Taylor's theorem with remainder
Calculus proved Taylor’s theorem one way, with one form of the remainder. Is there a proof that gives every form at once, how little smoothness does a Taylor polynomial need to approximate well, and what do the derivatives at a single point say about a maximum or a minimum?
Where you are 3 in Derivatives
Why this exists
The Taylor remainder and the Lagrange error bound proved Taylor’s theorem with the Lagrange remainder by applying Rolle’s theorem n + 1 times, and left three things open. Its bound cannot show that the Maclaurin series of \ln(1+x) converges to it for \displaystyle -1 < x < -\frac12, although it does. It says nothing when only f^{(n)}(a) exists. And the second derivative test (The first and second derivative tests) has a short proof from Taylor’s theorem that also covers f''(a) = 0.
This lesson proves Taylor’s theorem through Cauchy’s mean value theorem (The mean value theorem and its consequences), in a form whose free auxiliary function gives both the Lagrange and the Cauchy remainders; then the local (Peano) form, and from it the nth derivative test.
Definition
Throughout, I is an interval with more than one point, derivatives are those of Differentiability and the derivative rules, proved (one-sided at an endpoint of I), and f^{(0)} = f.
Higher derivatives. f : I \to \mathbb{R} is n times differentiable on I if f, f', \dots, f^{(n-1)} are differentiable on I; then f^{(n)} = \big(f^{(n-1)}\big)'. If also f^{(n)} is continuous on I, then f \in C^n(I). If this holds for every n, f is smooth, f \in C^{\infty}(I). For n \ge 1 and a \in I, f is n times differentiable at a if f^{(n-1)} exists on all of I and is differentiable at a.
Taylor polynomial and remainder. For such f (or for n = 0), as in Taylor polynomials and The Taylor remainder and the Lagrange error bound,
and R_n(x) = f(x) - P_n(x).
Strict relative extrema. f has a strict relative minimum at a \in I if some \delta > 0 has f(x) > f(a) for every x \in I with 0 < \lvert x - a \rvert < \delta; a strict relative maximum if instead f(x) < f(a) for all such x.
Analytic functions. f \in C^{\infty}(I) is analytic at a if for some r > 0 its Taylor series at a converges to f(x) for every x \in I with \lvert x - a \rvert < r.
Intuition
The proof in The Taylor remainder and the Lagrange error bound fixes the center a. The proof here fixes x and moves the center: let F(t) be the value at x of the Taylor polynomial centered at t. Then F(a) = P_n(x) and F(x) = f(x), so the remainder is the total change of F, and a mean value theorem turns it into one derivative, which telescopes to a single term. Cauchy’s theorem measures the change against any second function G; each choice of G is a form of the remainder.
The Peano form says not how big R_n(x) is, but that it vanishes faster than (x - a)^n as x \to a. So near a, f(x) - f(a) behaves like its first nonzero Taylor term, whose sign decides a maximum or a minimum.
Results
Theorem 1 (Taylor’s theorem, general remainder). Let n \ge 0, let a \ne x be points of I, and let J be the closed interval with endpoints a and x. Suppose f is n times differentiable on J with f^{(n)} continuous on J, and f^{(n+1)} exists at every point strictly between a and x. Let G : J \to \mathbb{R} be continuous, and differentiable with G' \ne 0 at every point strictly between a and x. Then some c strictly between a and x has
Theorem 2 (Lagrange and Cauchy forms). Under the hypotheses on f in Theorem 1, there are c and c' strictly between a and x with
The first is the Lagrange form, the second the Cauchy form. For n = 0 both are the mean value theorem.
Theorem 3 (local, or Peano, form). Let n \ge 1, a \in I, and f be n times differentiable at a. Then \displaystyle \displaystyle\lim_{x \to a} \frac{R_n(x)}{(x - a)^n} = 0.
