time-reversal method and cross-correlation techniques by normal mode theory: a three-point problem
TRANSCRIPT
submitted to Geophys. J. Int.
Time Reversal Method and Cross-Correlation techniques by
normal mode theory: a 3-point problem. DRAFT
J.-P. Montagner1, Y. Capdeville1, M. Fink2, C. Larmat3, H. Nguyen1, B. Romanowicz4
1 Seismology Laboratory, Institut de Physique du Globe UMR-CNRS 7154, 4 Place Jussieu 75252 Paris Cedex 05, France
2 L.O.A., E.S.P.C.I., Rue Vauquelin, 75005 Paris, France
3 Seismology Lab., U.C. Berkeley, Berkeley, CA, U.S.A.
4 L.A.N.L., Los Alamos, NM, U.S.A.
SUMMARY
Long-period time-reversal experiments in the Earth produce excellent focusing of seismic
waves at the earthquake location. Normal mode theory enables for a finite body such as the
Earth, to investigate why time–reversal technique works when applied to seismic waves
in an elastic medium. It is demonstrated how the informations given by the 3-components
of elastic waves can be optimally incorporated for the localization in time and space of
earthquake, for recovering the whole seismic moment tensor and the source time function.
A 3 point problem, with 2 fixed points and one variable point, can be defined. The similar-
ities (and differences) between time reversal imaging and cross-correlation techniques are
investigated. For a random surface distribution of seismic sources, the cross-correlation
of seismograms in 2 stations is not exactly the Green function between these 2 stations
We demonstrate the importance of a correct weighting of stations by Voronoi tessellation
on a global scale.
Key words: Time reversal, normal modes, cross-correlation, adjoint technique
2 J.-P. Montagner, Y. Capdeville, M. Fink, C. Larmat, H. Nguyen, B. Romanowicz
1 INTRODUCTION: TO BE COMPLETED
The normal mode approach is well suited for investigating the propagation of seismic waves in the
spherical Earth. So far, normal mode theory has been only applied to scalar acoustic waves (Draeger
& Fink (1997, 1999), Weaver & Lobkis (2002)). By using the normal mode formalism developed
by Capdeville (2005), Larmat et al. (2006) showed that it is very easy to back-propagate seismic
waves at long periods by using seismograms synthetized by normal mode summation to the Sumatra-
Andaman earthquake of 26 Dec. 2004. She showed that the focusing of seismic waves at the epicenter
is as good as with the Coupled Spectral Element Technique (Capdeville et al. (2002) where synthetic
seismograms are calculated in 3D-heterogeneous Earth. The time-reversal technique as implemented
by Larmat et al. (2006) has many similarities with the adjoint method recently detailed in Tromp et al.
(2005).
We show in this paper how to generalize the 1D-normal mode approach of Draeger & Fink (1999),
Weaver & Lobkis (2002)) to 3D-elastic medium. We also demonstrate why time reversal technique
works when applied to the Earth for seismic waves, and answer some questions such as how the infor-
mations given by the 3-components of elastic waves can be optimally incorporated for the localization
of earthquake, for recovering the whole seismic moment tensor and the source time function?
The potential applications of this technique are ongoing. In addition to automatically locate earth-
quakes and retrieve their seismic moment tensor, time-reversal techniques using normal modes can
be used for locate sources excited in the long-period range (period larger than 100s) where the effect
of lateral heterogeneities is not large. Larmat et al. (2008) applied time-reversal technique to locate
Glacial earthquakes. Other potential applications are the location of the source of seismic Hum (Suda
et al. (1998), Fukao et al. (2002)).
Time Reversal Method and Cross-Correlation techniques by normal mode theory: a 3-point problem 3
2 TIME REVERSAL WITH NORMAL MODE THEORY
2.1 Forward problem
The elasto-dynamics (Navier-Stokes) equation for a continuum medium can be written:
ρd2uidt2
=∑j
σij,j + fi + ρgi
fi et ρgi are external applied forces. E is the source, R receiver for the forward problem, M the
observation point of the time reversed field. Generally, when neglecting the advection term, this local
equation can be written under the form:
(ρ∂tt +H0)u(r, t) = F(r, t) (1)
whereH0 is an integro-differential operator calculated in a reference Earth model and F expresses
the ensemble of external forces applied in r at time t. Gravity forces are neglected and it is assumed
that F is zero for t < 0. Bra-ket notations from quantum mechanics (Cohen-Tannoudji et al., 1973)
are used for displacement eigenfunction uK(r, t):
uK = D|q,n, `,m〉 =n D`|K〉 = (nU`(r)er +n V`(r)∇1 +nW`(r)(−er ×∇1)Y ml (θ, φ) = (2)
(nU`er + 1√`(`+1)
nV`(eθ ∂∂θ +eφ1
sinθ
∂
∂φ) +
1√`(`+ 1)
nW`(eθ1
sinθ
∂
∂φ− eφ
∂
∂θ)Y ml (θ, φ)
uK = |K〉 and ωK are respectively the set of eigenfunctions and eigenfrequencies, solutions of
the equation without 2nd member (−ρω2K + H0)|K〉 = 0 where the indices K = (q, n, `,m) and
k = (q, n, `); q can take 2 values, one for spheroidal modes and two for the toroidal modes (q is
often emitted when there is no ambiguity); n (radial order), ` (angular order), m (azimuthal order) are
quantum numbers. For a spherically symmetric elastic non rotating earth model, there is a degeneracy
in m, the eigenfrequency ωK is independent of m, and can be noted ωk, and the eigenfunctions nU` ,
nV` , nW` are only functions of radius r. It is possible to expand any displacement u(r, t) on the set
of eigenfunctions uK(r, t) = |K〉e−iωKt:
u(r, t) =∑K
aK |K〉e−iωKt =∑K
aKuK(r)e−iωKt (3)
|K〉 are orthogonal, normalized functions, fulfilling the orthonormalization relationship:
〈K′|K〉 =∫Vρu∗K′(r)uK(r)dV = δKK′
Let us first introduce some useful notations, such as the dot product , u.v = ui.vi
4 J.-P. Montagner, Y. Capdeville, M. Fink, C. Larmat, H. Nguyen, B. Romanowicz
and the inner product
〈u|v〉 =∫Vρu∗(r)v(r)dV (4)
A very useful theorem is the addition theorem (also named closure relationship):∑
k |K(r)〉〈K(r′)| =
ΣkuK(r)u∗K(r′) = Iδ(r− r′) where I is the identity matrix.
