fx and fy are the partial derivatives at the
point (a, b), and dx and dy are the changes in x and y. The card takes the two partials and the two steps and
returns df. It is the change of the tangent plane between
(a, b) and (a + dx, b + dy), so df is also z − f(a, b) on that plane.
Holds as an estimate of the change Δf = f(a + dx, b + dy) − f(a, b) when f is differentiable at (a, b): then
Δf − df is small compared with the length of the step √(dx² + dy²), so the error of the estimate shrinks faster than the
step itself. If f has continuous partial derivatives near (a, b), it is differentiable there. The existence of the
two partials at (a, b) is not enough.
Worked example: x²y at (2, 3)
- The partials: fx = 2xy = 12 and fy = x² = 4 at (2, 3).
- Steps dx = 0.1 and dy = −0.05.
- df = 12 · 0.1 + 4 · (−0.05) = 1.2 − 0.2 = 1.0.
The actual change is Δf = 2.1² · 2.95 − 12 = 13.0095 − 12 = 1.0095. The differential is off by 0.0095, which is
what expanding (2 + dx)²(3 + dy) leaves out: 3dx² + 4 dx dy + dx² dy, terms of second and third order in the steps.
When the partials exist and the differential is wrong
Let f(x, y) = xy/(x² + y²) with f(0, 0) = 0. Along the axes f is 0, so both partial derivatives at the origin are 0,
and the formula gives df = 0 for every step. But f(h, h) = 1/2 for every h ≠ 0, so the actual change along the
diagonal is 1/2, however short the step. f is not even continuous at the origin, so it is not differentiable
there, and df is not an estimate of anything.
Differential or change
- Δf is the change in f itself, for a finite step.
- df is the change along the tangent plane, the first-order part.
- dx and dy are free steps. They are not the partial derivatives and are not found from f.
With dx and dy as measurement errors, the sum |fx| |dx| + |fy| |dy| is the
error propagation formula, with the signs removed to get the
worst case. Written for one variable, df = f′(x) dx is the linear approximation.
Common mistakes
- Reading df as Δf. They agree to first order only.
- Taking the partials at the end point. They belong at (a, b), where the step starts.
- Treating “the partials exist” as enough. The example above has both and no estimate.
- Mixing the steps. fx goes with dx and fy with dy.
Related: chain rule with two parameters,
directional derivative.