Nonlinear maps · local linear approximation

The Jacobian
Matrix

Every smooth map, seen closely enough, is linear. The Jacobian is that linear map — the derivative of a function that eats vectors and returns vectors.

The idea

A grid of derivatives

Take a map that bends the plane — F(x, y) = (u, v), sending points to points. Near any single point p, zoom in. The bending flattens out, and what remains is a plain linear transformation: a stretch, a rotation, a shear. That transformation is a 2×2 matrix of partial derivatives, one for each way an output can respond to each input:

J(p) =  [  ∂u/∂x  ∂u/∂y  ]
       [  ∂v/∂x  ∂v/∂y  ]

The one fact that makes it worth knowing: F(p + δ) ≈ F(p) + J(p) δ. Feed a small displacement into the matrix and you get the map's response — no calculus, just matrix × vector. Drag the point below and watch the matrix update.

Instrument I · deformation

Watch a circle become an ellipse

Left is the input plane; right is where F sends it. The gold point is yours to drag. Around it sits a small circle. The two arrows are the input axes — and their images on the right are exactly the two columns of the Jacobian. The gap between the wobbly true image and the clean ellipse is the nonlinearity the matrix throws away.

Input planep = (0, 0)
Image plane  F(·)u, v
x̂ → column 1 ŷ → column 2 circle, true image Jacobian ellipse (linear)

Jacobian at p

symbolic
=

Determinant  =  local area factor

0.00
signed area of the
column parallelogram
Instrument II · what "derivative" means

Zoom until the curve is a line

Here is the same neighborhood of p, centered and magnified. The gold grid is the map's true image; the light dashed grid is the Jacobian's flat approximation. Shrink the window: the two snap together. That convergence — true map meets linear map as scale → 0 — is what it means for J to be the derivative.

Neighborhood of F(p)center
true image of the grid linear approximation  F(p) + Jδ
Reading it

The determinant, three ways

The columns of J span a parallelogram — the image of the unit square. Its signed area is det J, and the sign and size each carry a distinct message:

Magnitude

How much space stretches

|det J| is the local area scale. Above 1, the map expands this patch; below 1, it squeezes it; exactly 1, it preserves area (a rotation, a shear, an incompressible flow).

Sign

Whether it flips

Positive keeps orientation intact. Negative means the map mirror-reverses the neighborhood — clockwise becomes counter-clockwise. Try the Fold map and cross the line x = 0.

Zero

Where it collapses

det J = 0 crushes the patch onto a line or point. The map is not locally invertible there — information is lost, and you can't run it backwards.

Shape

Circle to ellipse

A tiny circle maps to an ellipse whose axes are J's singular values σ₁, σ₂. Their product is |det J| — the ellipse's area over the circle's.

Where it shows up

The same matrix, everywhere

Once you can build the best local linear map, a surprising amount of mathematics is just choosing what to do with it.

Integration

Change of variables

∫∫F(R) g  dU = ∫∫R g ∘ F  |det J| dA

When you warp coordinates, area warps with them. The Jacobian determinant is the correction factor that keeps the integral honest — the reason polar coordinates carry that extra r.

Root finding

Newton's method, upgraded

xₙ₊₁ = xₙ − J(xₙ)⁻¹ F(xₙ)

Exactly the 1-D Newton step you know, with 1/f′ replaced by J⁻¹. Solve a nonlinear system by repeatedly solving its linear shadow.

Composition

Chain rule = matrix product

Jg∘f(x) = Jg(f(x)) · Jf(x)

Stack two maps and their Jacobians multiply — in order. This one line is the entire engine of backpropagation.

Robotics

Velocity kinematics

v₋₃₋ = J(θ)  θ̇

The Jacobian maps joint speeds to hand speed. Where det J = 0, the arm hits a singularity — a pose it can't move smoothly out of.