Diagonalization and Simultaneous Symmetrization of the Gas-Dynamic Matrices

Author(s): R. F. Warming, Richard M. Beam, B. J. Hyett

Source: Mathematics of Computation, Vol. 29, No. 132 (Oct., 1975), pp. 1037-1045
Published by: American Mathematical Society
Stable URL: http://www.jstor.org/stable/2005742 .
Accessed: 20/05/2011 16:01
Your use of the JSTOR archive indicates your acceptance of JSTOR's Terms and Conditions of Use, available at .
http://www.jstor.org/page/info/about/policies/terms.jsp. JSTOR's Terms and Conditions of Use provides, in part, that unless
you have obtained prior permission, you may not download an entire issue of a journal or multiple copies of articles, and you
may use content in the JSTOR archive only for your personal, non-commercial use.
Please contact the publisher regarding any further use of this work. Publisher contact information may be obtained at .
http://www.jstor.org/action/showPublisher?publisherCode=ams. .
Each copy of any part of a JSTOR transmission must contain the same copyright notice that appears on the screen or printed
page of such transmission.
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of
content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms
of scholarship. For more information about JSTOR, please contact support@jstor.org.

American Mathematical Society is collaborating with JSTOR to digitize, preserve and extend access to
Mathematics of Computation.

http://www.jstor.org

MATHEMATICSOF COMPUTATION, VOLUME 29, NUMBER 132

OCTOBER 1975, PAGES 1037-1045

Diagonalization and Simultaneous Symmetrization

of the Gas-Dynamic Matrices
By R. F. Wanning,RichardM. Beam and B. J. Hyett
Abstract. The hyperbolicity of the unsteady, inviscid, gas-dynamic equations implies the
existence of a similarity transformation which diagonalizes an arbitrary linear combination 1 kiA. of coefficient matrices Ai. The matrix T that accomplishes this transformation is given explicitly, and the spectral norms of T and T are computed. It is also
shown that the individual matrices Ai are simultaneously symmetrized by the same similarity transformation. Applications of the transformations T and T 1'and their norms include the well-posedness of the Cauchy problem, linear stability theory for finite-difference approximations, construction of difference schemes based on characteristic relations,
and simplification of the solution of block-tridiagonal systems that arise in implicit timesplit algorithms.

