***

    Indexless linear algebra algorithms


    Orthogonalization, linear equations, eigenvalues and eigenvectors
    Literate Haskell module Orthogonals.lhs

    Jan Skibinski, Numeric Quest Inc., Huntsville, Ontario, Canada

    1998.09.19, last modified 1998.12.28


    It has been argued that the functional paradigm offers more support for scientific computing than the traditional imperative programming, such as greater similarity of functional implementation to mathematical specification of a problem. However, efficiency of scientific algorithms implemented in Haskell is very low compared to efficiencies of C or Fortran implementations - notwithstanding the exceptional descriptive power of Haskell.

    It has been also argued that tradition and inertia are partially responsible for this sore state and that many functional algorithms are direct translations of their imperative counterparts.
    Arrays - with their indexing schemes and destructive updating are basic tools of imperative programming. But pure functional languages, which prohibit variable reassignments, cannot compete with imperative languages by using the same tools and following similar reasoning and patterns - unless the functional arrays themselves are designed with performance in mind. This is a case with Clean, where efficiency of one kind of their arrays -- strict unboxed array, approaches efficiency of C.
    But this has not been done for Haskell arrays yet. They are lazy, boxed and use auxiliary association lists (index, value) for initialization -- the latter being mostly responsible for low efficiency of those algorithms that create many interim arrays.
    It appears, that -- as long as indexing scheme is not used for lookups and updates -- Haskell lists are more efficient than arrays -- at least at the currents state of Haskell.

    With this in mind, we are attempting to demonstrate here that the indexing traps can be successfully avoided. This module implements afresh several typical problems from linear algebra. Standard Haskell lists are employed instead of arrays and not a single algorithm ever uses indices for lookups or updates.

    We do not claim high efficiency of these algorithms; consider them exploratory. However, we do claim that the clarity of these algorithms is significantly better than of those functionally similar algorithms that employ indexing schemes.

    Two major algorithms have been invented and implemented in Haskell: one for solving systems of linear equations and one for finding eigenvalues and eigenvectors of almost any type of a square matrix. This includes symmetric, hermitian, general complex or nonsymmetric matrices with real eigenvalues.

    Amazingly, both methods are based on the same factorization, akin to QR method, but not exactly the same as the standard QR one. A simple trick allows to extend this method to nonsymmetric real matrices with complex eigenvalues and thus one method applies to all types of matrices. It follows that the eigenvalue/eigenvector problem can be consistently treated all across the board. In addition, no administrative (housekeeping) boring trivia is required here and that helps to clearly explain the mechanisms employed.


    Contents

    • Notation

    • Scalar products and vector normalization
      • bra_ket, scalar product
      • sum_product, a cousin of bra_ket
      • norm, vector norm
      • normalized, vector normalized to one

    • Transposition and adjoining of matrices
      • transposed, transposed matrix
      • adjoint, transposed and conjugated matrix

    • Products involving matrices
      • matrix_matrix, product of two matrices as list of rows
      • matrix_matrix', product of two matrices as list of columns
      • triangle_matrix', upper triangular matrix times square matrix
      • matrix_ket, matrix times ket vector
      • bra_matrix, bra vector times matrix
      • bra_matrix_ket, matrix multiplied on both sides by vectors
      • scalar_matrix, scalar times matrix

    • Orthogonalization process
      • orthogonals, set of orthogonal vectors
      • gram_schmidt, vector perpendicular to a hyperplane

    • Solutions of linear equations by orthogonalization
      • one_ket_triangle, triangularization of one vector equation
      • one_ket_solution, solution for one unknown vector
      • many_kets_triangle, triangularization of several vector equations
      • many_kets_solution, solution for several unknown vectors

    • Matrix inversion
      • inverse, inverse of a matrix

    • QR factorization of matrices provided by "many_kets_triangle"
      • factors_QR, QR alike factorization of matrices
      • determinant, computation of the determinant based on the QR factorization

    • Similarity transformations and eigenvalues
      • similar_to, matrix obtained by similarity transformation
      • iterated_eigenvalues, list of approximations of eigenvalues
      • eigenvalues, final approximation of eigenvalues

    • Preconditioning of real nonsymmetric matrices
      • add_to_diagonal, simple preconditioning method

    • Examples of iterated eigenvalues
      • Symmetric real matrix
      • Hermitian complex matrix
      • General complex matrix
      • Nonsymmetric real matrix with real eigenvalues
      • Nonsymmetric real matrix with complex eigenvalues

    • Eigenvectors for distinct eigenvalues
      • eigenkets, eigenvectors for distinct eigenvalues

    • Eigenvectors for degenerated eigenvalues
      • eigenket', eigenvector based on a trial vector

    • Auxiliary functions
      • unit_matrix, a unit matrix with 1's on a diagonal
      • unit_vector, a vector with one non-zero component
      • diagonals, vector made of a matrix diagonal


    Notation

    What follows is written in Dirac's notation, as used in Quantum Mechanics. Matrices are represented by capital letters, while vectors come in two varieties:

    • Bra vector x, written < x |, is represented by one-row matrix

    • Ket vector y, written | y >, is represented by one-column matrix

    Bra vectors can be obtained from ket vectors by transposition and conjugation of their components. Conjugation is only important for complex vectors.

    Scalar product of two vectors | x > and | y > is written as

            < x | y >
    
    which looks like a bracket and is sometimes called a "bra_ket". This justifies "bra" and "ket" names introduced by Dirac. There is a good reason for conjugating the components of "bra-vector": the scalar product of
            < x | x >
    
    should be a square of the norm of the vector "x", and that means that it should be represented by a real number, or complex number but with its imaginary part equal to zero.


    
    > module Orthogonals where
    > import Data.Complex
    > import Data.Ratio
    > import qualified Data.List as List
    
    
    Scalar product and vector normalization

    The scalar product "bra_ket" is a basis of many algorithms presented here.

    
    > bra_ket :: (Scalar a, Num a) => [a] -> [a] -> a
    > bra_ket u v =
    >       --
    >       -- Scalar product of two vectors u and v,
    >       -- or < u | v > in Dirac's notation.
    >       -- This is equally valid for both: real and complex vectors.
    >       --
    >       sum_product u (map coupled v)
    
    
    Notice the call to function "coupled" in the above implementation of scalar product. This function conjugates its argument if it is complex, otherwise does not change it. It is defined in the class Scalar - specifically designed for this purpose mainly.
    This class also defines a norm of a vector that might be used by some algorithms. So far we have been able to avoid this.
    
