Orthogonal matrix

In linear algebra, an or thogonal matrix, or or thonormal matrix, is a real square mat rix whose
columns and rows are ort honormal vect ors.

One way t o express t his is

where QT is the transpose of Q and I is

the identity matrix.
This leads t o t he equivalent charact erizat ion: a mat rix Q is ort hogonal if it s t ranspose is equal
t o it s inverse:

where Q−1 is the inverse of Q.

An ort hogonal mat rix Q is necessarily invert ible (wit h inverse Q−1 = QT ), unit ary (Q−1 = Q∗),
where Q∗ is t he Hermit ian adjoint (conjugat e t ranspose) of Q, and t herefore normal
(Q∗Q = QQ∗) over t he real numbers. The det erminant of any ort hogonal mat rix is eit her +1 or
−1. As a linear t ransformat ion, an ort hogonal mat rix preserves t he inner product of vect ors,
and t herefore act s as an isomet ry of Euclidean space, such as a rot at ion, reflect ion or
rot oreflect ion. In ot her words, it is a unit ary t ransformat ion.
The set of n × n ort hogonal mat rices, under mult iplicat ion, forms t he group O(n), known as
t he ort hogonal group. The subgroup SO(n) consist ing of ort hogonal mat rices wit h
det erminant +1 is called t he special ort hogonal group, and each of it s element s is a special
or thogonal matrix. As a linear t ransformat ion, every special ort hogonal mat rix act s as a
rot at ion.

Overview

Visual understanding of multiplication by the

transpose of a matrix. If A is an orthogonal
matrix and B is its transpose, the ij-th element
of the product AAT will vanish if i≠j, because the
i-th row of A is orthogonal to the j-th row of A.

An ort hogonal mat rix is t he real specializat ion of a unit ary mat rix, and t hus always a normal
mat rix. Alt hough we consider only real mat rices here, t he definit ion can be used for mat rices
wit h ent ries from any field. However, ort hogonal mat rices arise nat urally from dot product s,
and for mat rices of complex numbers t hat leads inst ead t o t he unit ary requirement .
Ort hogonal mat rices preserve t he dot product ,[1] so, for vect ors u and v in an n-dimensional
real Euclidean space
where Q is an orthogonal matrix. To see
the inner product connection, consider a
vector v in an n-dimensional real
Euclidean space. Written with respect to
an orthonormal basis, the squared length
of v is vTv. If a linear transformation, in
matrix form Qv, preserves vector lengths,
then

Thus finit e-dimensional linear isomet ries—rot at ions, reflect ions, and t heir combinat ions—
produce ort hogonal mat rices. The converse is also t rue: ort hogonal mat rices imply ort hogonal
t ransformat ions. However, linear algebra includes ort hogonal t ransformat ions bet ween
spaces which may be neit her finit e-dimensional nor of t he same dimension, and t hese have no
ort hogonal mat rix equivalent .

Ort hogonal mat rices are import ant for a number of reasons, bot h t heoret ical and pract ical.
The n × n ort hogonal mat rices form a group under mat rix mult iplicat ion, t he ort hogonal group
denot ed by O(n), which—wit h it s subgroups—is widely used in mat hemat ics and t he physical
sciences. For example, t he point group of a molecule is a subgroup of O(3). Because float ing
point versions of ort hogonal mat rices have advant ageous propert ies, t hey are key t o many
algorit hms in numerical linear algebra, such as QR decomposit ion. As anot her example, wit h
appropriat e normalizat ion t he discret e cosine t ransform (used in MP3 compression) is
represent ed by an ort hogonal mat rix.
Examples
Below are a few examples of small ort hogonal mat rices and possible int erpret at ions.

(identity

transformation)

(rotation

about the origin)

(reflection across x-

axis)
 (permutation

of coordinate axes)

Elementary constructions

Lower dimensions
The simplest ort hogonal mat rices are t he 1 × 1 mat rices [1] and [−1], which we can int erpret
as t he ident it y and a reflect ion of t he real line across t he origin.