When using common spherical harmonics, the classical addition theorem expression is found:
m=∑m=−`
Y m` (θE , φE)Y m
` (θR, φR) =4π
2`+ 1P 0` (cos∆ER) (5)
where ∆ER is the angular distance between E and R. There are some other useful relationships, such
as generalization of classical addition theorem, applied to generalized spherical harmonics (Edmonds
(1960); Li & Romanowicz (1996); Capdeville et al. (2000)
m=∑m=−`
Y Nm` (θE , φE)Y Nm
` (θR, φR) =4π
2`+ 1eiN
′γERPNN′
` (cosβER)eiN′αER (6)
where PNN′
` (cosβER) are the generalized Legendre functions. The index E is related to the source
location, and the index R to the receiver location. The angle−αER is the back-azimuth at the receiver,
π− γER is the azimuth at the source, and βER is the epicentral distance between the source E and the
receiver R.
Bra and ket can be applied to the operator H0 such that:
〈K′(r)|H0|K(r′)〉 =∫VdV
∫V ′dV ′ u∗K′(r)H0uK(r′) = ω2
KδKK′ (7)
It is easy to calculate the excitation coefficient aK (Gilbert (1971)) by taking either the Fourier trans-
form or the Laplace transform of equation (2). At the receiver point R and for an earthquake at point
E, the displacement can be written:
u(rR, ω) =∑K
∫V ′dV ′ u∗K(rE)F(rE, ω)
uK(rR)−ω2 + ω2
K
=∑K
|KR〉〈KE|FE(ω)1
−ω2 + ω2K
(8)
where for sake of simplicity, we note |KE〉 = |K(rE)〉, |KR〉 = |K(rR)〉. The same notation will be
used later on for other points.
An inverse Fourier transform can be performed in order to come back in the time domain. For the
source of the earthquake, we can consider 2 simple cases. In the first one, the force system F(r, t) is a
delta function in time and space F(r, t) = FEδ(r− rE)δ(t). And in the second one, the force system
Time Reversal Method and Cross-Correlation techniques by normal mode theory: a 3-point problem 5
F(r, t) is a time Heaviside function H(t): F(r, t) = H(t)F0(r), which is much more realistic for an
earthquake.
For a delta function in time:
u(rR, t) =∑K
(∫VE
dVEu∗K(rE)FE(rE))uK(rR)H(t)sin(ωKt)ωK
=∑K
|KR〉〈KE|FE〉H(t)sin(ωKt)ωK(9)
Let us define by index γ the coordinates of vectors at point E, β at point R, and later on α for any
other point M in a Cartesian coordinate system. The Green function can be written:
Gγβ(rE , rR, t; 0) =∑K
u∗γK (rE)uβK(rR)H(t)sin(ωKt)ωK
=∑K
|KβR〉〈K
γE|w(t) (10)
where w(t) = H(t) sin(ωKt)ωK
. It is also useful to define the Green function of velocity u(rR, t) for
t > 0 as will be seen in section 3.2
G(rE, rR, t; 0) =∑K
|KR〉〈KE|H(t)sin(ωKt) (11)
For acoustic waves (scalar waves), the Green function is a scalar function, whereas for elastic
waves, it is a second order tensor. The spatial reciprocity GγβER = Gβγ
RE can be directly seen, but it
must be noted that it is more complex than for scalar waves.
For a source Heaviside Function in time, the general expression of displacement is given by:
u(rR, t) =∑K
(∫VE
dVE u∗K(rE)F0(rE))H(t)1− cos(ωKt)
ω2K
uK(rR) (12)
It is a remarkable expression, since the displacement is the result of the summation of all modes
calculated at time t. We can come back to the previous expression (10), for t > 0, by taking the time
derivative of equation (12). It can be shown that the contribution of eigenmodes is usually zero except
when there is a constructive interference corresponding to a stationnary phase (Romanowicz (1987)).
For t > 0, the corresponding acceleration Γ(rR, t) has a very simple expression:
Γ(rR, t) =∑K
uK(rR)H(t)cosωK(t)∫VE
dVEu∗K(rE)F0(rE)
For a point source in time and space, it is usual to introduce the seismic moment tensor Mmγ
such that F γ0 (rE) = −Mγm∂∂ξm
δ(ξ − rE).
u(rR, t) =∑K
ε∗Kγm(rE)MγmH(t)
1− cosωK(t)ω2K
uK(rR) =∑K
|KR〉〈KE|EMw′(t) (13)
This expression corresponds to Born approximation at zero order. Instead of the Green Function
6 J.-P. Montagner, Y. Capdeville, M. Fink, C. Larmat, H. Nguyen, B. Romanowicz
in displacement for a point source in time and space, we must use now (with respect to Mγm)):
Gβγ,mER (t) =∑K
uβK(rR)ε∗Kγm(rE)H(t)
1− cosωKtω2K
If the static term is discarded, and if the instrument vector v is introduced, the expressions derived
by Woodhouse & Girnius (1982) for a spherically non rotating earth model are retrieved (see table 1),
v.u(t) =∑k
(−cos(ωkt)
ω2k
∑m
Rmk (rR, θR, φR)Smk (rE , θE , φE)
)with
Rmk (rR, θR, φR) =N=1∑N=−1
RkN (rR)Y Nm` (θR, φR)
and
Smk (rE , θE , φE) =N=1∑N=−1
SkN (rE)Y Nm` (θE , φE)
For example, the vertical component is expressed as:
ur(rR, t) = vr∑n,`,m
nU`(rR)Y m` (θR, φR)(∂rU(rE)Mrr +
12F (rE)(Mθθ +Mφφ))Y m
` (θE , φE)eiωkt
If the addition theorem (6) is taken into account:
ur(rR, t) = vr∑n,`
nU`(rR)−2`+ 1
4πP 0` (cos(∆ER)(∂rU(rE)Mrr +
12F (rE)(Mθθ +Mφφ))eiωkt
Time Reversal Method and Cross-Correlation techniques by normal mode theory: a 3-point problem 7
2.2 Time reversal Application
Now u(rR, t) is the displacement field applied by time reversal recorded in different seismic stations
in rR to the Earth surface. The equivalent dynamic force is Γ(r) = −ω2u(rR, t). A force proportional
to displacement λ.u(rR, t) (or acceleration λ′.Γ(rR, t) ) can be applied.