    > class Eq a => Scalar a where
    >     coupled    :: a->a
    >     norm       :: [a] -> a
    >     almostZero :: a -> Bool
    >     scaled     :: [a] -> [a]
    
    > instance Scalar Double where
    >     coupled x    = x
    >     norm u       = sqrt (bra_ket u u)
    >     almostZero x = (abs x) < 1.0e-8
    >     scaled       = scaled'
    
    > instance Scalar Float where
    >     coupled x    = x
    >     norm u       = sqrt (bra_ket u u)
    >     almostZero x = (abs x) < 1.0e-8
    >     scaled       = scaled'
    
    > instance (Integral a) => Scalar (Ratio a) where
    >     coupled x    = x
    >     -- norm u    = fromDouble ((sqrt (bra_ket u u))::Double)
    >     -- Intended hack to silently convert to and from Double.
    >     -- But I do not know how to declare it properly.
    >     --
    >     -- Our type Fraction, when used instead of Ratio a, has its own
    >     -- definition of sqrt. No hack would be needed here.
    >     almostZero x = abs x < 1e-8
    >     scaled       = scaled'
    
    > instance (RealFloat a) => Scalar (Complex a) where
    >     coupled (x:+y) = x:+(-y)
    >     norm u         = sqrt (realPart (bra_ket u u)) :+ 0
    >     almostZero z   = (realPart (abs z)) < 1.0e-8
    >     scaled u       = [(x/m):+(y/m) | x:+y <- u]
    >        where m = maximum [max (abs x) (abs y) | x:+y <- u]
    
    > norm1 :: (Num a) => [a] -> a
    > norm1 = sum . map abs
    
    > norminf :: (Num a, Ord a) => [a] -> a
    > norminf = maximum . map abs
    
    > matnorm1 :: (Num a, Ord a) => [[a]] -> a
    > matnorm1 = matnorminf . transposed
    
    > matnorminf :: (Num a, Ord a) => [[a]] -> a
    > matnorminf = maximum . map norm1
    
    
    
    But we also need a slightly different definition of scalar product that will appear in multiplication of matrices by vectors (or vice versa): a straightforward accumulated product of two lists, where no complex conjugation takes place. We will call it a 'sum_product".
    
    > sum_product :: Num a => [a] -> [a] -> a
    > sum_product u v =
    >       --
    >       -- Similar to scalar product but without
    >       -- conjugations of | u > components
    >       -- Used in matrix-vector or vector-matrix products
    >       --
    >       sum (zipWith (*) u v)
    
    
    Some algorithms might need vectors normalized to one, although we'll try to avoid the normalizations due to its high cost or its inapplicability to rational numbers. Instead, we wiil scale vectors by their maximal components.
    
    > normalized :: (Scalar a, Fractional a) => [a] -> [a]
    > normalized u =
    >       map (/norm u) u
    
    > scaled' :: (Fractional t, Ord t) => [t] -> [t]
    > scaled' u =
    >       map (/norminf u) u
    
    

    Transposition and adjoining of matrices

    Matrices are represented here by lists of lists. Function "transposed" converts from row-wise to column-wise representation, or vice versa.

    When transposition is combined with complex conjugation the resulting matrix is called "adjoint".

    A square matrix is called symmetric if it is equal to its transpose

            A = AT
    
    It is called Hermitian, or self-adjoint, if it equals to its adjoint
            A = A+
    
    > transposed :: [[a]] -> [[a]]
    > transposed a
    >     | null (head a) = []
    >     | otherwise = ([head mi| mi <- a])
    >                   :transposed ([tail mi| mi <- a])
    
    > adjoint :: Scalar a => [[a]] -> [[a]]
    > adjoint a
    >     | null (head a) = []
    >     | otherwise = ([coupled (head mi)| mi <- a])
    >                   :adjoint ([tail mi| mi <- a])
    
    


    Linear combination and sum of two matrices

    One can form a linear combination of two matrices, such as

            C = alpha A + beta B
            where
                alpha and beta are scalars
    
    The most generic form of any combination, not necessary linear, of components of two matrices is given by "matrix_zipWith" function below, which accepts a function "f" describing such combination. For the linear combination with two scalars the function "f" could be defined as:
            f alpha beta a b = alpha*a + beta*b
    
    For a straightforward addition of two matrices this auxiliary function is simply "(+)".
    
    > matrix_zipWith :: (a -> b -> c) -> [[a]] -> [[b]] -> [[c]]
    > matrix_zipWith f a b =
    >     --
    >     -- Matrix made of a combination
    >     -- of matrices a and b - as specified by f
    >     --
    >     [zipWith f ak bk | (ak,bk) <- zip a b]
    
    > add_matrices :: (Num a) => t -> t1 -> [[a]] -> [[a]] -> [[a]]
    > add_matrices _ _ = matrix_zipWith (+)
    
    


    Products involving matrices

    Variety of products involving matrices can be defined. Our Haskell implementation is based on lists of lists and therefore is open to interpretation: sublists can either represent the rows or the columns of a matrix.

    The following definitions are somehow arbitrary, since one can choose alternative interpretations of lists representing matrices.

    C = A B

    Inner product of two matrices A B can be expressed quite simply, providing that matrix A is represented by a list of rows and B - by a list of columns. Function "matrix_matrix" answers list of rows, while "matrix_matrix'" - list of columns.

    Major algorithms of this module make use of "triangle_matrix'", which calculates a product of upper triangular matrix with square matrix and returns a rectangular list of columns.
    
    > matrix_matrix :: Num a => [[a]] -> [[a]] -> [[a]]
    > matrix_matrix a b
    > --
    > -- A matrix being an inner product
    > -- of matrices A and B, where
    > --     A is represented by a list of rows a
    > --     B is represented by a list of columns b
    > --     result is represented by list of rows
    > -- Require: length of a is equal of length of b
    > -- Require: all sublists are of equal length
    >
    >       | null a = []
    >       | otherwise = ([sum_product (head a) bi | bi <- b])
    >                  : matrix_matrix (tail a) b
    
    > matrix_matrix' :: (Num a) => [[a]] -> [[a]] -> [[a]]
    > matrix_matrix' a b
    >       --
    >       -- Similar to "matrix_matrix"
    >       -- but the result is represented by
    >       -- a list of columns
    >       --
    >       | null b = []
    >       | otherwise = ([sum_product ai (head b) | ai <- a])
    >                    : matrix_matrix' a (tail b)
    
    
    > triangle_matrix' :: Num a => [[a]] -> [[a]] -> [[a]]
    > triangle_matrix' r q =
    >       --
    >       -- List of columns of of a product of
    >       -- upper triangular matrix R and square
    >       -- matrix Q
    >       -- where
    >       --     r is a list of rows of R
    >       --     q is a list of columns of A
    >       --
    >       [f r qk | qk <- q]
    >       where
    >           f t u
    >               | null t = []
    >               | otherwise = (sum_product (head t) u)
    >                             : (f (tail t) (tail u))
    
    
    
    
    | u > = A | v >

    Product of a matrix and a ket-vector is another ket vector. The following implementation assumes that list "a" represents rows of matrix A.

    
    > matrix_ket :: Num a => [[a]] -> [a] -> [a]
    > matrix_ket a v = [sum_product ai v| ai <- a]
    
    
    < u | = < v | A

    Bra-vector multiplied by a matrix produces another bra-vector. The implementation below assumes that list "a" represents columns of matrix A. It is also assumed that vector "v" is given in its standard "ket" representation, therefore the definition below uses "bra_ket" instead of "sum_product".

