#linear #algebra #matrix #vector #math

nightly aljabar

A super generic, super experimental linear algebra library

9 releases

✓ Uses Rust 2018 edition

0.2.7 Jul 9, 2019
0.2.6 Jul 9, 2019
0.2.5 Jun 29, 2019
0.1.0 Jun 21, 2019

#29 in Math

Download history 15/week @ 2019-06-18 154/week @ 2019-06-25 33/week @ 2019-07-02

74 downloads per month

GPL-3.0 license

1.5K SLoC


Documentation Version Downloads

An experimental n-dimensional linear algebra and mathematics library for the Rust programming language that makes extensive use of unstable Rust features.

aljabar supports Vectors and Matrices of any size and will provide implementations for any mathematic operations that are supported by their scalars. Additionally, aljabar can leverage Rust's type system to ensure that operations are only applied to values that are the correct size.

aljabar relies heavily on unstable Rust features such as const generics and thus requires nightly to build.

Provided types

Currently, aljabar supports only Vector and Matrix types. Over the long term aljabar is intended to be feature compliant with cgmath.


Vectors can be constructed from arrays of any type and size. There are convenience constructor functions provided for the most common sizes:

let a = vec4( 0u32, 1, 2, 3 ); 
    Vector::<u32, 4>::from([ 0u32, 1, 2, 3 ])

Add, Sub, and Neg will be properly implemented for any Vector<Scalar, N> for any respective implementation of such operations for Scalar. Operations are only implemented for vectors of equal sizes.

let b = Vector::<f32, 7>::from([ 0.0f32, 1.0, 2.0, 3.0, 4.0, 5.0, 6.0, ]);
let c = Vector::<f32, 7>::from([ 1.0f32, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, ]) * 0.5; 
    b + c, 
    Vector::<f32, 7>::from([ 0.5f32, 1.5, 2.5, 3.5, 4.5, 5.5, 6.5 ])

If the scalar type implements Mul as well, then the Vector will be an InnerSpace and have the dot product defined for it, as well as the ability to find the squared distance between two vectors (implements MetricSpace) and the squared magnitude of a vector. If the scalar type is a real number then the distance between two vectors and the magnitude of a vector can be found in addition:

let a = vec2( 1i32, 1);
let b = vec2( 5i32, 5 );
assert_eq!(a.distance2(b), 32);       // distance method not implemented.
assert_eq!((b - a).magnitude2(), 32); // magnitude method not implemented.

let a = vec2( 1.0f32, 1.0 );
let b = vec2( 5.0f32, 5.0 );
const close: f32 = 5.65685424949;
assert_eq!(a.distance(b), close);       // distance is implemented.
assert_eq!((b - a).magnitude(), close); // magnitude is implemented.

// Vector normalization is also supported for floating point scalars.
    vec3( 0.0f32, 20.0, 0.0 )
    vec3( 0.0f32, 1.0, 0.0 )


Swizzling is supported for both the xyzw or rgba conventions. All possible combinations of the first four elements of a vector are supported. Single-element swizzle functions return scalar results, multi-element swizzle functions return vector results of the appropriate size based on the number of selected elements.

let a = vec4(0.0f32, 1.0, 2.0, 3.0 );

// A single element is returned as a scalar.
assert_eq!(1.0, a.y());

// Multiple elements are returned as a vector.
assert_eq!(vec3(2.0, 0.0, 3.0), a.zxw());

// The same element can be selected more than once.
assert_eq!(vec2(0.0f32, 0.0), a.rr());

Mixing xyzw and rgba swizzle conventions is not allowed.

let a = vec4(0.0f32, 1.0, 2.0, 3.0);
let b = a.rgzw(); // Does not compile.

Swizzling is supported on vectors of length less than 4. Attempting to access elements past the length of the vector is a compile error.

let a = vec3(0.0f32, 1.0, 2.0);
let b = a.xyz(); // OK, only accesses the first 3 elements.
let c = a.rgba(); // Compile error, attempts to access missing 4th element.


Matrices can be created from arrays of Vectors of any size and scalar type. As with Vectors there are convenience constructor functions for square matrices of the most common sizes:

let a = Matrix::<f32, 3, 3>::from( [ vec3( 1.0, 0.0, 0.0 ),
                                     vec3( 0.0, 1.0, 0.0 ),
                                     vec3( 0.0, 0.0, 1.0 ), ] );
let b: Matrix::<i32, 3, 3> =
            mat3x3( 0, -3, 5,
                    6, 1, -4,
                    2, 3, -2 );

All operations performed on matrices produce fixed-size outputs. For example, taking the transpose of a non-square matrix will produce a matrix with the width and height swapped:

    Matrix::<i32, 1, 2>::from( [ vec1( 1 ), vec1( 2 ) ] )
    Matrix::<i32, 2, 1>::from( [ vec2( 1, 2 ) ] )

As with Vectors, if the underlying scalar type supports the appropriate operations, a Matrix will implement element-wise Add and Sub for matrices of equal size:

let a = mat1x1( 1u32 );
let b = mat1x1( 2u32 );
let c = mat1x1( 3u32 );
assert_eq!(a + b, c);