Theorem 4 (uniqueness). Let n \ge 1, let f be n times differentiable at a, and let Q be a polynomial of degree at most n with \displaystyle \frac{f(x) - Q(x)}{(x - a)^n} \to 0 as x \to a. Then Q = P_n.
Theorem 5 (nth derivative test). Let n \ge 2, let a be an interior point of I, and let f be n times differentiable at a with f^{(k)}(a) = 0 for 1 \le k \le n - 1 and f^{(n)}(a) \ne 0. (a) If n is even, f has a strict relative minimum at a when f^{(n)}(a) > 0 and a strict relative maximum when f^{(n)}(a) < 0. (b) If n is odd, every interval around a holds points where f > f(a) and points where f < f(a), so a is no relative extremum.
The case n = 2 is the second derivative test. It needs f' on I and f'' at the one point a.
The integral form, stated without proof: if f^{(n+1)} is continuous on J, then \displaystyle R_n(x) = \frac{1}{n!}\int_a^x f^{(n+1)}(t)(x - t)^n\,dt. It follows by integrating by parts n times once the fundamental theorem is proved in The fundamental theorem of calculus, proved.
Why it is true
Theorem 1. Proof. For t \in J let
Then F(a) = P_n(x), and F(x) = f(x) since each term with k \ge 1 has the factor x - x. So F(x) - F(a) = R_n(x). Each f^{(k)} with k < n is differentiable on J, hence continuous (Theorem 3 of Differentiability and the derivative rules, proved), and f^{(n)} is continuous by hypothesis, so F is continuous on J. Strictly between a and x, F is differentiable, and for k \ge 1 the product rule gives
while the term k = 0 has derivative f'(t). In the sum, the second part of term k cancels the first part of term k - 1, leaving
Cauchy’s mean value theorem (Theorem 3 of The mean value theorem and its consequences) for F and G on J gives c strictly between a and x with \big(F(x) - F(a)\big)G'(c) = \big(G(x) - G(a)\big)F'(c); swapping the endpoints changes the sign of both differences, so this holds whether a < x or x < a. Substitute and divide by G'(c) \ne 0. \square
Theorem 2. Proof. For the Lagrange form take G(t) = (x - t)^{n+1}, so G'(t) = -(n+1)(x - t)^n \ne 0 strictly between a and x, and G(x) - G(a) = -(x - a)^{n+1}. Theorem 1 gives
which is the Lagrange form. For the Cauchy form take G(t) = t. \square Unlike the statement in The Taylor remainder and the Lagrange error bound, this one asks nothing of f^{(n+1)} at a and x.
Theorem 3. Proof. By induction on n, for all intervals and functions at once. For n = 1, \displaystyle \frac{R_1(x)}{x - a} = \frac{f(x) - f(a)}{x - a} - f'(a) \to 0 by the definition of f'(a). Let n \ge 2 and assume the theorem for n - 1. Put g = f'; then g^{(n-2)} = f^{(n-1)} exists on I and is differentiable at a, so g is n - 1 times differentiable at a. Differentiating term by term,
the Taylor polynomial of degree n - 1 of g at a. So R_n' = g - P_n' is the remainder of g, and the induction hypothesis gives, for \varepsilon > 0, a \delta > 0 with \lvert R_n'(t) \rvert < \varepsilon \lvert t - a \rvert^{n-1} for every t \in I with 0 < \lvert t - a \rvert < \delta. Now R_n is differentiable on I and R_n(a) = 0. For x \in I with 0 < \lvert x - a \rvert < \delta, the mean value theorem gives c strictly between a and x with R_n(x) = R_n'(c)(x - a), and as 0 < \lvert c - a \rvert < \lvert x - a \rvert,
\square
Theorem 4. Proof. Let D = P_n - Q. Substituting x = (x - a) + a and expanding, D(x) = \sum_{k=0}^{n} b_k (x - a)^k. Since D = (f - Q) - (f - P_n), the hypothesis and Theorem 3 give \displaystyle \frac{D(x)}{(x - a)^n} \to 0. Suppose D \ne 0 and let m be the least k with b_k \ne 0. Then \displaystyle \frac{D(x)}{(x - a)^m} = b_m + \sum_{k=m+1}^{n} b_k (x - a)^{k-m} \to b_m, but also \displaystyle \frac{D(x)}{(x - a)^m} = (x - a)^{n-m}\frac{D(x)}{(x - a)^n} \to 0. Uniqueness of limits (Theorem 1 of Limits of functions, epsilon and delta) gives b_m = 0, a contradiction. \square So a Taylor polynomial may be found by any route that produces a polynomial with this property (Example 3).
