I want to start with a confession. When I first heard the word “tensor,” I assumed it was one of those words that exists to make physicists sound clever. Something you learn in graduate school, surrounded by people who already know what it means, and everyone pretends they understood it the first time. It is not like that at all. A tensor is a completely natural object. It is what you are forced to invent when you ask a simple question about a crystal and discover that vectors are not enough to answer it.
Here is the question. You take a crystal, say a piece of calcite, the kind that makes double images when you look through it. You apply an electric field $\vec{E}$ to it. The field pushes the charges inside the crystal a little, and the crystal develops a dipole moment per unit volume, which we call the polarization $\vec{P}$. For a material like water or glass, which looks the same in every direction, the polarization is simply proportional to the field and points in the same direction: $\vec{P} = \alpha \vec{E}$, where $\alpha$ is a single number. One number, and you are done.
A crystal is different. Its atoms sit in a rigid lattice that has a definite structure, and that structure is not the same in every direction. Think of it mechanically. Imagine the atoms connected by springs, but the springs in one direction are stiff and the springs in a perpendicular direction are loose. If you push with a force at some angle, the charges move mostly along the loose direction and barely at all along the stiff direction. The displacement, which is the polarization, comes out pointing somewhere completely different from the force. The polarization is not parallel to the electric field. One number $\alpha$ cannot describe this, because one number cannot encode a relationship that depends on direction in this asymmetric way.
So what do you need? Let us work it out from scratch, without assuming anything. Suppose you apply a field purely in the $x$-direction, magnitude $E_x$. Where does the polarization go? There is no reason it has to go in the $x$-direction. It will have some $x$-component, some $y$-component, and some $z$-component. Since polarization is proportional to field strength (we are assuming small fields), each component of $\vec{P}$ is proportional to $E_x$, and we just have to give the proportionality constants names. Call them $\alpha_{xx}$, $\alpha_{yx}$, and $\alpha_{zx}$, where the first subscript tells you which component of $\vec{P}$ you are measuring and the second tells you which direction the field is in: $$P_x = \alpha_{xx}\, E_x, \qquad P_y = \alpha_{yx}\, E_x, \qquad P_z = \alpha_{zx}\, E_x.$$ Three numbers to describe what happens when you push in the $x$-direction. Now do the same for a field in the $y$-direction: $P_x = \alpha_{xy} E_y$, $P_y = \alpha_{yy} E_y$, $P_z = \alpha_{zy} E_y$. Three more numbers. And three more for a field in the $z$-direction. Nine coefficients in total.1The subscript ordering $\alpha_{ij}$ means: $i$ labels the component of $\vec{P}$ being produced, $j$ labels the direction of $\vec{E}$ that produces it. So $\alpha_{yx}$ is the coefficient that tells you how much $y$-polarization is produced by an $x$-field. Keep this straight from the beginning or the indices will confuse you forever.
Now suppose the field has all three components at once. Because polarization is linear in the field (doubling the field doubles the polarization), the contributions from each direction simply add. The $x$-component of $\vec{P}$ gets a contribution from the $x$-field (coefficient $\alpha_{xx}$), a contribution from the $y$-field (coefficient $\alpha_{xy}$), and a contribution from the $z$-field (coefficient $\alpha_{xz}$). Adding them up: $$\begin{aligned} P_x &= \alpha_{xx} E_x + \alpha_{xy} E_y + \alpha_{xz} E_z, \\ P_y &= \alpha_{yx} E_x + \alpha_{yy} E_y + \alpha_{yz} E_z, \\ P_z &= \alpha_{zx} E_x + \alpha_{zy} E_y + \alpha_{zz} E_z. \end{aligned}$$ These three equations can be written compactly as $$P_i = \sum_j \alpha_{ij} E_j$$ where $i$ and $j$ each run over $x$, $y$, $z$. The nine numbers $\alpha_{ij}$ are called the polarizability tensor. That is where tensors come from. Not from abstraction. From the forced recognition that describing direction-dependent proportionality requires not one number but nine.
Look at that table for a moment before moving on. If you apply an $x$-field, you read down the first column: you get some $x$-polarization ($\alpha_{xx}$), some $y$-polarization ($\alpha_{yx}$), and some $z$-polarization ($\alpha_{zx}$). If you apply a $y$-field, you read down the second column. The whole table encodes the complete directional response of the crystal. This is what a tensor is: a table of coefficients that tells you, for each possible direction of the input, what the output is in each possible direction. Not a mysterious geometric object. A table with a physical meaning for each entry.