And this is true for any type that implements Add, so therefore the following is possible as well:

let a = mat1x1(mat1x1(1u32));
let b = mat1x1(mat1x1(2u32));
let c = mat1x1(mat1x1(3u32));
assert_eq!(a + b, c);

For a given type T, if T: Clone and Vector<T, _> is an InnerSpace, then multiplication is defined for Matrix<T, N, M> * Matrix<T, M, P>. The result is a Matrix<T, N, P>:

let a: Matrix::<i32, 3, 3> =
    mat3x3( 0, -3, 5,
            6, 1, -4,
            2, 3, -2 );
let b: Matrix::<i32, 3, 3> =
    mat3x3( -1, 0, -3,
             4, 5, 1,
             2, 6, -2 );
let c: Matrix::<i32, 3, 3> =
    mat3x3( -2, 15, -13,
            -10, -19, -9,
             6, 3, 1 );
    a * b,

It is important to note that this implementation is not necessarily commutative. That is because matrices only need to share one dimension in order to be multiplied together in one direction. From this fact a somewhat pleasant trait bound appears: matrices that satisfy Mul<Self> must be square matrices. This is reflected by a matrix's trait bounds; if a Matrix is square, you are able to extract a diagonal vector from it:

    mat3x3( 1i32, 0, 0,
            0, 2, 0,
            0, 0, 3 )
    vec3( 1i32, 2, 3 ) 

    mat4x4( 1i32, 0, 0, 0, 
            0, 2, 0, 0, 
            0, 0, 3, 0, 
            0, 0, 0, 4 )
    vec4( 1i32, 2, 3, 4 ) 

Matrices can be indexed by either their native column major storage or by the more natural row major method. In order to use row-major indexing, call .index or .index_mut on the matrix with a pair of indices. Calling .index. with a single index will produce a Vector representing the appropriate column of the matrix.

let m: Matrix::<i32, 2, 2> =
           mat2x2( 0, 2,
                   1, 3 );

// Column-major indexing:
assert_eq!(m[0][0], 0);
assert_eq!(m[0][1], 1);
assert_eq!(m[1][0], 2);
assert_eq!(m[1][1], 3);

// Row-major indexing:
assert_eq!(m[(0, 0)], 0);
assert_eq!(m[(1, 0)], 1);
assert_eq!(m[(0, 1)], 2);
assert_eq!(m[(1, 1)], 3);

Provided traits

Zero and One

Defines the additive and multiplicative identity respectively. zero returns vectors and matrices filled with zeroes, while one returns the unit matrix and is unimplemented vectors.


As far as aljabar is considered, a Real is a value that has sqrt defined. By default this is implemented for f32 and f64.


If a Vector implements Add and Sub and its scalar implements Mul and Div, then that vector is part of a VectorSpace.


If two scalars have a distance squared, then vectors of that scalar type implement MetricSpace.


If a vector is a MetricSpace and its scalar is a real number, then .distance() is defined as an alias for .distance2().sqrt()


If a vector is a VectorSpace, a MetricSpace, and its scalar implements Clone, then it is also an InnerSpace and has the inner product defined for it. The inner product is known as the dot product and its respective method in aljabar takes that spelling.

A vector that is an InnerSpace also has .magnitude2() implemented, which is defined as the length of the vector squared.


If a vector is an InnerSpace and its scalar is a real number then it inherits a number of convenience methods:

  • .magnitude(): returns the length of the vector.
  • .normalize(): returns a vector with the same direction and a length of one.
  • .normalize_to(magnitude): returns a vector with the same direction and a length of magnitude.
  • .project_on(other): returns the vector projection of the current inner space onto the given argument.


A Matrix that implements Mul<Self>. Square matrices have the following methods defined for them:

  • .diagonal(): returns the diaganol vector of the matrix.
  • .determinant(): returns the determinant of the matrix. currently only supported for matrices of N <= 3
  • .invert(): returns the inverse of the matrix if one exists. currently unimplemented


Individual component access isn't implemented for vectors and most likely will not be as simple as a public field access when it is. For now, manually destructure the vector or use Index and IndexMut:

let a = Vector1::<u32>::from([ 5 ]);
assert_eq!(a[0], 5);
let mut b = Vector2::<u32>::from([ 1, 2 ]);
b[1] += 3;
assert_eq!(b[1], 5);

Truncation must be performed by manually destructuring as well, but this is due to a limitation of the current compiler.


Please feel free to submit pull requests of any nature.

Future work

There is a lot of work that needs to be done on aljabar before it can be considered stable or non-experimental.

  • Implement Matrix::invert.
  • Add optional Serde support.
  • Add and implement rotational and angular data structures.
  • Convert Vector ops to SIMD implementations.
  • Add more tests and get code coverage to 100%.
  • Implement better methods for accessing components of Vectors.
  • Add Point type.
  • Implement faster matrix multiplication algorithms.
  • Add procedural macro to create matrices of any width and height combination


~17K SLoC