Tensors as linear mappings of multidimensional arrays
Author thisrod
7 Stars
Updated Last
2 Years Ago
Started In
March 2020


This package exports a Tensar type, which can be identified with tensors in the mathematical sense, but represents them in a different way than mathematicians customarily define them.

A tensar is a linear mapping from m-dimensional arrays to n-dimensional ones. For example, let D2 be an 5×5 matrix that discretises the second derivative operator, and u be a 5×5×5 matrix that samples a field on a 3-dimensional grid. Then the Laplacian of the field can be computed as follows.

using Tensars, LinearAlgebra

julia> D2 = Tensar(D2)
5-vector → 5-vector Tensar{Float64}

julia> E = Tensar(Float64.(Matrix(I,5,5)))
5-vector → 5-vector Tensar{Float64}

julia> L = D2⊗E⊗E + E⊗D2⊗E + E⊗E⊗D2
5×5×5 → 5×5×5 Tensar{Float64}

julia> Lu = L*u;

julia> typeof(Lu)

The way to identify a Tensar with a mathematical tensor is specified below, in the section Mathematical tensors and tensor products. The motivation for defining it as a mapping of arrays is that those mappings are closed under composition.

This is research software, and the hypothesis is that tensars will be widely useful generalisation of matrices. Here is another example.

julia> using ForwardDiff

julia> v = randn(2,3,4);

julia> f(u) = [sum(u)]

julia> J_matrix = ForwardDiff.jacobian(f, v)

julia> J = reshape(J_matrix, size(f(v)), size(v))
2×3×4 → 1-vector Tensar{Float64}

julia> dv = ones(2,3,4);

julia> J*dv
1-element Array{Float64,1}:

This package could be regarded as a port of Sussman and Wisdom's up and down tuples, replacing the parentheses of Scheme with the brackets of Julia. No doubt I have botched it, in which case I apologise for messing up their design.

The current implementation is a proof of concept. A production version would implement eltype promotion, and treat I as a broadcastable identity operator, avoiding the need to specify E:

julia> L = D2⊗I⊗I + I⊗D2⊗I + I⊗I⊗D2

The storage required for a dense n-dimensional Tensar increases geometrically with n, so these will require special types for structured forms. One use might be to accumulate one-dimensional convolutions into a multidimensional operator, in order to evaluate them a cache-efficient way with a machine learning library. Another way to regard this package is as a refactoring of DiffEqOperators that went completely overboard.

The Julia tensor ecosystem

There are several other packages dealing with tensors. Eric Forgy has reviewed them in detail, so here we'll focus on how the relate to Tensars.

There are two major ways that Tensars is different:

  1. The usual approach is to take the textbook mathematical notation and semantics for tensors, and try to import as much of them into Julia as possible. Tensars goes the other way. The goal is to extend Julia's arrays and unilinear algebra in a consistent fashion, and make it convenient for programs to do multilinear algebra. The result can be identified with the mathematical treatment of tensors, but it looks quite different.

  2. Most tensor libraries are written with differential geometry, curvature and general relativity in mind, but this one was written with quantum mechanics in mind. This has some pros and cons. The benefit is that complex numbers and Hilbert spaces are more than an afterthought. They have just as natural an interface as do real numbers and Euclid or Lorentz spaces, and they're just as thoroughly tested. (Does anyone work with in complex spaces where vectors can have negative norms?)

The drawback is that index raising and lowering is less convenient. These operations are linear in the real case, but not in the complex case, and Tensars doesn't provide a convenient way to exploit that linearity. If you're au fait with relativistic index bashing, and you see an obvious way to incorporate it in the tensar formalism, please file an issue.

Construction and linear algebra

TODO Tensar(::scalar) will construct a UniformTensar. Tensar(::scalar, cs, rs) has a shape.

TODO What about 5×6 → 2×0×4 Tensar?