    
    > bra_matrix :: (Scalar a, Num a) => [a] -> [[a]] -> [a]
    > bra_matrix v a = [bra_ket v ai | ai <- a]
    
    
    alpha = < u | A | v >

    This kind of product results in a scalar and is often used to define elements of a new matrix, such as

            B[i,j] = < ei | A | ej >
    
    The implementation below assumes that list "a" represents rows of matrix A.
    
    > bra_matrix_ket :: (Scalar a, Num a) => [a] -> [[a]] -> [a] -> a
    > bra_matrix_ket u a v =
    >     bra_ket u (matrix_ket a v)
    
    
    B = alpha A

    Below is a function which multiplies matrix by a scalar:

    
    > scalar_matrix :: Num a => a -> [[a]] -> [[a]]
    > scalar_matrix alpha a =
    >       [[alpha*aij| aij <- ai] | ai<-a]
    
    


    Orthogonalization process

    Gram-Schmidt orthogonalization procedure is used here for calculation of sets of mutually orthogonal vectors.

    Function "orthogonals" computes a set of mutually orthogonal vectors - all orthogonal to a given vector. Such set plus the input vector form a basis of the vector space. Another words, they are the base vectors, although we cannot call them unit vectors since we do not normalize them for two reasons:
    • None of the algorithms presented here needs this -- quite costly -- normalization.
    • Some algorithms can be used either with doubles or with rationals. The neat output of the latter is sometimes desirable for pedagogical or accuracy reasons. But normalization requires "sqrt" function, which is not defined for rational numbers. We could use our module Fraction instead, where "sqrt" is defined, but we'll leave it for a future revision of this module.

    Function "gram_schmidt" computes one vector - orthogonal to an incomplete set of orthogonal vectors, which form a hyperplane in the vector space. Another words, "gram_schmidt" vector is perpendicular to such a hyperlane.

    
    > orthogonals :: (Scalar a, Fractional a) => [a] -> [[a]]
    > orthogonals x =
    >       --
    >       -- List of (n-1) linearly independent vectors,
    >       -- (mutually orthogonal) and orthogonal to the
    >       -- vector x, but not normalized,
    >       -- where
    >       --     n is a length of x.
    >       --
    >       orth [x] size (next (-1))
    >       where
    >           orth a n m
    >               | n == 1        = drop 1 (reverse a)
    >               | otherwise     = orth ((gram_schmidt a u ):a) (n-1) (next m)
    >               where
    >                   u = unit_vector m size
    >           size = length x
    >           next i = if (i+1) == k then (i+2) else (i+1)
    >           k = length (takeWhile (== 0) x) -- first non-zero component of x
    
    > gram_schmidt :: (Scalar a, Fractional a) => [[a]] -> [a] -> [a]
    > gram_schmidt a u =
    >       --
    >       -- Projection of vector | u > on some direction
    >       -- orthogonal to the hyperplane spanned by the list 'a'
    >       -- of mutually orthogonal (linearly independent)
    >       -- vectors.
    >       --
    >       gram_schmidt' a u u
    >       where
    >           gram_schmidt' [] _ w = w
    >           gram_schmidt' (b:bs) v w
    >               | all (== 0) b = gram_schmidt' bs v w
    >               | otherwise    = gram_schmidt' bs v w'
    >               where
    >                   w' = vectorCombination w (-(bra_ket b v)/(bra_ket b b)) b
    >           vectorCombination x c y
    >               | null x = []
    >               | null y = []
    >               | otherwise = (head x + c * (head y))
    >                             : (vectorCombination (tail x) c (tail y))
    
    


    Solutions of linear equations by orthogonalization

    A matrix equation for unknown vector | x >

            A | x > = | b >
    
    can be rewritten as
            x1 | 1 > + x2 | 2 > + x3 | 3 > + ... + xn | n > = | b >     (7.1)
            where
                    | 1 >, | 2 >... represent columns of the matrix A
    
    For any n-dimensional vector, such as "1", there exist n-1 linearly independent vectors "ck" that are orthogonal to "1"; that is, each satisfies the relation:
            < ck | 1 > = 0, for k = 1...m, where m = n - 1
    
    If we could find all such vectors, then we could multiply the equation (7.1) by each of them, and end up with m = n-1 following equations
            < c1 | 2 > x2 + < c1 | 3 > x3 + ... < c1 | n > xn = < c1 | b >
            < c2 | 2 > x2 + < c2 | 3 > x3 + ... < c2 | n > xn = < c2 | b >
            .......
            < cm | 2 > x2 + < cm | 3 > x3 + ... < cm | n > xn = < cm | b >
    
    But the above is nothing more than a new matrix equation
            A' | x' > = | b' >
            or
    
            x2 | 2'> + x3 | 3'> .... + xn | n'> = | b'>
            where
                primed vectors | 2' >, etc. are the columns of the new
                matrix A'.
    
    with the problem dimension reduced by one.
    Taking as an example a four-dimensional problem and writing down the successive transformations of the original equation we will end up with the following triangular pattern made of four vector equations:
            x1 | 1 > + x2 | 2 > + x3 | 3 >  + x4 | 4 >   = | b >
                       x2 | 2'> + x3 | 3'>  + x4 | 4'>   = | b'>
                                  x3 | 3''> + x4 | 4''>  = | b''>
                                              x4 | 4'''> = | b'''>
    
    But if we premultiply each vector equation by a non-zero vector of our choice, say < 1 | , < 2' |, < 3'' |, and < 4''' | - chosen correspondingly for equations 1, 2, 3 and 4, then the above system of vector equations will be converted to much simpler system of scalar equations. The result is shown below in matrix representation:
            | p11  p12   p13   p14 | | x1 | = | q1 |
            | 0    p22   p23   p24 | | x2 | = | q2 |
            | 0    0     p33   p34 | | x3 | = | q3 |
            | 0    0     0     p44 | | x4 | = | q4 |
    
    In effect, we have triangularized our original matrix A. Below is a function that does that for any problem size:
    
    > one_ket_triangle :: (Scalar a, Fractional a) => [[a]] -> [a] -> [([a],a)]
    > one_ket_triangle a b
    >     --
    >     -- List of pairs: (p, q) representing
    >     -- rows of triangular matrix P and of vector | q >
    >     -- in the equation P | x > = | q >, which
    >     -- has been obtained by linear transformation
    >     -- of the original equation A | x > = | b >
    >     --
    >     | null a = []
    >     | otherwise = (p,q):(one_ket_triangle a' b')
    >     where
    >         p    = [bra_ket u ak | ak <- a]
    >         q    = bra_ket u b
    >         a'   = [[bra_ket ck ai | ck <- orth] | ai <- v]
    >         b'   = [ bra_ket ck b  | ck <- orth]
    >         orth = orthogonals u
    >         u    = head a
    >         v    = tail a
    
    
    The triangular system of equations can be easily solved by successive substitutions - starting with the last equation.
    