1. Introduction.The inviscidgas-dynamicequationsin nonconservativeform in

three space variablescan be written as
-?
j A.'jaax.=?
-=0,
~~~~at
t
aI

(1.1)

(l.l)

j=1

where the vector u and the matricesA, are

p
ul
u =

U2

'

Al

U3

"I2 =

01

Ul ~p
0 ul

0
o

0
0

Ul
?

ul

pC2

U1i

_?

U2
?

U2

0
?

0
?

U3

0?

U2

llp
0

O
0

PC2

U2

A3
i0

O/P

0
1U3
O
0
0

0
0

p
0

U3

u13

i/p

PC2

U3

The elements of u are density, velocity components, and pressureand c = c(p, p) is the local sound speed. Associatedwith (1 .1) is a matrixP defined by
3
(1 .2)

p =

j=1

3
k.A

-,
j-l

kjA,(u),

- ?? < kj < ??.

Received December 16, 1974.

AMS(MOS)subject classifications (1970). Primary 65M10; Secondary 15A21, 15A18.
Key words and phrases. Hyperbolic partial differential equations, initial-value problems, inviscid
flows, construction and stability of finite-difference schemes, similarity transformation.
Copyrighti 1975, AmericanMathematicalSociety

1037

RICHARD M. BEAM AND B. J. HYETT

R. F. WARMING,

1038

The quasi-linearsystem (1.1) is called hyperbolicat the point (x,

singularmatrix T(k) such that
xi

u) if there exists a non-

X2
(1.3)

t,

T-'PT =3
0

X4

where the eigenvaluesX1of P are real and the norms of T and T-1 are uniformly
bounded in k = (kl, k2, k3)t, i.e.,
IITIIIIT-ll <K.

(1.4)

In Section 2 of this paper,we give explicit formulasfor the matrices T and T-1.
These matricesare rather simple in appearanceand, by careful normalizationof the columns of T, we have found that both TTt and its inverse are sparse(Tt denotes the
transposeof T). Consequently,the spectral norms of T and T-1 are readily computed
and we give explicit formulas for them. In addition, the similaritytransformation
T-'A1T simultaneouslysymmetrizesthe matricesAi for arbitraryk (Section 3). The
importanceof symmetrizingthe coefficient matriceshas been discussed in a recent paper by Turkel [14].
The relation between the matricesA, of the nonconservationform (1.1) and the
Jacobianmatricesof the conservationlaw form is given in Section 4. In fact, these matrices are similar [12, p. 363] and the transformationis given explicitly.
Finally, in the last section, severalapplicationsof the transformationsare indicated.
2. Bounded TransformationDiagonalizingP. The eigenvaluesof P are easily found
to be

xi

(2.1)
\4 =

= X2 =

k - u + c(k - k)112,

where

3 =
A5 =

k *u,
k *u - c(k *k)l /2,

k.k=

Zkj ,
j=1

k u=Zkuj.
j=1

Although the matrixP has an eigenvalueof multiplicity three, it has a complete set of
linearly independenteigenvectorsand, consequently, the matrix T is constructed using
these eigenvectorsas columns. Hence, we find

k1
O

(2.2)

0k
- k2

LO

k2
kk3
ki
O

k3
k
12

p/(c)
p/(c)
k /N
-~k /N~
k1/~~k/'

-k

k2/V/

-k2/v'k

k 3 /N2

-k3 /N2

pc/V2

j'

pc//5i2I

wherek= k(k k)l/2 The jth column of T is a right eigenvector correspondingto

[he eigenvalueX.. The inverseof T was formed from the left eigenvectorsof P:

GAS-DYNAMIC MATRICES

-k
(2.3)

k2
k3

1=

kTIV

kl /C
ki1c
2

k2
k2

k3

0
-k3
k2

1039

7Z/

k3/
- k3/V/2

k2/k

o0 -k //2-

k2/

l/(pc)
l/(V2ipc)

The determinantsof T and T-1 are

(2.4)

det T = (det T-1)-1 = pc.

The spectral norm of the matrix T can be computed from the formula
1TII= [r(TTt)l1/2

(2.5)

where r(TTt) is the spectral radiusof TTt. The eigenvectorscomprisingthe matrix T

as given by (2.2) were judiciously normalizedso that the product TTt led to the sparse
matrix
7t = I + Q,
(2.6)
where I is the identity matrix and Q is a matrix whose elements are all zero except for
the four corner elements:
q11 = (plc)2

(2.7)

= p2,

q15 =q51

q55 = p2C2 -1.

The singularvalues of T (i.e., the eigenvaluesa of TTt) are readily found to be

(2.8)

ff =

1 1

? +

+ c2 + p2c4

? (q>2-2

2c2
and consequently,
(2.9)

11
T12 = max=

Likewise, the product (T-1)tT-l

(2

_ 4p2c
2c 2

leads to a matrix with the following simple structure:

(2.10)
(T-1)tT-l =I+ R,
where R is a matrix whose elements are all zero except for the three corner elements:
(2.11)

r1= = =r

-1/c2,

r5

+ c

- p c

The eigenvaluesK of (I + R) are

(2.12)

K = 1, 1, 1,

2p2c4

and hence,
(2.13)
Finally, we note that
(2.14)

IIT1 112

max K =

p2c6)112

p? (-

? (

4p2c6)11/2

2p 2c4

11
Thl= (det T)l1T-1 11.

