[Intermediate → Advanced] — The language the whole field is written in. An afternoon spent here saves weeks of confusion later.

Why index notation

Vector notation runs out. You can write and , but there’s no clean symbol for the double contraction of a fourth-order stiffness tensor with a strain tensor.

Index notation writes components and lets bookkeeping rules do the work:

The second one is the general form of Hooke’s law. Vector notation can write it as , but that hides which indices pair up with which, and that pairing is exactly what you need when you work with it.

Tensor algebra is the language of continuum mechanics. Stress, strain, the deformation gradient, the stiffness of a material: all of them are tensors. This note builds that language from the bottom: vectors and bases first, then the shorthand, then the three ways to multiply vectors, then tensors.

Vectors, bases and linear independence

A vector is a sum of components

A vector is an arrow: it has a length and a direction, and it exists whether or not you draw axes. Once you do pick axes, you can describe the arrow as a recipe: “go this far along the first axis, this far along the second, this far along the third.”

In the familiar Cartesian orthonormal basis, with unit vectors :

The numbers are the components (the scaling factors). The unit vectors are the basis. The vector is the sum of each component times its basis vector.

“Orthonormal” packs two conditions into one word. Ortho: the basis vectors are at right angles to each other. Normal: each one has length 1.

Linear independence

A set of vectors is linearly independent when you can’t build any one of them out of the others. No amount of stretching and adding and will give you , so are independent.

The formal version: the only way to make

is to set every to zero. If some non-zero choice of ‘s gives zero, then one vector can be written using the others, and the set is dependent.

Examples in 2D:

  1. and are independent. One points along x, one along y, and neither is a stretched copy of the other.
  2. and are dependent, because the second is just twice the first. They point along the same line.
  3. Any three vectors in 2D are dependent. There’s only room for two independent directions in a plane, so the third can always be built from the first two.

Independence defines dimension

That last example is the key idea. The dimension of a space is the maximum number of linearly independent vectors it can hold.

  1. A 1D space (a line) holds at most 1 independent vector.
  2. A 2D space (a plane) holds at most 2.
  3. A 3D space holds at most 3, for example .

A basis is any set of that maximum size that is independent. Any vector in the space can then be written as a sum of basis vectors, in exactly one way.

Renaming the basis: parsimony

Now change the naming. Instead of , call the basis vectors . Instead of , call the components :

The trick is mixing a letter with a number: the letter says what kind of thing it is (a component , or a basis vector ), and the subscript number says which one. This achieves parsimony, which means saying the same thing with fewer symbols.

  1. Fewer symbols. The Cartesian form needs 6 different letters (). The indexed form needs 2 ( and ).
  2. It scales. In a 10D space, the old way needs 10 component letters plus 10 basis letters, about 20 symbols, and you’d run out of Greek. The indexed way still needs just 2: .
  3. It works for any basis. The don’t have to be . They can be any independent set, which matters as soon as you rotate the axes (see “Changing basis” below). From here on, the basis is always orthonormal (rotated or not), because the shortcuts below, especially , only hold for orthonormal bases.

Standard summation, before Einstein

What we have above is standard summation: the sign is written out. The same style writes a matrix times a vector like this:

Read it as: “for each from 1 to 3, add up over .” That’s three equations:



Watch the left-hand side. It’s , not . Once has been summed away it no longer exists, so it can’t appear on the left. This is the first of the index rules below, and it’s an easy slip when copying from the board.

Notice something about every one of these sums: the index being summed always appears twice in the term, and the other index appears once. Einstein noticed the same thing and dropped the sign.

The summation convention

A repeated index in a term is summed over. Einstein’s convention, and it removes an enormous amount of clutter. It is the standard notation for vectors and tensors, and every text in this field assumes it without saying so.

So becomes just , and becomes .

What a “term” is

The convention says “repeated in a term”, so the word matters. A term is one piece of an expression separated from the others by , or . Everything multiplied together inside that piece belongs to the same term.

  1. is one term. appears twice in it, so it’s summed.
  2. has two terms. is repeated in each one separately, so each term is its own sum: .
  3. has two terms, and appears once in each, so there’s no sum. is free, and it’s the same free index in both terms, which is required.

Two kinds of index:

Dummy (repeated) — summed over, and its name is arbitrary. .

Free (appears once) — ranges over 1,2,3, and must match on both sides of an equation.

Here is dummy (summed), is free. This is three equations, one per value of — a matrix–vector product written in four characters.

The rules that catch errors:

  1. Free indices must match on both sides. ✓. ✗ — on the right, is repeated (so summed) and is left free, so the left’s free index is but the right’s is
  2. An index may appear at most twice per term. is meaningless
  3. Rename dummy indices freely, but never to collide with a free index

Index-matching is your unit check. If the free indices don’t balance, the equation is wrong — no physics required to see it. This catches most algebra errors in the field.