    > one_ket_solution :: (Scalar a, Fractional a) => [[a]] -> [a] -> [a]
    > one_ket_solution a b =
    >     --
    >     -- List representing vector |x>, which is
    >     -- a solution of the matrix equation
    >     --     A |x> = |b>
    >     -- where
    >     --     a is a list of columns of matrix A
    >     --     b is a list representing vector |b>
    >     --
    >     solve' (unzip (reverse (one_ket_triangle a b))) []
    >     where
    >         solve' (d, c) xs
    >             | null d  = xs
    >             | otherwise = solve' ((tail d), (tail c)) (x:xs)
    >             where
    >                 x = (head c - (sum_product (tail u) xs))/(head u)
    >                 u = head d
    
    
    The triangularization procedure can be easily extended to a list of several ket-vectors | b > on the right hand side of the original equation A | x > = | b > -- instead of just one:
    
    > many_kets_triangle :: (Scalar a, Fractional a) => [[a]] -> [[a]] -> [([a],[a])]
    > many_kets_triangle a b
    >     --
    >     -- List of pairs: (p, q) representing
    >     -- rows of triangular matrix P and of rectangular matrix Q
    >     -- in the equation P X = Q, which
    >     -- has been obtained by linear transformation
    >     -- of the original equation A X = B
    >     -- where
    >     --     a is a list of columns of matrix A
    >     --     b is a list of columns of matrix B
    >     --
    >     | null a = []
    >     | otherwise = (p,q):(many_kets_triangle a' b')
    >     where
    >         p    = [bra_ket u ak   | ak <- a]
    >         q    = [bra_ket u bk   | bk <- b]
    >         a'   = [[bra_ket ck ai | ck <- orth] | ai <- v]
    >         b'   = [[bra_ket ck bi | ck <- orth] | bi <- b]
    >         orth = orthogonals u
    >         u    = head a
    >         v    = tail a
    
    
    Similarly, function 'one_ket_solution' can be generalized to function 'many_kets_solution' that handles cases with several ket-vectors on the right hand side.
    
    > many_kets_solution :: (Scalar a, Fractional a) => [[a]] -> [[a]] -> [[a]]
    > many_kets_solution a b =
    >     --
    >     -- List of columns of matrix X, which is
    >     -- a solution of the matrix equation
    >     --     A X = B
    >     -- where
    >     --     a is a list of columns of matrix A
    >     --     b is a list of columns of matrix B
    >     --
    >     solve' p q emptyLists
    >     where
    >         (p, q) = unzip (reverse (many_kets_triangle a b))
    >         emptyLists = [[] | _ <- [1..(length (head q))]]
    >         solve' a' b' x
    >             | null a'  = x
    >             | otherwise = solve' (tail a') (tail b')
    >                                 [(f vk xk):xk  | (xk, vk) <- (zip x v)]
    >             where
    >                 f vk xk = (vk - (sum_product (tail u) xk))/(head u)
    >                 u = head a'
    >                 v = head b'
    
    
    


    Matrix inversion

    Function 'many_kets_solution' can be used to compute inverse of matrix A by specializing matrix B to a unit matrix I:

    
            A X = I
    
    It follows that matrix X is an inverse of A; that is X = A-1.
    
    > inverse :: (Scalar a, Fractional a) => [[a]] -> [[a]]
    > inverse a = many_kets_solution a (unit_matrix (length a))
    >       --
    >       -- List of columns of inverse of matrix A
    >       -- where
    >       --     a is list of columns of A
    
    


    QR factorization of matrices

    The process described above and implemented by 'many_kets_triangle' function transforms the equation

            A X = B
    
    into another equation for the same matrix X
            R X = S
    
    where R is an upper triangular matrix. All operations performed on matrices A and B during this process are linear, and therefore we should be able to find a square matrix Q that describes the entire process in one step. Indeed, assuming that matrix A can be decomposed as a product of unknown matrix Q and triangular matrix R and that Q-1 is an inverse of matrix Q we can reach the last equation by following these steps:
            A X       = B
            (Q R) X   = B
            Q-1 Q R X = Q-1 B
            R X       = S
    
    It follows that during this process a given matrix B transforms to matrix S, as delivered by 'many_kets_triangle':
            S = Q-1 B
    
    from which the inverse of Q can be found:
            Q-1 = S B-1
    
    Having a freedom of choice of the right hand side matrix B we can choose the unit matrix I in place of B, and therefore simplify the definition of Q-1:
            Q-1 = S,  if B is unit matrix
    
    It follows that any non-singular matrix A can be decomposed as a product of a matrix Q and a triangular matrix R
            A = Q R
    
    where matrices Q-1 and R are delivered by "many_kets_triangle" as a result of triangularization process of equation:
            A X = I
    
    The function below extracts a pair of matrices Q and R from the answer provided by "many_kets_triangle". During this process it inverts matrix Q-1 to Q. This factorization will be used by a sequence of similarity transformations to be defined in the next section.
    
    > factors_QR :: (Scalar a, Fractional a) => [[a]] -> ([[a]],[[a]])
    > factors_QR a =
    >       --
    >       -- A pair of matrices (Q, R), such that
    >       -- A = Q R
    >       -- where
    >       --     R is upper triangular matrix in row representation
    >       --     (without redundant zeros)
    >       --     Q is a transformation matrix in column representation
    >       --     A is square matrix given as columns
    >       --
    >       (inverse (transposed q1),r)
    >       where
    >           (r, q1) = unzip (many_kets_triangle a (unit_matrix (length a)))
    
    


    Computation of the determinant

    
    > determinant :: (Scalar a, Fractional a) => [[a]] -> a
    > determinant a =
    >    let (q,r) = factors_QR a
    >    -- matrix Q is not normed so we have to respect the norms of its rows
    >    in  product (map norm q) * product (map head r)
    
    
    Naive division-free computation of the determinant by expanding the first column. It consumes n! multiplications.
    
    > determinantNaive :: (Num a) => [[a]] -> a
    > determinantNaive [] = 1
    > determinantNaive m  =
    >    sum (alternate
    >       (zipWith (*) (map head m)
    >           (map determinantNaive (removeEach (map tail m)))))
    
    
    Compute the determinant with about n^4 multiplications without division according to the clow decomposition algorithm of Mahajan and Vinay, and Berkowitz as presented by Günter Rote: Division-Free Algorithms for the Determinant and the Pfaffian: Algebraic and Combinatorial Approaches.
    
    > determinantClow :: (Num a) => [[a]] -> a
    > determinantClow [] = 1
    > determinantClow m =
    >    let lm = length m
    >    in  parityFlip lm (last (newClow m
    >           (nest (lm-1) (longerClow m)
    >               (take lm (iterate (0:) [1])))))
    
    
    Compute the weights of all clow sequences where the last clow is closed and a new one is started.
    
    > newClow :: (Num a) => [[a]] -> [[a]] -> [a]
    > newClow a c =
    >    scanl (-) 0
    >          (sumVec (zipWith (zipWith (*)) (List.transpose a) c))
    
    
    Compute the weights of all clow sequences where the last (open) clow is extended by a new arc.
    
