Symmetric 3×3 matrices

with repeated eigenvalues

Olaf Dietrich∗, Munich

Version: 2016-06-02

Abstract
Real symmetric 3 × 3 matrices have 6 independent entries (3 diagonal elements
and 3 off-diagonal elements) and they have 3 real eigenvalues (λ0 , λ1 , λ2 ). If 2 of
these 3 eigenvalues are the same, i. e., λ0 6= λ1 = λ2 , then the number of independent
entries of this matrix is reduced from 6 to 4. In this article, some relations between
the matrix elements of such a symmetric 3 × 3 matrix with a repeated eigenvalue are
presented, which reduce the number of degrees of freedom of the matrix from 6 to 4.

Contents
1 Introduction 2

2 Repeated eigenvalues – results 3

2.1 Main result . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2.2 Corollaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2.3 More corollaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.4 Summarized results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

3 How to check the result 6

3.1 Parametrizing the matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
3.2 Putting things together . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

References 8

∗
Please send comments, corrections, or suggestions to mailto:[email protected].

1
1 Introduction
A general real symmetric matrix S = ST is given by
 
s00 s01 s02
S=  s01 s11 s12  , smn = snm ∈ R. (1)
s02 s12 s22

Such a matrix has 6 independent entries: 3 diagonal elements (s00 , s11 , s22 ) and 3 off-
diagonal elements (s01 , s02 , s12 ).
Each such matrix can be diagonalized [1, 2], i. e., there exists an orthogonal matrix R
(with R−1 = RT ) such that
 
λ0 0 0
RT S R =  0 λ1 0  = diag(λ0 , λ1 , λ2 ), λn ∈ R. (2)
0 0 λ2

The orthogonal matrix R contains the elements of the eigenvectors (v0 , v1 , v2 ) (written as
column vectors) in its columns, i. e., R = v0 v1 v2 . The eigenvectors and eigenvalues


satisfy
Svm = λm vm , 0 ≤ m < 3, (3)
which is basically Eq. (2) multiplied (from the left) by R; Eq. (3) is then given columnwise
by the resulting matrix equation.
The 3 eigenvectors (v0 , v1 , v2 ) build an orthonormal basis of R3 [1, 2], i. e., they have
the scalar products vm · vn = δmn , where δnm is the Kronecker delta (δmn = 1 for m = n
and δmn = 0 for m 6= n). The latter property is of course equivalent to RT R = 1, which
defines an orthogonal matrix.
To describe an orthonormal basis (i. e., the orthogonal matrix R), 3 parameters are
required; different parametrizations exist such as:

• the 3 Euler angles (α, β, γ), which relate the orientation of an arbitrary orthonormal
basis (v0 , v1 , v2 ) to the standard basis (e0 , e1 , e2 ) (where the n-th component (em )n
of em is (em )n = δmn );

• two angles θ, φ (spherical coordinates) that define the orientation of v0 in space

plus one more angle ξ that defines the orientation of v1 in the plane orthogonal to
v0 ;

• any other parametrization of the rotation group in 3 dimensions (i. e., of the special1
orthogonal group) SO(3), which is 3-dimensional itself (in general, SO(n) has
dimension n(n − 1)/2).

These 3 parameters defining R together with the 3 eigenvalues correspond to the 6

degrees of freedom of the original symmetric matrix S as defined in Eq. (1).
1
The restriction to a determinant of 1 (i. e., to SO(3) instead of O(3)) corresponds to the choice of an
appropriate order of the eigenvectors within R.

2
2 Repeated eigenvalues – results
If 2 of the 3 eigenvalues in Eq. (2) are the same, i. e., (without loss of generality)
λ0 6= λ1 = λ2 , then the symmetric matrix S can be written as (inverting Eq. (2)):
 
λ0 0 0
S = R diag(λ0 , λ1 , λ1 ) RT = R  0 λ1 0  RT . (4)
0 0 λ1
Now, the 6 entries (cf. Eq. (1)) of the matrix S are no longer independent. Indeed, by
setting λ1 = λ2 , the number of parameters that are required to describe S is reduced
from 6 to 4, since (a) only two eigenvalues are remaining and (b) the orientation of the
eigenvector v1 (defined by ξ above) can now be chosen arbitrarily in the plane orthogonal
to v0 (since, with v1 and v2 , every linear combination of v1 and v2 is also an eigenvector
to the eigenvalue λ1 ). Hence, 4 possible parameters describing S are λ0 , λ1 , and the two
angles, θ, φ, describing the orientation of v0 .
So, if S can be parametrized by only 4 parameters, then there must exist 2 (independent)
equations relating the matrix elements s00 , s11 , s22 , s01 , s02 , s12 to each other.

2.1 Main result

One such set of equations is
 
s01 s12
s00 = s22 + s02 − (5)
s12 s01
 
s01 s02
s11 = s22 + s12 − (6)
s02 s01
 
s02 s12
s00 = s11 + s01 − . (7)
s12 s02
Only 2 of these 3 equations are independent; e. g., the last equation can be found by
subtracting the second one from the first one.
With any 2 of these 3 equations, 2 of the 3 diagonal elements (s00 , s11 , s22 ) can be
expressed by the 3 off-diagonal elements (s01 , s02 , s12 ) plus the remaining third diagonal
element (if the given off-diagonal elements are not zero).