How to read an index equation

Take

Read it in three steps.

  1. Find the repeated index. appears twice in the left-hand term, so there is a sum over .
  2. Find the free indices. and each appear once, on both sides. They balance, so the equation is well-formed.
  3. Count the equations. Each free index takes the values 1, 2, 3. Two free indices give equations, one for each pair .

Expanded, one of the nine (pick , ) is:

That’s row 1 of multiplied entry by entry with column 2 of , then added up. It is exactly how you’d compute entry of the matrix product . So the index equation is the matrix equation , written one entry at a time.

The general pattern: a repeated index on the inside (the second index of the first factor and the first index of the second factor, like the in ) is a matrix product. Move the repeated index and the meaning changes. sums down column against column , which is .

The Kronecker delta

Where it comes from

Take two basis vectors from an orthonormal basis and dot them together. Two different ones are at right angles, so they give 0. A basis vector with itself has length 1, so it gives 1:

That’s 9 facts. The Kronecker delta is just a name for all 9 at once:

Write the 9 values in a grid and you get the identity matrix. is the identity matrix written in index form. That’s all it is.

What it does: index substitution

Its useful property is that it swaps one index for another:

Why? is repeated, so expand the sum:

Pick . Then , , , and the whole thing collapses to . Whatever is, only the term where equals survives. So the delta “kills” the sum and replaces with .

Think of as a filter: “keep only the term where the two indices are equal.” The same reasoning gives:

That last one is a common slip — it’s summed, so it’s 3 in three dimensions, not 1.

The Levi-Civita (alternating) symbol

What it is

The Levi-Civita symbol , also called the alternating symbol or permutation symbol, is a three-index symbol. Each index runs over 1, 2, 3, so there are possible combinations. Every one of them is , or , by this rule:

Cyclic and acyclic arrangements

Write 1, 2, 3 around a circle, clockwise.

        1
      ↗   ↘
     3  ←  2
  1. Cyclic means reading with the arrows: 1→2→3, 2→3→1, 3→1→2. These are , , , and for each.
  2. Acyclic means reading against the arrows: , , . for each.
  3. Any repeat gives 0: , and so on.

So only 6 of the 27 are non-zero: three s and three s. The other 21 are zero. Another way to say the same thing: swapping any two indices flips the sign. ; swap the first two and .

Why it exists: the cross product

The delta captured what the dot product does to basis vectors. Epsilon captures what the cross product does to them. With a right-handed basis:

Cyclic order gives , reversed gives , a vector crossed with itself gives zero. That’s exactly epsilon’s rule, so all 9 facts collapse to one line:

It also gives the determinant of a matrix:

Products of alternating symbols

When two epsilons multiply, the result can always be written with deltas. The general rule is a determinant of deltas:

You rarely need that full form. Sharing indices (contracting) gives the versions you actually use:

  1. One shared index (the - identity):

    A way to remember it: with the shared index first in both, pair “first with first, second with second”, minus “first with second, second with first”.
  2. Two shared indices: . (Set in the line above: .)
  3. All three shared: . That’s just counting: the 6 non-zero entries, each squared to 1.

Check one by hand to trust it. Put into the first identity. Left: , and only survives, giving . Right: . ✓

Memorise the first one — it’s how you prove , which appears in both elasticity and fluid dynamics.

Three ways to multiply two vectors

These three get mixed up constantly, so here they are side by side. Each one is worked out by writing both vectors as sums of components and letting the basis vectors do the work.

ProductSymbolResultComponent formAlso called
Dotscalar (a number)scalar product, inner product
Crossvector (component )vector product
Dyadsecond-order tensordyadic product, outer product, tensor product

Keep the outputs straight: the dot product gives a scalar and is the inner product; the cross product gives a vector. When the two are chained, order matters — see the scalar triple product below.

Dot product in component form

Two different dummy letters ( and ) are used on purpose, so no index appears more than twice. The delta then substitutes . Result: , a single number. Geometrically, : how much the two vectors point the same way.

Cross product in component form

So component is (using , since moving the last index to the front is a cyclic shift). Expand component 1 to see it is the formula you already know:

Every other term has a repeated index, so it’s zero. The result is a vector at right angles to both and . The cross product only exists in 3D.

The scalar triple product

Chain the two products: cross first, then dot.

  1. Cross product first. gives a vector, with components .
  2. Then the inner (dot) product. Dotting with that vector gives a scalar.

In index form the whole thing is one term:

All three indices are repeated, so nothing is free and the result is a number. Geometrically, it’s the signed volume of the box (parallelepiped) with edges : positive when the three form a right-handed set, negative when left-handed, and zero when they lie in one plane (that is, when they are linearly dependent). It equals the determinant of the matrix with those three vectors as rows.

