Eigenvalues of a Matrix

Question

Given a square matrix, output the matrix's eigenvalues. Each eigenvalue should be repeated a number of times equal to its algebraic multiplicity.

The eigenvalues of a matrix A are scalar values λ such that, for some column vector v, A*v = λ*v. They are also the solutions to the characteristic polynomial of A: det(A - λ*I) = 0 (where I is the identity matrix with the same dimensions as A).

Outputs must be accurate to 3 significant digits. All inputs and outputs will be within the representable range of numerical values for your chosen language.

Builtins are acceptable, but you are encouraged to include solutions that do not use builtins.

Test Cases

In these test cases, I represents the imaginary unit. Complex numbers are written in the form a + b*I. All outputs have 3 significant digits of precision.

[[42.0]] -> [42.0]
[[1.0, 0.0], [0.0, 1.0]] -> [1.00, 1.00]
[[1.0, 2.0, 3.0], [4.0, 5.0, 6.0], [7.0, 8.0, 9.0]] -> [16.1, -1.12, -1.24e-15]
[[1.2, 3.4, 5.6, 7.8], [6.3, 0.9, -5.4, -2.3], [-12.0, -9.7, 7.3, 5.9], [-2.5, 7.9, 5.3, 4.4]] -> [7.20 + 5.54*I, 7.20 - 5.54*I, -4.35, 3.75]
[[-3.22 - 9.07*I, 0.193 + 9.11*I, 5.59 + 1.33*I, -3.0 - 6.51*I, -3.73 - 6.42*I], [8.49 - 3.46*I, -1.12 + 6.39*I, -8.25 - 0.455*I, 9.37 - 6.43*I, -6.82 + 8.34*I], [-5.26 + 8.07*I, -6.68 + 3.72*I, -3.21 - 5.63*I, 9.31 + 3.86*I, 4.11 - 8.82*I], [-1.24 + 9.04*I, 8.87 - 0.0352*I, 8.35 + 4.5*I, -9.62 - 2.21*I, 1.76 - 5.72*I], [7.0 - 4.79*I, 9.3 - 2.31*I, -2.41 - 7.3*I, -7.77 - 6.85*I, -9.32 + 2.71*I]] -> [5.18 + 16.7*I, -24.9 - 2.01*I, -5.59 - 13.8*I, 0.0438 - 10.6*I, -1.26 + 1.82*I]
[[-30.6 - 73.3*I, 1.03 - 15.6*I, -83.4 + 72.5*I, 24.1 + 69.6*I, 52.3 + 2.68*I, 23.8 + 98.0*I, 96.8 + 49.7*I, -26.2 - 5.87*I, -52.4 + 98.2*I, 78.1 + 6.69*I], [-59.7 - 66.9*I, -26.3 + 65.0*I, 5.71 + 4.75*I, 91.9 + 82.5*I, -94.6 + 51.8*I, 61.7 + 82.3*I, 54.8 - 27.8*I, 45.7 + 59.2*I, -28.3 + 78.1*I, -59.9 - 54.5*I], [-36.0 + 22.9*I, -51.7 + 10.8*I, -46.6 - 88.0*I, -52.8 - 32.0*I, -75.7 - 23.4*I, 96.2 - 71.2*I, -15.3 - 32.7*I, 26.9 + 6.31*I, -59.2 + 25.8*I, -0.836 - 98.3*I], [-65.2 - 90.6*I, 65.6 - 24.1*I, 72.5 + 33.9*I, 1.47 - 93.8*I, -0.143 + 39.0*I, -3.71 - 30.1*I, 60.1 - 42.4*I, 55.6 + 5.65*I, 48.2 - 53.0*I, -3.9 - 33.0*I], [7.04 + 0.0326*I, -12.8 - 50.4*I, 70.1 - 30.3*I, 42.7 - 76.3*I, -3.24 - 64.1*I, 97.3 + 66.8*I, -11.0 + 16.5*I, -40.6 - 90.7*I, 71.5 - 26.2*I, 83.1 - 49.4*I], [-59.5 + 8.08*I, 74.6 + 29.1*I, -65.8 + 26.3*I, -76.7 - 83.2*I, 26.2 + 99.0*I, -54.8 + 33.3*I, 2.79 - 16.6*I, -85.2 - 3.64*I, 98.4 - 12.4*I, -27.6 - 62.3*I], [82.6 - 95.3*I, 55.8 - 73.6*I, -49.9 + 42.1*I, 53.4 + 16.5*I, 80.2 - 43.6*I, -43.3 - 3.9*I, -2.26 - 58.3*I, -19.9 + 98.1*I, 47.2 + 62.4*I, -63.3 - 54.0*I], [-88.7 + 57.7*I, 55.6 + 70.9*I, 84.1 - 52.8*I, 71.3 - 29.8*I, -3.74 - 19.6*I, 29.7 + 1.18*I, -70.6 - 10.5*I, 37.6 + 99.9*I, 87.0 + 19.0*I, -26.1 - 82.0*I], [69.5 - 47.1*I, 11.3 - 59.0*I, -84.3 - 35.1*I, -3.61 - 35.7*I, 88.0 + 88.1*I, -47.5 + 0.956*I, 14.1 + 89.8*I, 51.3 + 0.14*I, -78.5 - 66.5*I, 2.12 - 53.2*I], [0.599 - 71.2*I, 21.7 + 10.8*I, 19.9 - 97.1*I, 20.5 + 37.4*I, 24.7 + 40.6*I, -82.7 - 29.1*I, 77.9 + 12.5*I, 94.1 - 87.4*I, 78.6 - 89.6*I, 82.6 - 69.6*I]] -> [262. - 180.*I, 179. + 117.*I, 10.3 + 214.*I, 102. - 145.*I, -36.5 + 97.7*I, -82.2 + 89.8*I, -241. - 104.*I, -119. - 26.0*I, -140. - 218.*I, -56.0 - 160.*I]