Theorem 5. Proof. Let \displaystyle b = \frac{f^{(n)}(a)}{n!} \ne 0. By hypothesis P_n(x) = f(a) + b(x - a)^n, so for x \ne a
By Theorem 3, and as a is interior, some \delta > 0 has (a - \delta, a + \delta) \subseteq I and \displaystyle \left\lvert \frac{R_n(x)}{(x - a)^n} \right\rvert < \lvert b \rvert for 0 < \lvert x - a \rvert < \delta, so there the bracket has the sign of b. (a) For even n, (x - a)^n > 0, so f(x) - f(a) has the sign of f^{(n)}(a). (b) For odd n, (x - a)^n has the sign of x - a, so f(x) - f(a) has one sign on (a, a + \delta) and the other on (a - \delta, a). \square
Worked examples
Two forms, two points
For f(x) = x^4 with a = 1, n = 2 and x = 3, find the point c of the Lagrange form and the point c' of the Cauchy form.
Solutionf'(1) = 4 and f''(1) = 12, so P_2(x) = 1 + 4(x - 1) + 6(x - 1)^2, P_2(3) = 33 and R_2(3) = 81 - 33 = 48. Also f'''(t) = 24t.
Lagrange: \displaystyle \frac{24c}{6} \cdot 2^3 = 32c = 48, so \displaystyle c = \frac{3}{2}.
Cauchy: \displaystyle \frac{24c'}{2}(3 - c')^2 \cdot 2 = 48, that is c'^3 - 6c'^2 + 9c' - 2 = 0, or (c' - 2)(c'^2 - 4c' + 1) = 0. The roots \displaystyle 2 \pm \sqrt{3} lie outside (1, 3), so c' = 2, not c.
Zoe’s headlamp reflector
At a headlamp plant, Zoe cuts a reflector from a sphere of radius R = 8 inches, with half-width x = 2 inches across the opening. Its depth is \displaystyle s = R - \sqrt{R^2 - x^2}, while Isaac’s drawing uses the parabola’s depth \displaystyle \frac{x^2}{2R}. Show that for 0 < x < R, with \displaystyle v = \frac{x^2}{R^2},
and evaluate both bounds for this reflector.
SolutionWrite s = R\,g(v) with g(v) = 1 - (1 - v)^{1/2}, smooth on (-\infty, 1), and 0 < v < 1. Here \displaystyle g'(v) = \frac12(1 - v)^{-1/2}, \displaystyle g''(v) = \frac14(1 - v)^{-3/2} and \displaystyle g'''(v) = \frac38(1 - v)^{-5/2} > 0. The Lagrange form with n = 2 gives \displaystyle g(v) = \frac{v}{2} + \frac{v^2}{8} + \frac{g'''(c)}{6}v^3 > \frac{v}{2} + \frac{v^2}{8}. With n = 1 it gives \displaystyle g(v) = \frac{v}{2} + \frac{v^2}{8(1 - c)^{3/2}} for some c \in (0, v), and (1 - c)^{-3/2} < (1 - v)^{-3/2}. Multiply by R, using \displaystyle R\frac{v}{2} = \frac{x^2}{2R} and \displaystyle R\frac{v^2}{8} = \frac{x^4}{8R^3}.
