Push on a lump of clay and it squashes — but not always straight back the way you pushed. Shear a pack of cards and the top slides sideways relative to the bottom. Spin an asymmetric object, like a wonky spinning top or a smartphone flipped in the air, and it can wobble about an axis different from the one you spun it around. In every one of these situations a familiar rule breaks: push in one direction and the response points in a different direction. A single number cannot capture that, and neither can a single vector. You need an object that turns one direction into another — a tensor.

This page is about tensors as physical objects: the machines that relate vectors to vectors — stress relating a surface's orientation to the force on it, the inertia tensor relating spin to angular momentum, the electromagnetic field tensor packing \mathbf{E} and \mathbf{B} into one relativistic object. We build on the abstract picture from what is a tensor; here the goal is a physicist's fluency — reading the indices, feeling what each slot does, and never again confusing "an array of numbers" with "a tensor."

Scalars, vectors, tensors: a ladder of directions

Line up the objects physics uses by how many directions they carry:

A rank-2 tensor is a machine that eats a vector and spits out a vector, w_i = T_{ij}\,v_j (summing over the repeated j — the Einstein summation convention). Because it can rotate and rescale the vector as it passes through, the output need not be parallel to the input — which is exactly the "push here, respond there" behaviour we started with.

The stress tensor: force per area, but which force on which area?

Stress is the tensor to cut your teeth on. Inside a loaded solid, imagine slicing through a tiny internal surface. The material on one side pulls on the other with a force per unit area — the traction \mathbf{t}. But the traction depends on how you orient the cut: slice horizontally and you feel one force, slice vertically and you feel another. So you need one number for the face's direction and one for the force's direction — two indices:

where \mathbf{n} is the unit normal picking the face and \sigma_{ij} is the Cauchy stress tensor. The component \sigma_{ij} reads as "the i-component of force on the face whose normal points along j." The diagonal entries \sigma_{xx}, \sigma_{yy} are normal stresses (push/pull, pressure and tension); the off-diagonal \sigma_{xy} are shear stresses (dragging sideways). Reveal the figure step by step to see the tractions on each face of a little material element.

For a body in equilibrium the stress tensor is symmetric, \sigma_{ij} = \sigma_{ji} (otherwise the element would spin up under a net torque). Symmetry means it can be diagonalised: rotate to the right axes — the principal axes — and all shear vanishes, leaving only pure pushes and pulls along three perpendicular directions. Those diagonal values are the principal stresses, and this is just the eigenvalue problem again, wearing an engineer's hard hat.

More tensors you already half-know

Once you spot the pattern — a machine turning one vector into another, not-necessarily-parallel vector — tensors are everywhere:

• The inertia tensor I_{ij} relates angular velocity to angular momentum: L_i = I_{ij}\,\omega_j. For a lopsided body \mathbf{L} is not parallel to \boldsymbol{\omega} — which is precisely why a poorly balanced wheel shakes. Spin it about a principal axis and the wobble disappears.
• The strain tensor \varepsilon_{ij} encodes how a material deforms — stretch on the diagonal, shear off it. Hooke's law in 3-D links it to stress through a rank-4 tensor, the elastic stiffness \sigma_{ij} = C_{ijkl}\,\varepsilon_{kl}.
• The conductivity tensor \sigma_{ij} in an anisotropic crystal: J_i = \sigma_{ij}E_j, so current can flow at an angle to the applied field.
• The electromagnetic field tensor F^{\mu\nu} — a rank-2, four-dimensional, antisymmetric tensor that folds all six components of \mathbf{E} and \mathbf{B} into one object. Maxwell's four equations collapse to two beautifully compact tensor statements, and the way \mathbf{E} and \mathbf{B} mix when you change frames is just F^{\mu\nu} transforming its indices.

Index gymnastics with a physicist's eye

The index notation is not decoration — it is a calculus you can compute with. Three moves cover most of second-year physics:

• Summation (contraction). A repeated index means "sum over it," and it disappears: T_{ij}v_j has a free index i left, so it is a vector. Contract both indices, T_{ii} = T_{11}+T_{22}+T_{33}, and you get the trace — a scalar (for stress, related to pressure).
• Free vs. dummy indices. A free index appears once per term and must match on both sides of an equation (w_i = T_{ij}v_j — i free, j summed). A dummy (repeated) index is summed and can be renamed at will. If an index appears three times in a term, you have made a mistake.
• Symmetry. Split any tensor into a symmetric and an antisymmetric part, T_{ij} = \tfrac12(T_{ij}+T_{ji}) + \tfrac12(T_{ij}-T_{ji}). Stress and inertia are symmetric; the field tensor F^{\mu\nu} is antisymmetric. The symmetry tells you how many independent numbers there really are (a symmetric 3\times3 tensor has 6, not 9).

And the deepest point, the one that separates a tensor from a mere grid of numbers: a tensor's components transform in a definite way when you rotate your axes, T'_{ij} = R_{ik}R_{jl}T_{kl}. That transformation law is what guarantees the physics — the traction on a real surface, the torque on a real body — comes out the same whatever coordinates you happened to choose.

Applying a rank-2 tensor, in code

Watch the defining behaviour directly: apply a stress tensor to a face normal and see that the traction generally points in a different direction. Here a pure-shear stress (\sigma_{xy}=\sigma_{yx}=1, zero on the diagonal) acts on the x-face normal — the resulting traction comes out along y, perpendicular to the push:

The input pointed along x; the output points along y. That turning of one direction into another is the whole reason stress must be a tensor and not a vector.