Answer 1

Haskell, 576 554 532 507 bytes

No built-ins!

import Data.Complex
s=sum
l=length
m=magnitude
i=fromIntegral
(&)=zip
t=zipWith
(x!a)b=x*a+b
a#b=[[s$t(*)x y|y<-foldr(t(:))([]<$b)b]|x<-a]
f a|let c=[1..l a];g(u,d)k|m<-[t(+)a b|(a,b)<-a#u&[[s[d|x==y]|y<-c]|x<-c]]=(m,-s[s[b|(n,b)<-c&a,n==m]|(a,m)<-a#m&c]/i k)=snd<$>scanl g(0<$c<$c,1)c
p?x|let f=foldl1(x!);c=l p-1;n=i c;q p=init$t(*)p$i<$>[c,c-1..];o=f(q p)/f p;a|d<-sqrt$(n-1)*(n*(o^2-f(q$q p)/f p)-o^2)=n/last(o-d:[o+d|m(o-d)<m(o+d)])=last$p?(x-a):[x|m a<1e-9]
z[a,b]=[-b/a]
z p=p?0:z(init$scanl1(p?0!)p)

Try it online!

Many thanks @ØrjanJohansen for a total of -47 bytes!

Explanation

First this computes the characteristic polynomial with the Faddeev–LeVerrier algorithm which is the function f. Then the function z computes all roots of that polynomial by iterating g which implements Laguerre's Method for finding a root, once a root is found it is removed and g gets called again until the polynomial has degree 1 which is trivially solved by z[a,b]=[-b/a].

Ungolfed

I re-inlined the functions sum,length,magnitude, fromIntegral, zipWith and (&) as well as the little helper (!). The function faddeevLeVerrier corresponds to f, roots to z and g to laguerre respectively.

