The formula
f'(x) = df/dx = lim [ f(x + h) - f(x) ] / h
h→0
The slope of the secant line between two points, taken to the limit as those points collapse onto one. That limit — if it exists — is the slope of the tangent right there.
dy/dx at x = a — the rate of change of y with respect to x, evaluated at a
Leibniz's notation, and it earns its keep: it reads like a ratio and behaves like one under the chain rule.
The rules you actually use at the bench:
Power: d/dx (x^n) = n·x^(n-1)
Product: d/dx (u·v) = u'·v + u·v'
Quotient: d/dx (u/v) = (u'·v - u·v') / v²
Chain: d/dx f(g(x)) = f'(g(x)) · g'(x)
The chain rule is the one that matters most in real work: it says rates multiply through a cascade. Temperature changes with time, resistance changes with temperature, voltage changes with resistance — multiply the three slopes and you have how fast the voltage moves.
Common derivatives worth memorizing:
d/dx sin(x) = cos(x) d/dx e^x = e^x
d/dx cos(x) = -sin(x) d/dx ln(x) = 1/x
The second derivative — the derivative of the derivative — is the rate of change of the rate:
f''(x) = d²f/dx²
Position → velocity → acceleration. Each arrow is one derivative.
Where you meet it
Slew rate on a scope. A step command asks an amplifier or actuator to jump instantly. It can't. The output ramps, and the steepest slope it can hit is the slew rate — a derivative with units of volts per microsecond. A 2 V edge that takes 1 µs is a slew rate of 2,000,000 V/s. If the input demands a faster dV/dt than the part can deliver, you get slewing distortion, and the review board wants to know why the settling time blew the spec.
Thermal ramp on a test stand. An environmental profile calls for 5 °C per minute — that's dT/dt, a derivative, and it's 0.0833 °C/s. Push the ramp rate past what the fixture's thermal mass can follow and the part lags the chamber, so the thermocouple and the setpoint disagree. The controller's PID loop is itself running on the error's derivative to anticipate overshoot.
Sensitivity analysis. You want to know how much a bolt-preload output shifts when torque coefficient drifts. That's a partial derivative ∂(output)/∂(input) — the sensitivity coefficient. Every uncertainty budget and every error stack-up in a calibration lab is built from these. A big derivative means a touchy parameter you'd better control tightly.
Gain slope on a Bode plot. The roll-off of a filter — 20 dB per decade for a single pole — is a derivative of gain with respect to log-frequency. The phase margin you read off to check stability comes from where those slopes land.
How it works
The derivative is a local statement. It tells you the slope at one point and nothing about two inches down the road. A function can be perfectly smooth here and have a kink an instant later. The mistake people make is treating a derivative computed at the operating point as if it held everywhere — that's linearization, and it's only trustworthy in a small neighborhood. Push far from the point and the real curve walks away from your tangent line.
Differentiability is stricter than continuity. A function can be continuous and still have no derivative at a point — the absolute-value corner at x = 0 is the classic case. The left slope is -1, the right slope is +1, and there's no single tangent. Any sharp corner, any switch-over, any place a signal changes rule — that's where the derivative doesn't exist, and where numerical code that assumes it does will misbehave.
Differentiation amplifies noise. This is the big one on real hardware. Taking a derivative divides a small change by a small time step, and any high-frequency wiggle in your data has a large slope even if its amplitude is tiny. Differentiate a noisy sensor stream raw and you get garbage — the noise dominates the signal. In the frequency domain, an ideal differentiator has gain proportional to frequency, so it boosts exactly the high-frequency junk you don't want. The fix is to low-pass filter first, or fit a smooth curve and differentiate that. Integration averages noise out; differentiation piles it up. Respect the asymmetry.
Numerical derivatives: use the centered difference. When you only have sampled data, approximate the slope. The naive forward difference [f(x+h) - f(x)] / h carries error proportional to h. The centered difference [f(x+h) - f(x-h)] / (2h) carries error proportional to h² — vastly better for the same step. Concretely, for sin(x) at x = 1 with h = 0.001, the true slope is cos(1) = 0.54030. The forward difference is off by about 4×10⁻⁴; the centered difference is off by about 9×10⁻⁸ — four orders of magnitude tighter from the same data. Default to centered unless you're stuck at a boundary. And don't chase h to zero: too small a step and floating-point roundoff in the subtraction takes over, so there's a sweet spot — roughly the square root of machine epsilon for a forward difference, and roughly the cube root for the centered one. In double precision that puts the centered-difference step near 10⁻⁵ times the scale of x.
Units carry through. A derivative's units are always the output's units over the input's units. dV/dt is volts per second. d(stress)/d(strain) is Young's modulus. If your units don't come out right, the derivative is wrong before you check the arithmetic.
History
Nobody invented the derivative in a single stroke. The tangent problem — find the slope touching a curve at one point — was chewed on for a long time before it had a machine to grind it. In the late 1620s Pierre de Fermat had a working method for locating maxima and minima and for drawing tangents: he nudged the variable by a tiny amount, formed the difference, divided, and then discarded the leftover small term. Lagrange later thought so much of it that he named Fermat the real inventor of the calculus [1]. Fermat had the mechanism; he didn't have the general theory tying it to areas.
That tie came from two men working apart. Isaac Newton wrote a tract on what he called fluxions in October 1666, at 23 — his fluxions were rates of change flowing in time, and the same manuscript held the first clear statement of the fundamental theorem linking derivatives to integrals [1][2]. Newton, being Newton, sat on it and didn't publish. Gottfried Wilhelm Leibniz came at the same ideas from Paris in the mid-1670s, and his notebooks show the dy/dx and ∫ notation taking shape by 1675 [2]. Leibniz published first, in a 1684 paper titled Nova Methodus pro Maximis et Minimis — "a new method for maxima and minima" [2].
Then it turned ugly. British and Continental camps fought for decades over who got there first, and in 1712 the Royal Society convened a committee — quietly steered by Newton himself — that ruled for Newton and against Leibniz for plagiarism [2]. The modern verdict is that both got there independently. The lasting irony is that Leibniz lost the fight but won the notation war: every engineer today writes dy/dx his way, not Newton's dotted fluxion, because Leibniz's symbols simply worked better for calculation [1][2].
Related tools
- /tools/stress-strain — Young's modulus is the derivative of stress with respect to strain
- /tools/beam-deflection — slope and curvature of a beam are successive derivatives of deflection
- /tools/rc-filter — the roll-off slope on a Bode plot is a derivative of gain vs log-frequency
- /tools/doppler-shift — frequency shift tracks the derivative of range, the closing velocity
- /tools/projectile-range — velocity is the derivative of position, acceleration the second derivative
- /tools/vibration-natural-freq — equations of motion are written in derivatives of displacement