Gradient, divergence, and curl

The formula

Everything runs off one operator, del (also called nabla), which is just the vector of partial derivatives:

∇ = ( ∂/∂x , ∂/∂y , ∂/∂z )

Point it at a scalar field f(x,y,z) and you get the gradient, a vector:

∇f = (∂f/∂x) x̂ + (∂f/∂y) ŷ + (∂f/∂z) ẑ

Reading: the gradient points in the direction f increases fastest, and its length is that steepest slope.

Dot it into a vector field F = (Fx, Fy, Fz) and you get the divergence, a scalar:

∇·F = ∂Fx/∂x + ∂Fy/∂y + ∂Fz/∂z

Reading: net outflow per unit volume — positive where the field is a source, negative where it's a sink.

Cross it into F and you get the curl, a vector:

∇×F = ( ∂Fz/∂y − ∂Fy/∂z ) x̂
    + ( ∂Fx/∂z − ∂Fz/∂x ) ŷ
    + ( ∂Fy/∂x − ∂Fx/∂y ) ẑ

Reading: the axis and strength of the local rotation — stick a tiny paddlewheel in the flow, and curl is twice its spin rate about each axis.

The directional derivative falls straight out of the gradient. The rate of change of f as you walk in a unit direction û is:

D_û f = ∇f · û

Reading: project the gradient onto the direction you're heading. Walk along a contour (perpendicular to ∇f) and it reads zero — you're going neither up nor down.

Chain divergence onto gradient and you get the Laplacian, the workhorse of heat and potential problems:

∇²f = ∇·(∇f) = ∂²f/∂x² + ∂²f/∂y² + ∂²f/∂z²

Where you meet it

Maxwell's equations, on any RF or EMC review board. All four are written in this vocabulary: ∇·D = ρ (charge is the source of the electric displacement), ∇·B = 0 (no magnetic monopoles, field lines close on themselves), ∇×E = −∂B/∂t and ∇×H = J + ∂D/∂t (a changing field curls up the other one). If you size a shielded enclosure, model an antenna, or argue about a ground plane, you are arguing about divergence and curl whether you write them or not.
Heat conduction on a thermal test stand. Fourier's law says heat flux is q = −k ∇T — heat flows down the temperature gradient, from hot to cold, and the minus sign is the whole physics. Set the divergence of that flux to the stored-energy rate and you get the heat equation, ρc ∂T/∂t = k ∇²T, which is what your thermal FEA is actually solving under the hood.
Fluid flow through a duct or pump loop. For an incompressible fluid, conservation of mass is exactly ∇·v = 0 — what flows into a control volume flows out. Vorticity, ω = ∇×v, is the field the CFD post-processor colors red when it wants to show you a separation bubble or a shed vortex off a strut.
Any structural or aero pressure field. Force per volume from a pressure field is −∇p. Wind loads, hydrostatic loads, the lift distribution on a fin — the load is the gradient of a scalar, pointing from high pressure toward low.

How it works

The three operators are not three unrelated tricks. They're the same ∇ applied three ways: straight (gradient), dotted (divergence), crossed (curl). If you can remember one vector of partial derivatives and how dot and cross products work, you can reconstruct all three at the bench without a reference card.

The gradient is always perpendicular to level surfaces. That's the fact people forget and then re-derive the hard way: on a contour map, the steepest path is straight across the contour lines, never along them. In optimization and in field problems alike, "perpendicular to the equipotentials" and "along the gradient" are the same statement.

Two identities are worth burning into memory because they show up as sanity checks constantly:

∇×(∇f) = 0        curl of a gradient is always zero
∇·(∇×F) = 0       divergence of a curl is always zero

The first says any field that is a pure gradient (a "conservative" field, like a static E-field or a gravity field) has no curl — no swirl, and the work you do going around any closed loop is zero. The second says any field written as a curl (like the magnetic field B = ∇×A) is automatically divergence-free, which is why ∇·B = 0 comes for free once you introduce the vector potential. Verified numerically: build f = sin(x)·y² + z·x, take the curl of its gradient at a random point, and you get (0, 0, 0); build a general F, take the divergence of its curl, and you get 8×10⁻¹⁷ — zero to machine precision.

A couple of quick numerical anchors so the definitions aren't abstract. The radial field F = (x, y, z) spreads outward everywhere, and its divergence is 1 + 1 + 1 = 3 — a uniform source, confirmed by finite differences at 3.0000. A rigid rotation about the z-axis, F = (−y, x, 0), has curl (0, 0, 2) — twice the rotation rate, which is exactly why curl equals twice the paddlewheel's angular velocity, not once.