2.2 Corollaries
Three other sets of equations, each of which expresses 2 elements (1 diagonal and 1 off-
diagonal element) of S by the 4 other elements, can be derived from the results above by
simple transformations. E. g., an expression for s02 can be easily obtained from Eq. (5),
and a second equation for s11 can be derived by inserting this result for s02 into Eq. (6):
s01 s12 (s00 − s22 )
s02 = (8)
s201 − s212
s2 − s212 s00 − s22
s11 = s22 + 01 − s212 2 (9)
s00 − s22 s01 − s212

3
(expressing s02 and s11 by s00 , s22 , s01 , s12 ). Analog results derived from Eqs. (6) and (7)
are
s01 s02 (s11 − s22 )
s12 = (10)
s201 − s202
s2 − s202 s11 − s22
s00 = s11 + 01 − s201 2 (11)
s11 − s22 s01 − s202

(expressing s12 and s00 by s11 , s22 , s01 , s02 ). Finally, based on Eqs. (7) and (5), one finds

s02 s12 (s00 − s11 )

s01 = (12)
s202 − s212
s2 − s212 s00 − s11
s22 = s00 + 02 − s202 2 (13)
s00 − s11 s02 − s212

(expressing s01 and s22 by s00 , s11 , s02 , s12 ).

Again, the results require that certain sub-expressions are not zero such as s201 − s212
and s00 − s22 in the first set of equations.

2.3 More corollaries

Some relations for the off-diagonal entries of S can be expressed as quadratic equations,
which can be easily solved; these results are found by rearranging the first equations
(Eqs. (8), (10), (12)) of the previous subsection:
s01
s212 + (s00 − s22 ) s12 − s201 = 0, (14)
s02
s02
s201 + (s22 − s11 ) s01 − s202 = 0, (15)
s12
s12
s202 + (s11 − s00 ) s02 − s212 = 0. (16)
s01
Finally, quadratic equations for the squared off-diagonal elements depending on all 3
diagonal elements(s00 , s11 , s22 ) and only 1 other off-diagonal element can be found by
combining 2 of the 3 Eqs. (8), (10), and (12). Inserting Eq. (10) into Eq. (8) yields for
s202 :
   
(s202 )2 + (s00 − s11 )(s11 − s22 ) − 2s201 s202 + s201 s201 − (s00 − s22 )(s11 − s22 ) = 0 (17)

and inserting Eq. (8) into Eq. (10) yields for s212 :
   
(s212 )2 + (s00 − s11 )(s22 − s00 ) − 2s201 s212 + s201 s201 − (s00 − s22 )(s11 − s22 ) = 0, (18)

so, by solving these two quadratic equations, the off-diagonal elements s02 and s12 can
obtained from the 3 diagonal elements and the remaining off-diagonal element s01 .

4
 Similarly, inserting first Eq. (12) into Eq. (8) and then Eq. (10) into Eq. (12) yields:
   
(s212 )2 + (s00 − s22 )(s11 − s00 ) − 2s202 s212 + s202 s202 + (s00 − s11 )(s11 − s22 ) = 0, (19)
   
(s201 )2 + (s11 − s22 )(s22 − s00 ) − 2s202 s201 + s202 s202 + (s00 − s11 )(s11 − s22 ) = 0; (20)

and inserting first Eq. (12) into Eq. (8) and then Eq. (8) into Eq. (12) yields:
   
(s201 )2 + (s00 − s22 )(s22 − s11 ) − 2s212 s201 + s212 s212 − (s00 − s11 )(s00 − s22 ) = 0, (21)
   
(s202 )2 + (s00 − s11 )(s11 − s22 ) − 2s212 s202 + s212 s212 − (s00 − s11 )(s00 − s22 ) = 0. (22)

2.4 Summarized results

The following table summarizes which of the previous results are relevant in different
situations depending on the given and dependent entries in the symmetric matrix S.
(Obviously, not all possible combinations with two given diagonal elements are included.)

Table 1: Summary of results

given elements dependent required
diagonal off-diagonal elements equations

s00 s01 , s02 , s12 s11 , s22 (5), (7)

s11 s01 , s02 , s12 s00 , s22 (6), (7)
s22 s01 , s02 , s12 s00 , s11 (5), (6)
s00 , s11 s02 , s12 s22 , s01 (12), (13)
s00 , s22 s01 , s12 s11 , s02 (8), (9)
s11 , s22 s01 , s02 s00 , s12 (10), (11)
s00 , s11 , s22 s01 s02 , s12 (17), (18)
s00 , s11 , s22 s02 s01 , s12 (19), (20)
s00 , s11 , s22 s12 s01 , s02 (21), (22)

