Taylor's remainder theorem — error estimation: Visual walkthrough

Step 1 — What is an approximation, and what is the "error"?

WHAT. We have a curvy function $f$ . Near a chosen point $a$ we replace it by a simpler curve, a polynomial $P_{n}$ — a curve built only from $+, -, \times$ of $x$ (no square roots, no $sin$ , nothing exotic). The two curves touch at $a$ and drift apart as you walk away.

WHY. Polynomials are the only functions a calculator can genuinely compute. So we trade the true curve for a polynomial and ask the honest question: how far apart are they?

PICTURE. Below, the burnt-orange curve is the true $f$ ; the teal curve is the polynomial $P_{n}$ . They kiss at $x = a$ . The plum vertical gap at a target point $x$ is the thing we want to bound.

Step 2 — What makes the gap open up? Bending.

WHAT. Two curves that touch at $a$ separate only if one bends away from the other. The amount of bending is measured by derivatives — a derivative is just "how fast a quantity is changing".

WHY this tool. We could ask "how far apart are the curves?" directly, but we don't know $f$ 's value (that's the whole reason we approximated). What we can often control is how curvy $f$ is — bounding a derivative like " $∣ f^{''} ∣$ never exceeds $5$ ". So we route the whole argument through derivatives, because bending is the thing we can pin down.

PICTURE. A gently bending curve (small next derivative) stays close to its polynomial; a violently bending curve peels away fast. Same polynomial degree, wildly different gaps.

Step 3 — Turn the gap into a function we can slide

WHAT. Freeze the target point $x$ . Now let the centre of the polynomial move; call the moving centre $t$ . Define $g (t) = f (x) - \sum_{k = 0}^{n} \frac{f ^{(k)} ( t )}{k !} (x - t)^{k} .$

WHY. This is a trick to make the unknown gap appear as the value of a function at one end of an interval, so we can hit it with a Mean-Value tool later. Read $g (t)$ as: "the leftover error if I had centred my sketch at $t$ instead of $a$ ."

PICTURE. Slide the centre $t$ from $a$ up to $x$ . When $t$ reaches $x$ , the sketch centred at $x$ is perfect there — the gap slams shut.

Step 4 — Differentiate $g$ : a magic telescope

WHAT. Ask how fast $g$ changes as the centre $t$ slides. Every term in the sum has two moving pieces ( $f^{(k)} (t)$ changes, and $(x - t)^{k}$ changes), so the product rule gives two pieces per term. Consecutive terms cancel in a chain — a telescope — and only the very last survives: $g^{'} (t) = - \frac{f ^{(n + 1)} ( t )}{n !} (x - t)^{n} .$

WHY. This is the payoff of the Taylor structure: the messy sum has a clean derivative. All the "book-keeping" derivatives cancel, leaving only the next derivative $f^{(n + 1)}$ — the very bending we identified in Step 2.

PICTURE. Imagine a row of dominoes where each term's " $+$ " cancels the next term's " $-$ ". Only the far-right piece is left standing.

Step 5 — A partner function, and the Mean Value comparison

WHAT. Introduce a simple companion $h (t) = (x - t)^{n + 1}, h (a) = (x - a)^{n + 1}, h (x) = 0, h^{'} (t) = - (n + 1) (x - t)^{n} .$ Now compare the rates of $g$ and $h$ using the Cauchy Mean Value Theorem: on the interval between $a$ and $x$ there is some hidden point $c$ where $\frac{g ( a ) - g ( x )}{h ( a ) - h ( x )} = \frac{g ^{'} ( c )}{h ^{'} ( c )} .$

WHY this tool and not the ordinary Mean Value Theorem. The plain MVT says a single curve has some point whose slope equals its average slope. Here we have two quantities racing side by side and we want to relate their net changes to their instantaneous rates at the same instant $c$ . That "same instant for both" is exactly what Cauchy's version delivers — it is the two-runner generalisation of MVT.

PICTURE. Two runners $g$ and $h$ over the same track $[a, x]$ . Somewhere at time $c$ their speed ratio equals their total-distance ratio. That shared instant $c$ is where our unknown lives.

Step 6 — Substitute and watch the cancellation

WHAT. Feed the four anchor values in. The left side is $\frac{R _{n} ( x ) - 0}{( x - a ) ^{n + 1} - 0}$ . The right side is $\frac{g ^{'} ( c )}{h ^{'} ( c )} = \frac{- \frac{f ^{(n + 1)} ( c )}{n !} ( x - c ) ^{n}}{- ( n + 1 ) ( x - c ) ^{n}} .$

WHY. The two minus signs cancel; the whole factor $(x - c)^{n}$ cancels top and bottom; and $n! \cdot (n + 1) = (n + 1)!$ . All the unknown-position clutter evaporates.

