Test matrices
[A,B,C,...] = gallery(matname,P1,P2,...)
[A,B,C,...] = gallery(matname,P1,P2,...,classname)
gallery(3)
gallery(5)
[A,B,C,...] = gallery(matname,P1,P2,...)
returns
the test matrices specified by matname
. The matname
input
is the name of a matrix family selected from the table below. P1,P2,...
are
input parameters required by the individual matrix family. The number
of optional parameters P1,P2,...
used in the calling
syntax varies from matrix to matrix. The exact calling syntaxes are
detailed in the individual matrix descriptions below.
[A,B,C,...] = gallery(matname,P1,P2,...,classname)
produces
a matrix of class classname
. The classname
input
must be either 'single'
or 'double'
(unless matname
is 'integerdata'
,
in which case 'int8'
, 'int16'
, 'int32'
, 'uint8'
, 'uint16'
,
and 'uint32'
are also allowed). If classname
is
not specified, then the class of the matrix is determined from those
arguments among P1,P2,...
that do not specify dimensions
or select an option. If any of these arguments is of class single
then
the matrix is single
; otherwise the matrix is double
.
gallery(3)
is a badly
conditioned 3by3 matrix and gallery(5)
is an
interesting eigenvalue problem.
The gallery holds over fifty different test matrix functions useful for testing algorithms and other purposes.
A = gallery('binomial',n)
returns
an n
byn
matrix,with integer
entries such that A^2 = 2^(n1)*eye(n)
.
Thus, B = A*2^((1n)/2)
is involutory, that
is, B^2 = eye(n)
.
C = gallery('cauchy',x,y)
returns
an n
byn
matrix, C(i,j)
= 1/(x(i)+y(j))
. Arguments x
and y
are
vectors of length n
. If you pass in scalars for x
and y
,
they are interpreted as vectors 1:x
and 1:y
.
C = gallery('cauchy',x)
returns
the same as above with y = x
. That is, the command
returns C(i,j) = 1/(x(i)+x(j))
.
Explicit formulas are known for the inverse and determinant
of a Cauchy matrix. The determinant det(C)
is nonzero
if x
and y
both have distinct
elements. C
is totally positive if 0
< x(1) <... < x(n)
and 0 < y(1) < ... < y(n)
.
C = gallery('chebspec',n,switch)
returns
a Chebyshev spectral differentiation matrix of order n
.
Argument switch
is a variable that determines the
character of the output matrix. By default, switch
= 0
.
For switch
= 0
("no
boundary conditions"), C
is nilpotent (C^n
= 0
) and has the null vector ones(n,1)
.
The matrix C
is similar to a Jordan block of size n
with
eigenvalue zero.
For switch
= 1
, C
is
nonsingular and wellconditioned, and its eigenvalues have negative
real parts.
The eigenvector matrix of the Chebyshev spectral differentiation matrix is illconditioned.
C = gallery('chebvand',p)
produces
the (primal) Chebyshev Vandermonde matrix based on the vector of points p
,
which define where the Chebyshev polynomial is calculated.
C = gallery('chebvand',m,p)
where m
is
scalar, produces a rectangular version of the above, with m
rows.
If p
is a vector, then C(i,j)
= T_{i – 1}(p(j))
where T_{i –
1} is the Chebyshev polynomial of degree i –
1. If p
is a scalar, then p
equally
spaced points on the interval [0,1]
are used to
calculate C
.
A = gallery('chow',n,alpha,delta)
returns A
such
that A = H(alpha) + delta*eye(n)
, where H_{i,j}(α)
= α^{(i – j +
1)} and argument n
is the order of
the Chow matrix. Default value for scalars alpha
and delta
are 1
and 0
, respectively.
H(alpha)
has p = floor(n/2)
eigenvalues
that are equal to zero. The rest of the eigenvalues are equal to 4*alpha*cos(k*pi/(n+2))^2,
k=1:np
.
C = gallery('circul',v)
returns
the circulant matrix whose first row is the vector v
.
A circulant matrix has the property that each row is obtained from the previous one by cyclically permuting the entries one step forward. It is a special Toeplitz matrix in which the diagonals "wrap around."
If v
is a scalar, then C = gallery('circul',1:v)
.
The eigensystem of C
(n
byn
)
is known explicitly: If t
is an n
th
root of unity, then the inner product of v
and w =
[1 t t^{2} ... t^{(n –
1)}] is an eigenvalue of C
and w(n:1:1)
is
an eigenvector.
A = gallery('clement',n,k)
returns
an n
byn
tridiagonal matrix
with zeros on its main diagonal and known eigenvalues. It is singular
if n
is odd. About 64 percent of the entries of
the inverse are zero. The eigenvalues include plus and minus the numbers n1
, n3
, n5
, ...
,
(1 or 0).
For k=0
(the default), A
is
nonsymmetric. For k=1
, A
is
symmetric.
gallery('clement',n,1)
is diagonally similar
to gallery('clement',n)
.
For odd N = 2*M+1, M+1
of the singular values
are the integers sqrt((2*M+1)^2  (2*K+1).^2), K = 0:M
.
Note:
Similar properties hold for 
A = gallery('compar',A,1)
returns A
with
each diagonal element replaced by its absolute value, and each offdiagonal
element replaced by minus the absolute value of the largest element
in absolute value in its row. However, if A
is
triangular compar(A,1)
is too.
gallery('compar',A)
is diag(B) 
tril(B,1)  triu(B,1)
, where B
= abs(A)
. compar(A)
is
often denoted by M(A) in the literature.
gallery('compar',A,0)
is the same as gallery('compar',A)
.
A = gallery('condex',n,k,theta)
returns
a "counterexample" matrix to a condition estimator.
It has order n
and scalar parameter theta
(default
100).
The matrix, its natural size, and the estimator to which it
applies are specified by k
:
 4by4  LINPACK 
 3by3  LINPACK 
 arbitrary  LINPACK ( 

