Higher-Order Derivatives

Suppose a car's position at time $t$ is $s(t)$ . Its velocity is $s'(t)$ . But what tells you whether the car is speeding up or slowing down? You need the derivative of the velocity itself — a "derivative of a derivative." This simple idea, iterating the operation of differentiation, is the subject of this article. It turns out to carry enormous structural weight in analysis. Higher-order derivatives let us classify functions by how much smoothness they possess. Some functions are differentiable only once. Others are differentiable any finite number of times. Some are differentiable infinitely often. This hierarchy is organized into the classes $C^n$ and $C^\infty$ . It is not a pedantic bookkeeping device. It is exactly the hypothesis needed to control the error in polynomial approximations (via Taylor's Theorem with Remainder). It also helps detect concavity and convexity of a graph. It lets us distinguish local maxima from minima beyond the first-derivative test. And it helps make sense of physical quantities like acceleration and jerk. The construction is inductive. Define $f''$ as the derivative of $f'$ wherever that derivative exists. Then define $f'''$ as the derivative of $f''$ , and so on. Each step requires the previous derivative to exist on an open set (or at least a punctured neighborhood). This ensures differentiating it again makes sense. This recursive dependency is the source of every subtlety in the theory. A function can be differentiable without its derivative being differentiable. The resulting gaps in smoothness are exactly what the classes $C^n$ are designed to measure. This article gives the precise inductive definition of $f^{(n)}$ . It proves the basic algebraic facts (linearity, the product rule in its Leibniz form). It defines and studies the smoothness classes $C^n(I)$ and $C^\infty(I)$ . It also exhibits sharp counterexamples showing why each hypothesis in these definitions is load-bearing.

Key takeaways

The $n$ -th derivative $f^{(n)}$ is defined inductively: $f^{(0)} = f$ , and $f^{(n)} = (f^{(n-1)})'$ wherever the right side exists.
Existence of $f^{(n)}(a)$ requires $f^{(n-1)}$ to exist and be differentiable at $a$ , hence $f^{(n-1)}$ must exist on a neighborhood of $a$ .
$f \in C^n(I)$ means $f^{(n)}$ exists on $I$ and is continuous there; $f \in C^\infty(I)$ means derivatives of all orders exist.
The Leibniz rule generalizes the product rule: $(fg)^{(n)} = \sum_{k=0}^n \binom{n}{k} f^{(k)} g^{(n-k)}$ .
Smoothness classes are strictly nested: $C^{n+1}(I) \subsetneq C^n(I)$ , witnessed by explicit functions like $x^n|x|$ .
Key idea: differentiation is an operator you can iterate, and the n-th time you can iterate it and stay continuous measures a function's smoothness.

Where this lives today

Higher-order derivatives are the backbone of Taylor approximation: the coefficients of the Taylor polynomial of $f$ at a point are literally $f^{(k)}(a)/k!$ , and the size of the remainder term is controlled by a bound on $f^{(n+1)}$ — the entire subject of Taylor's Theorem with Remainder is a direct continuation of this article. In geometry and optimization, the sign of $f''$ characterizes Convexity and the Second Derivative of a function, which underlies convex optimization algorithms throughout machine learning and operations research. In physics, $f'', f'''$ , and beyond appear as acceleration, jerk, snap, and higher kinematic quantities used in robotics motion planning and aerospace trajectory design, where smoothness of a planned path (often demanded to be $C^2$ or higher) prevents mechanically damaging jerks in actuators. In numerical analysis, finite-difference schemes for solving differential equations require explicit error bounds built from higher derivatives (via Taylor expansions), and the stability of these schemes depends on which $C^n$ class the true solution belongs to. In differential geometry and PDE theory, the classes $C^k$ and $C^\infty$ (and their infinite-dimensional generalizations, Sobolev spaces) are the standard language for stating exactly how much regularity a solution to a differential equation is assumed or proved to have.

Ask the tutor: Why does $f \in C^\infty$ at a point not guarantee that $f$ equals its Taylor series near that point?