In this 3 point problem, E(θE , φE), R(θR, φR),M(θM , φM ), the time-reversed signal will be
successively calculated at any point at the surface of the earth M and then, at the particular case of
point E (where M = E). Since the problem is linear, the time-reversed field at any point M due a
distribution of receivers will be also consider . Finally, the effect of a random distribution of sources
will be calculated. At the receiver point rR, the signal is found to be the result of the convolution of
the Green function with the source time function f(t) ∗GER(t). The time reversed seismogram in rR
is f(−t) ∗GER(−t) and in point rM the wavefield S(rM , t) will be:
S(rM, t) = f(−t) ∗GER(−t) ∗GRM (t) (14)
or in terms of deformation:
Ξ(rM, t) = M(−t) ∗EER(−t) ∗ERM (t)
where EER = 1/2(∇+∇T )GER
So the time-reversed signal is the cross-correlation of 2 Green functions. As already noted by
Draeger & Fink (1999) the signal in point rE is the autocorrelation function of GER which is known
to have its maximum at t = 0. At other points M, this signal is the cross-correlation of the 2 Green
Functions GER and GRM .
GER(−t) ∗GRM (t) =∫dτGER(τ + t)GRM (τ)
This result can be used as well to understand the cross-correlation of noise signals in 2 different
stations (see section 3.2). This correspondence between cross-correlation and time reversal was al-
ready noted in several papers since Draeger & Fink (1999) (see for example Derode et al. (2003)). But
this property has been primarily used for getting the Green function between 2 points, not for studying
the source, as done for example in seismology by Gajewski & Tessmer (2005) for local studies and by
Larmat et al. (2006), Larmat et al. (2008) at regional and global scale.
For the elastic case, we can send back the expression of the displacement found in section 2.1 and
following the approach of Capdeville (2005), all expressions are written in the Fourier domain. For
a point source in time F0(rE, ω) = eFδ(rE), the displacement field in the Fourier domain resulting
8 J.-P. Montagner, Y. Capdeville, M. Fink, C. Larmat, H. Nguyen, B. Romanowicz
from the forward propagation is according to equation (8):
uR(ω) =∑K
|KR〉〈KE|FE1
−ω2 + ω2k
(15)
And in the time domain (delta function in time): uR(t) =∑
K |KR〉〈KE|FEsinωktωk
For a point force Heaviside function H(t):
uR(ω) =∑K
|KR〉〈KE|FE1
iω(−ω2 + ω2k)
In point M , the time-reversed field for a delta-function in time is:
SM (ω) =∑J
|JM〉〈JR|uR(ω)1
(−ω2 + ω2j )
Usinjg the expression of uR(ω), it is found for a source delta function that:
SM (ω) =∑J
∑K
|JM〉〈JR|KR〉〈KE|FE1
(−ω2 + ω2k)
1(−ω2 + ω2
j )
It is necessary to bear in mind that in this expression, 〈JR|KR〉 does not correspond to the inner
product as defined in equation (5) but only to the scalar product of 2 eigenfunctions at point R. Be
careful since the convolution F ∗GER(−t) ∗GRM (t) is written from left (source) to right (point M)
whereas equation (15) is written in the opposite sense, which might induce some confusion. In terms
of cartesian coordinates, this equation can be written:
SαM (ω) =∑
J
∑K Σβγ
|JαM〉〈JβR|K
βR〉〈K
γE|F
γE
(−ω2+ω2k)(−ω2+ω2
j )
=∑J
Σβ|JαM〉〈JR
β|(−ω2 + ω2
j )︸ ︷︷ ︸BαβRM (ω)
∑K
Σγ|Kβ
R〉〈KγE|F
γE
(−ω2 + ω2k)︸ ︷︷ ︸
CβER(ω)
For sake of simplicity, let us continue the calculation by assuming that the source function in E is
a delta function. The case of a source Heaviside function in E is detailed in Appendix A. Sums over
J and K can be separated:
SM (ω) =∑J
|JM〉〈JR|(−ω2 + ω2
j )︸ ︷︷ ︸BRM (ω)
∑K
|KR〉〈KE|FE
(−ω2 + ω2k)︸ ︷︷ ︸
CER(ω)
(16)
Since the spectrum of the time-reversed field SR is the product of the spectra of 2 functions BRM
and CER, the field in the time domain is the convolution of these 2 fonctions. Since the time-reversed
of uR(t) is sent back, the cross-correlation between the 2 functions must be calculated (the cross-
Time Reversal Method and Cross-Correlation techniques by normal mode theory: a 3-point problem 9
correlation is the convolution of the time-reversed signal) . The Fourier transform of a time reversed
function f(−t) is F (−ω). Consequently,
BRM (t) =∑J
|JM〉〈JR|sinωjt
ωjH(t)
CER(t) =∑K
|KR〉〈KE|FEsinωkt
ωkH(t)
And finally, if the time-reversed field is calculated for time duration ∆T = [t1; t2] between times
−t2, and −t1, we introduce the time variable tR so that refocusing occurs at tR = 0.
SM (t) = BRM (t) ∗CER(−t) =∑J
∑K
|JM〉〈JR|KR〉〈KE|FEωk
1ωj
∫ t2
t1
dτsinωjτsinωk(tR + τ)︸ ︷︷ ︸=jk
(17)
The case with a Heaviside source time function can be done in a similar way (see Appendix A)
2.2.1 Calculation of =jk
=jk =∫ t2
t1
dτsin(ωk(tR + τ))sin(ωjτ)
=jk = −12sinωktR
∫ t2
t1
dτ [sin(ωk+ωj)τ+sin(ωk−ωj)τ ]+12cosωktR
∫ t2
t1
dτ [cos(ωk+ωj)τ−cos(ωk−ωj)τ ]
The terms with (ωj +ωk)τ are negligible since 1/(ωj +ωk)� ∆T , provided that a long enough time
series is used.=jk = 1
2∆TcosωktR if ωk = ωj
=jk = 12sinωktRωk−ωj (cos(ωk − ωj)t2 − cos(ωk − ωj)t1) if ωk 6= ωj
+12cosωktRωk−ωj (sin(ωk − ωj)t2 − sin(ωk − ωj)t1)
The first term is dominant provided that ∆T is long enough. The second term (when ωk 6= ωj) is
only important when this difference is small. The time 1∆ω corresponds to the Heisenberg time. At very
long period, the effect of heterogeneities (and rotation) is small, and their effect can be neglected. For
the earth normal modes, 1∆ω ≈ 2.103s for eigenperiods around 200s, which means that if δT � 1
∆ω ,
the non-diagonal term of Ijk will tend to interfere destructively and the contribution of non-diagonal
terms will be small compared to the diagonal contribution. However, that is only a zeroth-order ap-
proximation valid at very long-period. These expressions make it possible to incorporate the coupling
between different modes, but this study is beyond the scope of this paper (see Capdeville et al. (2000)
for a complete discussion of such a coupling). Therefore, if ∆T � 1∆ω , =jk ≈ 1
2δkj∆TcosωktR.
