The Derivative: Definition via Difference Quotients

How fast is a quantity changing at a single instant? This differs from changing on average over an interval. A car's speedometer answers this question every second. Yet "speed at an instant" is a strange idea. An instant has zero duration. So the ordinary notion of "distance divided by time" seems to collapse into $0/0$ . The derivative is the precise resolution of this paradox. It is built entirely from the machinery of Limits of Functions. That machinery already handles delicate limiting processes rigorously. The derivative matters because it converts an intuition into a number. This intuition is geometric and physical — the slope of a tangent line, the instantaneous rate of change. It becomes a single well-defined real number, computable by a limit. This number underlies almost every quantitative model of change. Examples include velocity and acceleration in mechanics, marginal cost in economics, and growth rates in biology. It also underlies the local linear approximation. This makes calculus a tool for approximation rather than just description. The big picture is this: near a point $a$ , a differentiable function looks almost like a straight line. This holds when you zoom in far enough. The derivative $f'(a)$ is the slope of that line. We capture "almost like a straight line" in a specific way. We compare the average slope over a shrinking interval — the difference quotient. Then we ask whether these average slopes converge to a limiting value. This happens as the interval shrinks to a point. This article gives the formal $\varepsilon$ – $\delta$ definition of the derivative. It defines the derivative as a limit of difference quotients. It proves that this limit, when it exists, is unique. It interprets this limit as the slope of the tangent line. It works through fully rigorous examples. One example shows a case where the derivative fails to exist. It sets the stage for showing differentiability forces continuity. It also prepares for the algebraic differentiation rules, and for the chain rule.

Key takeaways

The difference quotient $\frac{f(x)-f(a)}{x-a}$ measures the average rate of change of $f$ over the interval between $a$ and $x$ .
The derivative $f'(a)$ is defined as the limit of this quotient as $x \to a$ , when that limit exists.
Geometrically, $f'(a)$ is the slope of the tangent line to the graph of $f$ at the point $(a, f(a))$ .
The limit, when it exists, is unique — this is inherited directly from uniqueness of limits of functions.
Not every continuous function is differentiable: sharp corners (like $|x|$ at $0$ ) produce one-sided quotients that disagree.
The derivative is only ever defined at points where $f$ is defined on some open interval around $a$ , or at least in a punctured neighborhood permitting the limit to make sense.
Key idea: the derivative is the limit of average slopes as the interval of averaging shrinks to a single point.

Where this lives today

The derivative as a limit of difference quotients is the bedrock computational tool of physics, where $f'(t)$ for a position function $f$ gives instantaneous velocity, and the second derivative gives acceleration, feeding directly into Newton's laws of motion. In economics, marginal cost and marginal revenue are derivatives of cost and revenue functions with respect to quantity produced. In numerical analysis and scientific computing, finite-difference approximations $\frac{f(a+h)-f(a)}{h}$ for small fixed $h$ are literally truncated difference quotients, used to approximate derivatives when a closed form is unavailable, with the approximation error controlled by how close the discrete quotient is to the true limit. In machine learning, the gradient used in gradient descent is built from partial derivatives, each defined via exactly this one-variable limiting process along each coordinate direction. In control theory and signal processing, the derivative operator models rates of change of signals, and difference quotients underlie discrete-time approximations (digital differentiation) implemented in embedded systems. The idea of a "local linear approximation" coming out of this definition — see Algebraic Rules of Differentiation and eventually The Chain Rule — is the seed of the entire framework of local linearization used throughout applied mathematics, from Newton's method for root-finding to perturbation theory in physics.

Ask the tutor: Can you show me why differentiability at a point must imply continuity there, using this same difference-quotient definition?

f

Intuition

Fix a point $a$ on the graph of $f$ and a second point $x$ nearby. The straight line through $(a, f(a))$ and $(x, f(x))$ is a secant line, and its slope is exactly the difference quotient $Q_a(x) = \frac{f(x)-f(a)}{x-a}$ . This is an average rate of change: if $f$ measures position and $x - a$ measures elapsed time, $Q_a(x)$ is average velocity over that time interval. The derivative asks: what happens to the slope of this secant line as $x$ slides toward $a$ , so the interval shrinks to nothing? If the secant slopes converge to a single number $f'(a)$ , the secant lines themselves converge (in a precise sense, via their slopes and the common point $(a,f(a))$ ) to a limiting line — the tangent line at $a$ :

y = f(a) + f'(a)(x-a).

This is the line through $(a,f(a))$ that best approximates $f$ near $a$ , in the sense that the error $f(x) - \big[f(a) + f'(a)(x-a)\big]$ vanishes even after dividing by $x-a$ .