The 2 × 2 mat rices have t he form

which orthogonality demands satisfy the

three equations
In considerat ion of t he first equat ion, wit hout loss of generalit y let p = cos θ, q = sin θ; t hen
eit her t = −q, u = p or t = q, u = −p. We can int erpret t he first case as a rot at ion by θ (where
θ
θ = 0 is t he ident it y), and t he second as a reflect ion across a line at an angle of 2 .

The special case of t he reflect ion mat rix wit h θ = 90° generat es a reflect ion about t he line at
45° given by y = x and t herefore exchanges x and y; it is a permut at ion mat rix, wit h a single 1
in each column and row (and ot herwise 0):

The ident it y is also a permut at ion mat rix.

A reflect ion is it s own inverse, which implies t hat a reflect ion mat rix is symmet ric (equal t o it s
t ranspose) as well as ort hogonal. The product of t wo rot at ion mat rices is a rot at ion mat rix,
and t he product of t wo reflect ion mat rices is also a rot at ion mat rix.

Higher dimensions
Regardless of t he dimension, it is always possible t o classify ort hogonal mat rices as purely
rot at ional or not , but for 3 × 3 mat rices and larger t he non-rot at ional mat rices can be more
complicat ed t han reflect ions. For example,
represent an inversion t hrough t he origin and a rotoinversion, respect ively, about t he z-axis.

Rot at ions become more complicat ed in higher dimensions; t hey can no longer be complet ely
charact erized by one angle, and may affect more t han one planar subspace. It is common t o
describe a 3 × 3 rot at ion mat rix in t erms of an axis and angle, but t his only works in t hree
dimensions. Above t hree dimensions t wo or more angles are needed, each associat ed wit h a
plane of rot at ion.

However, we have element ary building blocks for permut at ions, reflect ions, and rot at ions t hat
apply in general.

Primitives
The most element ary permut at ion is a t ransposit ion, obt ained from t he ident it y mat rix by
exchanging t wo rows. Any n × n permut at ion mat rix can be const ruct ed as a product of no
more t han n − 1 t ransposit ions.

A Householder reflect ion is const ruct ed from a non-null vect or v as

Here t he numerat or is a symmet ric mat rix while t he denominat or is a number, t he squared
magnit ude of v. This is a reflect ion in t he hyperplane perpendicular t o v (negat ing any vect or
component parallel t o v). If v is a unit vect or, t hen Q = I − 2vvT suffices. A Householder
reflect ion is t ypically used t o simult aneously zero t he lower part of a column. Any ort hogonal
mat rix of size n × n can be const ruct ed as a product of at most n such reflect ions.
A Givens rot at ion act s on a t wo-dimensional (planar) subspace spanned by t wo coordinat e
axes, rot at ing by a chosen angle. It is t ypically used t o zero a single subdiagonal ent ry. Any
n(n − 1)
rot at ion mat rix of size n × n can be const ruct ed as a product of at most 2
such
rot at ions. In t he case of 3 × 3 mat rices, t hree such rot at ions suffice; and by fixing t he
sequence we can t hus describe all 3 × 3 rot at ion mat rices (t hough not uniquely) in t erms of
t he t hree angles used, oft en called Euler angles.

A Jacobi rot at ion has t he same form as a Givens rot at ion, but is used t o zero bot h off-
diagonal ent ries of a 2 × 2 symmet ric submat rix.

Properties

Matrix properties
A real square mat rix is ort hogonal if and only if it s columns form an ort honormal basis of t he
Euclidean space Rn wit h t he ordinary Euclidean dot product , which is t he case if and only if it s
rows form an ort honormal basis of Rn. It might be t empt ing t o suppose a mat rix wit h
ort hogonal (not ort honormal) columns would be called an ort hogonal mat rix, but such
mat rices have no special int erest and no special name; t hey only sat isfy MT M = D, wit h D a
diagonal mat rix.

