Endpoint And Singularity Pitfalls In Error Bounds

Why the Conditions Matter

A Taylor error estimate is a guarantee only when its hypotheses hold. A plausible-looking number for a derivative bound, or an alternating pattern in a few terms, is not enough. If the function has a singularity between the center and the target, or if the relevant terms do not decrease, the familiar estimate may give a false conclusion.

This tutorial focuses on three diagnostic questions: Is the required derivative bounded on the whole interval? Does that interval stay inside the function’s domain? If an alternating-series bound is used, do the term magnitudes decrease to zero? Checking these conditions is part of the estimate, not an optional preliminary.

What a Lagrange Error Bound Requires

For a function \(f\), its degree-\(n\) Taylor polynomial centered at \(a\) is

P_n(x)=\sum_{k=0}^{n}\frac{f^{(k)}(a)}{k!}(x-a)^k.

The Lagrange form of the remainder says that, if \(f^{(n+1)}\) is continuous on the whole closed interval between \(a\) and \(x\), then for some \(\xi\) between them,

f(x)-P_n(x) = \frac{f^{(n+1)}(\xi)}{(n+1)!}(x-a)^{n+1}.

In particular, if \(\lvert f^{(n+1)}(t)\rvert\le M\) everywhere on that interval, then

\lvert f(x)-P_n(x)\rvert \le \frac{M\lvert x-a\rvert^{n+1}}{(n+1)!}.

The word everywhere is essential. A derivative value at the center, or a few checked values inside the interval, cannot establish a bound on the whole interval. A continuous derivative on a closed bounded interval does have a finite maximum in absolute value; a derivative that becomes unbounded near an endpoint or an interior singularity does not.

Pitfall 1: A Derivative Becomes Singular at an Endpoint

Worked Example: A Square Root at the Edge of Its Domain

Consider \(f(x)=\sqrt{1+x}\), with Taylor center \(a=0\), and estimate its value at \(x=-1\) using the degree-one Taylor polynomial.

At the center, \(f(0)=1\) and \(f'(x)=1/(2\sqrt{1+x})\), so \(f'(0)=1/2\). Therefore

P_1(x)=1+\frac{x}{2}.

At the target, \(f(-1)=0\) and \(P_1(-1)=1-1/2=1/2\). Thus the actual remainder is

f(-1)-P_1(-1)=0-\frac12=-\frac12, \qquad \lvert f(-1)-P_1(-1)\rvert=\frac12.

The second derivative on \((-1,\infty)\) is \(f''(x)=-1/[4(1+x)^{3/2}]\). Its magnitude grows without bound as \(x\) approaches \(-1\) from the right. For example, a proposed value \(M=1/4\) would give the claimed bound \(M\lvert-1\rvert^2/2=1/8\), smaller than the actual error \(1/2\). But \(M=1/4\) is not a valid bound for \(\lvert f''\rvert\) on \([-1,0]\): \(f''\) is unbounded there. The Lagrange estimate cannot be applied on that interval.

As a check that the method works away from the singular endpoint, take \(x=-3/4\) instead. On \([-3/4,0]\), \(\lvert f''(x)\rvert\le 2\). The polynomial gives \(P_1(-3/4)=1-3/8=5/8\), while \(f(-3/4)=1/2\), so the error magnitude is \(1/8\). The Lagrange bound is \(2(3/4)^2/2=9/16\), and indeed \(1/8\le 9/16\).

This example also shows why “the function is defined at the endpoint” is not sufficient. The value \(f(-1)\) exists, but the second derivative required for this remainder estimate does not remain bounded up to that endpoint.

Pitfall 2: The Interval Crosses a Domain Boundary

The Taylor estimate concerns the segment joining the center \(a\) and the target \(x\), not just the two endpoint values. Even when \(f(a)\) and \(f(x)\) both exist, that segment might pass through a point where \(f\) is undefined. Then the hypotheses for a remainder formula on the segment are not satisfied.

Worked Example: Two Defined Endpoints Separated by a Pole

Let \(f(x)=1/(x-1)\), choose center \(a=0\), and target \(x=2\). The degree-zero Taylor polynomial is \(P_0(x)=f(0)=-1\). The target value is \(f(2)=1\), so direct subtraction gives

$$ f(2)-P_0(2)=1-(-1)=2. $$

Here \(f'(t)=-1/(t-1)^2\), which is undefined at \(t=1\) and unbounded as \(t\) approaches 1. The interval \([0,2]\) crosses that pole. Consequently, there is no finite \(M\) that bounds \(\lvert f'\rvert\) throughout the interval, and the degree-zero Lagrange estimate cannot be used.

For instance, inserting \(M=1/2\) into the formula would produce \(M\lvert2-0\rvert=1\), which is less than the actual error \(2\). The problem is not a surprising exception to a valid theorem: \(M=1/2\) is not a derivative bound, and the interval does not satisfy the theorem’s conditions. Substituting the exact endpoint values confirms the error and exposes why the proposed estimate is invalid.

Before looking for a derivative maximum, check that the entire segment is in the domain and that the required derivatives exist there. A derivative bound cannot bridge a point where the function itself has a pole.

Pitfall 3: Alternating Does Not Automatically Mean Decreasing

The alternating-series remainder bound has its own hypotheses. For a series whose terms alternate in sign and have magnitudes \(b_k\), the standard bound after a partial sum applies when the magnitudes decrease (from the relevant point onward) and tend to zero. Under those conditions, the error is no greater than the magnitude of the first omitted term. Sign alternation alone does not provide this guarantee.