10 J.-P. Montagner, Y. Capdeville, M. Fink, C. Larmat, H. Nguyen, B. Romanowicz
S(rM , tR) =∑J
1ωj|JM〉〈JR|
∑K
|KR〉〈KE|F (rE)ωk
12δkj∆TcosωktR (18)
2.2.2 The observation point is at the source point E: Cavity Equation
The time reversal of acoustic waves (scalar case) has been extensively investigated by Draeger & Fink
(1999). If FEδ(t) = F , the convolution F.GER(−t) ∗GRE(t) if J = K, according to equation (18),
is:
SE(tR) =∑K
1ω2k
|KE〉〈KR|KR〉〈KE|12
∆TcosωktR
For acoustic waves, the bras and kets are scalar functions. It can be easily seen that this expression
is the same as the convolution of GEE(−t) by GRR(t) if J = K. As a matter of fact, F.GEE(t) =∑J
1ωk|KE〉〈KE|F sinωkt
ωk, GRR(t) =
∑J
1ωj|JR〉〈JR|
sinωjtωk
, and
F.GEE(−t) ∗GRR(t) =∑J
1ωj|JR〉〈JR|
∑K
1ωk|KE〉〈KE|F
1ωk=jk
Since there is no difference for scalar eigenfunctions between bra and ket, for J = K, Draeger & Fink
(1999) obtained the remarkable cavity equation for acoustic waves:
F.GER(−t) ∗GRE(t) = F.GEE(−t) ∗GRR(t) =∑K
1ωk2|KR〉〈KR|KE〉〈KE|
12
∆TcosωktR
In the elastic case, the equivalent of the cavity equation is more complex, since it is necessary to
take account of the tensorial character of Green functions and of the difference between bra and ket.
Considering the equation (18) and if point M is located in E,
SαE(tR) =∑K
∑J
1ωkωj
Σβγ |JαE〉〈JβR|K
βR〉〈K
γE|F
γE =jk
If only the dominant diagonal elements of =jk is taken into account, the sum over J can be sup-
pressed and by reordering the different terms:
SαE(tR) =∑K
Σβγ1ω2k
〈KβR|K
βR〉|K
αE〉〈K
γE|F
γE =kk (19)
If only the convolution product CC1 = GER(−t) ∗GRE(t) is considered:
CCαγ1 = GER(−t) ∗GRE(t)(tR) =∑K
Σβ1ω2k
〈KβR|K
βR〉|K
αE〉〈K
γE| =kk (20)
So, the effect of the source is separated from the effect of the receiver and the 2 Green functions
GEE(t) and GRR(t) can be introduced:
Time Reversal Method and Cross-Correlation techniques by normal mode theory: a 3-point problem 11
GαγEE(t) =∑K
|KαE〉〈K
γE|H(t)
sinωkt
ωk
GβζRR(t) =∑J
|JβR〉〈JζR|H(t)
sinωjt
ωj
The cross-correlation CC2 of GEE(−t) ∗ GRR(t) involves the same integral =kj(t) as in equation
(16)
CCαβγζ2 = GαγEE(−t) ∗GβζRR(t) =∑K
∑J
|JβR〉〈JζR|
ωj
|KαE〉〈K
γE|
ωk=kj(t) (21)
This expression of CCαβγζ2 can be contracted by taking ζ = α in order to have a cross-correlation
similar to CCαγ1 , and by summing over β,
CCβγ2 = GαγEE(−t) ∗GβαRR(t) =∑K
∑J
Σα|JβR〉〈JαR|
ωj
|KαE〉〈K
γE|
ωk=kj(t)
If J = K
CCβγ2 = GαγEE(−t) ∗GβαRR(t) =∑K
Σα1ω2k
|KβR〉〈K
αR|Kα
E〉〈KγE|=kk(t)
In order to facilitate the comparison of CC1 and CC2, we interchange the role of E and R, α and β,
we obtain:
CCαγ2 = GβγRR(−t) ∗GαβEE(t) =∑K
Σβ1ω2k
|KαE〉〈K
βE|K
βR〉〈K
γR|=kk(t) (22)
The expression of CCαγ2 in equation (22) looks similar to the one in equation (20), but it is
not. Actually, the scalar product 〈KβR|K
βR〉 appears in equation (20) whereas it is the scalar product
〈KβR|K
βE〉 which comes up in equation (22). However, if we apply to CCαγ1 and CCαγ2 a vector force
proportional to |KR〉, then the same expression is obtained:
CCαγ1 |KγR〉 = CCαγ2 |K
γR〉 =
∑K
Σβγ1ω2k
|KαE〉〈K
βR|K
βR〉〈K
γE|K
γR〉=kk(t)
If we reverse the roles of E and R, we have to apply a vector force 〈KR|.
Equation (20) has the interesting property that it displays the contribution of one receiver and it
can be easily generalized to a complete global distribution of earthquakes as it will be seen in section
3.
12 J.-P. Montagner, Y. Capdeville, M. Fink, C. Larmat, H. Nguyen, B. Romanowicz
2.2.3 Degenerate modes in azimuthal number m
In order to go further, the expression of the eigenfunction provided in equation (2) can be taken into
account. Some particular cases will be detailed when the eigenfunctions are degenerate.∑
k is now
written explicitly∑
k =∑
n,`,m as well as the eigenvector |kR〉 = nD`(r)Y m` (θ, φ)
In the general case, the equation (18) can be rewritten
SM (tR) =∑
n′,`′,m′n′D`′Y
m′`′ (rM)n′D∗`′Y
m′∗`′ (rR)
∑n,`,m
nD`Ym` (rR)nD∗`Y
m∗` (rE)FE
=jknω`n′ω`′
(23)
For example, the vertical component of the displacement, assuming also that the force is also a
vertical force uK(r, θ, φ) = nU`(r)Y m` (θ, φ). In that case, the addition theorem (or Closure relation-
ship) for spherical harmonics (equation 5) can be used:
SM (tR) =∑
n′,`′,m′n′U`′Y
m′`′ (rM)n′U`′Y
m′∗`′ (rR)
∑n,`,m
nU`Ym` (rR)nU`Y m∗
` (rE)FE=jk
nω`n′ω`′
=∑n′,`′
n′U`′(rM )P 0`′(cos∆RM )n′U`′(rR)
∑n,`
nU`(rR)P 0` (cos∆ER)nU`(rE)FE
=jknω`n′ω`′
For the 3- components of displacement, the same kind of calculation can be performed, but the
vectorial operators nD` impose to use the more general addition theorem as defined by equation
(6) (Capdeville et al. (2000)), which introduces generalized Legendre functions PNN′
` (cos∆) with
Nε[−2,+2] The source term nD∗`Ym∗` (rE)FE can be transformed if the source force is the gradient
of the moment tensor and is equal to nE`M
SM (tR) =∑n′,`′
n′D`′(rM)P 0`′(cos∆RM )n′D∗`′(rR)
∑n,`
nD`(rR)nE∗`M(rE)P 0` (cos∆ER)
=jknω`n′ω`′
(24)
For sake of simplicity, the notation of Woodhouse & Girnius (1982) can be used.