The simplest way to construct a Tensar is to reshape an array with a column size and a row size. Following the convention for vectors and matrices, x and y are arrays, while A and B are tensars.

julia> A_matrix = rand(2*3*4, 5*6);

julia> A = reshape(A_matrix, (2,3,4), (5,6))
5×6 → 2×3×4 Tensar{Float64}

julia> typeof(A)

julia> Array(A)
2×3×4×5×6 Array{Float64,5}:

The linear mapping A*x is what anyone who works with 3D fields will expect it to be.

julia> x = rand(5, 6);

julia> A*x ≈ reshape(A_matrix*x[:], (2,3,4))

By convention, when a function that returns some property of an Array is also defined for for Tensar, it returns the same property of the arrays that form the column and row spaces of the Tensar as a 2-tuple.

julia> colsize(A)
(2, 3, 4)

julia> colsize(A,2)

julia> rowsize(A)
(5, 6)

julia> size(A)
((2, 3, 4), (5, 6))

julia> ncols(A)

julia> nrows(A)

julia> ndims(A)
(3, 2)

Following this convention, the length of a Tensar is the size of the matrix for the linear transformation that it represents.

julia> length(A)
(24, 30)

julia> ans == size(A_matrix)

The constructor Tensar(::AbstractArray) usually constructs a column tensar, that maps a scalar to a scalar multiple of the array.

julia> ket = Tensar(rand(2,3,4))
scalar → 2×3×4 Tensar{Float64}

The exception is that Matrix is treated as a linear transformation.

julia> M = rand(3,4);

julia> Tensar(M)
4-vector → 3-vector Tensar{Float64}

So the general way to construct a column tensar is:

julia> Tensar(M, ndims(M), 0)
scalar → 3×4 Tensar{Float64}

The constructor also accepts nrows and ncols.

julia> Tensar(rand(4,5,6), 1, 2)
5×6 → 4-vector Tensar{Float64}

The adjoint of a column tensar is a row tensar, which maps arrays to scalars according to LinearAlgebra.dot.

julia> bra = ket'
2×3×4 → scalar Tensar{Float64}

julia> x = rand(2,3,4);

julia> bra*x ≈ Array(ket)⋅x

julia> Tensar(rand(5)')
5-vector → scalar Tensar{Float64}

Just as Julia supports 0-dimensional arrays with a single element, tensars can have shape 0,0. This requires some choices about which products return scalars, arrays and tensars. The rules are stated below.

For now, indexing tensars simply indexes into the array of their elements. This is the simplest way to do it, but I'm not convinced it is the right way, and it might change.

It is common to think of a matrix shape as “rows × columns”, but this becomes confusing when generalised to tensars. A matrix maps its row space to its column space, and Tensar{T,2,3} maps 3-dimensional arrays to 2-dimensional ones, so it has 3 row dimensions and 2 column ones. It can be particularly confusing that size(A) == ((2, 3, 4), (5, 6)), while A displays as a 5×6 → 2×3×4 mapping, but the alternatives all seem worse. Users are advised to toss “number of rows × number of columns” down the nearest memory hole, and start thinking “column length × row length”.

Although the elements of a Tensar form an array, and Array(Tensar(x)) == x identically, Tensar is not a subtype of AbstractArray. Every matrix can be identified with a 1,1-dimensional tensar, but there are 2,0- and 0,2-dimensional tensars that have the same elements but represent different linear transformations. Similarly, every vector can be identified with a column tensar, but there is a distinct row tensar with the same elements.

Tensars form a linear algebra over their element type in exactly the same way that matrices do. They can be added just like matrices, and both * and reduce to the scalar product when one operand is a scalar.

The rules on array and tesnar shapes

When an array is reshaped to a tensar, there are two rules on the array and tesnar shapes. Suppose the tensar shape is a×b×c←d×(e*f)×g.

  1. The tensor sizes divide the array sizes. This tensar could be formed from an (a,b,c,d,e*f,g) array, an (a*b*c, d*e*f*g) array, or any intermediate choice of commas and multiplications. But not from an (a,b,c,d,e,f,g) array, or a (1,a*b*c,d*e*f*g) array.

  2. The array size can be split into a prefix that divides the column length of the tensar, and a suffix that divides its row length. This excludes an (a*b,c*d,e*f*g) array, for example.

