1)^{m-1}}{(m-1)!} f^{(m-1)}(x_1) + \frac{(b-x_1)(b-\xi)^{m-1}}{(m-1)!} f^{(m)}(\xi), \end{aligned}

where x 1 < ξ < b . Now all terms of the sum on the right are non-negative; therefore f ( b ) is greater than or equal to any of them. In particular,

\frac{(b-x_1)^n}{n!} f^{(n)}(x_1) \le f(b), \tag{1}

and this is true for any n .

Now consider the function f''(x). Since the derivatives of f''(x) are also derivatives of f ( x ) , they all are non-negative in a x b . Therefore we can use f''(x) in place of f ( x ) in (1), obtaining

\frac{(b-x_1)^n}{n!} f^{(n+2)}(x_1) \le f''(b), \quad \text{or}f^{(n+2)}(x_1) \le \frac{n!}{(b-x_1)^n} f''(b) \cdot \tag{2}

This last inequality also holds for any n ; in particular, it holds for n 2 , where n is as in (1). Then it becomes

f^{(n)}(x_1) \le \frac{(n-2)!}{(b-x_1)^{n-2}} f''(b) \cdot \tag{3}

This is true for any point x 1 of the interval.

Now let x 0 be some point of the interval, and consider the Taylor expansion about x 0 for f ( x ) , where x is also in the interval. The Cauchy remainder after n terms has absolute value

| x x 0 | | x ξ | n 1 ( n 1 ) ! f ( n ) ( ξ ) ,

where ξ is between x and x 0 . By (3), this is less than or equal to

\begin{aligned} &\frac{|x-x_0|\ |x-\xi|^{n-1}}{(n-1)!} \cdot \frac{(n-2)!}{(b-\xi)^{n-2}} f''(b) \\ &= \frac{|x-x_0|\ |x-\xi| \cdot f''(b)}{n-1} \left( \frac{|x-\xi|}{b-\xi} \right)^{n-2}. \end{aligned}

This will approach 0 as n becomes large if | x ξ | b ξ . If x 0 < x , then | x ξ | = x ξ b ξ . If x < x 0 , and if | x x 0 | b x 0 , then | x ξ | | x x 0 | b x 0 b ξ . Therefore we know that the remainder approaches zero whenever | x x 0 | b x 0 . Let us summarize the result we have proved, which is the Bernstein convergence theorem.

Theorem 1

Let f ( x ) be such that f(x) \ge 0, f'(x) \ge 0, f''(x) \ge 0, \dots for all x such that a x b . Let a < x 0 < b , and let x be a point of the interval such that | x x 0 | b x 0 . Then the Taylor's series

f(x_0) + \frac{(x-x_0)}{1!} f'(x_0) + \frac{(x-x_0)^2}{2!} f''(x_0) + \cdots

converges to f ( x ) .

For instance, this theorem tells us that the Taylor's series for e x formed about any point at all converges to e x for all x . The result can be extended to a considerably larger class of functions if we make a few observations. First suppose there is some number M such that f(x) \ge M, f'(x) \ge M, f''(x) \ge M, \dots whenever a x b . If M 0 , then the Taylor's series of f ( x ) converges to f ( x ) by the Bernstein theorem. If M < 0 , let

g ( x ) = f ( x ) M e ( x a )

Then g ( x ) M M e x a = M ( 1 e x a ) 0 , since both factors are less than or equal to zero. Likewise, g'(x) = f'(x) - M e^{x-a} \ge 0, g''(x) \ge 0, \dots. Therefore the Taylor's series for g ( x ) converges to g ( x ) . As we have remarked, the Taylor's series for e x a converges to e x a ( = e a e x ) . Since the operation of forming the Taylor's series is a linear one, the Taylor's series for f ( x ) is that for g ( x ) plus M times that for e x a ; the remainder in each of these goes to zero, and therefore the remainder for f ( x ) , which is the sum of these remainders, goes to zero. Thus we see that the Taylor's series for f ( x ) converges to f ( x ) if all derivatives of f ( x ) are bounded from below.

Next suppose f ( n ) ( x ) 0 for a x b , whenever n > N for some fixed N . Then let M be the smallest of the numbers

0, \min f(x), \min f'(x), \dots, \min f^{(N)}(x) \quad \text{on } a \le x \le b \cdot

Then f ( i ) ( x ) M for all i 0 , and we see as above that the Taylor's series for f ( x ) converges to f ( x ) . Since the Taylor's series for f ( x ) has as terms the negatives of those in the Taylor's series for f ( x ) , we can use f ( x ) to prove that the Taylor's series for f ( x ) converges to f ( x ) whenever all f ( n ) ( x ) are bounded from above, in particular if all f ( n ) ( x ) 0 .