Since the norms of T and T-1 are independent of the real parameterskI, they
are uniformlybounded in k as requiredby (1.4).

R. F. WARMING, RICHARD M. BEAM AND B. J. HYETT

1040

In a curvilinearcoordinate system, the first-ordersystem (1.1) is altered by the addition of a nonhomogeneousterm, and the matricesAi are replacedby A1/hj where hj
is a scale factor. The transformationsof this section are still valid since the kj's are arbitrary real numberswhich can be redefined to absorb the scale factors.
3. Symmetrizationof the MatricesAj. The matrix T, given by (2.2), was constructed to diagonalizethe matrix P defined by (1.2). In addition, this same transforming
matrix T simultaneouslysymmetrizesthe individualmatricesAj. This result can be written in partitioned form as follows:

(3.1)

T-1AT =[

j1=1, 2, 3,

]
B. i5

where Dj and Cj are the diagonal matrices

Dj=

uj
O uj

uj

uj + ck

0
.

Cj
_

uj -ckj_

and

B1

=74=k3
[i)~~~~~
~

k3]
32 21 B2=+[ X2t~

0]

'3

B3=~~[k

kj.
k

It may be of interest to comment on how we obtained the above symmetric form.

In general, a transformationT which diagonalizesP will not symmetrizethe individual
matricesAi. In fact, the original T, say To, which we constructeddid not symrpetrize
the Aj's. But when we made the individualcalculationsTo 1AjTo, the results were
nearly symmetric and it was more or less obvious how to renormalizethe columns of
To to achieve the symmetric form (3.1). This same renormalizationalso led to a sparse
form for TTt.
A more direct approachwould be to first find some particularmatrix S that
would simultaneouslysymmetrizethe A,'s. Then S1PS would be symmetric and, consequently, this matrix could be diagonalizedby an orthogonal matrix U. If we define
T = SU, then this matrix would simultaneouslydiagonalizeP and symmetrizethe Aj's.
However,we are unawareof any systematic procedureof simultaneouslysymmetrizing
a set of noncommutingmatrices such as those defined by Eq. (1.1).
The symmetric form (3.1) is, of course, not unique, since any orthogonal similarity transformationon (3.1) yields a symmetric matrix. For example, if U is the 5 x S
orthogonal matrix

the3n

then

GAS-DYNAMIC MATRICES

[F.

Ut(Tl-A 1T) U

H.]

1041

j=1, 2,3,

HtU.

where E = uI 4,In

is the n x n identity matrix, and

Hf=c[O,-k3,k2,72

],
Hk

= c[k3,O,-k

,k2],

H3 =c[-k2,k1,O,k3].

4. Transformationsfor ConservationLaw Matrices.The gas-dynamicequations

(1.1) can also be written in conservationlaw form as
(4.1)
~ ~ ~ ~ ~~i"" aU 3 3F.
(4.1)
+ : Iax = 0,
at j=1

jx

where

p~ ~

pu1
+ P651

pu1pulu1
~
U =pU2

'Fi

PU2Uj + P52 X

pu3

p)
uj(eT +P631

_eT_

P"-

eT=

pU3Uj

'Y

1 2 q2q2

3
E Ui2

j=1

50 is the Kroneckerdelta and 'y is the ratio of specific heats. Eq. (4.1) can be reexpressed as
(4.2)

A aZ

aU+

where the A 's are the Jacobian matrices aF./aiu.

The matricesA1 of the nonconservativeform (1.1) and the matricesAi of the conservativeform (4.1) are related by the similaritytransformation
(4.3)

Ai

AM

where M is the Jacobian matrix 3iu/au. Computingthe requisite derivatives,we find

F1
~ I
~~M
au =
M = au~IU

(4 4)
(4.4)

U1

0?

U2

U3

2/2

Pu1 PU2 pU3

1/('y-

1)

and
1
~
(4U5/p
)
45 M-

U2/P

- U3/p
Ths

fq2
o(re_x)Uap-l(e,)U

Thus, for example,

0
O

l/p

= au =

0?
0

l/p
0

0
lI/p
(y-

O)3

0
0
(y-

1)

1042

R. F. WARMING, RICHARD M. BEAM AND B. 1. HYETT

A1 =MA 1M-1
0
(72lq2
-

1
- u2

- (y

U2

-U1u3

U3

a1

-(,y - l)u2

3)ul

- u1 u2

0
-(

0
-(e

)3
0

U1

a52

U0

- l)ulu2

(y-1

- (Y- 1)Ulu3

Yu1

where
a51 = K( - 1)Ulq

a52 = [yeT/p

'YeTUl/pI,

- (Y -

1)(u2 + q2/2)].

The matricesA, are quite complicatedcomparedwith the simple form of the matricesA
of Eq. (1.1). If one is concerned with constructingthe transformationwhich diagonalizes P =_ j= k1Aj or if one wants to symmetrizethe matricesAi, it is far simpler to
use the.similarity transform(4.3) and work directly with the A1's. For example, to
symmetrizeAp, we write simply
(4.6)

T-1M-1A.'MT=
T-1A.T,
i
I

and the symmetric result is given by the previouscalculation (3.1).

5. Applications.
(a) The CauchyProblem. It is known that if a first-ordersystem
(5.1)

au +A
at

au =0

'ai X.

(with constant coefficients) is hyperbolic, then the Cauchy problem is well-posed (see,
e.g., [7]). It is of interest to see how the results of Section 2 fit into a more general