    > extendClow :: (Num a) => [[a]] -> [[a]] -> [[a]]
    > extendClow a c =
    >    map (\ai -> sumVec (zipWith scaleVec ai c)) a
    
    
    Given the matrix of all weights of clows of length l compute the weight matrix for all clows of length (l+1). Take the result of 'newClow' as diagonal and the result of 'extendClow' as lower triangle of the weight matrix.
    
    > longerClow :: (Num a) => [[a]] -> [[a]] -> [[a]]
    > longerClow a c =
    >    let diagonal = newClow a c
    >        triangle = extendClow a c
    >    in  zipWith3 (\i t d -> take i t ++ [d]) [0 ..] triangle diagonal
    
    
    Auxiliary functions for the clow determinant.
    
    > {- | Compositional power of a function,
    >      i.e. apply the function n times to a value. -}
    > nest :: Int -> (a -> a) -> a -> a
    > nest 0 _ x = x
    > nest n f x = f (nest (n-1) f x)
    >
    > {- successively select elements from xs and remove one in each result list -}
    > removeEach :: [a] -> [[a]]
    > removeEach xs =
    >    zipWith (++) (List.inits xs) (tail (List.tails xs))
    >
    > alternate :: (Num a) => [a] -> [a]
    > alternate = zipWith id (cycle [id, negate])
    >
    > parityFlip :: Num a => Int -> a -> a
    > parityFlip n x = if even n then x else -x
    >
    > {-| Weight a list of numbers by a scalar. -}
    > scaleVec :: (Num a) => a -> [a] -> [a]
    > scaleVec k = map (k*)
    >
    > {-| Add corresponding numbers of two lists. -}
    > {- don't use zipWith because it clips to the shorter list -}
    > addVec :: (Num a) => [a] -> [a] -> [a]
    > addVec x [] = x
    > addVec [] y = y
    > addVec (x:xs) (y:ys) = x+y : addVec xs ys
    >
    > {-| Add some lists. -}
    > sumVec :: (Num a) => [[a]] -> [a]
    > sumVec = foldl addVec []
    
    


    Similarity transformations and eigenvalues

    Two n-square matrices A and B are called similar if there exists a non-singular matrix S such that:

            B = S-1 A S
    
    It can be proven that:
    • Any two similar matrices have the same eigenvalues
    • Every n-square matrix A is similar to a triangular matrix whose diagonal elements are the eigenvalues of A.

    If matrix A can be transformed to a triangular or a diagonal matrix Ak by a sequence of similarity transformations then the eigenvalues of matrix A are the diagonal elements of Ak.

    Let's construct the sequence of matrices similar to A

            A, A1, A2, A3...
    
    by the following iterations - each of which factorizes a matrix by applying the function 'factors_QR' and then forms a product of the factors taken in the reverse order:
            A = Q R          = Q (R Q) Q-1    = Q A1 Q-1        =
              = Q (Q1 R1) Q-1 = Q Q1 (R1 Q1) Q1-1 Q-1 = Q Q1 A2 Q1-1 Q-1 =
              = Q Q1 (Q2 R2) Q1-1 Q-1 = ...
    
    
    We are hoping that after some number of iterations some matrix Ak would become triangular and therefore its diagonal elements could serve as eigenvalues of matrix A. As long as a matrix has real eigenvalues only, this method should work well. This applies to symmetric and hermitian matrices. It appears that general complex matrices -- hermitian or not -- can also be handled this way. Even more, this method also works for some nonsymmetric real matrices, which have real eigenvalues only.
    The only type of matrices that cannot be treated by this algorithm are real nonsymmetric matrices, whose some eigenvalues are complex. There is no operation in the process that converts real elements to complex ones, which could find their way into diagonal positions of a triangular matrix. But a simple preconditioning of a matrix -- described in the next section -- replaces a real matrix by a complex one, whose eigenvalues are related to the eigenvalues of the matrix being replaced. And this allows us to apply the same method all across the board.
    It is worth noting that a process known in literature as QR factorization is not uniquely defined and different algorithms are employed for this. The algorithms using QR factorization apply only to symmetric or hermitian matrices, and Q matrix must be either orthogonal or unitary.
    But our transformation matrix Q is not orthogonal nor unitary, although its first row is orthogonal to all other rows. In fact, this factorization is only similar to QR factorization. We just keep the same name to help identify a category of the methods to which it belongs.
    The same factorization is used for tackling two major problems: solving the systems of linear equations and finding the eigenvalues of matrices.
    Below is the function 'similar_to', which makes a new matrix that is similar to a given matrix by applying our similarity transformation.
    Function 'iterated_eigenvalues' applies this transformation n times - storing diagonals of each new matrix as approximations of eigenvalues.
    Function 'eigenvalues' follows the same process but reports the last approximation only.
    
    
    > similar_to :: (Scalar a, Fractional a) => [[a]] -> [[a]]
    > similar_to a =
    >       --
    >       -- List of columns of matrix A1 similar to A
    >       -- obtained by factoring A as Q R and then
    >       -- forming the product A1 = R Q = (inverse Q) A Q
    >       -- where
    >       --     a is list of columns of A
    >       --
    >       triangle_matrix' r q
    >       where
    >           (q,r) = factors_QR a
    
    > iterated_eigenvalues :: (Scalar a1, Fractional a1, Eq a, Num a) => [[a1]] -> a -> [[a1]]
    > iterated_eigenvalues a n
    >       --
    >       -- List of vectors representing
    >       -- successive approximations of
    >       -- eigenvalues of matrix A
    >       -- where
    >       --     a is a list of columns of A
    >       --     n is a number of requested iterations
    >       --
    >       | n == 0 = []
    >       | otherwise = (diagonals a)
    >                     : iterated_eigenvalues (similar_to a) (n-1)
    
    > eigenvalues :: (Scalar a1, Fractional a1, Eq a, Num a) => [[a1]] -> a -> [a1]
    > eigenvalues a n
    >       --
    >       -- Eigenvalues of matrix A
    >       -- obtained by n similarity iterations
    >       -- where
    >       --     a are the columns of A
    >       --
    >       | n == 0    = diagonals a
    >       | otherwise = eigenvalues (similar_to a) (n-1)
    
    


    Preconditioning of real nonsymmetric matrices

    As mentioned above, our QR-like factorization method works well with almost all kind of matrices, but with the exception of a class of real nonsymmetric matrices that have complex eigenvalues.

    There is no mechanism in that method that would be able to produce complex eigenvalues out of the real components of this type of nonsymmetric matrices. Simple trivial replacement of real components of a matrix by its complex counterparts does not work because zero-valued imaginary components do not contribute in any way to production of nontrivial imaginary components during the factorization process.
    What we need is a trick that replaces real nonsymmetric matrix by a nontrivial complex matrix in such a way that the results of such replacements could be undone when the series of similarity transformations finally produced the expected effect in a form of a triangular matrix.
    The practical solution is surprisingly simple: it's suffice to add any complex number, such as "i", to the main diagonal of a matrix, and when triangularization is done -- subtract it back from computed eigenvalues. The explanation follows.