Now something wonderful happens. Of the nine numbers, not all are independent. Here is the argument. Think about the energy you spend polarizing the crystal. The work done per unit volume when you polarize it is $u_P = \frac{1}{2}\vec{E}\cdot\vec{P}$. Substitute $P_i = \sum_j \alpha_{ij} E_j$ and you get $$u_P = \frac{1}{2}\sum_{i,j} \alpha_{ij} E_i E_j.$$ Now imagine a thermodynamic cycle. Turn on an $x$-field, then turn on a $y$-field, then turn off the $x$-field, then turn off the $y$-field. The crystal is back where it started. The net work done must be zero. When you compute the work around this cycle carefully, the terms proportional to $\alpha_{xx}$ and $\alpha_{yy}$ cancel out cleanly. The net work is proportional to $\alpha_{xy} – \alpha_{yx}$. For this to vanish, you need $\alpha_{xy} = \alpha_{yx}$. The same argument applies to any pair of indices, so the tensor is symmetric: $\alpha_{ij} = \alpha_{ji}$.2This symmetry is a physical consequence of energy conservation, not a mathematical assumption. There exist tensors in physics that are not symmetric, for example the angular momentum flux tensor, and they genuinely need all nine components. Always ask which physical law enforces symmetry before counting independent components. Symmetry cuts nine components to six independent ones. That is not a small reduction. In higher-rank tensors, symmetry can reduce hundreds of components to just a handful, which is why physicists care about it so much.
Six is still more than one. But here is where the geometry becomes beautiful. The energy formula $u_P = \frac{1}{2}\sum_{i,j} \alpha_{ij} E_i E_j$ is a quadratic function of the field components. In two dimensions it would look like $\alpha_{xx} E_x^2 + 2\alpha_{xy} E_x E_y + \alpha_{yy} E_y^2 = 2u_0$. Ask: which electric field vectors $\vec{E}$ produce exactly the same energy density $u_0$? The set of all such vectors traces out a curve in the $(E_x, E_y)$ plane. Since the energy is always positive and finite for any nonzero field, that curve is always an ellipse, never a parabola or hyperbola. In three dimensions, the set of all $\vec{E}$ producing the same $u_0$ traces out an ellipsoid. The shape and orientation of this ellipsoid encodes everything about the tensor.
Every ellipsoid has three perpendicular axes called the principal axes, the directions along which it is longest, shortest, and intermediate. When you align your coordinate system with those axes, something remarkable happens to the tensor: all the off-diagonal components vanish. The tensor becomes diagonal: $$\alpha_{ij} = \begin{pmatrix} \alpha_{aa} & 0 & 0 \\ 0 & \alpha_{bb} & 0 \\ 0 & 0 & \alpha_{cc} \end{pmatrix}.$$ Along the principal axes, the polarization is parallel to the field, with one proportionality constant per axis. The crystal may be complicated in any arbitrary direction, but it has three special directions where it behaves like an isotropic material, each with its own effective $\alpha$. This is always possible for any symmetric tensor of rank two, in any number of dimensions. It is one of the deepest structural facts about this type of object, and it has nothing to do with the specific physics of crystals. It is purely geometric.
If all three principal values are equal, $\alpha_{aa} = \alpha_{bb} = \alpha_{cc} = \alpha$, the ellipsoid is a sphere and the material is isotropic. The tensor reduces to $\alpha_{ij} = \alpha \delta_{ij}$, where $\delta_{ij}$ is one if $i = j$ and zero otherwise. This is called the Kronecker delta, and it plays the role of the identity: $P_i = \alpha \sum_j \delta_{ij} E_j = \alpha E_i$, which is just $\vec{P} = \alpha \vec{E}$. The tensor framework contains the simple isotropic case as a special instance.3The Kronecker delta $\delta_{ij}$ has the remarkable property that it looks identical in every coordinate system. You can rotate your axes by any angle and its components stay the same: ones on the diagonal, zeros everywhere else. This is because it is proportional to the metric of flat Euclidean space, which has no preferred direction. The identity tensor is the one tensor that genuinely has no directional content.
The polarizability of a crystal is one example, but the same structure appears everywhere in physics. The conductivity of a crystal is a tensor: $j_i = \sum_j \sigma_{ij} E_j$ because the current density $\vec{j}$ and the electric field $\vec{E}$ are generally not parallel in a crystal. The moment of inertia is a tensor. You may have seen the scalar moment of inertia $I$ in a basic mechanics course, with $L = I\omega$. But this only works when you spin an object about one of its symmetry axes. For a general axis, the angular momentum $\vec{L}$ and the angular velocity $\vec{\omega}$ are not parallel, in exactly the same way that $\vec{P}$ and $\vec{E}$ are not parallel in an anisotropic crystal.4The inertia tensor has the explicit form $I_{ij} = \sum_\text{particles} m(r^2 \delta_{ij} – r_i r_j)$, where the sum is over all particles in the body and $r^2 = x^2+y^2+z^2$. The diagonal component $I_{xx} = \sum m(y^2+z^2)$ is exactly the moment of inertia about the $x$-axis that you saw in introductory mechanics. The off-diagonal components $I_{xy} = -\sum m\, xy$ are the products of inertia, and they vanish when your axes align with the principal axes of the body.