 LAPACK (RCOND) (default). It is the inverse of this matrix that is a counterexample. 
If n
is not equal to the natural size of
the matrix, then the matrix is padded out with an identity matrix
to order n
.
A = gallery('cycol',[m n],k)
returns
an m
byn
matrix with cyclically
repeating columns, where one "cycle" consists of randn(m,k)
.
Thus, the rank of matrix A
cannot exceed k
,
and k
must be a scalar.
Argument k
defaults to round(n/4)
,
and need not evenly divide n
.
A = gallery('cycol',n,k)
,
where n
is a scalar, is the same as gallery('cycol',[n
n],k)
.
[c,d,e] = gallery('dorr',n,theta)
returns
the vectors defining an n
byn
,
row diagonally dominant, tridiagonal matrix that is illconditioned
for small nonnegative values of theta
. The default
value of theta
is 0.01
. The
Dorr matrix itself is the same as gallery('tridiag',c,d,e)
.
A = gallery('dorr',n,theta)
returns
the matrix itself, rather than the defining vectors.
A = gallery('dramadah',n,k)
returns
an n
byn
matrix of 0
's
and 1
's for which mu(A) = norm(inv(A),'fro')
is
relatively large, although not necessarily maximal. An antiHadamard
matrix A
is a matrix with elements 0
or 1
for which mu(A)
is maximal.
n
and k
must both be scalars.
Argument k
determines the character of the output
matrix:
k = 1  Default. 
k = 2 