    Consider the eigenproblem for real and nonsymmetric matrix A.

            A | x > = a | x >
    
    Let us now define a new complex matrix B, such that:
            B = A + alpha I
            where
                I is a unit matrix and alpha is a complex scalar
    
    It is obvious that matrices A and B commute; that is:
            A B = B A
    
    It can be proven that if two matrices commute then they have the same eigenvectors. Therefore we can use vector | x > of matrix A as an eigenvector of B:
            B | x > = b | x >
            B | x > = A | x > + alpha I | x >
                    = a | x > + alpha | x >
                    = (a + alpha) | x >
    
    It follows that eigenvalues of B are related to the eigenvalues of A by:
            b = a + alpha
    
    After eigenvalues of complex matrix B have been succesfully computed, all what remains is to subtract "alpha" from them all to obtain eigenvalues of A. And nothing has to be done to eigenvectors of B - they are the same for A as well. Simple and elegant!

    Below is an auxiliary function that adds a scalar to the diagonal of a matrix:

    
    > add_to_diagonal :: Num a => a -> [[a]] -> [[a]]
    > add_to_diagonal alpha a =
    >       --
    >       -- Add constant alpha to diagonal of matrix A
    >       --
    >       [f ai ni | (ai,ni) <- zip a [0..(length a -1)]]
    >       where
    >           f b k = p++[head q + alpha]++(tail q)
    >               where
    >                   (p,q) = splitAt k b
    >
    
    
    


    Examples of iterated eigenvalues

    Here is an example of a symmetric real matrix with results of application of function 'iterated_eigenvalues'.

            | 7  -2  1 |
            |-2  10 -2 |
            | 1  -2  7 |
    
             [[7.0,     10.0,    7.0],
              [8.66667, 9.05752, 6.27582],
              [10.7928, 7.11006, 6.09718],
              [11.5513, 6.40499, 6.04367],
              [11.7889, 6.18968, 6.02142],
              [11.8943, 6.09506, 6.01068],
              [11.9468, 6.04788, 6.00534],
              [11.9733, 6.02405, 6.00267],
              [11.9866, 6.01206, 6.00134],
              [11.9933, 6.00604, 6.00067],
              [11.9966, 6.00302, 6.00034],
              [11.9983, 6.00151, 6.00017],
              [11.9992, 6.00076, 6.00008],
              [11.9996, 6.00038, 6.00004],
              [11.9998, 6.00019, 6.00002],
              [11.9999, 6.00010, 6.00001],
              [11.9999, 6.00005, 6.00001]]
    
              The true eigenvalues are:
              12, 6, 6
    
    
    Here is an example of a hermitian matrix. (Eigenvalues of hermitian matrices are real.) The algorithm works well and converges fast.
            | 2   0     i|
            [ 0   1   0  |
            [ -i  0   2  |
    
            [[2.8     :+ 0.0, 1.0 :+ 0.0, 1.2     :+ 0.0],
             [2.93979 :+ 0.0, 1.0 :+ 0.0, 1.06021 :+ 0.0],
             [2.97972 :+ 0.0, 1.0 :+ 0.0, 1.02028 :+ 0.0],
             [2.9932  :+ 0.0, 1.0 :+ 0.0, 1.0068  :+ 0.0],
             [2.99773 :+ 0.0, 1.0 :+ 0.0, 1.00227 :+ 0.0],
             [2.99924 :+ 0.0, 1.0 :+ 0.0, 1.00076 :+ 0.0],
             [2.99975 :+ 0.0, 1.0 :+ 0.0, 1.00025 :+ 0.0],
             [2.99992 :+ 0.0, 1.0 :+ 0.0, 1.00008 :+ 0.0],
             [2.99997 :+ 0.0, 1.0 :+ 0.0, 1.00003 :+ 0.0],
             [2.99999 :+ 0.0, 1.0 :+ 0.0, 1.00001 :+ 0.0],
             [3.0     :+ 0.0, 1.0 :+ 0.0, 1.0     :+ 0.0],
             [3.0     :+ 0.0, 1.0 :+ 0.0, 1.0     :+ 0.0],
             [3.0     :+ 0.0, 1.0 :+ 0.0, 1.0     :+ 0.0]]
    
    
    Here is another example: this is a complex matrix and it is not even hermitian. Yet, the algorithm still works, although its fluctuates around true values.
            | 2-i   0      i |
            | 0     1+i  0   |
            |   i   0    2-i |
    
            [[2.0     :+ (-1.33333), 1.0 :+ 1.0, 2.0     :+ (-0.666667)],
             [1.89245 :+ (-1.57849), 1.0 :+ 1.0, 2.10755 :+ (-0.421509)],
             [1.81892 :+ (-1.80271), 1.0 :+ 1.0, 2.18108 :+ (-0.197289)],
             [1.84565 :+ (-1.99036), 1.0 :+ 1.0, 2.15435 :+ (-0.00964242)],
             [1.93958 :+ (-2.07773), 1.0 :+ 1.0, 2.06042 :+ 0.0777281],
             [2.0173  :+ (-2.06818), 1.0 :+ 1.0, 1.9827  :+ 0.0681793],
             [2.04357 :+ (-2.02437), 1.0 :+ 1.0, 1.95643 :+ 0.0243654],
             [2.03375 :+ (-1.99072), 1.0 :+ 1.0, 1.96625 :+ (-0.00928429)],
             [2.01245 :+ (-1.97875), 1.0 :+ 1.0, 1.98755 :+ (-0.0212528)],
             [1.99575 :+ (-1.98307), 1.0 :+ 1.0, 2.00425 :+ (-0.0169263)],
             [1.98938 :+ (-1.99359), 1.0 :+ 1.0, 2.01062 :+ (-0.00640583)],
             [1.99145 :+ (-2.00213), 1.0 :+ 1.0, 2.00855 :+ 0.00212504],
             [1.9968  :+ (-2.00535), 1.0 :+ 1.0, 2.0032  :+ 0.00535265],
             [2.00108 :+ (-2.00427), 1.0 :+ 1.0, 1.99892 :+ 0.0042723],
             [2.00268 :+ (-2.00159), 1.0 :+ 1.0, 1.99732 :+ 0.00158978],
             [2.00213 :+ (-1.99946), 1.0 :+ 1.0, 1.99787 :+ (-0.000541867)],
             [2.00079 :+ (-1.99866), 1.0 :+ 1.0, 1.9992  :+ (-0.00133514)],
             [1.99973 :+ (-1.99893), 1.0 :+ 1.0, 2.00027 :+ (-0.00106525)],
             [1.99933 :+ (-1.9996) , 1.0 :+ 1.0, 2.00067 :+ (-0.000397997)],
             [1.99947 :+ (-2.00013), 1.0 :+ 1.0, 2.00053 :+ 0.000134972]]
    