Now let f(x) \ge 0, f'(x) \le 0, f''(x) \ge 0, f'''(x) \le 0, \dots; in other words, the derivatives of f ( x ) alternate in sign. Let a < x 0 < b . Then b < x 0 < a , and if b ξ a , let g ( ξ ) = f ( ξ ) . Then

g(\xi) \ge 0, g'(\xi) = -f'(-\xi) \ge 0, g''(\xi) \ge 0, \dots ,

or g ( n ) ( ξ ) 0 for all n whenever b ξ a . Thus we know by the Bernstein theorem that the Taylor's series for g ( ξ ) about x 0 converges to g ( ξ ) whenever | ξ ( x 0 ) | a ( x 0 ) , or whenever | ξ + x 0 | x 0 a . The Taylor's series for g ( ξ ) = f ( ξ ) about x 0 is:

\begin{aligned} g(-x_0) + \frac{(\xi+x_0)}{1!} g'(-x_0) + \frac{(\xi+x_0)^2}{2!} g''(-x_0) + \dots \\ = f(x_0) - \frac{(\xi+x_0)}{1!} f'(x_0) + \frac{(\xi+x_0)^2}{2!} f''(x_0) - \dots \cdot \end{aligned}

Now let x = ξ , ξ = x . Then whenever | x + x 0 | x 0 a , or | x x 0 | x 0 a , we have, in the limit,

\begin{aligned} f(x) = g(\xi) &= f(x_0) - \frac{(-x+x_0)}{1!} f'(x_0) + \frac{(-x+x_0)^2}{2!} f''(x_0) - \dots \\ &= f(x_0) + \frac{(x-x_0)}{1!} f'(x_0) + \frac{(x-x_0)^2}{2!} f''(x_0) + \dots \cdot \end{aligned}

But this is exactly the Taylor's series for f ( x ) . Note that in this case it converges to f ( x ) whenever x is closer to x 0 than x 0 is to a , while in the original theorem we had convergence to f ( x ) whenever x was closer to x 0 than x 0 was to b . Observe also that as before it is enough that the signs alternate from some N on.

We shall apply this last observation to investigate the function f ( x ) = x α for x > 0 , α any number. Then

f'(x) = \alpha x^{\alpha-1},\quad f''(x) = \alpha(\alpha - 1)x^{\alpha-2},\quad \dots ,

and eventually

f ( N ) ( x ) = α ( α 1 ) ( α N + 1 ) x α N , where  α < N

Then f ( N + 1 ) ( x ) = α ( α 1 ) ( α N + 1 ) ( α N ) x α N 1 has the opposite sign from f ( N ) ( x ) for all x > 0 , since α N < 0 , f ( N + 2 ) ( x ) has the opposite sign from f ( N + 1 ) ( x ) , and the signs of f ( n ) ( x ) alternate for n > N . Therefore if ϵ is any small positive number and x 0 > ϵ , the Taylor's series for f ( x ) about x 0 converges to f ( x ) provided | x x 0 | x 0 ϵ . If | x x 0 | < x 0 , then we are sure that | x x 0 | x 0 ϵ for some positive ϵ . Therefore the series converges to f ( x ) provided | x x 0 | < x 0 .

Now let us change notation a little and write x 0 = a , x = a + b . Then we expand f ( x ) = ( a + b ) α in powers of ( x x 0 ) = b . (Here a > 0 .) Thus

\begin{aligned} x^\alpha = (a+b)^\alpha = a^\alpha + \alpha a^{\alpha-1}b + \frac{\alpha(\alpha-1)}{2} a^{\alpha-2}b^2 + \cdots\\ + \frac{\alpha(\alpha-1) \dots (\alpha-n)}{n!} a^{\alpha-n}b^n + \cdots \end{aligned}

in the sense that the series converges to ( a + b ) α whenever | ( a + b ) a | < a , or whenever | b | < a . This is the so-called binomial series, and the coefficients

α ( α 1 ) ( α n ) n ! = ( α n )

are called the binomial coefficients. Observe that if α is a positive integer, all the coefficients are 0 from some point on, and we have the ordinary binomial theorem.

NOTE: In early investigations of Taylor's series, people often simply assumed that the series approached the function whenever the series approached any kind of limit. That this is false is seen by considering the function

f(x) = \left\{\begin{aligned} &e^{-\frac{1}{x^2}}, && x \neq 0 \\ &0, && x = 0 \end{aligned}\right\}\cdot

Then f(0) = 0 = f'(0) = f''(0) = \dots = 0, so the Taylor's series about x 0 = 0 is simply a series of 0's, having limit 0. However, the function is never 0 except at x = 0 .

EXERCISES

Exercise 1.
  1. Apply the Bernstein theorem to show that the Taylor's series for sin x and cos x about any x 0 converge to the respective functions for all x .
Exercise 2.
  1. Apply the theorem to show that the Taylor's series for log x about x 0 = 1 converges to log x if | x 1 | < 1 .
Exercise 3.
  1. Use a = 32 , b = 1 in the binomial series to estimate 33 5 to four decimal places.
Exercise 4.
  1. Use a = 100 , b = 2 to estimate 98 . Since this is also 2 7 , obtain an estimate of 2 to five places.