5
3 How to check the result
3.1 Parametrizing the matrix
The diagonalized form of a symmetric 3 × 3 matrix with a repeated eigenvalue can be
written as    
λ0 0 0 1 0 0
0 λ1 0  = λ1 1 + (λ0 − λ1 ) 0 0 0 . (23)
0 0 λ1 0 0 0
The general (i. e., not diagonalized) form of such a symmetric matrix can be found by
multiplying this expression by R and RT as in Eq. (4):
 
1 0 0
S = λ1 R 1 RT + (λ0 − λ1 ) R 0 0 0 RT . (24)
0 0 0

With R = [v0 v1 v2 ], one finds

   T
1 0 0  v0T
0 R = v0 0 0 v1  = v0 v0T = v0 ⊗ v0
T
(25)

R 0 0
 
0 0 0 v2T

where v0 ⊗ v0 is the dyadic (or outer) product of the vector v0 with itself. Using
R 1 RT = 1 and combining the last two equations yields

S = λ1 1 + (λ0 − λ1 )(v0 ⊗ v0 ). (26)

Writing v0 = (r, s, t)T , the dyadic product is

 
rr rs rt
v0 ⊗ v0 = rs ss st (27)
rt st tt

and since
√ v0 is normalized to length 1, i. e., v0 · v0 = r + s + t = 1, one can substitute
2 2 2

t = ± 1 − r2 − s2 to obtain:
√
r2 ±r√1 − r2 − s2
 
rs
S = λ1 1 + (λ0 − λ1 )  √ rs √ s
2 ±s 1 − r2 − s2  (28)
2 2
±r 1 − r − s ±s 1 − r − s 2 2 2
1−r −s 2

or, combined2 :
√
λ1 + (λ0 − λ1 )r2
!
(λ0 − λ1 )rs ±(λ0 − λ1 )r√1 − r2 − s2
S= (λ0 −√λ1 )rs λ1 + (λ0√− λ1 )s2 ±(λ0 − λ1 )s 1 − r2 − s2 (29)
±(λ0 − λ1 )r 1 − r2 − s2 ±(λ0 − λ1 )s 1 − r2 − s2 λ1 + (λ0 − λ1 )(1 − r2 − s2 )
s02
2
Some helpful relations, which can be obtained from Eq. (29), are r
s
= s12
as well as λ1 = s00 − s01 ss02
12
,
λ1 = s11 − s01 ss02
12
, and λ1 = s22 − s02 s12
s01
.

6
3.2 Putting things together
The main result (Eqs. (5), (6), (7)) can now be checked by substituting the expressions
from Eq. (29). For Eq. (5), one finds (starting with the right-hand side of the equation):
 
s01 s12
s22 + s02 −
s12 s01
s2 − s212
= s22 + s02 01
s12 s01
(λ0 − λ1 )2 r2 s2 − (λ0 − λ1 )2 s2 (1 − r2 − s2 )
= s22 + s02 √
±(λ0 − λ1 )s 1 − r2 − s2 (λ0 − λ1 )rs
2r2 − 1 + s2
= s22 + s02 √
±r 1 − r2 − s2 (30)
 p  2
2r − 1 + s 2
= s22 + (λ0 − λ1 )r 1 − r2 − s2 √
r 1 − r2 − s2
2 2
= s22 + (λ0 − λ1 )(2r − 1 + s )
 
= λ1 + (λ0 − λ1 )(1 − r2 − s2 ) + (λ0 − λ1 )(2r2 − 1 + s2 )
= λ1 + (λ0 − λ1 )r2
= s00 .

Similarly, for Eq. (6):

 
s01 s02
s22 + s12 −
s02 s01
s2 − s202
= s22 + s12 01
s02 s01
(λ0 − λ1 )2 r2 s2 − (λ0 − λ1 )2 r2 (1 − r2 − s2 )
= s22 + s12 √
±(λ0 − λ1 )r 1 − r2 − s2 (λ0 − λ1 )rs
2s2 − 1 + r2
= s22 + s12 √
±s 1 − r2 − s2 (31)
 p  2s2 − 1 + r2
= s22 + (λ0 − λ1 )s 1 − r2 − s2 √
s 1 − r2 − s2
= s22 + (λ0 − λ1 )(2s2 − 1 + r2 )
 
= λ1 + (λ0 − λ1 )(1 − r2 − s2 ) + (λ0 − λ1 )(2s2 − 1 + r2 )
= λ1 + (λ0 − λ1 )s2
= s11 .

As mentioned above, Eq. (7) can be obtained by subtracting Eq. (6) from Eq. (5).

7
References
[1] Anton H. Elementary Linear Algebra. 10th ed. John Wiley & Sons, Hoboken 2010.

[2] Datta KB. Matrix and Linear Algebra. Prentice-Hall of India, New Delhi 1991.

Acknowledgement
Thanks to Orodruin at physicsforms.com, who suggested to use Eq. (23).

Manuscript version history

2016-05-19 first version

2016-05-20 update:
• minor corrections in the introduction
• removed trivial solutions to quadratic equations
• added more quadratic equations to calculate 2 off-diagonal elements if all 3
diagonal elements are given

2016-06-02 update:
• re-arranged Eqs. (30) and (31) to clarify the reasoning
• added auxiliary relations in footnote 2
• added Table 1