             The true eigenvalues are
             2 - 2i, 1 + i, 2
    
    Some nonsymmetric real matrices have all real eigenvalues and our algorithm still works for such cases. Here is one such an example, which traditionally would have to be treated by one of the Lanczos-like algorithms, specifically designed for nonsymmetric real matrices. Evaluation of
    iterated_eigenvalues [[2,1,1],[-2,1,3],[3,1,-1::Double]] 20
    gives the following results
            [[3.0,     -0.70818,-0.291815],
             [3.06743, -3.41538, 2.34795],
             [3.02238, -1.60013, 0.577753],
             [3.00746, -2.25793, 1.25047],
             [3.00248, -1.88764, 0.885154],
             [3.00083, -2.06025, 1.05943],
             [3.00028, -1.97098, 0.970702],
             [3.00009, -2.0148,  1.01471],
             [3.00003, -1.99268, 0.992648],
             [3.00001, -2.00368, 1.00367],
             [3.0,     -1.99817, 0.998161],
             [3.0,     -2.00092, 1.00092],
             [3.0,     -1.99954, 0.99954],
             [3.0,     -2.00023, 1.00023],
             [3.0,     -1.99989, 0.999885],
             [3.0,     -2.00006, 1.00006],
             [3.0,     -1.99997, 0.999971],
             [3.0,     -2.00001, 1.00001],
             [3.0,     -1.99999, 0.999993],
             [3.0,     -2.0,     1.0]]
    
             The true eigenvalues are:
             3, -2, 1
    
    Finally, here is a case of a nonsymmetric real matrix with complex eigenvalues:
            | 2 -3 |
            | 1  0 |
    
    The direct application of "iterated_eigenvalues" would fail to produce expected eigenvalues:
            1 + i sqrt(2) and 1 - i sqrt (2)
    
    But if we first precondition the matrix by adding "i" to its diagonal:
            | 2+i  -3|
            | 1     i|
    
    and then compute its iterated eigenvalues:
    iterated_eigenvalues [[2:+1,1],[-3,0:+1]] 20
    then the method will succeed. Here are the results:
    
            [[1.0     :+ 1.66667, 1.0     :+   0.333333 ],
            [0.600936 :+ 2.34977, 1.39906 :+ (-0.349766)],
            [0.998528 :+ 2.59355, 1.00147 :+ (-0.593555)],
            [1.06991  :+ 2.413,   0.93009 :+ (-0.412998)],
            [1.00021  :+ 2.38554, 0.99979 :+ (-0.385543)],
            [0.988004 :+ 2.41407, 1.012   :+ (-0.414074)],
            [0.999963 :+ 2.41919, 1.00004 :+ (-0.419191)],
            [1.00206  :+ 2.41423, 0.99794 :+ (-0.414227)],
            [1.00001  :+ 2.41336, 0.99999 :+ (-0.413361)],
            [0.999647 :+ 2.41421, 1.00035 :+ (-0.414211)],
            [0.999999 :+ 2.41436, 1.0     :+ (-0.41436) ],
            [1.00006  :+ 2.41421, 0.99993 :+ (-0.414214)],
            [1.0      :+ 2.41419, 1.0     :+ (-0.414188)],
            [0.99999  :+ 2.41421, 1.00001 :+ (-0.414213)],
            [1.0      :+ 2.41422, 1.0     :+ (-0.414218)],
            [1.0      :+ 2.41421, 0.99999 :+ (-0.414213)],
            [1.0      :+ 2.41421, 1.0     :+ (-0.414212)],
            [1.0      :+ 2.41421, 1.0     :+ (-0.414213)],
            [1.0      :+ 2.41421, 1.0     :+ (-0.414213)],
            [1.0      :+ 2.41421, 1.0     :+ (-0.414213)]]
    
    After subtracting "i" from the last result, we will get what is expected.


    Eigenvectors for distinct eigenvalues

    Assuming that eigenvalues of matrix A are already found we may now attempt to find the corresponding aigenvectors by solving the following homogeneous equation

            (A - a I) | x > = 0
    
    for each eigenvalue "a". The matrix
            B = A - a I
    
    is by definition singular, but in most cases it can be triangularized by the familiar "factors_QR" procedure.
            B | x > = Q R | x > = 0
    
    It follows that the unknown eigenvector | x > is one of the solutions of the homogeneous equation:
            R | x > = 0
    
    where R is a singular, upper triangular matrix with at least one zero on its diagonal.
    If | x > is a solution we seek, so is its scaled version alpha | x >. Therefore we have some freedom of scaling choice. Since this is a homogeneous equation, one of the components of | x > can be freely chosen, while the remaining components will depend on that choice. To solve the above, we will be working from the bottom up of the matrix equation, as illustrated in the example below:
            | 0     1     1     3     | | x1 |
            | 0     1     1     2     | | x2 |      /\
            | 0     0     2     4     | | x3 | = 0  ||
            | 0     0     0     0     | | x4 |      ||
    
    Recall that the diagonal elements of any triangular matrix are its eigenvalues. Our example matrix has three distinct eigenvalues: 0, 1, 2. The eigenvalue 0 has degree of degeneration two. Presence of degenerated eigenvalues complicates the solution process. The complication arises when we have to make our decision about how to solve the trivial scalar equations with zero coefficients, such as
            0 * x4 = 0
    
    resulting from multiplication of the bottom row by vector | x >. Here we have two choices: "x4" could be set to 0, or to any nonzero number 1, say. By always choosing the "0" option we might end up with the all-zero trivial vector -- which is obviously not what we want. Persistent choice of the "1" option, might lead to a conflict between some of the equations, such as the equations one and four in our example.

    So the strategy is as follows.

    If there is at least one zero on the diagonal, find the topmost row with zero on the diagonal and choose for it the solution "1". Diagonal zeros in other rows would force the solution "0". If the diagonal element is not zero than simply solve an arithmetic equation that arises from the substitutions of previously computed components of the eigenvector. Since certain inaccuracies acumulate during QR factorization, set to zero all very small elements of matrix R.

    By applying this strategy to our example we'll end up with the eigenvector

            < x | = [1, 0, 0, 0]
    

    If the degree of degeneration of an eigenvalue of A is 1 then the corresponding eigenvector is unique -- subject to scaling. Otherwise an eigenvector found by this method is one of many possible solutions, and any linear combination of such solutions is also an eigenvector. This method is not able to find more than one solution for degenerated eigenvalues. An alternative method, which handles degenerated cases, will be described in the next section.

    The function below calculates eigenvectors corresponding to distinct selected eigenvalues of any square matrix A, provided that the singular matrix B = A - a I can still be factorized as Q R, where R is an upper triangular matrix.