theory developed by Kreiss [6]. He proved the following theorem: The Cauchy problem
for the system (5.1) is well-posed if and only if there exist real constants C > 0 and a,
and a positive definite Hermitianmatrix H(k) for which
(5.2a)

C1'

<

H(k)

<

CI

and
(5.2b)
where

H(k)P(ik) + P*(ik)H(k) < 2aH(k),

P(ik) = iZ kj1Aj.
Here Kreiss'theorem has been specializedto the particularcase of the system (5.1).
For the (linearized)gas-dynamicequations, a positive definite Hermitianmatrix that fulfills conditions (5.2) is the sparsematrix given by (2.10). The constant C is the maximum of the eigenvalues(2.9) and (2.13), and a is zero.
The general theory of well-posednessdeveloped by Kreiss [6] includes Eq. (5.1)
with variablecoefficients where the matricesA1 = A,(x, t) are smooth functions of x
and t. In the nonlinear case, A, = Aj(u, x, t) and the Cauchy problem is only wellposed in a sufficiently small interval0 < t < T [7], [12].

GAS-DYNAMIC MATRICES

1043

(b) Stability Theory. Linear stability theory for finite-differenceapproximations

to the hyperbolic system (1.1) or (4.2) is based on linearizedversionswhere the matrices
A, (or A,) are assumedto be constant. The simplest situation occurs when the amplification matrix G [12] is a polynomial in the matrix P defined by (1.2). In this case, G
is diagonalizedby a similaritytransformationwhere the matrix that accomplishesthe
similarityis given by (2.2). Since T and T-1 are bounded independent of k, the von
Neumann condition is sufficient as well as necessary for stability. Examplesof twostep, Lax-Wendroffmethods where G is a polynomial in P are the algorithmsproposed
by Rubin and Preiser[13], [2] and Zwas [15].
In a recent paper, Turkel [14] gave a similaritytransformationthat symmetrizes
the matricesA, (see Eq. (4.2)) for the case of two spatial dimensions. (The same transformation was given earlierby MacCormack[8].) The extension to three spatial dimensions is given by the similaritytransformation(4.6). Turkel used the resulting sparse
symmetric forms to simplify the linear stability analysis in situations where it is necessary to compute the spectral radius of G *G to provide a sufficient stability analysis.
(c) Difference Schemes Based on CharacteristicRelations. There are severalmethods of constructingdifference schemes for multidimensionalhyperbolic systems which
utilize characteristicrelations in their derivation. The eigenvaluesand eigenvectorsof
the matrix P, defined by (1.2), play an essential role in these methods. The method described by Magomedovand Kholodov [10] uses the eigenvaluesof the matricesA, and
the similaritytransformationsthat diagonalizeeach A1. For the gas-dynamicequations
(1.1), the requisite eigenvaluesand transformationsare given by Eqs. (2.1), (2.2), and
(2.3) by a proper choice of k.
Johnston and Pal [5] have derived a difference method which makes use of the bicharacteristicsof the system. In this technique, one needs the eigenvaluesof the matrix P. Appropriatederivativesof these eigenvaluesyield the bicharacteristicsof the system. The difference scheme then follows by an approximateintegrationalong a bicharacteristic. The resultingdifference schemes [5] are such that the amplification matrix
is a linear function of P. Hence, the von Neumann condition is necessaryand sufficient
for stability, as describedin the first paragraphof Section 5(b).
(d) Implicit Split Algorithms. Efficient explicit algorithmsfor the hyperbolic system (1.1) or (4.1) can be constructedwith the aid of the method of fractional steps or
time splitting (see, e.g., [9], [3]). There is some motivation for consideringimplicit
methods because of the time-step limitations inherent in explicit algorithms. Although
multidimensionalimplicit methods requirethe solution of large sparsematrices, the
method of fractional steps provides an efficient solution procedure. For example, a
time-split, implicit, (Crank-Nicolson)algorithmfor the system (1.1) is the following:
(5.3a)

(Li)un +1/3

(5.3b)

(L2)un+2/3 = (L2)un+1/3

(5.3c)

(L3)un+l = (L3)Un+2/3

(L )un,

1044

R. F. WARMING, RICHARD M. BEAM AND B. J. HYETT

where
(L)

V[I? AtAj(Do)x.]

and
Uij1+ 1,12,j3 -u

(Do)x1u

1-1

j3

2Ax1

with similardefinitions for (Do)X2 (Do)x3.