When the array is 0-dimensional, or the tensar is a row, column or scalar, only () and scalar are deemed to divide () and scalar. A 0-dimensional array has a single element, but it can only form a scalar ← scalar tensar, not scalar ← 1-vector or 1×1 ← scalar or whatever.

These rules are relaxed slightly for constructing TensArray. Scalar row or column sizes can correspond to a single array dimension, in order that every shape of Tensar can be a TensArray matrix.

Generalising the matrix product

There are two product operators that act on Tensar. The tensor product has its usual mathematical meaning, which will be discussed below. Note that neither * nor is not commutative.

The operator product * is identical to the matrix product, when matrices and tensars are both identified with linear mappings. The product A*x is the image of the array x under the linear mapping A. The product A*B is the composition of the linear mappings A and B. If x is a vector, then x'*A is the same as x'*Array(A).

Here are the full rules on which products return scalars, arrays and tensars. The tensar library treats anything other than an AbstractArray or AbstractTensar as a scalar.

  1. The product of two tensars is always a tensar. In particular, this returns a scalar ← scalar Tensar instead of a scalar.

  2. The product of a scalar and a tensar is a scalar product, and returns a tensar.

  3. The product of a tensar and an array follows the Julia rules for matrix multiplication. The result is usually an array, except that a scalar is returned instead of a 0-dimensional array.


Tensars have adjoints, like any linear mappings. Row and column Tensars have the same adjoint relationship as row and column vectors, or bras and kets in Dirac notation.

Tensar(A, n, 0)' == Tensar(conj.(A), 0, n)

In general, A' is the unique tensar that satisfies the identity

V'*A'*U' = conj(U*A*V)

where U is any appropiately shaped row tensar, and V any appropiately shaped column tensar.

Tensors and tensor products

It was mentioned at the beginning that tensars can be identified with tensors. The time has come to describe this correspondence and the tensor product , which acts on tensars the same way it acts on the corresponding tensors.

Mathematicians (well, OK, physicists) conventionally define an m,n-tensor as a multilinear mapping from m dual vectors and n vectors to a scalar. This will be identified with a certain m,n-dimensional tensar. A symbol such as A could denote either of these, but the tensor mapping will be written as a function application A(u1, ..., um; v1, ..., vn) and the tensar mapping as A*x.

The tensor product has a very simple definition in terms of tensors. Tensors are scalar valued, and just multiplies those scalars.

(A⊗B)(u1, ..., um; v1, ..., vn) =
    A(u1, ..., uj; v1, ..., vk) × B(u[j+1], ..., um; v[k+1], ..., vn)

The correspondence between tensors and tensars can be built up using the tensor product, starting from basis vectors, then progressing to row and column tensars, and finally to general tensars.

A dual vector v' can be identified with a 0,1-tensor in an obvious way, as the mapping u -> v'⋅u. Similarly, a vector u can be identified with the 1,0-tensor v -> v⋅u. These tensors are identified with Tensar(u) and Tensar(v').

A vector is identified with a tensor that takes a dual vector argument, so a tensor product of vectors is a tensor with only dual vector arguments. If e[j] is the usual basis vector, an array can be formed by tabulating

A[j, k, ...] = (e[j]⋅v1)×(e[k]⋅v2)×... = (e[j]⊗e[k]⊗...)(v1, v1, ...)

The tensor e[j]⊗e[k]⊗... corresponds to Tensar(A).

The product of 0,1-dimensional tensars can be written in terms of their 1,0-dimensional adjoints:

(v1'⊗v2'⊗...) = (v1⊗v2⊗...)'

This is consistent with the definition of in terms of tensors. We now have the machinery to identify a general Tensar with a tensor:

A(u₁, ..., u_m, v₁, ..., v_n) = (u₁⊗...⊗u_m)*A*(v₁⊗...⊗v_n)

The action of on general tensars is now determined by identifying them with tensors.