The mistakes that bite:

Type errors. Gradient eats a scalar and gives a vector. Divergence eats a vector and gives a scalar. Curl eats a vector and gives a vector. Writing ∇·f on a scalar or ∇×(scalar) is nonsense, and it usually means you've lost track of what your field actually is. Check the types before you check the algebra.
Treating ∇ like an ordinary vector. It isn't — it's an operator, so order matters and it doesn't commute. F·∇ (a scalar operator that differentiates whatever sits to its right) is not the same object as ∇·F (the divergence, a plain scalar). The advective term in Navier–Stokes, (v·∇)v, is the first one, and swapping it for the second is a classic derailment.
Forgetting the coordinate system. The clean forms above are Cartesian. In cylindrical or spherical coordinates the operators pick up extra terms and metric factors — divergence in spherical coordinates has a 1/r² and an r² buried in it. If your duct is round or your antenna pattern is spherical and your hand-calc disagrees with the solver by a factor of r, you almost certainly used the Cartesian form where you shouldn't have.
Assuming curl-free means safe to use a potential. A field can have zero curl locally and still not admit a single-valued potential if the domain has a hole in it (think of the field around a current-carrying wire). The topology of the region matters, not just the local derivative.

The limit of validity is smoothness: these are derivatives, so they assume the field is differentiable. At a material interface, a shock, or a point charge, the field isn't smooth and the operators either blow up or have to be interpreted in the integral (weak) sense. That integral form is the deeper truth — the divergence theorem and Stokes' theorem tie these local operators to net flux and circulation through real surfaces — but the point-wise formulas are what you compute with day to day.

History

The operator started life inside quaternions. William Rowan Hamilton, in Dublin, built the quaternion algebra in the 1840s and wrote a differential operator that carried the seed of everything here ^[1]^[2]. The upside-down triangle symbol got its odd name almost as a joke: in an 1870 letter, the encyclopedist William Robertson Smith told Peter Guthrie Tait that the shape looked like a nabla, an ancient Hellenistic-Greek harp with a triangular frame. Tait and James Clerk Maxwell used the word in private correspondence, half in fun, and it stuck ^[1]^[2].

Maxwell did some of the naming — but only some, and not the way the folklore has it. Casting about for words to describe what ∇ does, he sent Tait a list of "rough-hewn names" in a November 1870 letter: "slope" for what we now call the gradient, "convergence" for the scalar part of ∇ on a vector, and for the vector part "twist" — which he fretted suggested a screw — then the "sufficiently racy" "twirl," before landing on "curl" ^[6]. That last one he aired in his 1871 paper "On the Mathematical Classification of Physical Quantities" ^[9] and put into his 1873 Treatise on Electricity and Magnetism "with great diffidence" ^[6] — so of our three modern names, Maxwell coined exactly one. His "convergence" even carries the opposite sign: "divergence" is William Kingdon Clifford's word, introduced in his Elements of Dynamic (1878) as the negative of Maxwell's quantity, so that a source reads positive ^[7]. "Gradient" came last of all. Maxwell stuck with "slope" in the Treatise and swapped it for "space-variation" in the second edition; the modern word only settled in around the turn of the century, with Horace Lamb's 1897 calculus text the earliest known use ^[8].

Two men, working apart, did the cleanup and gave us modern vector calculus ^[4]. Josiah Willard Gibbs at Yale printed lecture notes titled Elements of Vector Analysis privately for his students in 1881 and 1884 — "a pamphlet printed for private distribution," as Gibbs himself later described it, never formally published ^[3]^[5]. Oliver Heaviside, a self-taught Englishman whose schooling stopped at sixteen, independently assembled the same grad-div-curl toolkit in his electrical papers of the 1880s and turned it loose on Maxwell's theory ^[3]^[4]. The two systems met in 1888, when Heaviside received a copy of Gibbs's pamphlet and found in it, as he later recounted, essentially the system he had already been led to on his own ^[3]. The quaternion loyalists did not take it well — Tait damned Gibbs's pamphlet as "a sort of hermaphrodite monster, compounded of the notations of Hamilton and of Grassmann" ^[3]. The modern textbook form was locked in by Edwin Bidwell Wilson's 1901 Vector Analysis, built from Gibbs's lectures, which also argued for pronouncing ∇ as the short, convenient "del" ^[1]^[2]^[5]. That's the name most engineers say at the whiteboard today.

Related tools

/tools/heat-conduction
/tools/heat-exchanger-duty
/tools/thermal-expansion
/tools/rf-power-density
/tools/antenna-gain-beamwidth
/tools/bernoulli-pressure
/tools/dynamic-pressure
/tools/pipe-flow-velocity
/tools/pressure-vessel-hoop

Sources

Written by HE in our own words from the cited sources — engineering judgment included, your stamp still required. All entries →