The det erminant of any ort hogonal mat rix is +1 or −1. This follows from basic fact s about
det erminant s, as follows:

The converse is not t rue; having a det erminant of ±1 is no guarant ee of ort hogonalit y, even
wit h ort hogonal columns, as shown by t he following count erexample.
Wit h permut at ion mat rices t he det erminant mat ches t he signat ure, being +1 or −1 as t he
parit y of t he permut at ion is even or odd, for t he det erminant is an alt ernat ing funct ion of t he
rows.

St ronger t han t he det erminant rest rict ion is t he fact t hat an ort hogonal mat rix can always be
diagonalized over t he complex numbers t o exhibit a full set of eigenvalues, all of which must
have (complex) modulus 1.

Group properties
The inverse of every ort hogonal mat rix is again ort hogonal, as is t he mat rix product of t wo
ort hogonal mat rices. In fact , t he set of all n × n ort hogonal mat rices sat isfies all t he axioms
n(n − 1)
of a group. It is a compact Lie group of dimension 2
, called t he ort hogonal group and
denot ed by O(n).

The ort hogonal mat rices whose det erminant is +1 form a pat h-connect ed normal subgroup of
O(n) of index 2, t he special ort hogonal group SO(n) of rot at ions. The quot ient group
O(n)/SO(n) is isomorphic t o O(1), wit h t he project ion map choosing [+1] or [−1] according t o
t he det erminant . Ort hogonal mat rices wit h det erminant −1 do not include t he ident it y, and so
do not form a subgroup but only a coset ; it is also (separat ely) connect ed. Thus each
ort hogonal group falls int o t wo pieces; and because t he project ion map split s, O(n) is a
semidirect product of SO(n) by O(1). In pract ical t erms, a comparable st at ement is t hat any
ort hogonal mat rix can be produced by t aking a rot at ion mat rix and possibly negat ing one of it s
columns, as we saw wit h 2 × 2 mat rices. If n is odd, t hen t he semidirect product is in fact a
direct product , and any ort hogonal mat rix can be produced by t aking a rot at ion mat rix and
possibly negat ing all of it s columns. This follows from t he propert y of det erminant s t hat
negat ing a column negat es t he det erminant , and t hus negat ing an odd (but not even) number
of columns negat es t he det erminant .

Now consider (n + 1) × (n + 1) ort hogonal mat rices wit h bot t om right ent ry equal t o 1. The
remainder of t he last column (and last row) must be zeros, and t he product of any t wo such
mat rices has t he same form. The rest of t he mat rix is an n × n ort hogonal mat rix; t hus O(n) is
a subgroup of O(n + 1) (and of all higher groups).
Since an element ary reflect ion in t he form of a Householder mat rix can reduce any ort hogonal
mat rix t o t his const rained form, a series of such reflect ions can bring any ort hogonal mat rix t o
t he ident it y; t hus an ort hogonal group is a reflect ion group. The last column can be fixed t o
any unit vect or, and each choice gives a different copy of O(n) in O(n + 1); in t his way
O(n + 1) is a bundle over t he unit sphere Sn wit h fiber O(n).

Similarly, SO(n) is a subgroup of SO(n + 1); and any special ort hogonal mat rix can be
generat ed by Givens plane rot at ions using an analogous procedure. The bundle st ruct ure
persist s: SO(n) ↪ SO(n + 1) → Sn. A single rot at ion can produce a zero in t he first row of t he
last column, and series of n − 1 rot at ions will zero all but t he last row of t he last column of an
n × n rot at ion mat rix. Since t he planes are fixed, each rot at ion has only one degree of
freedom, it s angle. By induct ion, SO(n) t herefore has

degrees of freedom, and so does O(n).