k = 3 

A = gallery('fiedler',c)
,
where c
is a length n
vector,
returns the n
byn
symmetric
matrix with elements abs(n(i)n(j))
. For scalar c
, A = gallery('fiedler',1:c)
.
Matrix A
has a dominant positive eigenvalue
and all the other eigenvalues are negative.
Explicit formulas for inv(A)
and det(A)
are
given in [Todd, J., Basic Numerical Mathematics,
Vol. 2: Numerical Algebra, Birkhauser, Basel, and Academic Press,
New York, 1977, p. 159] and attributed to Fiedler. These indicate
that inv(A)
is tridiagonal except for nonzero (1,n)
and (n,1)
elements.
A = gallery('forsythe',n,alpha,lambda)
returns
the n
byn
matrix equal to the
Jordan block with eigenvalue lambda
, excepting
that A(n,1) = alpha
. The default values of scalars alpha
and lambda
are sqrt(eps)
and 0
,
respectively.
The characteristic polynomial of A
is given
by:
det(At*I) = (lambdat)^N  alpha*(1)^n.
F = gallery('frank',n,k)
returns
the Frank matrix of order n
. It is upper Hessenberg
with determinant 1
. If k = 1
,
the elements are reflected about the antidiagonal (1,n)
— (n,1)
.
The eigenvalues of F
may be obtained in terms of
the zeros of the Hermite polynomials. They are positive and occur
in reciprocal pairs; thus if n
is odd, 1
is
an eigenvalue. F
has floor(n/2)
illconditioned
eigenvalues — the smaller ones.
A = gallery('gcdmat',n)
returns
the n
byn
matrix with (i,j)
entry gcd(i,j)
.
MatrixA
is symmetric positive definite, and A.^r
is
symmetric positive semidefinite for all nonnegative r
.
A = gallery('gearmat',n,i,j)
returns
the n
byn
matrix with ones
on the sub and superdiagonals, sign(i)
in the (1,abs(i))
position, sign(j)
in
the (n,n+1abs(j))
position, and zeros everywhere
else. Arguments i
and j
default
to n
and n
, respectively.
Matrix A
is singular, can have double and
triple eigenvalues, and can be defective.
All eigenvalues are of the form 2*cos(a)
and
the eigenvectors are of the form [sin(w+a), sin(w+2*a), ...,
sin(w+n*a)]
, where a
and w
are
given in Gear, C. W., "A Simple Set of Test Matrices for Eigenvalue
Programs," Math. Comp., Vol. 23 (1969),
pp. 119125.
A = gallery('grcar',n,k)
returns
an n
byn
Toeplitz matrix with 1
s
on the subdiagonal, 1
s on the diagonal, and k
superdiagonals
of 1
s. The default is k = 3
.
The eigenvalues are sensitive.
A = gallery('hanowa',n,d)
returns
an n
byn
block 2
by2
matrix
of the form:
[d*eye(m) diag(1:m) diag(1:m) d*eye(m)]
Argument n
is an even integer n=2*m
.
Matrix A
has complex eigenvalues of the form d
± k*i
,
for 1 <= k <= m
. The default value of d
is 1
.
[v,beta,s] = gallery('house',x,k)
takes x
,
an n
element column vector, and returns V
and beta
such
that H*x = s*e1
. In this expression, e1
is
the first column of eye(n)
, abs(s) = norm(x)
,
and H = eye(n)  beta*V*V'
is a Householder matrix.
k
determines the sign of s
:






If x
is complex, then sign(x) =
x./abs(x)
when x
is nonzero.
If x = 0
, or if x = alpha*e1
(alpha
>= 0
) and either k = 1
or k
= 2
, then V = 0
, beta = 1
,
and s = x(1)
. In this case, H
is
the identity matrix, which is not strictly a Householder matrix.

