The Derivative

The speedometer paradox#

Your car's speedometer says 60 km/h. Ask what that number means and the obvious answer is: if you kept this up for an hour, you'd cover 60 kilometers. But you didn't keep it up for an hour. The needle reports a speed right now, at this instant, and it will read something different a second from now.

Here's the trouble. Speed is distance divided by time. At a single instant, no time passes and no distance is covered. The honest computation is $0 / 0$, which is not a number at all. Yet the needle points somewhere, and if you brake hard it moves, so it is clearly measuring something real.

Zeno noticed a version of this two and a half thousand years ago: an arrow in flight, considered at any one instant, is motionless — so how does it ever move? The paradox stood because the tools to answer it did not exist. Newton and Leibniz built those tools in the seventeenth century, and the answer they found is the derivative.

The resolution is not to divide by zero. It is to never quite get there.

Average rate, then squeeze#

Start with something you can compute without controversy. If a car is at position $s(t)$ at time $t$, then over the interval from $t$ to $t + h$ it moves $s(t+h) - s(t)$ in time $h$. The average speed over that stretch is

$$\frac{s(t+h) - s(t)}{h}$$

This is completely unproblematic as long as $h \neq 0$. Take $h = 1$ hour and you get the average over an hour. Take $h = 1$ second and you get the average over a second — much closer to "right now", but still an average.

The move that unlocks everything is to stop asking for the value at $h = 0$ and start asking what the averages approach as $h$ shrinks. Take $h = 0.1$, then $0.01$, then $0.001$. If those numbers settle down toward one particular value, that value is what the speedometer means. It is a number no single interval delivers, but every shrinking sequence of intervals points at.

For a general function $f$, the ratio

$$\frac{f(x+h) - f(x)}{h}$$

is called the difference quotient. Geometrically, it is the slope of the straight line through the two points $(x, f(x))$ and $(x+h, f(x+h))$ on the graph — a line that cuts the curve in two places, called a secant. Shrinking $h$ slides the second point toward the first, and the secant pivots.

Watching the secant become a tangent#

Drag the h slider from right to left (or press Shrink h) and watch the gold secant line rotate. The two dots start far apart and the secant is visibly a compromise — it splits the difference between the steep part of the curve and the shallow part. As $h$ shrinks the dots converge, the secant swings, and it settles onto the dashed violet line: the tangent, the line that touches the curve at one point and matches its direction there.

Two things to try. First, watch the readout at the bottom rather than the picture: the difference quotient marches toward $f'$ and the gap collapses, even though $h$ never reaches zero. Second, drag the base x slider to move the point along the curve. On $x^3/6 - x$, park it near $x = -1.4$ where the curve peaks and the tangent goes flat — the difference quotient converges to zero. Slide right past the dip and the same construction returns a positive number. The slope is not one number belonging to the function; it is a different number at every point.

Switch the curve to $\sin(x)$ and repeat: the tangent is steepest where the sine crosses the axis and flat at the crests. That pattern is the whole subject in miniature.

The definition, and the notation#

What the animation shows, written down, is the definition of the derivative:

$$f'(x) = \lim_{h \to 0} \frac{f(x+h) - f(x)}{h}$$

Read it carefully, because the subtlety is in the $\lim$. It does not say "plug in $h = 0$" — that gives $0/0$. It says: as $h$ gets arbitrarily close to zero (from either side, and never equal to zero), the quotient gets arbitrarily close to some number $L$, and $f'(x)$ is that number $L$.

Why should such an $L$ exist? Because for a smooth curve, zooming in far enough makes the graph indistinguishable from a straight line. This is called local linearity, and it is the real content of differentiability: a function is differentiable at $x$ precisely when it looks like a line under sufficient magnification. Once the curve is locally a line, every secant through two nearby points on it has essentially that line's slope, so the quotients agree with each other and converge.

Work one example by hand, with $f(x) = x^2$:

$$\frac{(x+h)^2 - x^2}{h} = \frac{x^2 + 2xh + h^2 - x^2}{h} = \frac{2xh + h^2}{h} = 2x + h$$

The cancellation is the crucial step: the offending $h$ in the denominator divides out, and what remains, $2x + h$, is perfectly well behaved as $h \to 0$. So $f'(x) = 2x$. No division by zero ever occurred.

Notation varies by tradition and all of it means this same limit:

• $f'(x)$ — Lagrange's prime, compact and standard.
• $\dfrac{dy}{dx}$ — Leibniz's, which keeps the "rise over run" ancestry visible and makes the chain rule look like fractions cancelling.
• $\dot{x}$ — Newton's dot, still used in physics for derivatives with respect to time.