PICTURE. The plum $(x - c)^{n}$ blocks vanish from numerator and denominator, leaving a bare stack.

Step 7 — From exact to usable: replace $c$ by a ceiling $M$

WHAT. We can't locate $c$ , but we can say "wherever it is, $∣ f^{(n + 1)} ∣$ never exceeds some number $M$ on the interval." Replacing $f^{(n + 1)} (c)$ by its worst-case size $M$ can only inflate the estimate, never shrink it: $∣ R_{n} (x) ∣ \leq \frac{M}{( n + 1 )!} ∣ x - a ∣^{n + 1} .$

WHY. A guaranteed bound must survive the worst possible $c$ . Using the maximum $M$ makes the inequality airtight for every hidden $c$ at once. (Picking a single guessed $c$ — a classic mistake — could undershoot and give a false guarantee.)

PICTURE. The true $∣ f^{(n + 1)} ∣$ wiggles under a flat teal ceiling $M$ . Whatever $c$ is, its bar cannot poke above the ceiling.

Step 8 — Edge & degenerate cases (never leave the reader stranded)

WHAT / WHY / PICTURE, three scenarios the bound must survive:

$x = a$ (degenerate, zero walk). Then $∣ x - a ∣^{n + 1} = 0$ , so the bound is $0$ : the sketch and the truth coincide exactly at the centre. No error, as it must be.
$∣ x - a ∣ > 1$ (far walk). The power grows, not shrinks. Here the factorial in the denominator must out-race the growing power; for a fixed $x$ it eventually does (factorials beat any fixed power), which is the seed of the Radius of Convergence idea $R_{n} \to 0$ .
Alternating series (extra help). If the tail alternates in sign (like $sin$ , $cos$ ), consecutive errors partially cancel, so the true gap is often far below the Lagrange ceiling — that slack is exactly what the Alternating Series Estimation Theorem captures.

The one-picture summary

Everything on this page compresses to one image: the true curve, the polynomial, the gap $R_{n}$ , the bending ceiling $M$ , the walk $∣ x - a ∣$ , and the factorial crusher — combining into the boxed bound. In Big-O and asymptotic error language this is $R_{n} = O (∣ x - a ∣^{n + 1})$ .

Recall Feynman: tell it to a 12-year-old

You sketch a curvy hill with a few pencil strokes; that sketch is the polynomial. The gap between your sketch and the real hill is the "error". To pin the gap down we play a game: slide the centre of the sketch from your starting spot toward the target — when it reaches the target the gap snaps to zero. Watching how fast that gap changes, almost everything cancels like a row of dominoes, leaving only one thing: how sharply the hill bends. A clever "two-runners" comparison (Cauchy MVT) turns that into an exact formula, but it hides the bending at a mystery spot $c$ we can't find. So we cheat honestly: we say "the bending is never bigger than $M$ anywhere here" and use that worst case. Multiply the bending ceiling by how far you walked (raised to one-more power) and divide by a big factorial that grows insanely fast — and out pops a guaranteed "your sketch is off by at most this much."

Recall

Which function do we slide to make the gap appear at an endpoint?

g (t) = f (x) - \sum_{k = 0}^{n} \frac{f ^{(k)} ( t )}{k !} (x - t)^{k}

, with

g (a) = R_{n} (x)

and

g (x) = 0

After differentiating

g

, why does only one term survive?

The product-rule terms of consecutive

k

telescope (cancel), leaving

g^{'} (t) = - \frac{f ^{(n + 1)} ( t )}{n !} (x - t)^{n}

Why Cauchy MVT rather than the ordinary MVT?

We compare two racing quantities

g

and

h

and need their rates at the same instant

c

; Cauchy's version handles two functions at once.

What cancels in the final substitution?

The two minus signs and the whole

(x - c)^{n}

factor, leaving

\frac{f ^{(n + 1)} ( c )}{( n + 1 )!} (x - a)^{n + 1}

Why replace

f^{(n + 1)} (c)

M

c

is unknown; using the maximum

M

makes the bound airtight for every possible

c

What happens to the bound when

x = a

∣ x - a ∣^{n + 1} = 0

, so the bound is

0

— sketch and truth coincide at the centre.

Connections

Parent · Hinglish — the stated result this page draws.
Taylor & Maclaurin Series — the polynomial $P_{n}$ being bounded.
Mean Value Theorem / Cauchy Mean Value Theorem — the engine of Steps 5–6.
Radius of Convergence — the far-walk case and $R_{n} \to 0$ .
Alternating Series Estimation Theorem — tighter bound when the tail alternates.
Big-O and asymptotic error — $R_{n} = O (∣ x - a ∣^{n + 1})$ .