Tangent Planes & Linear Approximation

Kneel down in a field and the ground looks perfectly flat, even though you are standing on a giant curved ball spinning through space. Zoom in on any smooth surface — a sand dune, a soap bubble, the Earth itself — and close enough to one point it becomes almost indistinguishable from a flat sheet touching it there. That is exactly why an ordinary paper map works beautifully for finding your way around town, even though the very same flat sheet would be hopeless stretched over the whole curved planet. Calculus turns that "close enough to flat" feeling into an exact formula.

For a one-variable curve, the tangent line is the best straight-line approximation near a point. A surface z = f(x, y) has two partial derivatives, one slope in the x direction and one in the y direction, and together they pin down the best flat approximation — a tangent plane. At a point (a, b) it is

z = f(a, b) + f_x(a, b)\,(x - a) + f_y(a, b)\,(y - b).

Read it as a recipe: start at the height f(a, b), then add the rise from moving (x - a) in the x direction at slope f_x, and the rise from moving (y - b) in the y direction at slope f_y. Each x-slice of this plane is exactly the tangent line to the corresponding slice of the surface.

Building the plane, step by step

Build the tangent plane to the paraboloid f(x, y) = x^2 + y^2 at the point (a, b) = (1, 1), then use it to estimate f(1.1, 0.9) without squaring anything.

Step 1 — the partials. Differentiate in each variable, holding the other constant:

f_x = 2x, \qquad f_y = 2y.

Step 2 — evaluate at the point (1, 1). The height and the two slopes are

f(1, 1) = 1 + 1 = 2, \qquad f_x(1, 1) = 2, \qquad f_y(1, 1) = 2.

Step 3 — assemble the plane. Substitute into the template:

z = 2 + 2\,(x - 1) + 2\,(y - 1).

Step 4 — read it as a linearization. The right-hand side is the linear approximation L(x, y) of f near (1, 1):

L(x, y) = 2 + 2(x - 1) + 2(y - 1).

Step 5 — estimate f(1.1, 0.9). Here x - 1 = 0.1 and y - 1 = -0.1, so the two corrections cancel:

L(1.1, 0.9) = 2 + 2(0.1) + 2(-0.1) = 2 + 0.2 - 0.2 = 2.

Step 6 — check against the truth. The exact value is

f(1.1, 0.9) = 1.1^2 + 0.9^2 = 1.21 + 0.81 = 2.02.

The estimate 2 is off by only 0.02 — the error is second-order in the small step, exactly as the tangent-line story promised, now in two variables.

It's worth seeing exactly where that 0.02 comes from. The differential predicted df = f_x\,dx + f_y\,dy = 2(0.1) + 2(-0.1) = 0, which is precisely L(1.1, 0.9) - f(1, 1) — the plane's own change. The true change, f(1.1, 0.9) - f(1, 1) = 0.02, differs from that by exactly dx^2 + dy^2 = 0.01 + 0.01 = 0.02. That squared leftover is the second-order term the differential is built to ignore — and it's also why the trick works so well here: the two first-order corrections happened to cancel, leaving only the (tiny) curvature term visible.

The total differential

Writing dx = x - a and dy = y - b for the small displacements, the change in height predicted by the plane is the total differential

df = f_x\, dx + f_y\, dy.

It is the workhorse of error propagation: a small wiggle dx in one input and dy in the other produce, to first order, a change df in the output. Drag the point a below to watch the tangent line track the slice of the paraboloid and read off the linearization there.

Let f be differentiable at (a, b). Then:

Worked example: tolerance on a measured volume

The total differential turns into a practical tolerance check whenever a quantity is assembled from measured, error-prone ingredients. Suppose a sheet-metal box has a fixed height h = 5 cm (cut exactly), but its base is a length l = 8 cm and a width w = 6 cm, each set by a machine accurate only to dl = 0.1 cm and dw = 0.05 cm. How far off could the finished volume be?

Step 1 — write volume as a function of the two measured variables. With h fixed, V(l, w) = lwh is a function of exactly two variables, just like f(x, y) above.

Step 2 — the partials.

V_l = wh, \qquad V_w = lh.

Step 3 — evaluate at (l, w, h) = (8, 6, 5):

V_l = 6 \times 5 = 30, \qquad V_w = 8 \times 5 = 40.

Step 4 — the total differential.

dV = V_l\, dl + V_w\, dw = 30(0.1) + 40(0.05) = 3 + 2 = 5\ \text{cm}^3.

The nominal volume is V = 8 \times 6 \times 5 = 240\ \text{cm}^3, so the two measurement tolerances could push the true volume off by roughly 5\ \text{cm}^3 — about 2\% — without ever having to multiply out a worst case exactly. This is precisely how engineers turn a spec sheet of individual tolerances into one overall error budget, and it works because dV is linear in dl and dw: double a tolerance and its contribution to the error simply doubles.

Having partial derivatives is not enough to guarantee a tangent plane — the classic culprit is a cone. Take z = \sqrt{x^2 + y^2}, an ice-cream-cone shape with its point sitting at the origin. Walk away from the tip in any direction and the surface climbs at the same steady rate — every directional slope you measure there looks identical. And yet there is obviously no single flat plane you could balance on that sharp point: it's a corner, not a smooth cap. (The partials f_x(0,0) and f_y(0,0) don't even exist in the ordinary two-sided sense for this cone — the slope flips sign depending which way you approach the tip from.)

Even functions whose partials do exist everywhere can still fail to have a tangent plane. In one variable, "the derivative exists" and "the function is differentiable" are the same statement. In two variables they come apart, and the gap is a genuine trap. A function can have both partial derivatives at a point and yet not be differentiable there — not even continuous. The standard specimen is

f(x, y) = \frac{xy}{x^2 + y^2} \quad (f(0, 0) = 0).

Along the axes f is identically zero, so f_x(0, 0) = f_y(0, 0) = 0 both exist. But we already saw this function has no limit at the origin — so it is not even continuous there, let alone differentiable. The partials see only two directions; differentiability demands the tangent plane approximate f from every direction at once.

The clean sufficient condition: if f_x and f_y exist and are continuous near (a, b), then f is differentiable there (such an f is called C^1). For everyday functions — polynomials, exponentials, sines — this always holds, which is why the tangent-plane formula can be applied without anxiety.

The Earth is a sphere, and a sphere has no tangent plane that fits it globally — try to flatten an orange peel and it always tears or wrinkles somewhere. But zoom in on any one spot, and the curved surface and its tangent plane are almost the same thing: the gap between them shrinks faster than the distance you move away, exactly as it does for f(x, y) = x^2 + y^2 above. That is why a street map, accurate over a few kilometres, can pretend the Earth is flat and get away with it — while a map of the whole globe squashed onto one flat sheet must distort something (Greenland famously balloons to the size of Africa on a Mercator projection).

Surveyors and cartographers exploit this on purpose: pick a plane tangent to the Earth at the town you care about, project onto it, and the map stays honest over a useful radius around that point — the very same "match the height, match both slopes" idea behind the tangent-plane formula, just scaled up from a paraboloid to a planet.

See the plane kiss the surface

Near a point the surface is almost flat — the tangent plane is the flat sheet that just kisses it there. Here is the bowl z = \tfrac12(x^2 + y^2) as a rotatable surface, with its tangent plane laid across a small patch at (a, b) = (0.8, -0.6). Drag to spin it: the shaded sheet touches the surface at the marked point and matches both of its slopes there, hugging the bowl closely nearby and drifting away only as you move off — exactly the second-order error the linearization is built to ignore.