Permut at ion mat rices are simpler st ill; t hey form, not a Lie group, but only a finit e group, t he
order n! symmet ric group Sn. By t he same kind of argument , Sn is a subgroup of Sn + 1. The
even permut at ions produce t he subgroup of permut at ion mat rices of det erminant +1, t he
n!
order 2 alt ernat ing group.
Canonical form
More broadly, t he effect of any ort hogonal mat rix separat es int o independent act ions on
ort hogonal t wo-dimensional subspaces. That is, if Q is special ort hogonal t hen one can always
find an ort hogonal mat rix P, a (rot at ional) change of basis, t hat brings Q int o block diagonal
form:

where t he mat rices R1, ..., Rk are 2 × 2 rot at ion mat rices, and wit h t he remaining ent ries zero.
Except ionally, a rot at ion block may be diagonal, ±I. Thus, negat ing one column if necessary,
and not ing t hat a 2 × 2 reflect ion diagonalizes t o a +1 and −1, any ort hogonal mat rix can be
brought t o t he form
The mat rices R1, ..., Rk give conjugat e pairs of eigenvalues lying on t he unit circle in t he
complex plane; so t his decomposit ion confirms t hat all eigenvalues have absolut e value 1. If n
is odd, t here is at least one real eigenvalue, +1 or −1; for a 3 × 3 rot at ion, t he eigenvect or
associat ed wit h +1 is t he rot at ion axis.

Lie algebra
Suppose t he ent ries of Q are different iable funct ions of t, and t hat t = 0 gives Q = I.
Different iat ing t he ort hogonalit y condit ion

yields

Evaluat ion at t = 0 (Q = I) t hen implies

In Lie group t erms, t his means t hat t he Lie algebra of an ort hogonal mat rix group consist s of
skew-symmet ric mat rices. Going t he ot her direct ion, t he mat rix exponent ial of any skew-
symmet ric mat rix is an ort hogonal mat rix (in fact , special ort hogonal).

For example, t he t hree-dimensional object physics calls angular velocit y is a different ial
rot at ion, t hus a vect or in t he Lie algebra t angent t o SO(3). Given ω = (xθ, yθ, zθ), wit h
v = (x, y, z) being a unit vect or, t he correct skew-symmet ric mat rix form of ω is

The exponent ial of t his is t he ort hogonal mat rix for rot at ion around axis v by angle θ; set t ing
θ θ
c = cos 2 , s = sin 2 ,

Numerical linear algebra

Benefits
Numerical analysis t akes advant age of many of t he propert ies of ort hogonal mat rices for
numerical linear algebra, and t hey arise nat urally. For example, it is oft en desirable t o comput e
an ort honormal basis for a space, or an ort hogonal change of bases; bot h t ake t he form of
ort hogonal mat rices. Having det erminant ±1 and all eigenvalues of magnit ude 1 is of great
benefit for numeric st abilit y. One implicat ion is t hat t he condit ion number is 1 (which is t he
minimum), so errors are not magnified when mult iplying wit h an ort hogonal mat rix. Many
algorit hms use ort hogonal mat rices like Householder reflect ions and Givens rot at ions for t his
reason. It is also helpful t hat , not only is an ort hogonal mat rix invert ible, but it s inverse is
available essent ially free, by exchanging indices.

Permut at ions are essent ial t o t he success of many algorit hms, including t he workhorse
Gaussian eliminat ion wit h part ial pivot ing (where permut at ions do t he pivot ing). However, t hey
rarely appear explicit ly as mat rices; t heir special form allows more efficient represent at ion,
such as a list of n indices.

Likewise, algorit hms using Householder and Givens mat rices t ypically use specialized
met hods of mult iplicat ion and st orage. For example, a Givens rot at ion affect s only t wo rows
of a mat rix it mult iplies, changing a full mult iplicat ion of order n3 t o a much more efficient
order n. When uses of t hese reflect ions and rot at ions int roduce zeros in a mat rix, t he space
vacat ed is enough t o st ore sufficient dat a t o reproduce t he t ransform, and t o do so robust ly.
(Following St ewart (1976), we do not st ore a rot at ion angle, which is bot h expensive and badly
behaved.)