Rather than run the limit every time, a handful of rules cover almost everything you meet:

$$\frac{d}{dx}x^n = nx^{n-1}, \qquad \frac{d}{dx}\sin x = \cos x, \qquad \frac{d}{dx}e^x = e^x$$ $$(fg)' = f'g + fg', \qquad \left(\frac{f}{g}\right)' = \frac{f'g - fg'}{g^2}, \qquad \big(f(g(x))\big)' = f'(g(x))\,g'(x)$$

Every one of these is proved by pushing the difference quotient through the algebra above. The rules are labor-saving devices; the limit is the definition.

The derivative is a function#

Notice that $f'(x) = 2x$ is not a number. It is a rule — feed it a point, get the slope there. This is the shift that makes calculus powerful: the derivative of a function is itself a function, with its own graph, its own shape, and its own derivative.

The top panel is $f$, the bottom is $f'$, and they share an x-axis. As the point sweeps left to right, the gold tangent tilts on top and its slope is plotted as a height below. Press Sweep, then watch three correspondences:

• Where $f$ rises, $f'$ sits above its axis. Where $f$ falls, $f'$ sits below. Sign of the derivative = direction of travel.
• Where $f$ has a peak or a valley, the tangent goes flat and $f'$ crosses zero (both dots turn green). Maxima and minima of $f$ are zeros of $f'$ — this single fact is the engine of all of optimization.
• Where $f$ is steepest, $f'$ is farthest from zero. Magnitude of the derivative = how fast things change.

Switch to $x^3/3 - x$ and step the slider by hand through $x = -1$ and $x = 1$: the derivative $x^2 - 1$ touches zero exactly at the hump and the dip. Switch to the bell curve $e^{-x^2/2}$ and note something easy to miss — $f'$ is zero at the summit, but it is most negative on the way down the right flank, at the inflection point where the curve stops bending one way and starts bending the other.

Continuity is not enough: corners#

Here is the misconception worth correcting early. It is tempting to think that if a graph has no breaks or jumps — if you can draw it without lifting your pen — it must have a slope everywhere. It does not.

Take $f(x) = |x|$. It is continuous at $0$; nothing jumps. But compute the difference quotient at $x = 0$:

$$\frac{|0+h| - |0|}{h} = \frac{|h|}{h} = \begin{cases} +1 & h > 0 \\ -1 & h < 0 \end{cases}$$

Approach from the right and the quotient is $+1$ forever; approach from the left and it is $-1$ forever. There is no single value the quotient converges to, so the limit does not exist and $f$ has no derivative at 0. Zoom in on the corner as far as you like and it never straightens out — it is a corner at every magnification. Local linearity fails.

So differentiability is strictly stronger than continuity. Differentiable $\Rightarrow$ continuous (a function with a well-defined tangent cannot jump), but not the reverse. The gap is bigger than a few corners, too: Weierstrass constructed a function that is continuous everywhere and differentiable nowhere — infinitely crinkled at every scale, all corner and no slope.

This matters practically. Machine learning uses $\mathrm{ReLU}(x) = \max(0, x)$, which has exactly the corner of $|x|$ at the origin; frameworks simply pick a subgradient there and move on. And it explains why "smooth" is a real hypothesis in theorems rather than politeness: Taylor series need infinitely many derivatives to exist, and Newton's method needs a tangent line to slide down.

Where this leads#

The derivative is not one technique. It is the vocabulary that most of quantitative science is written in.

Physics. Position differentiated gives velocity; velocity differentiated gives acceleration. Newton's second law, $F = m\,\ddot{x}$, is a statement about a second derivative, which is why the laws of motion are differential equations rather than formulas.

Optimization. Because maxima and minima of a smooth function occur where $f' = 0$, "find the best" becomes "solve an equation". When the equation is too hard to solve exactly, you follow the slope downhill instead — that is gradient descent, and it is how neural networks with billions of parameters are trained. The derivative points uphill; step the other way.

Root-finding. Newton's method locates solutions of $f(x) = 0$ by replacing the curve with its tangent line, solving the easy linear problem, and repeating. It converges startlingly fast, and it is nothing but the derivative used as a local stand-in for the function.

Approximation. The tangent line $f(x+h) \approx f(x) + f'(x)h$ is the first-order approximation to $f$ near $x$. Keep going with higher derivatives and you get the Taylor series, which is how a calculator actually evaluates $\sin$ and $e^x$.

Everywhere else. Marginal cost in economics, reaction rates in chemistry, the option Greeks in finance, edge detection in image processing — each is a derivative wearing a domain-specific name.

All of it rests on the small move made at the start: refuse to divide by zero, take the limit instead, and an instantaneous rate becomes a perfectly ordinary number.