[v, beta] = gallery('house',x)
takes x
,
a scalar or n
element column vector, and returns v
and beta
such
that eye(n,n)  beta*v*v'
is a Householder matrix.
A Householder matrix H
satisfies the relationship
H*x = sign(x(1))*norm(x)*e1
where e1
is the first column of eye(n,n)
.
Note that if x
is complex, then sign(x)
exp(i*arg(x))
(which equals x./abs(x)
when x
is
nonzero).
If x = 0
, then v = 0
and beta
= 1
.
A = gallery('integerdata',imax,[m,n,...],j)
returns
an m
byn
by... array A
whose
values are a sample from the uniform distribution on the integers 1:imax
. j
must
be an integer value in the interval [0, 2^321]
.
Calling gallery('integerdata', ...)
with different
values of J
will return different arrays. Repeated
calls to gallery('integerdata',...)
with the same imax
,
size vector and j inputs will always return the same array.
In any call to gallery('integerdata', ...)
you
can substitute individual inputs m
,n
,...
for the size vector input [m,n,...]
. For example, gallery('integerdata',7,[1,2,3,4],5)
is
equivalent to gallery('integerdata',7,1,2,3,4,5)
.
A = gallery('integerdata',[imin imax],[m,n,...],j)
returns
an m
byn
by... array A
whose
values are a sample from the uniform distribution on the integers imin:imax
.
[A,B,...] = gallery('integerdata',[imin imax],[m,n,...],j)
returns
multiple m
byn
by... arrays A
, B
,
..., containing different values.
A = gallery('integerdata',[imin imax],[m,n,...],j,classname)
produces
an array of class classname
. classname
must
be 'uint8'
, 'uint16'
, 'uint32'
, 'int8'
, 'int16'
, int32'
, 'single'
or 'double'
.
A = gallery('invhess',x,y)
,
where x
is a length n
vector
and y
is a length n1
vector,
returns the matrix whose lower triangle agrees with that of ones(n,1)*x'
and
whose strict upper triangle agrees with that of [1 y]*ones(1,n)
.
The matrix is nonsingular if x(1) ~= 0
and x(i+1)
~= y(i)
for all i
, and its inverse is
an upper Hessenberg matrix. Argument y
defaults
to x(1:n1)
.
If x
is a scalar, invhess(x)
is
the same as invhess(1:x)
.
A = gallery('invol',n)
returns
an n
byn
involutory (A*A
= eye(n))
and illconditioned matrix. It is a diagonally
scaled version of hilb(n)
.
B = (eye(n)A)/2
and B = (eye(n)+A)/2
are
idempotent (B*B = B)
.
[A,d] = gallery('ipjfact',n,k)
returns A
,
an n
byn
Hankel matrix, and d
,
the determinant of A
, which is known explicitly.
If k = 0
(the default), then the elements of A
are A(i,j)
= (i+j)!
If k = 1
, then the elements
of A
are A(i,j) 1/(i+j)
.
Note that the inverse of A
is also known
explicitly.
A = gallery('jordbloc',n,lambda)
returns
the n
byn
Jordan block with
eigenvalue lambda
. The default value for lambda
is 1
.
A = gallery('kahan',n,theta,pert)
returns
an upper trapezoidal matrix that has interesting properties regarding
estimation of condition and rank.
If n
is a twoelement vector, then A
is n(1)
byn(2)
;
otherwise, A
is n
byn
.
The useful range of theta
is 0 < theta
< pi
, with a default value of 1.2
.
To ensure that the QR factorization with column pivoting does
not interchange columns in the presence of rounding errors, the diagonal
is perturbed by pert*eps*diag([n:1:1])
. The default pert
is 25
,
which ensures no interchanges for gallery('kahan',n)
up
to at least n = 90
in IEEE arithmetic.
A = gallery('kms',n,rho)
returns
the n
byn
KacMurdockSzego
Toeplitz matrix such that A(i,j) = rho^(abs(ij))
,
for real rho
.
For complex rho
, the same formula holds except
that elements below the diagonal are conjugated. rho
defaults
to 0.5.
The KMS matrix A
has these properties:
An LDL' factorization with L
inv(gallery('triw',n,rho,1))'
, and D(i,i) (1abs(rho)^2)*eye(n)
, except D(1,1) = 1
.
Positive definite if and only if 0 < abs(rho)
< 1
.
The inverse inv(A)
is tridiagonal.
B = gallery('krylov',A,x,j)
returns
the Krylov matrix
[x, Ax, A^2x, ..., A^(j1)x]
where A
is an n
byn
matrix
and x
is a length n
vector.
The defaults are x ones(n,1)
, and j = n
.
B = gallery('krylov',n)
is
the same as gallery('krylov',randn(n))
.
A = gallery('lauchli',n,mu)
returns
the (n+1)
byn
matrix