Why the order is forced. is meaningless: is a scalar, and you can’t cross a scalar with a vector. So the cross product must happen first, and the brackets are often dropped: .

Because doesn’t change under a cyclic shift, neither does the triple product: . Swapping any two vectors flips the sign.

The dyad (dyadic product, outer product)

The third product keeps everything. Nothing is summed:

Two free indices means 9 components, so the result is a second-order tensor. It’s often written with no symbol between, which is where the name dyad (a pair) comes from.

What a dyad does. A dyad is a machine that eats a vector and returns a vector:

It measures how much lines up with (a number), then returns scaled by that number. Whatever you feed in, the output always points along .

Order matters. . Swapping them swaps the indices, versus , which gives the transpose.

Dyads and matrices

Write the 9 components of in a grid. Row , column holds :

A column times a row. Compare the dot product, which is a row times a column, , and gives a single number. Inner product: row × column → 1×1. Outer product: column × row → 3×3.

Every second-order tensor is a sum of dyads. The 9 basis dyads act as a basis for second-order tensors, just as are a basis for vectors:

Same idea as , one level up. A single dyad is a very special tensor (every row is a multiple of , so its matrix has rank 1), but adding dyads together builds any tensor at all, the stress tensor included.

Trace of a dyad

The trace is the sum of the diagonal entries, . For a dyad, the diagonal is , so:

The trace of the outer product is the inner product. That’s the link between the two: setting the two free indices of equal (a contraction) sums them away and leaves a scalar. The trace doesn’t change when you rotate the axes, which is why the pressure, , comes out the same in any coordinate system.

Changing basis: coordinate transformations

Stay in Euclidean space (ordinary flat 3D) and use two orthonormal bases: the old one and a rotated one . One arrow can be written in either:

To get the new components, dot both sides with :

Each is the cosine of the angle between new axis and old axis . is orthogonal, , which in matrix form says : rotations keep lengths and angles.

2D example. Rotate the axes by :

The vector with old components gets new components . The arrow didn’t move. Only the numbers describing it changed.

Dyads tell you how tensors transform. Each vector in a dyad picks up its own :

Since every second-order tensor is a sum of dyads, every one transforms the same way: , or . That’s where the “one per index” rule in the next section comes from.

Vectors are not matrices

Matrices are algebraic constructs. Vectors and tensors are geometric constructs. They are related, but they are not the same thing.

  1. A vector is an arrow. It exists without any axes.
  2. A column of three numbers is that arrow’s components in one particular basis.
  3. Change the basis and the column changes, while the arrow stays put (the 2D example above).

So when someone calls a matrix “a vector” or “a tensor”, they are quietly assuming a basis. If that assumption is wrong, everything goes haywire.

A concrete way it goes wrong. Let have components in the standard basis. Let have components in a basis rotated by 90°, whose first axis points along the old . Add the columns and you get , an arrow along of length 2. But actually points along , so the true sum is in the standard basis. The matrix arithmetic was flawless; the answer is wrong, because the two columns were in different bases.

The rule: you may only add, multiply or compare component arrays when they are written in the same basis. Index notation makes this safe by keeping the basis implied and fixed throughout an equation.

What a tensor actually is

Not “a multidimensional array”. That’s the representation, and confusing the two causes real trouble.

A tensor is a physical or geometric object whose components transform in a specific way when you change coordinate systems.

The defining property is the transformation rule. Under a rotation (an orthogonal matrix taking old axes to new):




One per index. That’s the whole pattern.

Why this matters physically: a tensor represents something real, so it must not depend on your choice of axes. The stress state at a point is a fact about the material; the nine numbers you write down depend on how you oriented your coordinate system. The transformation rule is what guarantees both descriptions mean the same thing.

A 3×3 array of numbers is not automatically a tensor. If it doesn’t transform correctly, it’s just an array — and any equation you write with it isn’t a physical law, because it would say something different in rotated coordinates.

This is also the deep reason “tensor” in machine learning means something looser: a PyTorch tensor is an n-dimensional array with no transformation law attached. Same word, different concept, and it’s worth keeping straight.

Operations

aᵢbᵢ                    inner product          → scalar
aᵢbⱼ                    outer product          → order 2
σᵢⱼnⱼ                   contraction            → vector
σᵢⱼεᵢⱼ                  double contraction     → scalar
σᵢᵢ                     trace                  → scalar (invariant)

Contraction lowers the order by two, once per summed pair. That’s why (order 4 contracted with order 2 over two indices) gives order 2.