-- Transpose a matrix/list
transpose a = foldr (zipWith(:)) (replicate (length a) []) a

-- Straight forward implementation for matrix-matrix multiplication
(#) :: [[Complex Double]] -> [[Complex Double]] -> [[Complex Double]]
a # b = [[sum $ zipWith (*) x y | y <- transpose b]|x<-a]


-- Faddeev-LeVerrier algorithm
faddeevLeVerrier :: [[Complex Double]] -> [Complex Double]
faddeevLeVerrier a = snd <$> scanl go (zero,1) [1..n]
  where n = length a
        zero = replicate n (replicate n 0)
        trace m = sum [sum [b|(n,b)<-zip [1..n] a,n==m]|(m,a)<-zip [1..n] m]
        diag d = [[sum[d|x==y]|y<-[1..n]]|x<-[1..n]]
        add as bs = [[x+y | (x,y) <- zip a b] | (b,a) <- zip as bs]
        go (u,d) k = (m, -trace (a#m) / fromIntegral k)
          where m = add (diag d) (a#u)


-- Compute roots by succesively removing newly computed roots
roots :: [Complex Double] -> [Complex Double]
roots [a,b] = [-b/a]
roots   p   = root : roots (removeRoot p)
  where root = laguerre p 0
        removeRoot = init . scanl1 (\a b -> root*a + b)

-- Compute a root of a polynomial p with an initial guess x
laguerre :: [Complex Double] -> Complex Double -> Complex Double
laguerre p x = if magnitude a < 1e-9 then x else laguerre p new_x
  where evaluate = foldl1 (\a b -> x*a+b)
        order' = length p - 1
        order  = fromIntegral $ length p - 1
        derivative p = init $ zipWith (*) p $ map fromIntegral [order',order'-1..]
        g  = evaluate (derivative p) / evaluate p
        h  = (g ** 2 - evaluate (derivative (derivative p)) / evaluate p)
        d  = sqrt $ (order-1) * (order*h - g**2)
        ga = g - d
        gb = g + d
        s = if magnitude ga < magnitude gb then gb else ga
        a = order /s
        new_x = x - a

Answer 2

Octave, 4 bytes

@eig

Try it online!

Only two bytes more than the MATL golfing language equivalent!

Defines an anonymous function handle to the eig built-in. Interestingly, the MATLAB design philosophy goes against many high-end languages, which like to use DescripteFunctionNamesTakingArguments(), whereas MATLAB and consequently Octave tend to get the shortest unambiguous function name possible. For example, to get a subset of eigenvalues (e.g., the smallest n in absolute magnitude), you use eigs.

As a bonus, here's a function (works in MATLAB, and in theory could work in Octave but their solve is not really up to the task) that doesn't use built-ins, but instead symbolically solves the eigenvalue problem, det(A-λI)=0, and converts it to numerical form using vpa

@(A)vpa(solve(det(A-sym('l')*eye(size(A)))))

Answer 3

MATL, 2 bytes

Yv

Try it online!

Explanation

I followed the usual advice in numerical linear algebra: instead of writing your own function, use a built-in that is specifically designed to avoid numerical instabilities.

Incidentally, it's shorter. ¯\_(ツ)_/¯

Answer 4

Mathematica, 11 bytes

Eigenvalues

Try it online!

Answer 5

R, 22 bytes

function(m)eigen(m)$va

Try it online!

Takes m as a matrix. Frustratingly, the eigen function in R returns an object of class eigen, which has two fields: values, the eigenvalues, and vectors, the eigenvectors.

However, more annoyingly, the optional argument only.values returns a list with two fields, values, containing the eigenvalues, and vectors, set to NULL, but since eigen(m,,T) is also 22 bytes, it's a wash.

Answer 6

Python + numpy, 33 bytes

from numpy.linalg import*
eigvals

Answer 7

Pari/GP, 19 bytes

m->mateigen(m,1)[1]

Try it online!

---

Pari/GP, 24 bytes

It seems that someone downvoted every answer that uses a built-in. OK, if you don't like built-in:

m->polroots(matdet(x-m))

Try it online!

Answer 8

Julia 0.6, 12 bytes

n->eig(n)[1]

Try it online!

Unfortunately, eig returns both the eigenvalues and eigenvectors, as a tuple, so we waste another 9 bytes to lambdify it and grab the first item.