For a Taylor series evaluated at a particular \(x\), the relevant term magnitudes include the powers of \(x-a\) as well as the Taylor coefficients. It is those evaluated magnitudes that must decrease; the signs of successive coefficients do not settle the question.

Worked Example: Alternating Taylor Contributions That Grow

Consider the polynomial \(f(x)=1-2x+\frac{1}{10}x^2-2x^3\), centered at \(a=0\), and use its degree-one Taylor polynomial at \(x=1\). Since \(f(0)=1\) and \(f'(0)=-2\),

P_1(x)=1-2x, \qquad P_1(1)=1-2=-1.

Evaluate the function at the target:

f(1)=1-2+\frac{1}{10}-2=-\frac{29}{10}.

Therefore the actual remainder and its magnitude are

f(1)-P_1(1) = -\frac{29}{10}-(-1) = -\frac{19}{10}, \qquad \lvert f(1)-P_1(1)\rvert=\frac{19}{10}.

The first omitted Taylor contribution is \(f''(0)/2!\,(1-0)^2=1/10\). Using it as an alternating-series error bound would claim an error at most \(1/10\), but the actual error is \(19/10\). The omitted contributions here are \(1/10\) and \(-2\); their magnitudes increase from \(1/10\) to \(2\). Thus the required decreasing-term condition fails. Direct substitution into the polynomial checks the calculation and shows that the proposed bound is false.

Why the Alternating-Series Bound Works

The decreasing-term condition is not a technical decoration. It is what prevents the later terms from accumulating into an error larger than the first omitted term. Here is a direct justification. Suppose a tail has alternating contributions with magnitudes \(c_1,c_2,\ldots\), where \(c_j\ge 0\), \(c_{j+1}\le c_j\), and \(c_j\to0\). After factoring out the sign of the first omitted contribution, the tail has the form

T=c_1-c_2+c_3-c_4+\cdots.

Group consecutive pairs.
Each pair \(c_{2j-1}-c_{2j}\) is nonnegative because \(c_{2j-1}\ge c_{2j}\). Hence the even partial sums are nonnegative.

Bound the partial sums above.
Each even partial sum is at most \(c_1\): after the first term, the subtracted terms can be grouped with following terms as \(c_{2j}-c_{2j+1}\ge0\), with any final subtracted term also nonnegative.

Pass to the tail limit.
The alternating series converges under these conditions, and its even and odd partial sums both lie between \(0\) and \(c_1\). Their common limit therefore satisfies \(0\le T\le c_1\).

Restoring the sign that was factored out shows that the remainder has the sign of its first omitted term and magnitude at most \(c_1\). If the magnitudes increase, the pairwise nonnegative differences used in this argument are no longer assured, so the conclusion need not hold.

Endpoints Need Diagnosis, Not a Blanket Rule

An endpoint does not automatically invalidate every error estimate. What matters is whether the particular method’s hypotheses hold at that endpoint. For example, the alternating expansion

\ln 2=1-\frac12+\frac13-\frac14+\cdots

has term magnitudes \(1,1/2,1/3,1/4,\ldots\), which decrease to zero. Its alternating-series remainder bound applies at this endpoint, even though the corresponding power series is being evaluated at the boundary of its interval of convergence. This contrasts with the square-root example: there, the derivative needed for the Lagrange estimate becomes unbounded at the endpoint.

For an endpoint or a point near a domain boundary, identify the proposed method first. Then check the conditions that method actually requires. A failure of the Lagrange derivative condition does not by itself disprove an alternating-series estimate, and a valid alternating pattern does not supply a missing derivative bound.

A Reliable Diagnostic Sequence

Before accepting a Taylor error estimate, work through these checks in order:

Specify the method. Is the claim based on a Lagrange remainder, an alternating-series remainder, or another justified result?
Inspect the full interval. Locate both the center and target, then check every point between them for domain boundaries or singularities.
Check the required derivatives. For the Lagrange estimate after degree \(n\), verify that \(f^{(n+1)}\) exists and has a finite bound on the entire segment.
Check the evaluated terms. For the alternating-series estimate, verify alternating signs and nonincreasing magnitudes tending to zero, beginning with the terms relevant to the remainder.
Verify the conclusion independently where possible. Substitute the target into the function and polynomial, or use a separately justified bound. A numerical result cannot repair missing hypotheses, but it can reveal a false claimed estimate.

The central habit is simple: do not compute a bound first and check its conditions afterward. Establish that the theorem applies, then calculate the bound it provides.

Key takeaway: A Lagrange bound needs control of the relevant derivative across the entire center-to-target interval. An alternating-series bound needs decreasing term magnitudes tending to zero. Singularities, crossed domain boundaries, and growing alternating terms each remove a condition essential to the estimate.

Check Your Understanding

For each question, identify the condition that must be checked before accepting the proposed error bound.

A function is defined at the Taylor center and target, but has a pole between them. Why is that not enough to apply the Lagrange remainder formula?
For a degree-\(n\) Taylor polynomial, which derivative must be bounded to obtain the usual Lagrange estimate?
In the square-root example, why does the existence of \(f(-1)\) fail to guarantee a finite degree-one Lagrange bound on \([-1,0]\)?
A sequence of Taylor contributions alternates in sign, but its magnitudes increase. Which hypothesis of the alternating-series remainder bound fails?
Why can an alternating-series bound still work at an endpoint where a Lagrange derivative bound fails?