The trace of a Tensar is a tensor contraction. It takes two arguments, the column and row dimensions to contract over. These must have equal lengths. Index raising works as follows. When permutedims is implemented, it will be possible to shuffle the raised index into an appropriate place.

julia> A = Tensar(rand(2,3,4,5,6,7), 3, 3)
5×6×7 → 2×3×4 Tensar{Float64}

julia> g = Tensar(rand(6,6),2,0)
scalar → 6×6 Tensar{Float64}

julia> g⊗A
5×6×7 → 6×6×2×3×4 Tensar{Float64}

julia> tr(g⊗A,1,2)
5×7 → 6×2×3×4 Tensar{Float64}

Contraction has always looked like a matrix product. The Tensar formalism makes apparent how this corresponds to a composition of linear mappings (except that the implementation currently has a bug, so these don't come out equal):

julia> M = Tensar(rand(4,5))
5-vector → 4-vector Tensar{Float64}

julia> N = Tensar(rand(5,6))
6-vector → 5-vector Tensar{Float64}

julia> M⊗N
5×6 → 4×5 Tensar{Float64}

julia> tr(M⊗N, 2, 1)
6-vector → 4-vector Tensar{Float64}

julia> M*N
6-vector → 4-vector Tensar{Float64}

I'm not quite satisfied with this. I think that, lurking somewhere on the edge of it, there are some interesting ideas about inner products of arrays and g^{ij} being the inverse of g_{ij}. If you know enough multilinear algebra to see that clearly, please explain it to me.


Arrays are the currency of Julia. This means that ForwardDiff.jacobian will continue to return an array for the forseeable future, and the only way for tensors to gain market share is if they can be bought and sold for arrays. Any tensor library has to like that, or lump it and accept that it tensors will be a walled garden. (Thanks to Michael Abbot for pointing that out at an early stage in the development of Tensars.)

The TensArrays package aims to provide tensors unbiquitously, disguised as arrays.

julia> J_matrix = ForwardDiff.jacobian(f, v)

julia> using TensArrays

julia> J_matrix = TensArray(J_matrix, size(f(v)), size(v)) 1×24 TensArray{Float64,2}: 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 … 1.0 1.0 1.0 1.0 1.0 1.0 1.0

julia> J_matrix isa AbstractMatrix

julia> using Tensars

julia> Tensar(J_matrix)
1-vector ← 2×3×4 Tensar{Float64}

This is a deliberately minimalist library, to reduce the performance cost and encourage functions like jacobian to return a TensArray. As a result, the tensor nature is fragile. J_matrix will revert to an Array if it is broadcast with anything except a scalar or TensArray, or if it is reshaped or broadcast in a way that is inconsistent with its tensor dimensions. The intention is that code that wishes to use the tensor information in a TensArray will convert it to a Tensar as soon as possible, or to the author's favorite way to represent tensors.

julia> 2J_matrix
1×24 TensArray{Float64,2}:
 2.0  2.0  2.0  2.0  2.0  2.0  2.0  2.0  …  2.0  2.0  2.0  2.0  2.0  2.0  2.0
julia> ans + rand(size(J_matrix)...)
1×24 Array{Float64,2}:
 2.634  2.40308  2.28292  2.75842  …  2.1715  2.74704  2.2228  2.66855
julia> reshape(J_matrix, 1, 2, 3, 4)
julia> reshape(J_matrix, 1, 12, 2)

I'm still thinking about how this should interact with lazy reshaping. It is unlikely to be implemented reliably until someone thinks harder about nested array wrappers.

Array dimensions can not be a mixture of row and column tensor dimensions.

julia> A_matrix = TensArray(A_matrix, (2,3,4), (5,6));

julia> size(A_matrix)
(24, 30)

julia> reshape(A_matrix, 6, 4, 30)
6×4×30 TensArray

julia> reshape(A_matrix, 48, 15)
48×15 Array

Future work

Generalise UniformScaling to axis expansion. These are the axes that the codimensions act along, these are the axes the contradimensions act along, all the other axes have unspecified length and the elements are diagonal along them. When the kernel is a scalar, all other axes means all axes, and it reduces to uniform scaling.