Decompositions
A number of import ant mat rix decomposit ions (Golub & Van Loan 1996) involve ort hogonal
mat rices, including especially:

QR decomposition
M = QR, Q orthogonal, R upper
triangular
Singular value decomposition
 M = UΣVT, U and V orthogonal, Σ
diagonal matrix
Eigendecomposition of a symmetric
matrix (decomposition according to
the spectral theorem)
S = QΛQT, S symmetric, Q orthogonal,
Λ diagonal
Polar decomposition
M = QS, Q orthogonal, S symmetric
positive-semidefinite

Examples
Consider an overdet ermined syst em of linear equat ions, as might occur wit h repeat ed
measurement s of a physical phenomenon t o compensat e for experiment al errors. Writ e
Ax = b, where A is m × n, m > n. A QR decomposit ion reduces A t o upper t riangular R. For
example, if A is 5 × 3 t hen R has t he form
The linear least squares problem is t o find t he x t hat minimizes ‖ Ax − b ‖, which is equivalent
t o project ing b t o t he subspace spanned by t he columns of A. Assuming t he columns of A
(and hence R) are independent , t he project ion solut ion is found from AT Ax = AT b. Now AT A is
square (n × n) and invert ible, and also equal t o RT R. But t he lower rows of zeros in R are
superfluous in t he product , which is t hus already in lower-t riangular upper-t riangular fact ored
form, as in Gaussian eliminat ion (Cholesky decomposit ion). Here ort hogonalit y is import ant not
only for reducing AT A = (RT QT )QR t o RT R, but also for allowing solut ion wit hout magnifying
numerical problems.

In t he case of a linear syst em which is underdet ermined, or an ot herwise non-invert ible mat rix,
singular value decomposit ion (SVD) is equally useful. Wit h A fact ored as UΣVT , a sat isfact ory
solut ion uses t he Moore-Penrose pseudoinverse, VΣ +UT , where Σ + merely replaces each non-
zero diagonal ent ry wit h it s reciprocal. Set x t o VΣ +UT b.

The case of a square invert ible mat rix also holds int erest . Suppose, for example, t hat A is a
3 × 3 rot at ion mat rix which has been comput ed as t he composit ion of numerous t wist s and
t urns. Float ing point does not mat ch t he mat hemat ical ideal of real numbers, so A has
gradually lost it s t rue ort hogonalit y. A Gram–Schmidt process could ort hogonalize t he
columns, but it is not t he most reliable, nor t he most efficient , nor t he most invariant met hod.
The polar decomposit ion fact ors a mat rix int o a pair, one of which is t he unique closest
ort hogonal mat rix t o t he given mat rix, or one of t he closest if t he given mat rix is singular.
(Closeness can be measured by any mat rix norm invariant under an ort hogonal change of basis,
such as t he spect ral norm or t he Frobenius norm.) For a near-ort hogonal mat rix, rapid
convergence t o t he ort hogonal fact or can be achieved by a "Newt on's met hod" approach due
t o Higham (1986) (1990), repeat edly averaging t he mat rix wit h it s inverse t ranspose. Dubrulle
(1999) has published an accelerat ed met hod wit h a convenient convergence t est .
For example, consider a non-ort hogonal mat rix for which t he simple averaging algorit hm t akes
seven st eps

and which acceleration trims to two

steps (with γ = 0.353553, 0.565685).

Gram-Schmidt yields an inferior solut ion, shown by a Frobenius dist ance of 8.28659 inst ead of
t he minimum 8.12404.

Randomization
Some numerical applicat ions, such as Mont e Carlo met hods and explorat ion of high-
dimensional dat a spaces, require generat ion of uniformly dist ribut ed random ort hogonal
mat rices. In t his cont ext , "uniform" is defined in t erms of Haar measure, which essent ially
requires t hat t he dist ribut ion not change if mult iplied by any freely chosen ort hogonal mat rix.
Ort hogonalizing mat rices wit h independent uniformly dist ribut ed random ent ries does not
result in uniformly dist ribut ed ort hogonal mat rices, but t he QR decomposit ion of independent
normally dist ribut ed random ent ries does, as long as t he diagonal of R cont ains only posit ive
ent ries (Mezzadri 2006). St ewart (1980) replaced t his wit h a more efficient idea t hat Diaconis
& Shahshahani (1987) lat er generalized as t he "subgroup algorit hm" (in which form it works just
as well for permut at ions and rot at ions). To generat e an (n + 1) × (n + 1) ort hogonal mat rix,
t ake an n × n one and a uniformly dist ribut ed unit vect or of dimension n + 1. Const ruct a
Householder reflect ion from t he vect or, t hen apply it t o t he smaller mat rix (embedded in t he
larger size wit h a 1 at t he bot t om right corner).