Here \displaystyle v = \frac{1}{16} and \displaystyle \frac{x^2}{2R} = 0.25 inch, so the gap lies between \displaystyle \frac{1}{256} \approx 0.00391 and \displaystyle \frac{1}{256}\left(\frac{16}{15}\right)^{3/2} \approx 0.00430 inch. Directly, \displaystyle s = 8 - \sqrt{60} \approx 0.25403, a gap of about 0.00403 inch.
The nth derivative test
Classify the critical point 0 of f(x) = 6\sin x - 6x + x^3 and of h(x) = 2 - 2\cos x - x\sin x.
SolutionFor f: f' = 6\cos x - 6 + 3x^2, f'' = 6x - 6\sin x, f''' = 6 - 6\cos x, f^{(4)} = 6\sin x, f^{(5)} = 6\cos x, which at 0 are 0, 0, 0, 0, 6. The first nonzero one has odd order 5, so by Theorem 5(b), 0 is not an extremum.
For h, use Theorem 4. By Theorem 3, \displaystyle \sin x = x - \frac{x^3}{6} + R_3(x) with \displaystyle \frac{R_3(x)}{x^3} \to 0, so \displaystyle x\sin x - \left(x^2 - \frac{x^4}{6}\right) = xR_3(x) is small against x^4, and Theorem 4 makes \displaystyle x^2 - \frac{x^4}{6} the Taylor polynomial of degree 4 of x\sin x. That of 2 - 2\cos x is \displaystyle x^2 - \frac{x^4}{12}, and coefficients subtract, so h has \displaystyle P_4(x) = \frac{x^4}{12}. Hence h'(0) = h''(0) = h'''(0) = 0 and \displaystyle h^{(4)}(0) = \frac{4!}{12} = 2 > 0: a strict relative minimum, by Theorem 5(a).
The second derivative test when f’’ is not continuous
Let \displaystyle f(x) = x^2 + 3x^4\sin\frac{1}{x} for x \ne 0 and f(0) = 0. Show that f has a strict relative minimum at 0, although f'' takes negative values in every interval around 0.
SolutionFor x \ne 0, \displaystyle f'(x) = 2x + 12x^3\sin\frac1x - 3x^2\cos\frac1x. At 0, \displaystyle \frac{f(h)}{h} = h + 3h^3\sin\frac1h \to 0 by squeezing with \displaystyle \lvert 3h^3\sin\frac1h \rvert \le 3\lvert h \rvert^3, so f'(0) = 0. Next,
by the same squeeze, so f''(0) = 2 > 0, and Theorem 5 with n = 2 gives a strict relative minimum. For x \ne 0,
At \displaystyle x_k = \frac{2}{(4k + 1)\pi}, k \ge 1, the sine is 1 and the cosine 0, so f''(x_k) = 36x_k^2 - 1 < 0, as \displaystyle x_k \le \frac{2}{5\pi} < \frac16. A proof through the Lagrange form needs f''(c) > 0 for all c near 0, which fails; Theorem 3 needs only f''(0).
The logarithm series on the far left
Prove that the Maclaurin series of \ln(1+x) converges to \ln(1+x) for every x \in (-1, 0), and find how many terms the proof guarantees for \ln 0.4 within 10^{-3}.
Solutionf(t) = \ln(1+t) is smooth on (-1, \infty) with \displaystyle f^{(n+1)}(t) = \frac{(-1)^n n!}{(1+t)^{n+1}}. Fix x \in (-1, 0) and let r = -x. The Cauchy form gives c \in (x, 0) with
Now 1 + c > 1 - r, and \displaystyle r - \frac{c + r}{1 + c} = \frac{c(r - 1)}{1 + c} > 0 because c < 0 and r < 1. So \displaystyle 0 < \frac{c + r}{1 + c} < r and
By the criterion of The Taylor remainder and the Lagrange error bound, the series converges to \ln(1+x). The Lagrange bound \displaystyle \frac{1}{n+1}\left(\frac{r}{1 - r}\right)^{n+1} grows without limit for \displaystyle r > \frac12, because it must allow c near x, where f^{(n+1)} is huge.