C^n

Intuition

Think of differentiation as a machine: feed it a function, it returns another function (the derivative), provided the limit defining that derivative exists everywhere you need it. Iterating the machine is exactly what $f''$ , $f'''$ , etc. record. The subtlety is that the machine can jam: if $f'$ itself is not differentiable at some point, the process simply stops there, and $f''$ fails to exist at that point — even though $f$ and $f'$ might both be perfectly well-behaved (continuous, even differentiable) nearby. A good mental picture is a tower of functions $f, f', f'', f''', \dots$ , each one built from the last by the same limiting operation, with the tower's height (how many rungs you can climb before the process breaks) measuring the smoothness of $f$ . The classes $C^n$ slice this tower at height $n$ and additionally demand that the top function you reach is continuous — not just that the limit exists pointwise, but that it exists and varies continuously. This extra continuity requirement is what makes $C^n$ well-behaved enough to support approximation theorems. Geometrically, $f''$ measures the curvature or bending of the graph of $f$ : where $f'' > 0$ the graph curves upward (convex), where $f'' < 0$ it curves downward (concave), foreshadowing the link between Second Derivative Test and convexity. Physically, if $f$ is position, $f'$ is velocity, $f''$ is acceleration, and $f'''$ is "jerk" — the rate of change of acceleration, which is what makes a smooth elevator ride feel smooth rather than jarring.

Figure

The tower of derivatives: each rung requires the one below to be differentiable

I \subseteq \mathbb{R}

n

n \geq 0

We prove the inclusion here; the strictness is exhibited explicitly in the Worked Examples and the Counterexample section below.

f \in C^{n+1}(I)

Worked examples

Example 1: A function that is exactly $C^2$ but not $C^3$ , computed explicitly. Let $f: \mathbb{R} \to \mathbb{R}$ be defined by

f(x) = x^3 |x| = \begin{cases} x^4 & x \geq 0 \\ -x^4 & x < 0. \end{cases}

We compute the derivatives directly and locate exactly where smoothness fails. Computing $f'$ . For $x > 0$ , $f(x) = x^4$ , so $f'(x) = 4x^3$ . For $x < 0$ , $f(x) = -x^4$ , so $f'(x) = -4x^3$ . At $x=0$ , using the definition of the derivative,

f'(0) = \lim_{h \to 0} \frac{f(h) - f(0)}{h} = \lim_{h\to 0} \frac{h^3|h|}{h} = \lim_{h \to 0} h^2 |h| = 0,

since $|h^2|h|| \le h^2 \to 0$ (squeeze). So $f'(x) = 4|x|x^2 = 4x^2|x|$ for all $x$ (check: at $x>0$ this is $4x^3$ ✓, at $x<0$ this is $-4x^3$ ✓, at $0$ it's $0$ ✓). Note $f'$ is continuous everywhere (it is $4x^2|x|$ , a product/composition of continuous functions), so $f \in C^1(\mathbb{R})$ . Computing $f''$ . For $x>0$ : $f'(x) = 4x^3 \Rightarrow f''(x) = 12x^2$ . For $x<0$ : $f'(x) = -4x^3 \Rightarrow f''(x) = -12x^2$ . At $x=0$ :

f''(0) = \lim_{h\to 0} \frac{f'(h) - f'(0)}{h} = \lim_{h \to 0} \frac{4h^2|h|}{h} = \lim_{h\to 0} 4h|h| = 0.

So $f''(x) = 12x|x|$ for all $x$ (check at $x>0$ : $12x^2$ ✓; at $x<0$ : $-12x^2$ ✓). This is continuous everywhere (again a product of continuous functions), so $f \in C^2(\mathbb{R})$ . Computing $f'''$ . For $x > 0$ : $f''(x) = 12x^2 \Rightarrow f'''(x) = 24x$ . For $x<0$ : $f''(x) = -12x^2 \Rightarrow f'''(x) = -24x$ . At $x = 0$ :

f'''(0) = \lim_{h \to 0} \frac{f''(h) - f''(0)}{h} = \lim_{h\to0} \frac{12h|h|}{h} = \lim_{h \to 0} 12|h| = 0.

So $f'''$ does exist everywhere, and equals $f'''(x) = 24|x|$ (check: $x>0$ gives $24x$ ✓, $x<0$ gives $-24x = 24|x|$ ✓, and at $0$ gives $0$ ✓). This function $24|x|$ is continuous, so in fact $f \in C^3(\mathbb{R})$ too! Computing $f^{(4)}$ . Now differentiate $f'''(x) = 24|x|$ at $x=0$ :

\lim_{h \to 0^+} \frac{24|h| - 0}{h} = \lim_{h\to 0^+} \frac{24h}{h} = 24, \qquad \lim_{h \to 0^-} \frac{24|h|-0}{h} = \lim_{h \to 0^-}\frac{-24h}{h} = -24.