BARBARA
According to Romanowicz (1987), it is possible to apply a phase stationary approximation. The
multiplication of Legendre polynomials P 0`′(cos∆RM ) by P 0
` (cos∆ER) is maximum for ∆ = 0 or π.
It means that points E, R and M must be on the same great circle in order to provide a maximum
contribution to the time-reversed displacement (see Figure 1). In addition, since in equation (24), the
contributions of each eigenfrequency will tend to cancel out and all sinusoids will only be in phase
(and to interfere constructively) when tR = 0.
There will be also some non null contributions to time-reversed seismograms for some stationary
phases, corresponding to particular group velocities. In that case some ghosts wavetrain will come up
Time Reversal Method and Cross-Correlation techniques by normal mode theory: a 3-point problem 13
E R
M
ΔER
ΔEM ΔRM
N
Figure 1. Earthquake epicenter E- receiver R geometry. M is the observation point of the time-reversed field
as demonstrated by Capdeville et al. (2000). A simple way to get rid of these ghost trains is to send
back as many seismograms as possible.
14 J.-P. Montagner, Y. Capdeville, M. Fink, C. Larmat, H. Nguyen, B. Romanowicz
3 FROM A 3-POINT PROBLEM TO A MULTIPLE POINT-PROBLEM
According to equation (18) and (24), it is very easy to sum the contributions of many receivers (or
many sources). A global coverage of the earth by seismic stations or seismic sources will be considered
3.1 Time-reversal Imaging: 1 source, n receivers
Let us consider that the earth is covered by a global seismic network at points rRi, i = (1, nR). Due
to the fact that the Earth covered by 2 thirds of oceans and that most broadband seismic stations are
installed in northern hemishere, the spatial coverage is obviously uneven. How can we take account
of the non-uniform spatial coverage of the Earth? It is necessary to weight the contribution of each
station to the time-reversed field. The simplest choice consists in weighting each station by a surface
ai such that∑nR
i=1 ai = Searth, Searth being the surface of the Earth.
The time-reversed field at any point M at the surface of the Earth is then:
SM (tR) =∑i
aiS∆T (tR)i =∑i
ai∑k
∑j
1ωkωj
|JM〉〈JR|KR〉〈KE|FE =jk
If we assume that the spatial coverage by receivers is good enough, the sum∑
i ai can be replaced
by the integral∫earth dS. Since |KR〉 = nD`(r=0)Y m
` (θ, φ), we can integrate over the whole surface
of the Earth:nR∑i=1
ai〈JR|KR〉 =∫dS nD`(r=0)Y m
` (θ, φ)n′D`′(r=0)Y m′l′ (θ, φ)nD`(r=0)n′D`′(r=0)δ``′δmm′
The eigenfunctions of receiver stations are considered at the surface of the Earth r = 0.
Finally, for M = E:
SE(tR) =∑k
1ω2k
|KE〉〈KE|FEnD`2(0)
∆TcosωktR2ω2
k
This expression is maximum for tR = 0 since it is the only time where all cosine terms add
constructively. Alternative demonstration by using equation (24), where point E is at the pole:∫dSP 0
`′(cos∆RE)P 0` (cos∆ER) =
12`+ 1
δ``′
SE(tR) =∑k
1nω`2
nD`(rE)(nD`(rE).FE)nD`2(0)
12
∆Tcosnω`tR
At any point M at the surface of the Earth, the calculation done in section 2.1 is still valid, except
that the eigenfunctions in point rM must be introduced.
S(rM , tR) =∑k
|KM〉〈KE|F (rE)ω2k
D`n(0)D`
n(0)12
∆TcosωktR (25)
Time Reversal Method and Cross-Correlation techniques by normal mode theory: a 3-point problem 15
Barbara: it must be shown now that the sum of the products |K(rM)〉〈K(rE)|ω2k
is in average close to
zero except at point E. May be, it is obvious.
3.2 Cross-correlations: n sources, 2 receivers
- Random distribution of seismic noise sources
Due to time reversal invariance and spatial reciprocity, it can be imagined that instead of a distribu-
tion of stations in rR, we have a distribution of random sources. In equations of the forward problem,
the roles of point E and R are symmetric, we can consider noise sources at point rR, with a random
distribution. Due to the linearity of the problem it is possible to stack all contributions from these dif-
ferent sources, and the equation (23) is very similar to the Green function defined at the end of section
(1). The 2 Green functions GER(t) and GRM (t) are given by equation (10) and the corresponding
seismograms of displacement due to a force system F γE are:
uβR(t) = ΣγFγEG
γβER(t) =
∑K Σγ |Kβ
R〉〈KγE|F
γEH(t) sinωktωk
uαM (t) = ΣγFγEG
γαEM (t) =
∑J Σγ |JαM〉〈J
γE|F
γEH(t) sinωjtωj
or in terms of velocity for t > 0: uβR(t) = ΣγFγEG
γβER(t) =
∑K Σγ |Kβ
R〉〈KγE|F
γEH(t)sinωkt
uαM (t) = ΣγFγEG
γαEM (t) =
∑J Σγ |JαM〉〈J
γE|F
γEH(t)sinωjt
The cross-correlation CC of GER(t) ∗ GEM (t) involves the same integral =kj(t) as in equation (17)
CCβα(τ) =∫ t2
t1
uβ?R (t)uαM (t+ τ)dt =∑J
Σζ |JαM〉〈JζE|F
ζE
∑K
ΣγFγE |K
γE〉〈K
βR|=jk(τ)
Let us consider the simple case where only one component for F γE is excited (for example the
vertical component) and equal to 1, the sum over γ disappear:
CCβα(τ) = F γ2E
∑K
∑J
|JαM〉〈JγE|K
γE〉〈K
βR|=jk(τ) (26)
And for the dominant term, when J = K:
CCβα(τ) = F γ2E
∑K
|KαM〉〈K
γE|K
γE〉〈K
βR|=kk(τ) (27)
The cross-correlation CCβα(τ) is the Green function of velocity as defined in equation (11) if and
only if 〈KγE|K
γE〉 = 1. If sources are only located at the surface of the Earth, this condition is not
fulfilled.