For \ln 0.4, x = -0.6 and the bound \displaystyle \frac{0.6^{n+1}}{0.4} is about 0.00118 at n = 14 and 0.00071 at n = 15. So P_{15}(-0.6) is guaranteed; its actual error is about 4.1 \times 10^{-5}.
Three functions the derivatives cannot tell apart
Let \psi(x) = e^{-1/x^2} for x \ne 0 and \psi(0) = 0, and let h(x) = \psi(x) for x \ge 0 and h(x) = -\psi(x) for x < 0. Prove that h is smooth with h^{(k)}(0) = 0 for every k, and compare \psi, -\psi and h at 0.
SolutionExample 7 of The Taylor remainder and the Lagrange error bound proved that \psi is smooth with every \psi^{(k)}(0) = 0, and that \displaystyle \frac{\psi^{(k)}(t)}{t} \to 0 as t \to 0.
Proof. By induction on k: h^{(k)} exists on \mathbb{R}, equals \operatorname{sgn}(x)\,\psi^{(k)}(x) for x \ne 0, and h^{(k)}(0) = 0. For k = 0 this is the definition. Assume it for k. Near x \ne 0, h^{(k)} equals \pm\psi^{(k)} on an open interval missing 0, so h^{(k+1)}(x) = \operatorname{sgn}(x)\,\psi^{(k+1)}(x) (Theorem 6 of Differentiability and the derivative rules, proved). At 0, \displaystyle \left\lvert \frac{h^{(k)}(t) - h^{(k)}(0)}{t} \right\rvert = \left\lvert \frac{\psi^{(k)}(t)}{t} \right\rvert \to 0, so h^{(k+1)}(0) = 0. \square
So all three are smooth with zero Taylor series at 0, none is analytic there, and yet 0 is a strict minimum of \psi, a strict maximum of -\psi, and no extremum of h, which has the sign of x (Figure 2). No test reading only the derivatives at 0 can classify 0.
Boundaries
- Theorem 2 asks for f^{(n)} continuous on J. For n = 0 this cannot be dropped: f = 0 on [0, 1), f(1) = 1 has f' = 0 on (0, 1), yet R_0(1) = 1. In Example 4, f'' is discontinuous at 0, so Theorem 2 with n = 2 does not apply on [0, x]; Theorem 3 does.
- The Peano form is a limit, not a bound. A number for R_n(x) needs Theorem 2 and a bound on f^{(n+1)}.
- The derivative tests can be silent. If every derivative at a is 0, nothing read from them decides (Example 6). If f is not n times differentiable at a for the needed n, Theorem 5 cannot start, though a may still be an extremum (Exercise 9).
- Smooth is not analytic (Example 6, Exercise 10), and no form of the remainder wins everywhere: the Cauchy form settles \ln(1+x) on \displaystyle \left(-1, -\frac12\right), while in Example 2 the Lagrange form gives both bounds.
Where this goes next
The fundamental theorem of calculus, proved supplies the fundamental theorem behind the integral form of the remainder. Power series, proved differentiates a power series term by term, so it is the Taylor series of its own sum at the center. Pointwise and uniform convergence asks when P_n \to f uniformly on an interval.
Exercises
The two forms of the remainder
- For f(x) = x^3 with a = 0, n = 1 and x = 2, find every point of the Lagrange form and every point of the Cauchy form.
- For f(x) = \dfrac{1}{1 - x} with a = 0, n = 1 and \displaystyle x = \frac12, find the point of each form.
- Using the bound of Example 5, find the least n for which \displaystyle P_n\left(-\frac34\right) is guaranteed within 10^{-2} of \displaystyle \ln\frac14. What does the Lagrange bound give at \displaystyle x = -\frac34?
- Zoe’s second reflector has R = 5 inches and half-width x = 1.2 inches. Bound the gap \displaystyle s - \frac{x^2}{2R} as in Example 2, and compare with its actual value.