In order that the above algorithmbe uniformly second order in the time and spatial increments At, Ax,, the matrix coefficients Ai must be evaluatedwithin some predictor-correctorsequence and the order in which the operatorsL1, L2, L3, appearmust
be reversedwhen the solution is advancedin the next time step, i.e., n + 1 -* n + 2.
We will not go into these details which are discussed elsewhere [11].
Each fractional step of the algorithm(5.3) requiresthe solution of a block-tridiagonal system whose submatrices are each of order 5. There is a well-known direct method [4] for solving tridiagonalmatrices and a generalizationfor block tridiagonals. For
the present application, 5 x 5 matrices must be inverted in the sequence of steps required for the direct solution. However,we can take advantageof the fact that the matrices Ai are each diagonalizedby the similarity transformationT- AiT by appropriate
choice of the parametersk,. If the A1's were spatially constant, we could change variables by introducingw = T-lu and obtain an uncoupled set of equations for the components of w. Otherwise,we can still effect a considerablesimplificationif we properly
take into account the dependence of the Ai's upon the vector of dependent variablesu.
For example, if we multiply Eq. (5.3a) by T-1(k) with kt = (1, 0, 0) and write out the
resultingequations, we find we can invert the operator on the left of (5.3a) by solving
one block-tridiagonalsystem with 2 x 2 submatrices(which are easily inverted analytically) followed by three scalartridiagonalinversions. An equivalent reduction can be
achievedby reordering(multiplyingby a permutationmatrix) the elements of the vector u in each fractional step of algorithm(5.3). A more complete description of the algorithm and numericalcomputations for unsteady aerodynamicflows will be reported
elsewhere [1].
Computational Fluid Dynamics Branch
Ames Research Center, NASA
Moffett Field, California 94035
1. R. BEAM & R. F. WARMING, "Implicit, time-split difference schemes for the transient solution of hyperbolic equations." (In preparation.)
2. B. EILON, "A note concerning the two-step Lax-Wendroff method in three dimensions,"
Math. Comp., v. 26, 1972, pp. 41-43. MR 45 #9503.
3. D. GOTTLIEB, "Strang-type difference schemes for multidimensional problems," SIAM J.
Numer. Anal., v. 9, 1972, pp. 650-661. MR 47 #2826.
4. E. ISAACSON & H. B. KELLER, Analysis of Numerical Methods, Wiley, New York, 1966.
MR 34 #924.
5. R. L. JOHNSTON & S. K. PAL, "The numerical solution of hyperbolic systems using bicharacteristics," Math. Comp., v. 26, 1972, pp. 377-392. MR 46 #4758.
6. H.-O. KREISS, "Uber sachgemasse Cauchyprobleme," Math. Scand., v. 13, 1963, pp. 109128. MR 29 #6177.
7. H.-O. KREISS & J. OLIGER, Methods for the Approximate Solution of Time Dependent
Problems, Global Atmospheric Research Programme Publications Series, no. 10, 1973.

GAS-DYNAMIC MATRICES

1045

8. R. W. MacCORMACK, The Effect of Viscosity in Hypervelocity

Impact Cratering, AIAA
paper no. 69-354,
1969.
9. R. W. MacCORMACK & A. J. PAULLAY, Computational Efficiency Achieved by Time
Splitting of Finite Difference Operators, AIAA paper no. 72-154,
1972.
10. K. M. MAGOMEDOV & A. S. HOLODOV, "The construction of difference schemes for
hyperbolic equations based on characteristic relations," U.S.S.R. Comput. Math. and Math. Phys.,
v. 9, 1969, pp. 158-176.
MR 42 #2696.
11. G. I. MARCUK, "On the theory of the splitting-up method," in Numerical Solution of Partial Differential Equations, II (SYNSPADE 1970) (Proc. Sympos., Univ. of Maryland, College Park,
Md., 1970), Academic Press, New York, 1971, pp. 469-500.
MR 44 #1234.
12. R. D. RICHTMYER & K. W. MORTON, Difference Methods for Initial-Value Problems,
2nd ed., Interscience Tracts in Pure and Appl. Math., no. 4, Interscience, New York, 1967. MR 36
#3515.
13. E. L. RUBIN & S. PREISER, "Three-dimensional
second-order accurate difference
schemes for discontinuous
hydrodynamic flows," Math. Comp., v. 24, 1970, pp. 57-63.
MR 41
#9494.
14. E. TURKEL, "Symmetrization
of the fluid dynamic matrices with applications," Math.
Comp., v. 27, 1973, pp. 729-736.
15. G. ZWAS, "On two-step Lax-Wendroff methods in several dimensions," Numer. Math., v.
20, 1972/73, pp. 350-355.
MR 48 #1484.