Now for the question that most introductions skip. The nine components of $\alpha_{ij}$ were measured with respect to some particular set of coordinate axes. If you rotate your axes, the components change. Does the tensor change? No. The crystal is the same crystal. The physical relationship between $\vec{P}$ and $\vec{E}$ is unchanged. What changes is only the numerical values of the components, in a specific and predictable way determined by how you rotated the axes. The components are not the tensor. They are the tensor’s representation in a particular coordinate system, just as the components $(E_x, E_y, E_z)$ are not the electric field but its representation in a particular coordinate system.
The transformation law for a rank-2 tensor under a rotation is $$\alpha’_{\mu\nu} = \sum_{\rho,\sigma} R_{\mu\rho}\, R_{\nu\sigma}\, \alpha_{\rho\sigma},$$ where $R_{\mu\rho}$ is the rotation matrix. Each index picks up exactly one factor of the rotation matrix. A rank-1 tensor (a vector) picks up one factor: $v’_\mu = \sum_\rho R_{\mu\rho} v_\rho$. A rank-0 tensor (a scalar) picks up no factors: it is unchanged. This is the general pattern. The rank of a tensor tells you how many rotation-matrix factors appear in its transformation law, and therefore how many indices it carries.
This transformation law is not just bookkeeping. It is the guarantee that tensor equations are coordinate-independent. If $P_i = \sum_j \alpha_{ij} E_j$ holds in one coordinate system, it holds in every coordinate system, because both sides transform in exactly the same way. You write the physics once and it is automatically valid for any observer, any orientation of the laboratory, any choice of axes. This is why tensors are indispensable in general relativity, where there is no preferred coordinate system at all, and in continuum mechanics, where the material has no reason to know which way you pointed your $x$-axis.
Tensors can have more than two indices. The elastic constants of a crystal, which relate the stress tensor $S_{ij}$ (internal forces) to the strain tensor $T_{ij}$ (deformations), form a rank-4 tensor $\gamma_{ijkl}$ with $3^4 = 81$ components in principle. Symmetry reduces this to 21 independent constants for the least symmetric crystal. For a cubic crystal, 3. For an isotropic material, just 2. The rank-4 tensor must have the form $\gamma_{ijkl} = a\,\delta_{ij}\delta_{kl} + b\,(\delta_{ik}\delta_{jl} + \delta_{il}\delta_{jk})$ in the isotropic case, because those are the only rank-4 combinations of $\delta_{ij}$ that have the required symmetry and pick out no preferred direction. Two constants, $a$ and $b$, and you have described the full elastic behaviour of any isotropic solid. The same pattern of symmetry reducing an enormous number of apparent components to a small set of physical ones repeats at every rank.
So here is the summary, stated as plainly as possible. A tensor is what you get when one physical vector quantity depends linearly on another, and the proportionality is not the same in every direction. The components of the tensor tell you, for each possible input direction, what the output is in each possible output direction. That is why a rank-2 tensor in three dimensions has $3 \times 3 = 9$ components: three possible input directions times three possible output directions. The components change when you rotate your axes, but the physical relationship they encode does not. For a symmetric tensor, you can always find principal axes where the off-diagonal components vanish, reducing the description to three numbers. And a tensor equation, once you have written it correctly, is automatically true in every coordinate system without any further work.
None of this required linear algebra. It required only the idea of proportionality, the fact that polarization is linear in the electric field, the simple geometry of ellipses, and the physical demand that energy be conserved. The full formalism of linear algebra makes the same ideas faster to compute with, but the ideas themselves are prior to the formalism. A tensor is a physical object. The indices are just how we write it down.
References and Footnotes
- 1The subscript ordering $\alpha_{ij}$ means: $i$ labels the component of $\vec{P}$ being produced, $j$ labels the direction of $\vec{E}$ that produces it. So $\alpha_{yx}$ is the coefficient that tells you how much $y$-polarization is produced by an $x$-field. Keep this straight from the beginning or the indices will confuse you forever.
- 2This symmetry is a physical consequence of energy conservation, not a mathematical assumption. There exist tensors in physics that are not symmetric, for example the angular momentum flux tensor, and they genuinely need all nine components. Always ask which physical law enforces symmetry before counting independent components.
- 3The Kronecker delta $\delta_{ij}$ has the remarkable property that it looks identical in every coordinate system. You can rotate your axes by any angle and its components stay the same: ones on the diagonal, zeros everywhere else. This is because it is proportional to the metric of flat Euclidean space, which has no preferred direction. The identity tensor is the one tensor that genuinely has no directional content.
- 4The inertia tensor has the explicit form $I_{ij} = \sum_\text{particles} m(r^2 \delta_{ij} – r_i r_j)$, where the sum is over all particles in the body and $r^2 = x^2+y^2+z^2$. The diagonal component $I_{xx} = \sum m(y^2+z^2)$ is exactly the moment of inertia about the $x$-axis that you saw in introductory mechanics. The off-diagonal components $I_{xy} = -\sum m\, xy$ are the products of inertia, and they vanish when your axes align with the principal axes of the body.