Derivative tests and limits
- Classify the critical point 0 of g(x) = 24\cos x - 24 + 12x^2 - x^4.
- Find and classify the critical points of k(x) = x^4 e^{-x}.
- For f(x) = \cos(x^2), find f^{(4)}(0), f^{(6)}(0) and f^{(8)}(0) using Theorem 4, and classify 0.
- Use Theorem 3 to find \displaystyle \displaystyle\lim_{x \to 0} \frac{x - \sin x}{x(1 - \cos x)}.
Where the tests stop
- Let \displaystyle f(x) = x^4\left(2 + \sin\frac{1}{x}\right) for x \ne 0 and f(0) = 0. Show that f'(0) = f''(0) = 0 and that f'''(0) does not exist, so Theorem 5 cannot be applied, yet 0 is a strict relative minimum.
- Let \varphi(t) = e^{-1/t} for t > 0 and \varphi(t) = 0 for t \le 0, and take as known that \varphi is smooth with every \varphi^{(j)}(0) = 0 (the argument of Example 7 of The Taylor remainder and the Lagrange error bound adapts). Prove that \beta(x) = \varphi(1 - x^2) is smooth on \mathbb{R}, positive on (-1, 1), zero outside it, and not analytic at 1.
Proofs
- Let x > a, p > 0, and let f satisfy the hypotheses of Theorem 1. Using G(t) = (x - t)^p, prove that some c \in (a, x) has R_n(x) = \dfrac{f^{(n+1)}(c)}{n!\,p}(x - c)^{n+1-p}(x - a)^p, and show that p = n + 1 and p = 1 give the two forms of Theorem 2.
- Prove that if f is n + 1 times differentiable on I and f^{(n+1)} = 0 on I, then f is a polynomial of degree at most n on I. Deduce that a polynomial of degree d equals its Taylor polynomial of degree d at every center.
- Let n \ge 2, let a be an interior point of I, let f be n times differentiable at a with f^{(k)}(a) = 0 for 1 \le k \le n - 1 and f^{(n)}(a) \ne 0, and let f have a relative minimum at a. Prove that n is even and f^{(n)}(a) > 0. Deduce that if f is twice differentiable at a with a relative minimum there, then f''(a) \ge 0.
- Let n \ge 1 and f \in C^{n+1}(I) on an open interval I containing a. Prove that f^{(k)}(a) = 0 for 0 \le k \le n if and only if \displaystyle \displaystyle\lim_{x \to a} \frac{f(x)}{(x - a)^{n+1}} exists.
Answers
- f(0) = f'(0) = 0, so P_1 = 0 and R_1(2) = 8. Lagrange: \displaystyle \frac{6c}{2} \cdot 2^2 = 12c = 8, so \displaystyle c = \frac{2}{3}. Cauchy: 6c(2 - c) \cdot 2 = 8, so 3c^2 - 6c + 2 = 0 and \displaystyle c = 1 \pm \frac{1}{\sqrt{3}}, about 0.423 and 1.577, both in (0, 2). The point need not be unique.
- \displaystyle R_1\left(\frac12\right) = 2 - \frac32 = \frac12 and \displaystyle f''(t) = \frac{2}{(1 - t)^3}. Lagrange: \displaystyle \frac{1}{(1 - c)^3} \cdot \frac14 = \frac12, so c = 1 - 2^{-1/3} \approx 0.206. Cauchy: \displaystyle \frac{2}{(1 - c)^3}\left(\frac12 - c\right)\frac12 = \frac12; with u = 1 - c this is u^3 - 2u + 1 = 0, or (u - 1)(u^2 + u - 1) = 0. Only \displaystyle u = \frac{\sqrt5 - 1}{2} puts c in \displaystyle \left(0, \frac12\right), so \displaystyle c = \frac{3 - \sqrt5}{2} \approx 0.382.