    
    > eigenkets :: (Scalar a, Fractional a) => [[a]] -> [a] -> [[a]]
    > eigenkets a u
    >       --
    >       -- List of eigenkets of a square matrix A
    >       -- where
    >       --     a is a list of columns of A
    >       --     u is a list of eigenvalues of A
    >       --     (This list does not need to be complete)
    >       --
    >       | null u        = []
    >       | not (null x') = x':(eigenkets a (tail u))
    >       | otherwise     = (eigenket_UT (reverse b) d []):(eigenkets a (tail u))
    >       where
    >           a'  = add_to_diagonal (-(head u)) a
    >           x'  = unit_ket a' 0 (length a')
    >           b   = snd (factors_QR a')
    >           d   = discriminant [head bk | bk <- b] 1
    >           discriminant v n
    >               | null v    = []
    >               | otherwise = x : (discriminant (tail v) m)
    >               where
    >                   (x, m)
    >                       | (head u) == 0     = (n, 0)
    >                       | otherwise         = (n, n)
    >           eigenket_UT c e xs
    >               | null c    = xs
    >               | otherwise = eigenket_UT (tail c) (tail e) (x:xs)
    >               where
    >                   x = solve_row (head c) (head e) xs
    >
    >           solve_row (v:vs) n x
    >               | almostZero v = n
    >               | otherwise    = q/v
    >               where
    >                   q
    >                       | null x = 0
    >                       | otherwise = -(sum_product vs x)
    >
    >           unit_ket b' m n
    >               | null b'              = []
    >               | all (== 0) (head b') = unit_vector m n
    >               | otherwise            = unit_ket (tail b') (m+1) n
    
    


    Eigenvectors for degenerated eigenvalues

    Few facts:

    • Eigenvectors of a general matrix A, which does not have any special symmetry, are not generally orthogonal. However, they are orthogonal, or can be made orthogonal, to another set of vectors that are eigenvectors of adjoint matrix A+; that is the matrix obtained by complex conjugation and transposition of matrix A.
    • Eigenvectors corresponding to nondegenerated eigenvalues of hermitian or symmetric matrix are orthogonal.
    • Eigenvectors corresponding to degenerated eigenvalues are - in general - neither orthogonal among themselves, nor orthogonal to the remaining eigenvectors corresponding to other eigenvalues. But since any linear combination of such degenerated eigenvectors is also an eigenvector, we can orthogonalize them by Gram-Schmidt orthogonalization procedure.
    Many practical applications deal solely with hermitian or symmetric matrices, and for such cases the orthogonalization is not only possible, but also desired for variety of reasons.
    But the method presented in the previous section is not able to find more than one eigenvector corresponding to a degenerated eigenvalue. For example, the symmetric matrix
                |  7  -2   1 |
            A = | -2  10  -2 |
                |  1  -2   7 |
    
    has two distinct eigenvalues: 12 and 6 -- the latter being degenerated with degree of two. Two corresponding eigenvectors are:
            < x1 | = [1, -2, 1] -- for 12
            < x2 | = [1,  1, 1] -- for 6
    
    It happens that those vectors are orthogonal, but this is just an accidental result. However, we are missing a third distinct eigenvector. To find it we need another method. One possibility is presented below and the explanation follows.
    
    > eigenket' :: (Scalar a, Fractional a) => [[a]] -> a -> a -> [a] -> [a]
    > eigenket' a alpha eps x' =
    >       --
    >       -- Eigenket of matrix A corresponding to eigenvalue alpha
    >       -- where
    >       --     a is a list of columns of matrix A
    >       --     eps is a trial inaccuracy factor
    >       --         artificially introduced to cope
    >       --         with singularities of A - alpha I.
    >       --         One might try eps = 0, 0.00001, 0.001, etc.
    >       --     x' is a trial eigenvector
    >       --
    >       scaled [xk' - dk | (xk', dk) <- zip x' d]
    >       where
    >           b = add_to_diagonal (-alpha*(1+eps)) a
    >           d = one_ket_solution b y
    >           y = matrix_ket (transposed b) x'
    
    
    Let us assume a trial vector | x' >, such that
            | x' > = | x > + | d >
            where
                | x > is an eigenvector we seek
                | d > is an error of our estimation of | x >
    
    We first form a matrix B, such that:
            B = A - alpha I
    
    and multiply it by the trial vector | x' >, which results in a vector | y >
            B | x' > = |y >
    
    On another hand:
            B | x' > = B | x > + B | d > = B | d >
            because
                B | x > = A | x > - alpha | x > = 0
    
    Comparing both equations we end up with:
            B | d > = | y >
    
    that is: with the system of linear equations for unknown error | d >. Finally, we subtract error | d > from our trial vector | x' > to obtain the true eigenvector | x >.

    But there is some problem with this approach: matrix B is by definition singular, and as such, it might be difficult to handle. One of the two processes might fail, and their failures relate to division by zero that might happen during either the QR factorization, or the solution of the triangular system of equations.

    But if we do not insist that matrix B should be exactly singular, but almost singular:

            B = A - alpha (1 + eps) I
    
    then this method might succeed. However, the resulting eigenvector will be the approximation only, and we would have to experiment a bit with different values of "eps" to extrapolate the true eigenvector.

    The trial vector | x' > can be chosen randomly, although some choices would still lead to singularity problems. Aside from this, this method is quite versatile, because:

    • Any random vector | x' > leads to the same eigenvector for nondegenerated eigenvalues,
    • Different random vectors | x' >, chosen for degenerated eigenvalues, produce -- in most cases -- distinct eigenvectors. And this is what we want. If we need it, we can the always orthogonalize those eigenvectors either internally (always possible) or externally as well (possible only for hermitian or symmetric matrices).
    It might be instructive to compute the eigenvectors for the examples used in demonstration of computation of eigenvalues. We'll leave to the reader, since this module is already too obese.


    Auxiliary functions

    The functions below are used in the main algorithms of this module. But they can be also used for testing. For example, the easiest way to test the usage of resources is to use easily definable unit matrices and unit vectors, as in:

            one_ket_solution (unit_matrix n::[[Double]])
                             (unit_vector 0 n::[Double])
            where n = 20, etc.
    
    
    > unit_matrix :: Num a => Int -> [[a]]
    > unit_matrix m =
    >       --
    >       -- Unit square matrix of with dimensions m x m
    >       --
    >       [ [ if j==k then 1 else 0 | j <- [0 .. m-1] ] | k <- [0 .. m-1]]
    
    > unit_vector :: Num a => Int -> Int -> [a]
    > unit_vector i m =
    >       --
    >       -- Unit vector of length m
    >       -- with 1 at position i, zero otherwise
    >       map (\k -> if k==i then 1 else 0) [0 .. m-1]
    
    > diagonals :: [[a]] -> [a]
    > diagonals a =
    >       --
    >       -- Vector made of diagonal components
    >       -- of square matrix a
    >       --
    >       diagonals' a 0
    >       where
    >           diagonals' b n
    >               | null b = []
    >               | otherwise =
    >                   (head $ drop n $ head b) : (diagonals' (tail b) (n+1))
    
    
    
    -----------------------------------------------------------------------------
    --
    -- Copyright:
    --
    --      (C) 1998 Numeric Quest Inc., All rights reserved
    --
    -- Email:
    --
    --      jans@numeric-quest.com
    --
    -- License:
    --
    --      GNU General Public License, GPL
    --
    -----------------------------------------------------------------------------