Evaluating the utility of gravity gradient tensor components
Mark Pilkington
Geological Survey of Canada
Tensor component choice
Txx Txy Txz
Tyz
Tyy
Tzz
Single components
Combinations
Concatenations
Which to use?
Qualitative interpretation
Quantitative interpretation
Tensor component choice
Quantitative interpretation [Inversions]
(Txx, Txy, Txz, Tyy, Tyz) Li, 2001(Tuv, Txy), Tzz Zhdanov et al., 2004(Txz, Tyz, Tzz, Tuv) Droujinine et al., 2007(Tuv, Txy) Li, 2010(Tuv, Txy), Tzz, (Tzz, Tuv, Txy) Martinez & Li, 2011Tzz, (Txz, Tyz, Tzz), (Txz, Tyz, Txz, Tyy, Txx) Martinez et al., 2013
Rating the solutions:goodness of fitsharp/smoothclose to geology
Inversion versus component combinations
Martinez et al., 2013
Tzz
Txz, Tyz, Tzz
Txz, Tyz, Txz, Tyy, Txx
Txz, Tyz, Txz, Tzz, Tyy, Txx
Components inverted:
RMS error Txx Txy Txz Tyy Tyz Tzz1-C 23.9 23.2 31.8 23.1 26.1 16.53-C 17.5 16.0 15.9 16.0 12.4 22.55-C 16.6 12.6 16.3 15.8 12.2 24.36-C 15.7 13.0 17.9 13.8 13.8 21.4
Outline
Aim: quantitative rating of component/combinations
Approach: inversion using a simple model – estimate parameter errors
Method: linear inverse theory – analyse model/data relations
Inversion method used
Inversion Parametric[underdetermined inversionproblem]
n datam parameters m >> n m << n
Model 3-D volume Specified shapequantity
Solution Physical property Parameters (density …) (depth, dip…)
Methodology Regularized inversion Overdeterminedleast – squares
Solution Resolution, covariance Parameter errorsappraisal
Prism model
z
t
bw
xcyc
Inverse theory
Forward problem: b = f (x) b = data
x =
parameters
(linearized) db = Adx A = Jacobian
[model dependent]
aij = dbi/dxj
Inverse problem : dx = A+db
A = UVT singular
value
decomposition
Inverse theory
A = UVT singular value decomposition
U = data eigenvectors
V = parameter eigenvectors
= singular values
R = VVT Resolution matrix (=I)
S = UUT Data information matrix
C = CdV-2VT Covariance matrix
Model parameter errors
C = CdV-2VT Parameter covariance matrix
Cd = Data covariance
=singular values
small large C large small C
Cd = e2I Equal data errorCd = D Variable data error
Variable component errors
Components have different error levels: e.g., e(Txx) = e(Txz) only relative levels requiredestimate based on FFT or equivalent source methodratio Tzz : Txz, Tyz : Txy : Txx, Tyy = 1 : 0.70 : 0.37 : 0.59
Component quantities are combined: e.g., H1 = sqrt(Txz2+Tyz2) combine errors: e(Tuv) = [0.5 (e(Txx)2+e(Tyy)2)]1/2
Component quantities tested
Single components:
Txx Tyy Tzz Txy Tyz Txz Tuv
Invariants:
I1 = TxxTyy+TyyTzz+TxxTzz-Txy2-Tyz2-Txz2
I2 = Txx(TyyTzz-Tyz2)+Txy(TyzTxz-TxyTzz)+Txz(TxyTyz-TxzTyy)H1 = sqrt(Txz2+Tyz2) H2 = sqrt[Txy2+0.25(Tyy-Txx)2]
Concatenations:
(Tuv, Txy)(Txz, Tyz, Tzz)
(Txy, Tyz, Txz)
(Txx, Tyy, Txy) (Txz, Tyz, Txz, Txy, Txx)(Tyy, Tyz, Txz, Txy, Txx)
Inversion tests
Procedure:
•Specify model and evaluate matrix A [db=Adx]•Calculate covariance matrix C•Get parameter standard deviations (p.s.d.)•Rank p.s.d. for each parameter versus component quantity
Models tested:
xc yc z t w b
32 32 4 1,3,6,13,43 12 12 0.2
32 32 4 13 0.1,0.5,2,6,9 12 0.2
32 32 0.1,1,3,6,12 40 12 12 0.2
32 32 2 1 1 1 0.2
32 32 2 4 4 4 0.2
32 32 2 8 8 8 0.2
32 32 0.5,1,2 4 1 1 0.2
32 32 0.5,1,2,4 1 8 8 0.2
32 32 0.5,1,2,4 2 2 2 0.2
Eigenvector matrix V
Eigenvector matrix V
Invariants:
I1 = TxxTyy+TyyTzz+TxxTzz-Txy2-Tyz2-Txz2
I2 = Txx(TyyTzz-Tyz2)+Txy(TyzTxz-TxyTzz) +Txz(TxyTyz-TxzTyy)
H1 = sqrt(Txz2+Tyz2)
H2 = sqrt[Txy2+0.25(Tyy-Txx)2]
Eigenvector matrix V
Correlation matrix
corrij = covij
[ covii covjj ]1/2
Parameter errors
xc,yc = locationz = deptht = thicknessw = widthb = breadth = density
Parameter errors
xc,yc = locationz = deptht = thicknessw = widthb = breadth = density
Parameter errors
xc,yc = locationz = deptht = thicknessw = widthb = breadth = density
Parameter error ranking [29 models]
error
high
low
Parameter errors versus averaging
No averagingcorrection
With averaging correction
Conclusions
Concatenated components produce smallest parameter errors
Invariants I1, I2 best performers in combined component category
Purely horizontal components poor performers
Tzz best single component
Parameter rankings
I1Txz
higher error higher error
Width error versus coordinate rotation
coordinateaxis
bodyaxis
Information density matrix
Information density versus eigenvector