- With \displaystyle r = \frac34, \displaystyle \lvert R_n \rvert < 4\left(\frac34\right)^{n+1}: about 0.0127 at n = 19 and 0.0095 at n = 20, so n = 20. The Lagrange bound takes \displaystyle M = \frac{n!}{(1/4)^{n+1}}, the largest \lvert f^{(n+1)} \rvert on \displaystyle \left[-\frac34, 0\right], and gives \displaystyle \frac{3^{n+1}}{n+1}, which grows without limit.
- v = 0.0576, \displaystyle \frac{x^2}{2R} = 0.144 and \displaystyle \frac{x^4}{8R^3} = 0.0020736, so \displaystyle 0.0020736 < s - 0.144 < \frac{0.0020736}{0.9424^{3/2}} \approx 0.002267. Directly, \displaystyle s = 5 - \sqrt{23.56} \approx 0.14614, a gap of about 0.00214.
- The derivatives of 24\cos x at 0 are 24, 0, -24, 0, 24, 0, -24; adding those of -24 + 12x^2 - x^4 gives g(0) = g'(0) = \cdots = g^{(5)}(0) = 0 and g^{(6)}(0) = -24. Even order, negative: a strict relative maximum.
- k'(x) = x^3(4 - x)e^{-x}, so the critical points are 0 and 4. k''(x) = (x^4 - 8x^3 + 12x^2)e^{-x} and k''(4) = -64e^{-4} < 0: a strict relative maximum at 4. At 0, \displaystyle \frac{k(x) - x^4}{x^4} = e^{-x} - 1 \to 0, so Theorem 4 gives P_4 = x^4: k'(0) = k''(0) = k'''(0) = 0 and k^{(4)}(0) = 24 > 0, a strict relative minimum.
- By Theorem 3, \displaystyle \cos u = 1 - \frac{u^2}{2} + \frac{u^4}{24} + S(u) with \displaystyle \frac{S(u)}{u^5} \to 0, so \displaystyle \frac{S(x^2)}{x^8} = x^2 \cdot \frac{S(x^2)}{x^{10}} \to 0 and Theorem 4 gives \displaystyle P_8(x) = 1 - \frac{x^4}{2} + \frac{x^8}{24}. Hence \displaystyle f^{(4)}(0) = -\frac{4!}{2} = -12, f^{(6)}(0) = 0 and \displaystyle f^{(8)}(0) = \frac{8!}{24} = 1680. With f', f'', f''' zero at 0, Theorem 5 with n = 4 gives a strict relative maximum.
- The limit is \displaystyle \frac13. By Theorem 3, \displaystyle x - \sin x = \frac{x^3}{6} - R(x) with \displaystyle \frac{R(x)}{x^3} \to 0, and \displaystyle 1 - \cos x = \frac{x^2}{2} - S(x) with \displaystyle \frac{S(x)}{x^2} \to 0. Dividing top and bottom by x^3, the quotient is \displaystyle \frac{1/6 - R(x)/x^3}{1/2 - S(x)/x^2} \to \frac{1/6}{1/2}.
- \displaystyle \frac{f(h)}{h} = h^3\left(2 + \sin\frac1h\right) \to 0, so f'(0) = 0. For x \ne 0, \displaystyle f'(x) = 8x^3 + 4x^3\sin\frac1x - x^2\cos\frac1x, and \displaystyle \frac{f'(h)}{h} \to 0 by squeezing, so f''(0) = 0. For x \ne 0, \displaystyle f''(x) = 24x^2 + 12x^2\sin\frac1x - 6x\cos\frac1x - \sin\frac1x. At \displaystyle x_k = \frac{1}{2k\pi + \pi/2}, f''(x_k) = 36x_k^2 - 1, so \displaystyle \frac{f''(x_k)}{x_k} \to -\infty; at \displaystyle y_k = \frac{1}{2k\pi + 3\pi/2}, f''(y_k) = 12y_k^2 + 1, so \displaystyle \frac{f''(y_k)}{y_k} \to \infty. So f'''(0) does not exist and Theorem 5 has no n to use. Still f(x) \ge x^4 > 0 = f(0) for x \ne 0.