Nearest orthogonal matrix

The problem of finding t he ort hogonal mat rix Q nearest a given mat rix M is relat ed t o t he
Ort hogonal Procrust es problem. There are several different ways t o get t he unique solut ion,
t he simplest of which is t aking t he singular value decomposit ion of M and replacing t he
singular values wit h ones. Anot her met hod expresses t he R explicit ly but requires t he use of a
mat rix square root :[2]

This may be combined wit h t he Babylonian met hod for ext ract ing t he square root of a mat rix
t o give a recurrence which converges t o an ort hogonal mat rix quadrat ically:

where Q0 = M.
These it erat ions are st able provided t he condit ion number of M is less t han t hree.[3]

Using a first -order approximat ion of t he inverse and t he same init ializat ion result s in t he
modified it erat ion:
Spin and pin
A subt le t echnical problem afflict s some uses of ort hogonal mat rices. Not only are t he group
component s wit h det erminant +1 and −1 not connect ed t o each ot her, even t he +1
component , SO(n), is not simply connect ed (except for SO(1), which is t rivial). Thus it is
somet imes advant ageous, or even necessary, t o work wit h a covering group of SO(n), t he spin
group, Spin(n). Likewise, O(n) has covering groups, t he pin groups, Pin(n). For n > 2, Spin(n) is
simply connect ed and t hus t he universal covering group for SO(n). By far t he most famous
example of a spin group is Spin(3), which is not hing but SU(2), or t he group of unit
quat ernions.

The Pin and Spin groups are found wit hin Clifford algebras, which t hemselves can be built
from ort hogonal mat rices.

Rectangular matrices
If Q is not a square mat rix, t hen t he condit ions QT Q = I and QQT = I are not equivalent . The
condit ion QT Q = I says t hat t he columns of Q are ort honormal. This can only happen if Q is an
m × n mat rix wit h n ≤ m (due t o linear dependence). Similarly, QQT = I says t hat t he rows of Q
are ort honormal, which requires n ≥ m.

There is no st andard t erminology for t hese mat rices. They are variously called "semi-
ort hogonal mat rices", "ort honormal mat rices", "ort hogonal mat rices", and somet imes simply
"mat rices wit h ort honormal rows/columns".

For t he case n ≤ m, mat rices wit h ort honormal columns may be referred t o as ort hogonal k-
frames and t hey are element s of t he St iefel manifold.
See also

Biorthogonal system

Notes

1. "Paul's online math notes" (http://tutorial.

math.lamar.edu/Classes/LinAlg/Orthogon
alMatrix.aspx) , Paul Dawkins, Lamar
University, 2008. Theorem 3(c)

2. "Finding the Nearest Orthonormal Matrix"

(http://people.csail.mit.edu/bkph/articles/
Nearest_Orthonormal_Matrix.pdf) ,
Berthold K.P. Horn, MIT.

3. "Newton's Method for the Matrix Square

Root" (http://www.maths.manchester.ac.u
k/~nareports/narep91.pdf) Archived (htt
ps://web.archive.org/web/201109291313
 30/http://www.maths.manchester.ac.uk/~
nareports/narep91.pdf) 2011-09-29 at
the Wayback Machine, Nicholas J.
Higham, Mathematics of Computation,
Volume 46, Number 174, 1986.