Figure

Secant lines through

(1,1)

y=x^2

converging to the tangent of slope

2

As $x$ moves from $2$ toward $1.5$ toward $1$ , the secant slopes $3, 2.5, \dots$ converge to $2$ , the tangent slope — matching the algebraic computation done below.

f

g

f

y = f(a) + m(x-a)

Worked examples

Example 1: $f(x) = x^3 - 2x$ at $a = 2$ (asymmetric polynomial). We compute $f'(2)$ directly from the definition. Here $f(2) = 2^3 - 2(2) = 8 - 4 = 4$ . For $x \neq 2$ ,

Q_2(x) = \frac{f(x) - f(2)}{x - 2} = \frac{(x^3 - 2x) - 4}{x - 2}.

We factor the numerator. Since $x=2$ is a root of $x^3 - 2x - 4$ (check: $8 - 4 - 4 = 0$ ), polynomial division gives

x^3 - 2x - 4 = (x-2)(x^2 + 2x + 2).

(Verification: $(x-2)(x^2+2x+2) = x^3 + 2x^2 + 2x - 2x^2 - 4x - 4 = x^3 - 2x - 4$ . ✓.) Hence for $x \ne 2$ ,

Q_2(x) = \frac{(x-2)(x^2+2x+2)}{x-2} = x^2 + 2x + 2,

where the cancellation is valid precisely because $x \ne 2$ in the domain of $Q_2$ . The function $x \mapsto x^2+2x+2$ is a polynomial, hence continuous at $2$ , so

\lim_{x \to 2} Q_2(x) = \lim_{x\to 2} (x^2+2x+2) = 4 + 4 + 2 = 10.

Therefore $f$ is differentiable at $a=2$ and $f'(2) = 10$ . Geometrically, the tangent line to $y = x^3-2x$ at $(2,4)$ has slope $10$ and equation $y = 4 + 10(x-2) = 10x - 16$ .

Example 2: $f(x) = \sqrt{x}$ at $a = 4$ (a non-polynomial case requiring conjugates). Here $f(4) = 2$ . For $x > 0$ , $x \neq 4$ ,

Q_4(x) = \frac{\sqrt{x} - 2}{x - 4}.

Multiply numerator and denominator by the conjugate $\sqrt{x}+2$ (valid since $\sqrt x + 2 \neq 0$ for $x>0$ ):

Q_4(x) = \frac{(\sqrt{x}-2)(\sqrt{x}+2)}{(x-4)(\sqrt{x}+2)} = \frac{x - 4}{(x-4)(\sqrt{x}+2)} = \frac{1}{\sqrt{x}+2},

again with cancellation legitimate because $x \ne 4$ means $x - 4 \ne 0$ . The function $x \mapsto \frac{1}{\sqrt x + 2}$ is continuous at $x=4$ (as a composition/quotient of continuous functions with nonvanishing denominator $\sqrt 4 + 2 = 4 \ne 0$ ), so

\lim_{x\to 4} Q_4(x) = \frac{1}{\sqrt 4 + 2} = \frac{1}{4}.

Thus $f'(4) = 1/4$ . The tangent line at $(4,2)$ is $y = 2 + \tfrac14(x-4)$ . Note the domain subtlety: $f(x)=\sqrt x$ is only defined for $x \ge 0$ , so strictly the "open interval containing $a$ " in the definition should be read as an interval within the domain; at $a=4>0$ we can certainly find an open interval $(4-\delta, 4+\delta) \subset (0,\infty)$ for small $\delta$ , so the definition applies without modification.

f'(a)

Scope, edge cases & common traps

Situation	What to check	Common error
$f$ defined only on a closed interval $[a,b]$ , evaluating at endpoint $a$	Only a one-sided limit makes sense; some texts define one-sided derivatives $f'(a^+)$	Silently treating a one-sided derivative as if it were the (two-sided) derivative
Corner points, e.g. $\|x\|$ at $0$	One-sided quotient limits must agree	Assuming continuity implies differentiability
Vertical tangent, e.g. $f(x)=x^{1/3}$ at $0$	Difference quotient $\to \pm\infty$ , not a finite limit	Declaring $f'(0) = \infty$ as if it were a legitimate derivative value (the derivative is undefined here, by our definition which requires a finite limit)
Removable discontinuities in $Q_a$	Simplify $Q_a(x)$ algebraically before taking the limit; cancellation is valid only for $x \ne a$	Plugging $x=a$ directly into the unsimplified quotient, getting $0/0$ and stopping
Domain not an open interval around $a$ (e.g. $a$ is an isolated point of the domain)	Definition requires a neighborhood of $a$ in the domain so $Q_a$ is defined on a punctured neighborhood	Applying the derivative definition where no genuine limit process is possible

The mathematician behind it

Origins · 1 minIsaac Newton and how this was discovered

Isaac Newton

1642–1727 · Cambridge and London, England; developed fluxions and the calculus of instantaneous rates of change

Newton developed his "method of fluxions" — his term for what we now call derivatives — around 1665–1666 while Cambridge University was closed due to plague, deriving instantaneous velocities as limits of ratios of vanishing increments. Independently and roughly contemporaneously, Gottfried Wilhelm Leibniz in Germany developed a parallel formalism using the notation $\frac{dy}{dx}$ and infinitesimals, which proved more flexible for later generations and is still in use today. The rigorous $\varepsilon$ – $\delta$ limit definition used in this article, making precise what both Newton and Leibniz meant informally by "vanishingly small," is due to 19th-century analysts, chiefly Augustin-Louis Cauchy and later Karl Weierstrass, who replaced infinitesimal intuition with the modern limit concept.