[ones(1,n); mu*eye(n)]
The Lauchli matrix is a wellknown example in least squares
and other problems that indicates the dangers of forming A'*A
.
Argument mu
defaults to sqrt(eps)
.
A = gallery('lehmer',n)
returns
the symmetric positive definite n
byn
matrix
such that A(i,j) = i/j
for j >= i
.
The Lehmer matrix A
has these properties:
A
is totally nonnegative.
The inverse inv(A)
is tridiagonal
and explicitly known.
The order n <= cond(A) <= 4*n*n
.
L = gallery('leslie',a,b)
is
the n
byn
matrix from the Leslie
population model with average birth numbers a(1:n)
and
survival rates b(1:n1)
. It is zero, apart from
the first row (which contains the a(i)
) and the
first subdiagonal (which contains the b(i)
). For
a valid model, the a(i)
are nonnegative and the b(i)
are
positive and bounded by 1, i.e., 0 <
b(i) <= 1
.
L = gallery('leslie',n)
generates
the Leslie matrix with a = ones(n,1)
, b
= ones(n1,1)
.
A = gallery('lesp',n)
returns
an n
byn
matrix whose eigenvalues
are real and smoothly distributed in the interval approximately [2*N3.5,
4.5]
.
The sensitivities of the eigenvalues increase exponentially
as the eigenvalues grow more negative. The matrix is similar to the
symmetric tridiagonal matrix with the same diagonal entries and with
offdiagonal entries 1, via a similarity transformation with D
= diag(1!,2!,...,n!)
.
A = gallery('lotkin',n)
returns
the Hilbert matrix with its first row altered to all ones. The Lotkin
matrix A
is nonsymmetric, illconditioned, and
has many negative eigenvalues of small magnitude. Its inverse has
integer entries and is known explicitly.
A = gallery('minij',n)
returns
the n
byn
symmetric positive
definite matrix with A(i,j) = min(i,j)
.
The minij
matrix has these properties:
The inverse inv(A)
is tridiagonal
and equal to 1
times the second difference matrix,
except its (n,n)
element is 1
.
Givens' matrix, 2*Aones(size(A))
,
has tridiagonal inverse and eigenvalues 0.5*sec((2*r1)*pi/(4*n))^2
,
where r=1:n
.
(n+1)*ones(size(A))A
has elements
that are max(i,j)
and a tridiagonal inverse.
A = gallery('moler',n,alpha)
returns
the symmetric positive definite n
byn
matrix U'*U
,
where U = gallery('triw',n,alpha)
.
For the default alpha
= 1
, A(i,j)
= min(i,j)2
, and A(i,i) = i
. One of
the eigenvalues of A
is small.
C = gallery('neumann',n)
returns
the sparse n
byn
singular,
row diagonally dominant matrix resulting from discretizing the Neumann
problem with the usual fivepoint operator on a regular mesh. Argument n
is
a perfect square integer n = m^{2} or
a twoelement vector. C
is sparse and has a onedimensional
null space with null vector ones(n,1)
.
A = gallery('normaldata',[m,n,...],j)
returns
an m
byn
by... array A
.
The values of A
are a random sample from the standard
normal distribution. j
must be an integer value
in the interval [0, 2^321]
. Calling gallery('normaldata',
...)
with different values of j
will
return different arrays. Repeated calls to gallery('normaldata',...)
with
the same size vector and j
inputs will always
return the same array.
In any call to gallery('normaldata', ...)
you
can substitute individual inputs m
,n
,...
for the size vector input [m,n,...]
. For example, gallery('normaldata',[1,2,3,4],5)
is
equivalent to gallery('normaldata',1,2,3,4,5)
.
[A,B,...] = gallery('normaldata',[m,n,...],j)
returns
multiple m
byn
by... arrays A
, B
,
..., containing different values.
A = gallery('normaldata',[m,n,...],j, classname)
produces
a matrix of class classname
. classname
must
be either 'single'
or 'double'
.
Generate the arbitrary 6by4 matrix of data from the standard
normal distribution N(0, 1)
corresponding to j
= 2
:.
x = gallery('normaldata', [6, 4], 2);
Generate the arbitrary 1by2by3 single array of data from
the standard normal distribution N(0, 1)
corresponding
to j = 17
:.
y = gallery('normaldata', 1, 2, 3, 17, 'single');
Q = gallery('orthog',n,k)
returns
the kth type of matrix of order n
, where k
> 0
selects exactly orthogonal matrices, and k
< 0
selects diagonal scalings of orthogonal matrices.
Available types are:

Symmetric eigenvector matrix for second difference matrix. This is the default. 

Symmetric. 

Unitary,
the Fourier matrix. Q^4 is the identity. This is essentially the same
matrix as 
 Helmert matrix: a permutation of a lower Hessenberg matrix,
whose first row is 

Symmetric matrix arising in the Hartley transform. 

Symmetric matrix arising as a discrete cosine transform. 

Chebyshev
Vandermondelike matrix, based on extrema of 

Chebyshev
Vandermondelike matrix, based on zeros of 
C = gallery('parter',n)
returns
the matrix C
such that C(i,j) = 1/(ij+0.5)
.
C
is a Cauchy matrix and a Toeplitz matrix.
Most of the singular values of C
are very close
to pi
.
A = gallery('pei',n,alpha)
,
where alpha
is a scalar, returns the symmetric
matrix alpha*eye(n) + ones(n)
. The default for alpha
is
1. The matrix is singular for alpha
equal to either 0
or
n
.
A = gallery('poisson',n)
returns
the block tridiagonal (sparse) matrix of order n^2
resulting
from discretizing Poisson's equation with the 5point operator on
an n
byn
mesh.
A = gallery('prolate',n,w)
returns
the n
byn
prolate matrix with
parameter w
. It is a symmetric Toeplitz matrix.
If 0 < w < 0.5
then A
is
positive definite
The eigenvalues of A
are distinct,
lie in (0,1)
, and tend to cluster around 0
and 1
.
The default value of w
is 0.25.
A = gallery('randcolu',n)
is
a random n
byn
matrix with
columns of unit 2norm, with random singular values whose squares
are from a uniform distribution.
A'*A
is a correlation matrix of the form
produced by gallery('randcorr',n)
.
gallery('randcolu',x)
where x
is
an n
vector (n
> 1), produces
a random n
byn
matrix having
singular values given by the vector x
. The vector x
must
have nonnegative elements whose sum of squares is n
.
gallery('randcolu',x,m)
where m
>= n
, produces an m
byn
matrix.
gallery('randcolu',x,m,k)
provides
a further option:


 The initial transformation is omitted. This is much faster, but the resulting matrix may have zero entries. 
For more information, see [4].
gallery('randcorr',n)
is
a random n
byn
correlation
matrix with random eigenvalues from a uniform distribution. A correlation
matrix is a symmetric positive semidefinite matrix with 1
s
on the diagonal (see corrcoef
).
gallery('randcorr',x)
produces
a random correlation matrix having eigenvalues given by the vector x
,
where length(x) > 1
. The vector x
must
have nonnegative elements summing to length(x)
.
gallery('randcorr',x,k)
provides
a further option:
 The diagonal matrix of eigenvalues is initially subjected to a random orthogonal similarity transformation, and then a sequence of Givens rotations is applied (default). 
 The initial transformation is omitted. This is much faster, but the resulting matrix may have some zero entries. 
For more information, see [3] and [4].
H = gallery('randhess',n)
returns
an n
byn
real, random, orthogonal
upper Hessenberg matrix.
H = gallery('randhess',x)
if x
is
an arbitrary, real, length n
vector with n > 1
, constructs H
nonrandomly
using the elements of x
as parameters.
Matrix H
is constructed via a product of n1
Givens
rotations.
A = gallery('randjorth', n)
, for a positive
integer n
, produces a random n
byn
J
orthogonal
matrix A
, where
J = blkdiag(eye(ceil(n/2)),eye(floor(n/2)))
cond(A) = sqrt(1/eps)
J
orthogonality means that A'*J*A = J. Such
matrices are sometimes called hyperbolic.
A = gallery('randjorth', n, m)
, for positive
integers n
and m
, produces a
random (n+m
)by(n+m
) J
orthogonal
matrix A
, where
J = blkdiag(eye(n),eye(m))
cond(A) = sqrt(1/eps)
A = gallery('randjorth',n,m,c,symm,method)
uses the following optional input arguments:
c
— Specifies cond(A)
to
be the scalar c
.
symm
— Enforces symmetry
if the scalar symm
is nonzero.
method
— calls qr
to
perform the underlying orthogonal transformations if the scalar method
is
nonzero. A call to qr
is much faster than the default
method for large dimensions
A = gallery('rando',n,k)
returns
a random n
byn
matrix with
elements from one of the following discrete distributions:
 
 
 