The expression in equation (26) is very similar to equation (18). The role of point R is played by
point E, but we have now to consider a 2x2 tensor instead of a vector. We can apply twice the addition
16 J.-P. Montagner, Y. Capdeville, M. Fink, C. Larmat, H. Nguyen, B. Romanowicz
theorem and an expression similar to equation (24) is found implying P 0` cos∆ME and P 0
` cos∆RE .
By using the same arguments as in section 2.2.3, it can be demonstrated that the most important
contributions to CCβα(τ) are provided by sources located along the great-circle between R and M .
The same kind of calculation as in section 3.1 can be performed as well. If a homogeneous dis-
tribution of sources is assumed, with a density distribution n(θ, φ) associated with a surface an(θ, φ)
such that∫S n(θ, φ)sinθdθdφ = 1, and if only the case J = K is considered, then it is possible to
integrate:
CCβα(τ) = F γ2E
∑n,`,m
∑n′,`′,m′
〈(n, `,m)βR|(n′, `′,m′)αM〉(
∫Sn(θ, φ)sinθdθdφ〈(n′, `′,m′)γE(|n, `,m)γE〉)=jk(τ)
By using orthonormalization of eigenfunctions:
CCβα(τ) = F γ2E
∑n,`,m
∑n′,`′,m′
〈(n, `,m)βR|(n′, `′,m′)αM〉nD
γ` (rE)n′Dγ
`′(rE)=jk(τ)
and the addition theorem for J = K:
CCβα(τ) = F γ2E
∑n,`
nDβ` (rR)nDα
` (rM)P 0` (cos∆RM )nD
γ2` (rE)=kk(τ)
It is directly proportional to the Green function of velocity between points M and R. In isotropic
media if β 6= α CCβα(τ) = 0
These calculations demonstrate that time reversal (adjoint field) approach, and cross-correlations
of seismograms are directly related and that the same kind of calculations apply in both cases.
Time Reversal Method and Cross-Correlation techniques by normal mode theory: a 3-point problem 17
4 DISCUSSION AND APPLICATIONS
4.1 Voronoi cells
We develop in previous sections the whole theory of time-reversal method by using the normal mode
formalism. An important result is that the focusing at the epicenter is largely dependent on the weight-
ing of stations as defined in section 3.1. If the contribution of the different stations is not correctly
taken into account, the focus point will not be correctly located in time and space. This weighting is
particularly important when the station distribution is non-uniform, which is the case for the coverage
of Earth. Most broadband stations are located in continents and primarily in the northern hemisphere.
A simple way to take account of the uneven distribution is to use a Voronoi tessellation. This kind of
mapping at the surface of the Earth was implemented for example for global tomography by Debayle
& Sambridge (2004) and for time-reversal earthquake localization by Larmat et al. (2006). An exam-
ple of Voronoi cell distribution is shown in Figure 2 for a global distribution of 115 stations. As it can
be seen the surface distribution is very heterogeneous and some stations in the Pacific, Atlantic and
Indian oceans may have a very large distribution.
Figure 3 shows an example of focusing of time-reversed seismograms (an example of seismogram
is given on figure 3a) for an isotropic source located close to Sumatra obtained by using the 115
stations of Figure 2. Figure 3c shows a perfect refocusing when using the Voronoi cell tessellation
and Figure 3e when no weighting is applied. In this case the localization in space is very poor and the
radiation pattern (a circle for an isotropic source) is deformed.
4.2 One-bit discretization
In most calculations, we have used arguments based on the stationnary phase approximation. Figure
3b shows binarized seismograms which suppress the information on the amplitude of seismograms.
Binarization of seismograms is routinely used when calculating Green functions by cross-correlation
between 2 stations (Shapiro et al. (2004)). When binarized seismograms are time-reversed with the
appropriate weighting scheme, the focusing is almost as good as the focusing obtained by time-reversal
of complete seismograms. This experiment demonstrates that the information is primarily conveyed
by the phase of seismograms. It is not really a surprise, since the location of the earthquake and the
focusing time correspond to a constructive interference of the whole parts of seismograms.
18 J.-P. Montagner, Y. Capdeville, M. Fink, C. Larmat, H. Nguyen, B. Romanowicz
Figure 2. Voronoi cells used for time-reversal of Sumatra-Andaman earthquake for a global network of 115
stations. Each cell is associated with a station. The area of each cell reflects the global coverage of the Earth.
4.3 Inversion of the force system
According to equation (25), it is possible to obtain a linear system of equations which enables to
retrieve the vector force FFE. We only remind the results. The Green function components between
points rE and rR are equal to
Gβγ(rE, rR, t) =∑
K |KβR〉〈K
γE|
H(t)sinωktωk
and uβR(ω) =∑
K
∑γ |K
βR〉〈K
γE|
F γEωk=jk(ω)
At point source rE, by reordering the different terms, and taking J = K:
SαE(tR) =∑K
∑βγ
|KαE〉〈KR
β|KβR〉 〈KE
γ |F γE=kk(tR)ω2k
If F γE is a scalar force, 〈KEγ |F γE |JαE〉 = F γEu
γK(rE)uαJ (rE), and for an uniform distribution of stations
Ri at the surface of the Earth, weighted by using for example a Voronoi tesselation, which means that
the contribution of each station is associated with an area ai,∑i ai〈JR
β|KβR〉 ≈
∑β n′Dβ∗
`′ (0)nDβ` (0)
∫S Y
m′∗`′ Y m
` sinθdθdφ =∑
β n′Dβ∗`′ (0)nD
β` (0)δ``′δmm′
Time Reversal Method and Cross-Correlation techniques by normal mode theory: a 3-point problem 19
[a]
0 2000 4000 6000 8000 10000Time (s)
-4e+14
-2e+14
0
2e+14
4e+14
Am
plitu
de (
m)
[b]
[c] [d]
[e: wrong weighting Huong ]
Figure 3. One-bit normalization versus Amplitude time-reversal experiment. a: real seismogram. b: one-bit
seismogram. c: real seismograms and weighting by Voronoi tessellation. d:one-bit seismograms and weighting
by Voronoi tessellation. e: real seismograms and no weighting
20 J.-P. Montagner, Y. Capdeville, M. Fink, C. Larmat, H. Nguyen, B. Romanowicz
a) Strike-slip fault b) Inverse fault c) Normal fault
Figure 4. Sensitivity of the time-reversed field to the focal mechanism
assuming J = K
SαE(tR) =∑γβ
F γE
∑n`
nDβ2` (0)nD
γ` (rE)nDα
` (rE)||Y m` (θE , φE)||2=kk(tR)
ω2k
(28)
All oscillations are in phase for tR = 0, which explains why we get focusing of energy at that
time.
