linear transformation
/ LIN-ee-ur trans-for-MAY-shun /
A linear transformation is a way of reshaping space that keeps everything orderly: straight lines stay straight, the origin (the zero point) stays put, and evenly spaced grid lines stay evenly spaced — they may be stretched, squashed, rotated, or sheared, but never bent or torn. Picture a sheet of stretchy graph paper that you can pull, spin, or slant, but never crumple. That smooth, even reshaping is exactly what 'linear' means here.
Every linear transformation can be captured by a single matrix, and applying the transformation to a vector is just multiplying that matrix by the vector. This is the deep reason matrices matter: a grid of numbers is really a compact recipe for moving and reshaping space. Two formal rules pin it down — scaling an input scales the output by the same factor, and the transform of a sum equals the sum of the transforms. In plain terms, the parts add up with no surprises.
In machine learning, each layer of a neural network starts by applying a linear transformation: multiply the inputs by a weight matrix to mix and reshape them. But here is the catch worth remembering — stacking linear transformations only ever gives you another linear transformation, no matter how many you pile up. That is why networks insert a nonlinear step (an activation function) between layers. Without it, a hundred-layer network would collapse into the power of a single layer, and deep learning would not be deep at all.
The matrix [ [2, 0], [0, 1]] stretches everything to twice its width while leaving height alone — a square becomes a wide rectangle. The matrix [ [0, -1], [1, 0]] rotates the whole plane a quarter-turn counterclockwise. In both cases the origin never moves and grid lines stay straight and evenly spaced: that is what makes them linear.
A matrix is a verb: [ [2,0],[0,1]] says 'stretch wide,' [ [0,-1],[1,0]] says 'rotate.' Same grid, reshaped without bending.
A strictly linear transformation must keep the origin fixed; adding a shift (a bias term) makes it 'affine,' not linear. The reason deep networks need activation functions is precisely that linear-on-top-of-linear is still just linear — nonlinearity is what unlocks their power.