- Proof. Let q(x) = 1 - x^2. By induction on k, \beta^{(k)}(x) = \sum_{j=0}^{k} p_{k,j}(x)\,\varphi^{(j)}(q(x)) for polynomials p_{k,j}: for k = 0, p_{0,0} = 1, and the product and chain rules (Theorems 4 and 5 of Differentiability and the derivative rules, proved) turn each term into p_{k,j}'\,\varphi^{(j)}(q) + p_{k,j}\,q'\,\varphi^{(j+1)}(q), of the same form. So \beta is smooth. As q(x) > 0 exactly when \lvert x \rvert < 1, \beta > 0 on (-1, 1) and \beta = 0 outside. Since q(1) = 0, every \beta^{(k)}(1) = \sum_j p_{k,j}(1)\varphi^{(j)}(0) = 0: the Taylor series at 1 is zero, while \beta > 0 just left of 1. So \beta is not analytic at 1. \square
- Proof. G(t) = (x - t)^p is continuous on [a, x], with G'(t) = -p(x - t)^{p-1} \ne 0 on (a, x) and G(x) - G(a) = -(x - a)^p. Theorem 1 gives c \in (a, x) with \displaystyle R_n(x) = \frac{f^{(n+1)}(c)}{n!}(x - c)^n \cdot \frac{(x - a)^p}{p(x - c)^{p-1}}, which is the formula. For p = n + 1 the power of x - c is 0 and n!(n+1) = (n+1)!: the Lagrange form. For p = 1 it is the Cauchy form. \square
- Proof. Fix a \in I and x \in I, x \ne a. f^{(n)} is differentiable on I, hence continuous, so Theorem 2 gives \displaystyle R_n(x) = \frac{f^{(n+1)}(c)}{(n+1)!}(x - a)^{n+1} = 0. So f = P_n on I. A polynomial p of degree d has p^{(d+1)} = 0, so p equals its P_d at every center. \square
- Proof. Let \delta_1 be as in the definition of the relative minimum. If n were odd, Theorem 5(b) would give x with \lvert x - a \rvert < \delta_1 and f(x) < f(a). If n were even with f^{(n)}(a) < 0, Theorem 5(a) would give \delta_2 with f(x) < f(a) for 0 < \lvert x - a \rvert < \delta_2, and as a is interior some x \in I has 0 < \lvert x - a \rvert < \min\{\delta_1, \delta_2\}. Both contradict the minimum, so n is even and f^{(n)}(a) > 0. For the second claim, f'(a) = 0 by Fermat’s lemma (Theorem 1 of The mean value theorem and its consequences), and if f''(a) \ne 0 the first part with n = 2 gives f''(a) > 0. \square
- Proof. Suppose f^{(k)}(a) = 0 for k \le n, so P_n = 0. For x \ne a in I, Theorem 2 gives c_x strictly between a and x with \displaystyle \frac{f(x)}{(x - a)^{n+1}} = \frac{f^{(n+1)}(c_x)}{(n+1)!}. Let \varepsilon > 0. By continuity of f^{(n+1)} at a, some \delta > 0 has \lvert f^{(n+1)}(t) - f^{(n+1)}(a) \rvert < \varepsilon for t \in I, \lvert t - a \rvert < \delta. For 0 < \lvert x - a \rvert < \delta, also \lvert c_x - a \rvert < \delta, so the quotient is within \varepsilon of \displaystyle \frac{f^{(n+1)}(a)}{(n+1)!}, which is the limit. Conversely, if the limit L exists, then \displaystyle \frac{f(x)}{(x - a)^n} = (x - a)\frac{f(x)}{(x - a)^{n+1}} \to 0, and Theorem 4 with Q = 0 gives P_n = 0. \square