BARBARA: by using arguments of stationary phase, SαE ,∑
n` ....
It is then obtained a matrix relationship S(tR=0) = QF and it is possible to retrieve the force
system by inversion of the matrix Q. This linear system with respect to the different components FE ,
can be easily inverted. The coefficients of the matrix Q can be calculated by normal mode summation.
Time Reversal Method and Cross-Correlation techniques by normal mode theory: a 3-point problem 21
4.4 Time reversal of deformation- Moment tensor inversion
Similarly, it is possible to calculate the spatial derivatives of the Green function at the source point,
such that:
Gβγ,m(rE, rR, t) =∑K
|KβR〉ε
K?γm(rE)H(t)(
1− cosωktω2k
)
where the deformation tensor components are given by εγm(rE) = 12(∂|K
γE〉
∂ξm+ ∂|Km
E 〉∂ξγ
).
In terms of operators, the operator deformation is E = 12(∇+∇T ).
And uβR(ω) =∑
K
∑γ |K
βR〉
12(∂〈K
γE|
∂ξm+ ∂〈Km
E |∂ξγ
)Mγm
Similarly,Gαβ(rM, rR, ω) =∑
K |KαM〉〈K
βR|=(ω), andGαβ,m(rM, rR, ω) =
∑K |Kα
M〉εK?βm=′(ω)
If the spatial derivatives of the Green functions are time-reversed, an homogeneous expression with a
2nd-order tensor is obtained, which will give access the seismic moment tensor (Kawakatsu & Mon-
tagner (2008))
Ξ(rM , t) = M(−t) ∗G′ER(−t) ∗G′RM (t)
where G′ER and G′RM are the deformations associated with Green functions GER and GRM . In
terms of cartesian coordinates:
ΞαmM (ω) =∑J
∑K
∑βγ
|JαM〉iω(−ω2 + ω2
j )〈JβR|
|KβR〉
(−ω2 + ω2k)
12
(∂〈Kγ
E|∂ξm
+∂〈Km
E |∂ξγ
)Mγm
Actually in order to ensure the symmetry of the calculation (Kawakatsu & Montagner (2008)) such
that ΞαmM = ΞmαM , we have to consider not only Gβγ,m but also Gmβ,γ , since the contribution of Mmγ
must be the same as the contribution of Mγm
Gmβ,α(rM, rR, t) =∑K
εK?βα |KmM〉=′(ω)
and
ΞmαM (ω) =∑J
∑K
∑βγ
|JmM〉
iω(−ω2 + ω2j )〈JβR|
|KβR〉
(−ω2 + ω2k)
12
(∂〈Kγ
E|∂ξα
+∂〈Kα
E|∂ξγ
)Mγα
In the time domain:
ΞαmM tR) =∑J
∑K
∑βγ
|JαM〉〈JβR|K
βR〉
12
(∂〈Kγ
E|∂ξm
+∂〈Km
E |∂ξγ
)Mγm=′jk(tR)
The same theorems (addition theorem, orthonormalization of eigenfunctions) as for retrieving the
22 J.-P. Montagner, Y. Capdeville, M. Fink, C. Larmat, H. Nguyen, B. Romanowicz
force system can be applied:
ΞαmM tR) =∑n`
nD`α(rE)P 0
` (cos∆ER)||nD`(0)||2P 0` (cos∆ER)nD`
γ(rE)EγmMγm
As in equation (28) for the force system, it is found a linear system with respect to the moment
tensor components. Appendix B details how to retrieve the 6 different components of and the practical
application is beyond the scope of this paper.
5 CONCLUSIONS
ACKNOWLEDGMENTS
A number of colleagues have helped with suggestions for the improvement of this material and I
would particularly like to thank Hitoshi Kawakatsu, Earthquake Research Insititute of Tokyo for his
criticisms and corrections.
REFERENCES
Capdeville, Y., 2005. An efficient born normal mode to compute sensitivity kernels and synthetic seismograms
in the earth., Geophys. J. Int., 163, 1–8.
Capdeville, Y., Stutzmann, E., & Montagner, J.-P., 2000. Effects of a plume on long period surface waves
computed with normal modes coupling., Phys. Earth. Planet. Inter., 119, 57–74.
Capdeville, Y., Larmat, C., Vilotte, J.-P., & Montagner, J.-P., 2002. A new coupled spectral element and modal
solution method for global seismology: A first application to the scattering induced by a plume–like anomaly.,
Geophys. Res. Lett., 29(9), 32.1,32.4.
Debayle, E. & Sambridge, M., 2004. Inversion of massive surface wave datasets: Model construction and
resolution assessment, J. Geophys. Res., 109, B02316.
Derode, A., Larose, E., Tanter, M., Campillo, M., & Fink, M., 2003. Recovering the Green’s function from
field–field correlations in an open scattering medium (l), J. A. S. A., , 113(6), 2973–2976.
Draeger, C. & Fink, M., 1997. One–channel Time Reversal of Elastic Waves in a 2D–Silicon Cavity, Phys.
Rev. Lett., 79(3), 407–410.
Draeger, C. & Fink, M., 1999. One–channel Time Reversal in chaotic cavities: Theoretical limits, J. A. S. A.,
, 105(2), 611–617.
Edmonds, A., 1960. Angular Momentum and Quantum Mechanics, Princeton University Press, Princeton, NJ.
Fukao, Y., Nishida, K., Suda, N., Nawa, K., & Kobayashi, N., 2002. A theory of earth’s background free
oscillations, Journal of Geophysical Research, 107, B9, 2206, 2206.
Time Reversal Method and Cross-Correlation techniques by normal mode theory: a 3-point problem 23
Gajewski, D. & Tessmer, E., 2005. Reverse modelling for seismic event characterization, Geophys. J. Int.,
163, 276–284.
Gilbert, F., 1971. Excitation of the normal modes of the earth by earthquakes sources, Geophys. J. R. Astr.
Soc., 22, 223–226.
Kawakatsu, H. & Montagner, J.-P., 2008. Time reversal seismic source imaging and moment-tensor inversion,
Geophys. J. Int, 175, 686–688.
Larmat, C., Montagner, J.-P., M., Capdeville, Y., Tourin, A., & Clevede, E., 2006. Time–Reversal Imaging of
seismic sources and application to the Great Sumatra Earthquake, Geophys. Res. Lett., p. L19312.
Larmat, C., Tromp, J., Liu, Q., & Montagner, J.-P., 2008. Time reversal location of glacial earthquakes, J.
Geophys. Res., p. B09314.