References

Diaconis, Persi; Shahshahani, Mehrdad

(1987), "The subgroup algorithm for
generating uniform random variables",
Probability in the Engineering and
Informational Sciences, 1: 15–32,
doi:10.1017/S0269964800000255 (htt
ps://doi.org/10.1017%2FS0269964800
000255) , ISSN 0269-9648 (https://ww
w.worldcat.org/issn/0269-9648) ,
S2CID 122752374 (https://api.semanti
cscholar.org/CorpusID:122752374)
Dubrulle, Augustin A. (1999), "An
Optimum Iteration for the Matrix Polar
Decomposition" (http://etna.mcs.kent.e
du/) , Electronic Transactions on
Numerical Analysis, 8: 21–25

Golub, Gene H.; Van Loan, Charles F.

(1996), Matrix Computations (3/e ed.),
Baltimore: Johns Hopkins University
Press, ISBN 978-0-8018-5414-9
Higham, Nicholas (1986), "Computing
the Polar Decomposition—with
Applications" (http://eprints.maths.ma
nchester.ac.uk/694/1/high86p.pdf)
(PDF), SIAM Journal on Scientific and
Statistical Computing, 7 (4): 1160–
1174, doi:10.1137/0907079 (https://do
i.org/10.1137%2F0907079) ,
ISSN 0196-5204 (https://www.worldca
t.org/issn/0196-5204)
Higham, Nicholas; Schreiber, Robert
(July 1990), "Fast polar decomposition
of an arbitrary matrix", SIAM Journal on
Scientific and Statistical Computing, 11
(4): 648–655,
CiteSeerX 10.1.1.230.4322 (https://cite
seerx.ist.psu.edu/viewdoc/summary?d
oi=10.1.1.230.4322) ,
doi:10.1137/0911038 (https://doi.org/
10.1137%2F0911038) , ISSN 0196-
5204 (https://www.worldcat.org/issn/0
196-5204) , S2CID 14268409 (https://a
pi.semanticscholar.org/CorpusID:1426
8409) [1] (https://web.archive.org/we
b/20051016153437/http://www.ma.m
an.ac.uk/~higham/pap-mf.html)
Stewart, G. W. (1976), "The Economical
Storage of Plane Rotations",
Numerische Mathematik, 25 (2): 137–
138, doi:10.1007/BF01462266 (https://
doi.org/10.1007%2FBF01462266) ,
ISSN 0029-599X (https://www.worldca
t.org/issn/0029-599X) ,
S2CID 120372682 (https://api.semanti
cscholar.org/CorpusID:120372682)
Stewart, G. W. (1980), "The Efficient
Generation of Random Orthogonal
Matrices with an Application to
Condition Estimators", SIAM Journal on
Numerical Analysis, 17 (3): 403–409,
Bibcode:1980SJNA...17..403S (https://
ui.adsabs.harvard.edu/abs/1980SJN
A...17..403S) , doi:10.1137/0717034 (h
ttps://doi.org/10.1137%2F0717034) ,
ISSN 0036-1429 (https://www.worldca
t.org/issn/0036-1429)
Mezzadri, Francesco (2006), "How to
generate random matrices from the
classical compact groups", Notices of
the American Mathematical Society, 54,
arXiv:math-ph/0609050 (https://arxiv.o
rg/abs/math-ph/0609050) ,
Bibcode:2006math.ph...9050M (http
s://ui.adsabs.harvard.edu/abs/2006m
ath.ph...9050M)
External links

Wikiversity introduces the orthogonal

matrix.

"Orthogonal matrix" (https://www.ency

clopediaofmath.org/index.php?title=Or
thogonal_matrix) , Encyclopedia of
Mathematics, EMS Press, 2001 [1994]

Tutorial and Interactive Program on

Orthogonal Matrix (http://people.revole
du.com/kardi/tutorial/LinearAlgebra/M
atrixOrthogonal.html)

Retrieved from
"https://en.wikipedia.org/w/index.php?
title=Orthogonal_matrix&oldid=1202509034"
This page was last edited on 2 February 2024, at
22:52 (UTC). •
Content is available under CC BY-SA 4.0 unless
otherwise noted.