The one-sided limits disagree ( $24 \neq -24$ ), so $f^{(4)}(0)$ does not exist. Hence $f \notin C^4(\mathbb{R})$ — in fact $f^{(4)}$ fails to exist at even a single point, so $f$ is not $4$ times differentiable on any open interval containing $0$ .

Conclusion: $f(x) = x^3|x|$ satisfies $f \in C^3(\mathbb{R})$ but $f \notin C^4(\mathbb{R})$ (indeed $f^{(4)}(0)$ does not exist at all). This shows the chain $C^0 \supsetneq C^1 \supsetneq C^2 \supsetneq C^3 \supsetneq C^4$ is strict at this level, generalizing the classic $x|x|$ example (which is the $n=1$ analogue) to a higher order.

Example 2: Leibniz's rule applied to an asymmetric product, $n=4$ . Let $f(x) = e^{2x}$ and $g(x) = x^3 - x$ on $I = \mathbb{R}$ ; we compute $(fg)^{(4)}$ two ways to cross-check Leibniz's formula. Direct route. $h(x) = e^{2x}(x^3 - x)$ . We could expand and differentiate four times directly, but let's instead verify via Leibniz and then spot-check numerically at $x=1$ . Derivatives of $f$ . $f(x) = e^{2x}$ , so $f^{(k)}(x) = 2^k e^{2x}$ for all $k \geq 0$ . Thus $f(1) = e^2$ , $f'(1) = 2e^2$ , $f''(1) = 4e^2$ , $f'''(1) = 8e^2$ , $f^{(4)}(1) = 16e^2$ . Derivatives of $g$ . $g(x) = x^3 - x$ : $g'(x) = 3x^2 - 1$ , $g''(x) = 6x$ , $g'''(x) = 6$ , $g^{(4)}(x) = 0$ . At $x=1$ : $g(1) = 0$ , $g'(1) = 2$ , $g''(1) = 6$ , $g'''(1) = 6$ , $g^{(4)}(1) = 0$ . Applying Leibniz's rule with $n=4$ :

(fg)^{(4)}(1) = \sum_{k=0}^{4} \binom{4}{k} f^{(k)}(1)\, g^{(4-k)}(1).

Compute each term:

$k=0$ : $\binom{4}{0} f(1) g^{(4)}(1) = 1 \cdot e^2 \cdot 0 = 0$ .
$k=1$ : $\binom{4}{1} f'(1) g^{(3)}(1) = 4 \cdot 2e^2 \cdot 6 = 48e^2$ .
$k=2$ : $\binom{4}{2} f''(1) g''(1) = 6 \cdot 4e^2 \cdot 6 = 144 e^2$ .
$k=3$ : $\binom{4}{3} f'''(1) g'(1) = 4 \cdot 8 e^2 \cdot 2 = 64e^2$ .
$k=4$ : $\binom{4}{4} f^{(4)}(1) g(1) = 1 \cdot 16 e^2 \cdot 0 = 0$ .

Summing: $(fg)^{(4)}(1) = 0 + 48e^2 + 144e^2 + 64e^2 + 0 = 256 e^2$ . Cross-check by direct expansion. Write $h(x) = e^{2x} g(x)$ with $g = x^3 - x$ . One can show by induction (or repeated product rule) that

h^{(4)}(x) = e^{2x}\sum_{k=0}^4 \binom{4}{k} 2^k g^{(4-k)}(x) = e^{2x}\Big[g^{(4)} + 8g^{(3)} + 24 g'' + 32 g' + 16 g\Big](x),

matching the Leibniz coefficients $\binom{4}{k}2^k$ for $k=0,\dots,4$ : namely $16,32,24,8,1$ paired with $g, g', g'', g^{(3)}, g^{(4)}$ respectively (same computation, reorganized). At $x=1$ : $g(1){=}0, g'(1){=}2, g''(1){=}6, g'''(1){=}6, g^{(4)}(1){=}0$ , giving

h^{(4)}(1) = e^2[0 + 8(6) + 24(6) + 32(2) + 16(0)] = e^2[48 + 144 + 64] = 256e^2,

matching exactly. This numerically confirms Leibniz's rule in a genuinely asymmetric case (different function types, non-palindromic coefficients all engaged).

f'