Argument n
may be a twoelement vector, in
which case the matrix is n(1)
byn(2)
.
A = gallery('randsvd',n,kappa,mode,kl,ku)
returns
a banded (multidiagonal) random matrix of order n
with cond(A)
= kappa
and singular values from the distribution mode.
If n
is a twoelement vector, A
is n(1)
byn(2)
.
Arguments kl
and ku
specify
the number of lower and upper offdiagonals, respectively, in A
.
If they are omitted, a full matrix is produced. If only kl
is
present, ku
defaults to kl
.
Distribution mode
can be:
 One large singular value. 
 One small singular value. 
 Geometrically distributed singular values (default). 
 Arithmetically distributed singular values. 
 Random singular values with uniformly distributed logarithm. 
 If 
Condition number kappa
defaults to sqrt(1/eps)
.
In the special case where kappa < 0
, A
is
a random, full, symmetric, positive definite matrix with cond(A)
= kappa
and eigenvalues distributed
according to mode
. Arguments kl
and ku
,
if present, are ignored.
A = gallery('randsvd',n,kappa,mode,kl,ku,method)
specifies
how the computations are carried out. method = 0
is
the default, while method = 1
uses an alternative method that is much faster for
large dimensions, even though it uses more flops.
A = gallery('redheff',n)
returns
an n
byn
matrix of 0
's
and 1
's defined by A(i,j) = 1
,
if j = 1
or if i
divides j
,
and A(i,j) = 0
otherwise.
The Redheffer matrix has these properties:
(nfloor(log2(n)))1
eigenvalues
equal to 1
A real eigenvalue (the spectral radius) approximately sqrt(n)
A negative eigenvalue approximately sqrt(n)
The remaining eigenvalues are provably "small."
The Riemann hypothesis is true if and only if $$\mathrm{det}(A)=O({n}^{1/2+\epsilon})$$ for every ε >
0
.
Barrett and Jarvis conjecture that "the small eigenvalues
all lie inside the unit circle abs(Z) = 1
,"
and a proof of this conjecture, together with a proof that some eigenvalue
tends to zero as n
tends to infinity, would yield
a new proof of the prime number theorem.
A = gallery('riemann',n)
returns
an n
byn
matrix for which the
Riemann hypothesis is true if and only if
$$\mathrm{det}(A)=O(n!{n}^{\text{\hspace{0.05em}}1/2+\epsilon})$$
for every ε > 0.
The Riemann matrix is defined by:
A = B(2:n+1,2:n+1)
where B(i,j) = i1
if i
divides j
,
and B(i,j) = 1
otherwise.
The Riemann matrix has these properties:
Each eigenvalue e(i)
satisfies abs(e(i))
<= m1/m
, where m = n+1
.
i <= e(i) <= i+1
with at
most msqrt(m)
exceptions.
All integers in the interval (m/3,
m/2]
are eigenvalues.
A = gallery('ris',n)
returns
a symmetric n
byn
Hankel matrix
with elements
A(i,j) = 0.5/(nij+1.5)
The eigenvalues of A
cluster around π/2
and –π/2. This matrix was invented
by F.N. Ris.
A = gallery('sampling',x)
, where x
is
an n
vector, is the n
byn
matrix
with A(i,j) = X(i)/(X(i)X(j))
for i ~=
j
and A(j,j)
the sum of the offdiagonal
elements in column j
. A has eigenvalues 0:n1
.
For the eigenvalues 0 and n
–1, corresponding
eigenvectors are X
and ones(n,1)
,
respectively.
The eigenvalues are illconditioned. A
has
the property that A(i,j) + A(j,i) = 1
for i
~= j
.
Explicit formulas are available for the left eigenvectors of A
.
For scalar n
, sampling(n)
is
the same as sampling(1:n)
. A special case of this
matrix arises in sampling theory.
A = gallery('smoke',n)
returns
an n
byn
matrix with 1
's
on the superdiagonal, 1
in the (n,1)
position,
and powers of roots of unity along the diagonal.
A = gallery('smoke',n,1)
returns
the same except that element A(n,1)
is zero.
The eigenvalues of gallery('smoke',n,1)
are
the n
th roots of unity; those of gallery('smoke',n)
are
the n
th roots of unity times 2^(1/n)
.
A = gallery('toeppd',n,m,w,theta)
returns
an n
byn
symmetric, positive
semidefinite (SPD) Toeplitz matrix composed of the sum of m
rank
2 (or, for certain theta
, rank 1) SPD Toeplitz
matrices. Specifically,
T = w(1)*T(theta(1)) + ... + w(m)*T(theta(m))
where T(theta(k))
has (i,j)
element cos(2*pi*theta(k)*(ij))
.
By default: m = n
, w = rand(m,1)
,
and theta = rand(m,1)
.
P = gallery('toeppen',n,a,b,c,d,e)
returns
the n
byn
sparse, pentadiagonal
Toeplitz matrix with the diagonals: P(3,1) = a
, P(2,1)
= b
, P(1,1) = c
, P(1,2) = d
,
