Principles

Numeri is a many-purpose yet lightweight numerical and tensor library. It aims to

• provide powerful, high-level operations like eigenvalue and eigenvector decompositions
• be entirely Javascript to support in-browser use
• provide an API that fits in with other JavaScript projects
• achieve nearly the fastest runtime possible for native Javascript code
• support operations on arbitrarily-shaped tensors
• require 0 (non-dev) dependencies to ensure a very small package size

Creating a Tensor

const scalar0 = numeri.scalar(3)
const vector0 = numeri.vector([3, 4])
const matrix0 = numeri.fromFlat([3, 4, 5, 6, 7, 8], [3, 2]) //a 3x2 matrix whose first row is `[3, 4]`
const matrix1 = numeri.fromNested([[3, 4], [5, 6], [7, 8]]) //the same 3x2 matrix
const tens4d0 = numeri.fromFlat([5, 7], [1, 2, 1, 1]) //a 1x2x1x1 tensor
const zeroMat = numeri.zeros([4, 5])
const fourMat = numeri.fill([4, 5], 4)
const z123456 = numeri.range(7)
const eyeMatr = numeri.identity(5) //5x5 identity matrix

Miscellaneous

matrix0.shape //returns [3, 2]
matrix0.getLength() //total number of elements; 6
matrix0.toNested() //returns as nested arrays

Slicing and Accessing

matrix0.get(2, 1)

matrix0.slice(0) //returns the 0th row vector
matrix0.slice([0, 2]) //returns the submatrix with only the first two rows
matrix0.slice(undefined, 0) //returns the 0th column vector
matrix0.slice(undefined, [0, 4, 2]) //returns the submatrix with every even column from the 0th up to (not including) the 4th

Note that .slice returns a view.

Reshaping

matrix0.transpose([1, 0]) //if matrix0 is n x m, this returns an m x n matrix
matrix0.transpose() //shorthand, only possible if the tensor is a matrix
tens4d0.transpose([0, 2, 3, 1]) //(a x b x c x d) -> (a x c x d x b)

matrix0.reshape([2, 3]) //returns a new matrix [[3, 4, 5], [6, 7, 8]]

matrix0.tile([[0, 3]]) //returns a tiled matrix 3x as big in the 0 axis
matrix0.tile([[0, 3]], {addDims: true}) //returns a 3 x m x n tiled tensor

Note that .transpose returns a view.

Combining

numeri.stack([matrix0, matrix1], 1) //returns a 3D tensor by stacking these along the 1 axis
numeri.concat([matrix0, matrix1], 1) //return a 2D tensor by concatenating these along the 1 axis

Setting

matrix0.set([1, 2], 3) //sets the [1, 2] entry to 3
matrix0.update([1, 2], (x) => x + 3) //increments the [1, 2] entry by 3
matrix0.setAll(11)
matrix0.slice(1).setAll(12)

Unary Operators

matrix0.exp() //returns a new matrix with elementwise exponentiation
matrix0.map((x) => x * x)
matrix0.mapInPlace((x) => x * 2)

Binary Operators

matrix0.plus(matrix1) //returns a new matrix by elementwise addition
matrix0.minus(matrix1)
matrix0.times(matrix1)
matrix0.div(matrix1)
matrix0.elemwiseBinaryOp(matrix1, (a, b) => Math.pow(a, b))
matrix0.elemwiseBinaryOpInPlace(matrix1, (a, b) => Math.pow(a, b))

Matrix Operators

const { matMul, outerProd, dot } = numeri
matMul(matrix0, numeri.fromFlat([1, 2, 3, 4, 5, 6], [2, 3])) //matrix multiplication; only works on 2-dimensional tensors
matMul(matrix0, vector0) //infer vector0 to be treated as a column vector
dot(vector0, vector0) //vector dot product
outerProd(vector0, vector0) //outer product; i.e. A_ij = v0_i * v1_j

Broadcasting

matrix0.plus(1) //returns a matrix with 1 added to each element
matrix0.plus(numeri.scalar(1).broadcastOn(0, 1)) //equivalent; explicitly broadcast scalar to both dimensions

matrix0.plus(numeri.vector([1, 2]).broadcastOn(0)) //returns a matrix with [1, 2] added to each row
matrix0.plus(numeri.vector([1, 2, 3]).broadcastOn(1)) //returns a matrix with [1, 2, 3] added to each column

Note that broadcasted tensors also work in many other operations such as stack.

Reducing (single axis only)

vector0.argmin() //returns a number
matrix0.argmax({axis: 1}) //returns a vector with as many entries as matrix0 has rows
vector0.argmin({keepScalarAsTensor: true}) //returns a scalar Tensor
const { arg, values } = vector0.argmin({includeValues: true}) //get both the argmin and min

Reducing (possibly multiple axes)

matrix0.sum() //returns a number
matrix0.sum({axes: [1]}) //returns a vector with as many entries as matrix0 has rows
matrix0.sum({keepScalarAsTensor: true}) //returns a scalar Tensor
//these same options can be applied to any tensor
matrix0.norm() //Eulidean norm
matrix0.lpNorm(3) //L3 norm

Eigen / Eigenvalue / Eigenvector / Hessenberg Operations

const { symEig, symTridiagEig, symHessenberg } = numeri
const symMat = numeri.fromFlat([1, 2, 3, 2, 4, 5, 3, 5, 6], [3, 3])
const tridiagonalMat = numeri.fromFlat([1, 2, 0, 2, 4, 5, 0, 5, 6], [3, 3])

symEig(symMat) //symmetric matrix eigen decomposition; returns {vals} as tensor
symEig(symMat, {includeVecs: true}) //return {vals, vecs} with vecs a tensor whose columns are the eigenvectors
symTridiagEig(tridiagonalMat, {includeVecs: true}) //symmetric tridiagonal matrix eigen decomposition
symHessenberg(symMat, {includeQ: true}) //returns {hessenberg, q}; Hessenberg of symmetric matrix is symmetric tridiagonal

A Word about Views

The .slice and .transpose operators return views into the original tensors. This is also standard behavior in other major numerical libraries, like numpy and Tensorflow. A view shares the same data as the original tensor, but may have different shape and access the data at different indices.

This makes slicing and transposition O(1) time with respect to the data size. It also saves memory by not copying the data into each tensor. However, sometimes having data laid out in order makes operations faster. For instance, if we define a = parentTensor.slice(...sliceArgs).transpose(permutation) and then do many evaluations of a.plus(tensor0), a.plus(tensor1), ..., it could save time do a copy with a = parentTensor.slice(...sliceArgs).transpose(permutation).copy(). This will make each of the following .plus operations faster.

More Info

blog post