Scope, edge cases & common traps

Situation	What holds	What can fail
$f^{(n)}(a)$ exists at a single point $a$	$f^{(n-1)}$ must exist on a neighborhood of $a$ and be differentiable at $a$	$f^{(n)}$ need not exist at nearby points, and need not be continuous even at $a$
$f \in C^n(I)$	$f^{(k)}$ exists and is continuous on $I$ for every $0 \le k \le n$	$f^{(n+1)}$ may fail to exist anywhere
$f, g \in C^n(I)$	$\alpha f + \beta g,\ fg \in C^n(I)$ (Theorem, parts a,c)	$f/g \in C^n$ only where $g \neq 0$ ; must exclude zeros of $g$
$f \in C^\infty(I)$	All derivatives exist and are continuous	Does not imply $f$ is analytic (equal to its Taylor series) — see FAQ
Composition $f\circ g$	If $f,g \in C^n$ , then $f \circ g \in C^n$ (provable by induction using the chain rule, stated here without full proof)	The formula for $(f\circ g)^{(n)}$ (Faà di Bruno's formula) is far more complex than Leibniz's rule and is easy to misstate

Common trapThe classic error: confusing "

n

times differentiable at a point" with "

n

times differentiable on a neighborhood"

A very common mistake is to compute $f^{(n-1)}$ formally by a closed-form expression valid off some point, plug it into a difference quotient at that point, get a finite answer, and declare " $f^{(n)}$ exists at $a$ ." As the counterexample shows, this reasoning is invalid unless you separately verify $f^{(n-1)}$ is continuous (indeed differentiable) at $a$ , not merely that a numerical limit involving a possibly-discontinuous function happens to converge. The correct logical order is: first confirm $f^{(n-1)}$ exists on a full open neighborhood of $a$ (not just at isolated points), then test differentiability of that specific function at $a$ . Skipping this and directly manipulating formulas is the single most common source of erroneous "higher derivative" computations.

The mathematician behind it

Origins · 1 minGottfried Wilhelm Leibniz and how this was discovered

Gottfried Wilhelm Leibniz

1646–1716 · Co-inventor of calculus, Hanover and Paris

Leibniz developed the differential calculus independently of Newton in the 1670s–80s, and it is his notation — $dy/dx$ , $\int$ , and the systematic use of $d^n y / dx^n$ for higher derivatives — that survives in nearly all of modern analysis. His 1695 correspondence with Johann Bernoulli contains the first explicit statement of what we now call Leibniz's rule for the $n$ -th derivative of a product, discovered while investigating fractional (non-integer order) derivatives, an idea decades ahead of its formal development. Leibniz's insistence on symbolic manipulation and rigorous notation, rather than purely geometric argument, is precisely what makes the inductive machinery of $f^{(n)}$ tractable and teachable today.

History

- 1675–1676: Leibniz introduces differential notation

dx

dy

and begins systematic symbolic calculus.

- 1695: Leibniz states the general product-rule formula for

(fg)^{(n)}

in a letter to Johann Bernoulli, prompted by a question about fractional derivatives.

- 1797: Lagrange, in Théorie des fonctions analytiques, popularizes the prime notation

f', f'', f'''

for successive derivatives.

- 1821–1823: Cauchy, in his Cours d'Analyse and related lectures, gives the first fully rigorous

\varepsilon

–

\delta

treatment of limits underlying the modern definition of

f^{(n)}

- 1872: Weierstrass exhibits a function continuous everywhere but differentiable nowhere, sharpening the community's understanding of exactly how fragile smoothness properties are — motivating the careful class hierarchy

C^0 \supsetneq C^1 \supsetneq C^2 \supsetneq \cdots

used today.

What is the difference between "

f

n

times differentiable" and "

f \in C^n

" $f$ is $n$ times differentiable at a point (or on a set)" only asserts that the limit defining $f^{(n)}$ exists there — nothing is said about whether $f^{(n)}$ varies continuously. " $f \in C^n(I)$ " additionally requires $f^{(n)}$ to be continuous on $I$ . These are genuinely different: the counterexample $f(x)=x^2\sin(1/x)$ (extended by $f(0)=0$ ) is differentiable everywhere (so "1 time differentiable" everywhere), yet $f'$ is discontinuous at $0$ , so $f \notin C^1(\mathbb{R})$ . Every $C^n$ function is $n$ times differentiable, but not conversely. This distinction matters because theorems like Taylor's theorem with the Lagrange remainder typically require the stronger $C^n$ (or $C^{n+1}$ ) hypothesis, not merely "n times differentiable," to control error terms rigorously.

Does

f \in C^\infty

mean

f

equals its own Taylor series?