and P(1,3) = e
, where a
, b
, c
, d
,
and e
are scalars.
By default, (a,b,c,d,e)
= (1,10,0,10,1)
,
yielding a matrix of Rutishauser. This matrix has eigenvalues lying
approximately on the line segment 2*cos(2*t) + 20*i*sin(t)
.
A = gallery('tridiag',c,d,e)
returns
the tridiagonal matrix with subdiagonal c
, diagonal d
,
and superdiagonal e
. Vectors c
and e
must
have length(d)1
.
A = gallery('tridiag',n,c,d,e)
,
where c
, d
, and e
are
all scalars, yields the Toeplitz tridiagonal matrix of order n
with
subdiagonal elements c
, diagonal elements d
,
and superdiagonal elements e
. This matrix has eigenvalues
d + 2*sqrt(c*e)*cos(k*pi/(n+1))
where k = 1:n
. (see [1].)
A = gallery('tridiag',n)
is
the same as A = gallery('tridiag',n,1,2,1)
, which
is a symmetric positive definite Mmatrix (the negative of the second
difference matrix).
A = gallery('triw',n,alpha,k)
returns
the upper triangular matrix with ones on the diagonal and alpha
s
on the first k >= 0
superdiagonals.
Order n
may be a 2element vector, in which
case the matrix is n(1)
byn(2)
and
upper trapezoidal.
Ostrowski ["On the Spectrum of a Oneparametric Family of Matrices," J. Reine Angew. Math., 1954] shows that
cond(gallery('triw',n,2)) = cot(pi/(4*n))^2,
and, for large abs(alpha)
, cond(gallery('triw',n,alpha))
is
approximately abs(alpha)^n*sin(pi/(4*n2))
.
Adding 2^(2n)
to the (n,1)
element
makes triw(n)
singular, as does adding 2^(1n)
to
all the elements in the first column.
A = gallery('uniformdata',[m,n,...],j)
returns
an m
byn
by... array A
.
The values of A
are a random sample from the standard
uniform distribution. j must be an integer value in the interval [0,
2^321]
. Calling gallery('uniformdata', ...)
with
different values of j
will return different arrays.
Repeated calls to gallery('uniformdata',...)
with
the same size vector and j
inputs will always
return the same array.
In any call to gallery('uniformdata', ...)
you
can substitute individual inputs m
,n
,...
for the size vector input [m,n,...]
. For example, gallery('uniformdata',[1,2,3,4],5)
is
equivalent to gallery('uniformdata',1,2,3,4,5)
.
[A,B,...] = gallery('uniformdata',[m,n,...],j)
returns
multiple m
byn
by... arrays A
, B
,
..., containing different values.
A = gallery('uniformdata',[m,n,...],j, classname)
produces
a matrix of class classname
. classname
must
be either 'single'
or 'double'
.
Generate the arbitrary 6by4 matrix of data from the uniform
distribution on [0, 1] corresponding to j = 2
.
x = gallery('uniformdata', [6, 4], 2);
Generate the arbitrary 1by2by3 single array of data from
the uniform distribution on [0, 1] corresponding to j =
17
.
y = gallery('uniformdata', 1, 2, 3, 17, 'single');
A = gallery('wathen',nx,ny)
returns
a sparse, random, n
byn
finite
element matrix where n = 3*nx*ny + 2*nx + 2*ny + 1
.
Matrix A
is precisely the "consistent
mass matrix" for a regular nx
byny
grid
of 8node (serendipity) elements in two dimensions. A
is
symmetric, positive definite for any (positive) values of the "density," rho(nx,ny)
,
which is chosen randomly in this routine.
A = gallery('wathen',nx,ny,1)
returns
a diagonally scaled matrix such that
0.25 <= eig(inv(D)*A) <= 4.5
where D = diag(diag(A))
for any positive
integers nx
and ny
and any densities rho(nx,ny)
.
gallery('wilk',n)
returns
a different matrix or linear system depending on the value of n
.
 Upper triangular system 
 Lower triangular system 




[1] The MATLAB^{®} gallery of test matrices is based upon the work of Nicholas J. Higham at the Department of Mathematics, University of Manchester, Manchester, England. Further background can be found in the books MATLAB Guide, Second Edition, Desmond J. Higham and Nicholas J. Higham, SIAM, 2005, and Accuracy and Stability of Numerical Algorithms, Nicholas J. Higham, SIAM, 1996.
[2] Wilkinson, J. H., The Algebraic Eigenvalue Problem, Oxford University Press, London, 1965, p.308.
[3] Bendel, R. B. and M. R. Mickey, "Population Correlation Matrices for Sampling Experiments," Commun. Statist. Simulation Comput., B7, 1978, pp. 163182.
[4] Davies, P. I. and N. J. Higham, "Numerically Stable Generation of Correlation Matrices and Their Factors," BIT, Vol. 40, 2000, pp. 640651.