Symmetric and antisymmetric parts. Any second-order tensor splits uniquely:

This split is physically meaningful and appears repeatedly: the velocity gradient splits into rate of deformation (symmetric — actual stretching) and spin (antisymmetric — rigid rotation). Only the symmetric part causes stress in a fluid. → Kinematics

A symmetric 3×3 tensor has 6 independent components, an antisymmetric one has 3 (which is why it can be represented as a vector — the axial vector).

The volumetric/deviatoric split is the other decomposition that matters:

Pressure changes volume; the deviator changes shape. Metals yield on the deviatoric part alone — hydrostatic pressure doesn’t cause plastic flow, which is why deep-sea pressure crushes a hollow sphere but doesn’t yield a solid one. → Failure and Yield

Eigenvalues, principal values, invariants

For a symmetric tensor, the eigenvalue problem

has three real eigenvalues and mutually orthogonal eigenvectors.

Physically: there exists an orientation in which the stress tensor is diagonal — no shear, only normal stresses. Those are the principal stresses and the eigenvectors are the principal directions.

That’s why a brittle material fails on a plane perpendicular to the maximum principal stress, and it’s the basis of most failure criteria.

The characteristic equation gives three invariants — quantities unchanged by rotation:

Invariants are how you write a physical law that doesn’t depend on orientation. An isotropic material’s response can only depend on invariants — that’s not a convenience, it’s forced by the symmetry. → Constitutive Models

The von Mises stress is built from the second deviatoric invariant, which is why it’s a single number you can compare against a yield stress regardless of how the load is oriented.

Tensor calculus

Derivatives in index notation. The comma convention is standard and worth adopting:





The divergence of the stress tensor is the one that matters — is the net internal force per unit volume, and it’s the heart of the momentum balance.

The divergence theorem converts surface integrals to volume integrals:

Every balance law in this field is derived by writing it as a surface plus volume integral, applying this, and arguing that since the volume is arbitrary the integrand must vanish. That derivation pattern appears four times in note 06 — learn it once. → Conservation Laws

Practical advice

  1. Check free indices on both sides. It’s free error detection
  2. , not 1
  3. Never let an index appear three times. Rename the dummy
  4. Learn the - identity — it resolves most curl manipulations
  5. Symmetric ⟹ 6 components, real eigenvalues, orthogonal eigenvectors. Almost every tensor here is symmetric
  6. When a law must be orientation-independent, write it in invariants
  7. Cartesian only, for now. These notes assume rectangular coordinates, where covariant and contravariant components coincide and you can ignore the distinction. Curvilinear coordinates (cylindrical, spherical) reintroduce it along with Christoffel symbols — necessary for shells and for general relativity, and out of scope here

Words used in this note

  1. A vector is a quantity with a length and a direction, like an arrow, which exists independently of any axes.
  2. A component is one of the numbers that says how far a vector goes along one basis direction.
  3. A basis is a set of linearly independent vectors, as many as the dimension of the space, that every other vector can be built from.
  4. An orthonormal basis is one whose vectors are at right angles to each other and each have length 1.
  5. Vectors are linearly independent when none of them can be built by stretching and adding the others.
  6. The dimension of a space is the largest number of linearly independent vectors it can hold.
  7. Parsimony means saying the same thing with fewer symbols, which is what indexed names like and achieve.
  8. A term is one piece of an expression separated from the rest by a plus, minus or equals sign.
  9. A dummy index is an index repeated within one term; it is summed over, and its letter can be renamed freely.
  10. A free index is an index that appears once in a term; it ranges over 1, 2, 3 and must appear in every term of the equation.
  11. The Kronecker delta is 1 when its two indices are equal and 0 otherwise; it is the identity matrix in index form.
  12. The Levi-Civita symbol is for a cyclic arrangement of 1, 2, 3, for an acyclic one, and 0 when an index repeats.
  13. The dot product (inner product) multiplies two vectors to give a scalar.
  14. The cross product (vector product) multiplies two 3D vectors to give a vector at right angles to both.
  15. A dyad (outer product, tensor product) multiplies two vectors to give a second-order tensor, without summing anything.
  16. A tensor is a geometric object whose components change by a fixed rule, one per index, when the basis is rotated.
  17. The trace of a second-order tensor is the sum of its diagonal components, and it does not change under rotation.
  18. A contraction sets two indices equal and sums over them, which lowers the order of a tensor by two.

Check yourself

Try these on paper before looking anything up.

  1. Are , and linearly independent? What does your answer say about the space they span?
  2. In , which index is free and which is dummy? Write out in full. Is this or ?
  3. What’s wrong with , and with ?
  4. Evaluate and , , .
  5. Use the - identity to show .
  6. For and , write the matrix of and check its trace equals .
  7. Explain in two sentences why a column of numbers is not, by itself, a vector.

↑ Back to Continuum Mechanics