History

- 1665–1666: Newton develops fluxions and fluents during his stay at Woolsthorpe, computing tangent slopes as limiting ratios.

- 1675–1676: Leibniz independently develops differential calculus in Paris, introducing the notation

dy/dx

- 1823: Cauchy's Résumé des leçons gives an early rigorous limit-based treatment of the derivative.

- 1861: Weierstrass, lecturing in Berlin, formalizes the

\varepsilon

–

\delta

definition of limit underlying the modern derivative.

- 1872: Weierstrass exhibits a function continuous everywhere but differentiable nowhere, sharply separating continuity from differentiability.

Why do we need a limit at all — why not just define the derivative as

\frac{f(x)-f(a)}{x-a}

evaluated at

x=a

Because plugging $x=a$ directly gives $\frac{f(a)-f(a)}{a-a} = \frac{0}{0}$ , which is undefined — division by zero is never permitted. The entire content of the derivative is that even though this ratio is undefined exactly at $x=a$ , it can still approach a definite, finite value as $x$ gets arbitrarily close to (but never equal to) $a$ . The limit concept, built from Limits of Functions, is precisely the tool for making sense of "approaches a value without ever reaching the problematic point," and that is exactly what is needed here.

Does differentiability at a point require the function to be defined on both sides of that point?

Yes, for the ordinary two-sided derivative as defined here: the definition requires $f$ to be defined on a full open interval containing $a$ , so that $x$ can approach $a$ from both the left and the right within the domain. If $f$ is only defined on one side (for instance at an endpoint of a closed interval), one instead defines a one-sided derivative using a one-sided limit, and many texts distinguish this explicitly from the full (two-sided) derivative treated in this article.

Is every continuous function differentiable?

No. Continuity is a strictly weaker condition than differentiability. The counterexample $f(x)=|x|$ at $a=0$ shows a continuous function that fails to be differentiable at a single point because of a corner. Even more dramatically, Weierstrass constructed a function that is continuous at every real number yet differentiable at none of them, showing continuity can hold everywhere while differentiability holds nowhere. The article Differentiability Implies Continuity proves the correct one-directional implication: differentiability does force continuity, but never the converse.

What is the geometric meaning of

f'(a)

being negative, zero, or undefined?

If $f'(a) > 0$ , the tangent line rises left to right, meaning $f$ is instantaneously increasing near $a$ . If $f'(a) < 0$ , the tangent falls, so $f$ is instantaneously decreasing. If $f'(a) = 0$ , the tangent is horizontal, which is necessary (though not sufficient) for $a$ to be a local extremum of a differentiable function. If $f'(a)$ is undefined, the graph either has a corner (like $|x|$ ), a vertical tangent (like $x^{1/3}$ at $0$ ), or some other failure of the two one-sided quotient limits to exist and agree, so no single tangent slope can be assigned.

How does the

h \to 0

form of the definition relate to the

x \to a

form?

They are the same limit, only reparametrized. Setting $x = a+h$ , the condition $x \to a$ (with $x \ne a$ ) is exactly equivalent to $h \to 0$ (with $h \ne 0$ ), and $\frac{f(x)-f(a)}{x-a}$ becomes $\frac{f(a+h)-f(a)}{h}$ under this substitution. Both forms define exactly the same number $f'(a)$ when the corresponding limit exists; the $h\to 0$ form is often more convenient in explicit computations and is the version most often generalized in later topics like the definition of higher derivatives and partial derivatives.

Can the derivative be

+\infty

-\infty

Under the definition given here, no — differentiability requires the limit of the difference quotient to be a finite real number, so if the quotient diverges to $\pm\infty$ , $f$ is simply not differentiable at that point in this sense. Some texts separately discuss the geometric phenomenon of a "vertical tangent line" (as with $f(x)=x^{1/3}$ at $x=0$ , where the quotient does tend to $+\infty$ ) as a distinct, informally-described case, but it is not assigned a derivative value under the standard finite-limit definition used throughout this article and its sequels.

Why is uniqueness of the derivative worth proving — isn't it obvious?

It is worth proving because "obvious" claims are exactly where hidden assumptions hide, and the entire rigor of calculus rests on such foundational facts being genuinely established rather than assumed. The uniqueness of $f'(a)$ is not a new fact requiring new machinery: it follows immediately, via a short and completely standard $\varepsilon$ – $\delta$ argument, from the already-established uniqueness of limits of functions. Demonstrating this explicitly shows how the whole edifice of differentiation is built as a direct logical consequence of the limit concept, with no extra unproven assumptions smuggled in.