No — this is one of the most important subtleties in analysis. A function all of whose derivatives exist and are continuous (smooth) need not be analytic (equal to the sum of its Taylor series in a neighborhood of each point). The standard counterexample is $f(x) = e^{-1/x^2}$ for $x \ne 0$ , $f(0)=0$ : one can show by induction that every derivative $f^{(n)}(0) = 0$ , yet $f$ is not identically zero near $0$ . So its Taylor series at $0$ is the zero function, which does not equal $f$ anywhere except at $x=0$ itself. Thus $C^\infty$ is a strictly larger class than the analytic functions $C^\omega$ ; this gap is exactly what Taylor's Theorem with Remainder is designed to quantify and control.

Why does the definition require

f^{(n-1)}

to exist on a whole neighborhood, not just at the point

a

Differentiability at a point is inherently a two-sided limiting statement: $f^{(n-1)}$ must be defined for $a+h$ for all sufficiently small $h$ (both positive and negative, unless working on a one-sided endpoint) in order for the difference quotient $\frac{f^{(n-1)}(a+h)-f^{(n-1)}(a)}{h}$ to even make sense as $h \to 0$ . If $f^{(n-1)}$ were defined only at $a$ itself, the difference quotient could not be formed for $h \ne 0$ , and no derivative could ever be extracted. This is not a technicality specific to higher derivatives — it is inherited directly from the ordinary definition of a single derivative, which already requires the function to be defined near the point in question.

Is every

C^n

function's

n

-th derivative automatically differentiable?

No, and this is exactly the gap between $C^n$ and $C^{n+1}$ . Being in $C^n(I)$ only guarantees $f^{(n)}$ is continuous, not differentiable. The function $f(x) = x^{n+1}|x|$ (generalizing Worked Example 1's $x^3|x|$ , which is the $n=2$ case) is a clean example: it lies in $C^n(\mathbb{R})$ but its $n$ -th derivative fails to be differentiable at $0$ , so $f \notin C^{n+1}(\mathbb{R})$ . This is precisely how one proves the class inclusions $C^{n+1}(I) \subsetneq C^n(I)$ are strict for every $n$ , not just an isolated low-order phenomenon.

How does Leibniz's rule relate to the binomial theorem?

The formula $(fg)^{(n)} = \sum_{k=0}^n \binom{n}{k} f^{(k)} g^{(n-k)}$ has exactly the same combinatorial structure as the binomial theorem $(x+y)^n = \sum_{k=0}^n \binom{n}{k}x^k y^{n-k}$ , and the proofs are structurally parallel: both proceed by induction using Pascal's identity $\binom{n-1}{k-1}+\binom{n-1}{k}=\binom{n}{k}$ to merge two sums into one with binomial coefficients. Formally, one can think of the differentiation operator $D$ as behaving like " $D(fg) = (Df)g + f(Dg)$ ," which is the same recursive structure that generates binomial coefficients when iterated $n$ times — this is why the same numbers $\binom{n}{k}$ appear in both settings.

Do higher-order derivatives make sense for functions of several variables?

Yes, but the theory becomes richer: for $f: \mathbb{R}^m \to \mathbb{R}$ , one first defines partial derivatives $\partial f/\partial x_i$ , then second-order partials $\partial^2 f/\partial x_i \partial x_j$ , and the analogue of $C^n$ smoothness additionally requires all mixed partials up to order $n$ to exist and be continuous. A celebrated theorem (Clairaut's/Schwarz's theorem) states that under continuity hypotheses, mixed partials commute: $\partial^2 f/\partial x_i\partial x_j = \partial^2 f/\partial x_j \partial x_i$ . This multivariable extension is essential for defining the Hessian matrix, which generalizes the single-variable second derivative test used to classify critical points.

Why does the problem statement distinguish

C^n

from merely writing

f^{(n)}

"exists everywhere"?

Because "exists everywhere" is a strictly weaker, less useful hypothesis for proving theorems. Many key results in analysis — including error bounds in Taylor's theorem, and criteria for convexity — require not just that $f^{(n)}$ exists pointwise but that it does not jump around unpredictably, which is exactly what continuity guarantees. Working with $C^n$ instead of the weaker pointwise-existence notion lets you invoke the Extreme Value Theorem and Intermediate Value Theorem on $f^{(n)}$ itself (since those theorems require continuity), which is essential machinery for bounding remainder terms and locating critical points of higher order.