Li, X. B. & Romanowicz, B., 1996. Global Mantle Shear-Velocity Model Developed Using Nonlinear Asymp-
totic Coupling Theory, J. Geophys. Res., 101(B10), 22,245–22,272.
Romanowicz, B., 1987. Multiplet–multiplet coupling due to lateral heterogeneity: asymptotic effects on the
amplitude and frequency of the Earth’s normal modes., Geophys. J. R. Astr. Soc., 90, 75–100.
Shapiro, N. M., Campillo, M., Stehly, L., & Ritzwoller, M., 2004. High-resolution Surface–wave tomography
from ambient seismic noise, Science, 307, 1615–1618.
Suda, N., Nawa, K., & Fukao, Y., 1998. Earth’s backgrouund free oscillations, Science, 279, 2089–2091.
Tromp, J., Tape, C., & Liu, Q., 2005. Seismic tomography, adjoint methods, time reversal and banana–
doughnyt kernels, Geophys. J. Int., 160, 195–216.
Weaver, R. & Lobkis, O., 2002. On the emergence of the green’s function in the correlations of a diffuse field:
pulse echo using thermal photon, Ultrasonics, 40, 435–439.
Woodhouse, J. & Girnius, T., 1982. Surface waves and free oscillations in a regionalized earth model, Geophys.
J. R. Astron. Soc., 68, 653–673.
24 J.-P. Montagner, Y. Capdeville, M. Fink, C. Larmat, H. Nguyen, B. Romanowicz
Table A1. The coefficients RkN and SkN
N RkN = RNk (0, 0) SkN = SNk (0, 0)
0 k0vr −k0(∂rUMrr + 12F (Mθθ +Mφφ))
±1 k1(V ± iW )(∓vθ − ivφ) k1(X ∓ iZ)(Mrθ − iMrφ)
±2 k2(V ∓ iW )((Mθθ −Mφφ)∓ iMθφ)
F = r−1E (2U − `(`+ 1)V ) kn = 1
2n
(2`+14π
(`+n)!(`−n)!
)1/2
X = ∂rV + r−1E (U − V )
Z = ∂rW + r−1E W
Mθθ,Mφφ... are components of the moment tensor, vr, vθ, vφ are components of the instrument vector.
APPENDIX A: TIME REVERSAL WITH A HEAVISIDE FUNCTION: CALCULATION OF
=′(TR)
For a source Heaviside function H(t):
uR(ω) =∑k
|kR〉iω(−ω2 + ω2
k)〈kE|FE (A.1)
SM (ω) =∑
j
∑k
|jM〉(−ω2+ω2
j )〈jR| |kR〉
iω(−ω2+ω2k)〈kE|FE for a Heaviside function in E. Sums over j
and k can be separated:
SR(ω) =∑j
|jM〉〈jR|(−ω2 + ω2
k)︸ ︷︷ ︸BRM (ω)
∑k
|kR〉〈kE|FE
iω(−ω2 + ω2k)︸ ︷︷ ︸
CER(ω)
(A.2)
For a Heaviside function, the function CER is
CER(t) =∑k
|kR〉〈kE|1− cosωkt
ω2k
H(t)
SM (ω) = BRM (t) ∗CER(−t) =∑j
∑k
|jM〉〈jR|kR〉〈kE|FEω2k
1ωj=′jk(tR) (A.3)
The integral=′jk(tR) =∫ t2t1dτ−cosωk(tR+τ)sinωjτ = 1
2sinωktR∫ t2t1dτ(cos(ωj+ωk)τ−cos(ωj−
ωk)τ)+ 12cosωktR
∫ t2t1dτ(sin(ωj +ωk)τ−sin(ωj−ωk)τ). The terms with (ωj +ωk)τ are negligible
since 1/(ωj + ωk)� ∆T
(A VERIFIER)
The integral=jk(tR) = −∫ t2t1dτsinωk(tR+τ)sinωjτ ≈ 1
2∆TsinωktRδjk. It is a similar integral
as previously calculated in the scalar case.
Time Reversal Method and Cross-Correlation techniques by normal mode theory: a 3-point problem 25
APPENDIX B: INVERSION OF SEISMIC MOMENT TENSOR
Coordinates x (resp y, z) corresponds to 1 (resp 2, 3). For order 2 tensors, such as moment tensor Mij ,
the classical transformation of coordinates is used 12 → 6, 13 → 5, 23 → 4. So the 6 independent
components of Mij can come up explicitly. In case a force system, the order 2 Green tensor is square.
It is no longer the case when considering the deformation tensor. Following Kawakatsu & Montagner
(2008), we can write:
u1 = G11,1M1 +G1
2,2M2 +G13,3M3+ (G1
2,3 +G13,2)M4 + (G1
1,3 +G13,1)M5 + (G1
1,2 +G12,1)M6
= G11M1 +G1
2M2 +G13M3+ G1
4M4 +G15M5 +G1
6M6
u2 = G21,1M1 +G2
2,2M2 +G23,3M3+ (G2
2,3 +G23,2)M4 + (G2
1,3 +G23,1)M5 + (G2
1,2 +G22,1)M6
= G21M1 +G2
2M2 +G23M3+ G2
4M4 +G25M5 +G2
6M6
u3 = G31,1M1 +G3
2,2M2 +G33,3M3+ (G3
2,3 +G33,2)M4 + (G3
1,3 +G33,1)M5 + (G3
1,2 +G32,1)M6
= G31M1 +G3
2M2 +G33M3+ G3
4M4 +G35M5 +G3
6M6
Conversely,
M1 = G11u1 + G2
1u2 + G31u3
G11,1 u1 + G2
1,1 u2 + G31,1 u3
M2 = G122u1 + G2
22u2 + G322u3
G12,2 u1 + G2
2,2 u2 + G32,2 u3
M3 = G13u1 + G2
3u2 + G33u3
G13,3 u1 + G2
3,3 u2 + G33,3 u3
M4 = G14u1 + G2
4u2 + G34u3
(G12,3 + G1
3,2)u1 + (G22,3 + G2
3,2)u2 + (G32,3 + G3
3,2)u3
M5 = G15u1 + G2
5u2 + G35u3
(G11,3 + G1
3,1)u1 + (G21,3 + G2
3,1)u2 + (G31,3 + G3
3,1)u3
M6 = G16u1 + G6u2 + G3
6u3
(G11,2 + G1
2,1)u1 + (G21,2 + G2
2,1)u2 + (G31,2 + G3
2,1)u3
So now, it is easy to replace the expressions of the moment tensor as defined in text.
This paper has been produced using the Blackwell Scientific Publications GJI LATEX2e class file.