/[escript]/release/3.4.2/doc/inversion/Regularization.tex
ViewVC logotype

Annotation of /release/3.4.2/doc/inversion/Regularization.tex

Parent Directory Parent Directory | Revision Log Revision Log


Revision 4919 - (hide annotations)
Wed Apr 30 06:25:55 2014 UTC (5 years, 3 months ago) by jfenwick
File MIME type: application/x-tex
File size: 8637 byte(s)
Because we've got to!

1 gross 4074 \chapter{Regularization}\label{Chp:ref:regularization}
2    
3 gross 4099 The general cost function $J^{total}$ to be minimized has some of the cost
4     function $J^f$ measuring the defect of the result from the
5     forward model with the data, and the cost function $J^{reg}$ introducing the
6 caltinay 4095 regularization into the problem and makes sure that a unique answer exists.
7     The regularization term is a function of, possibly vector-valued, level set
8     function $m$ which represents the physical properties to be represented and is,
9     from a mathematical point of view, the unknown of the inversion problem.
10     It is the intention that the values of $m$ are between zero and one and that
11     actual physical values are created from a mapping before being fed into a
12 gross 4099 forward model. In general the cost function $J^{reg}$ is defined as
13 gross 4074 \begin{equation}\label{EQU:REG:1}
14 gross 4099 J^{reg}(m) = \frac{1}{2} \int_{\Omega} \left(
15     \sum_{k} \mu_k \cdot ( \omega^{(0)}_k \cdot m_k^2 + \omega^{(1)}_{ki}m_{k,i}^2 )
16     + \sum_{l<k} \mu^{(c)}_{lk} \cdot \omega^{(c)}_{lk} \cdot \chi(m_l,m_k) \right) \; dx
17 gross 4074 \end{equation}
18 gross 4099 where summation over $i$ is performed. The additional trade--off factors
19     $\mu_k$ and $\mu^{(c)}_{lk}$ ($l<k$) are between zero and one and constant across the
20     domain. They are potentially modified during the inversion in order to improve the balance between the different terms
21     in the cost function.
22    
23     $\chi$ is a given symmetric, non-negative cross-gradient function\index{cross-gradient
24     }. We use
25 gross 4074 \begin{equation}\label{EQU:REG:4}
26 gross 4099 \chi(a,b) = ( a_{,i} a_{,i}) \cdot ( b_{,j} b_{,j}) - ( a_{,i} b_{,i})^2
27 gross 4074 \end{equation}
28 gross 4268 where summations over $i$ and $j$ are performed, see~\cite{GALLARDO2005a}. Notice that cross-gradient function
29 gross 4099 is measuring the angle between the surface normals of contours of level set functions. So
30     minimizing the cost function will align the surface normals of the contours.
31    
32     The coefficients $\omega^{(0)}_k$, $\omega^{(1)}_{ki}$ and $\omega^{(c)}_{lk}$ define weighting factors which
33     may depend on their location within the domain. We assume that for given level set function $k$ the
34     weighting factors $\omega^{(0)}_k$, $\omega^{(1)}_{ki}$ are scaled such that
35     \begin{equation}\label{ref:EQU:REG:5}
36 gross 4122 \int_{\Omega} ( \omega^{(0)}_k + \frac{\omega^{(1)}_{ki}}{L_i^2} ) \; dx = \alpha_k
37 gross 4074 \end{equation}
38 gross 4099 where $\alpha_k$ defines the scale which is typically set to one. $L_i$ is the width of the domain in $x_i$ direction.
39     Similarly we set for $l<k$ we set
40     \begin{equation}\label{ref:EQU:REG:6}
41 gross 4122 \int_{\Omega} \frac{\omega^{(c)}_{lk}}{L^4} \; dx = \alpha^{(c)}_{lk}
42 gross 4074 \end{equation}
43 gross 4099 where $\alpha^{(c)}_{lk}$ defines the scale which is typically set to one and
44     \begin{equation}\label{ref:EQU:REG:6b}
45     \frac{1}{L^2} = \sum_i \frac{1}{L_i^2} \;.
46 gross 4074 \end{equation}
47    
48 gross 4099 In some cases values for the level set functions are known to be zero at certain regions in the domain. Typically this is the region
49     above the surface of the Earths. This expressed using a
50     a characteristic function $q$ which varies with its location within the domain. The function $q$ is set to zero except for those
51     locations $x$ within the domain where the values of the level set functions is known to be zero. For these locations $x$
52     $q$ takes a positive value. for a single level set function one has
53     \begin{equation}\label{ref:EQU:REG:7}
54     q(x) = \left\{
55     \begin{array}{rl}
56     1 & \mbox{ if } m \mbox{ is set to zero at location } x \\
57     0 & \mbox{ otherwise }
58     \end{array}
59     \right.
60 gross 4074 \end{equation}
61 gross 4099 For multi-valued level set function the characteristic function is set componentwise:
62     \begin{equation}\label{ref:EQU:REG:7b}
63     q_k(x) = \left\{
64     \begin{array}{rl}
65     1 & \mbox{ if component } m_k \mbox{ is set to zero at location } x \\
66     0 & \mbox{ otherwise }
67     \end{array}
68     \right.
69 gross 4074 \end{equation}
70    
71    
72 gross 4099 \section{Usage}
73 gross 4074
74 gross 4099 \begin{classdesc}{Regularization}{domain
75     \optional{, w0=\None}
76     \optional{, w1=\None}
77     \optional{, wc=\None}
78     \optional{, location_of_set_m=Data()}
79     \optional{, numLevelSets=1}
80     \optional{, useDiagonalHessianApproximation=\False}
81     \optional{, tol=1e-8}
82     \optional{, scale=\None}
83     \optional{, scale_c=\None}
84     }
85 gross 4074
86 gross 4099
87     initializes a regularization component of the cost function for inversion.
88     \member{domain} defines the domain of the inversion. \member{numLevelSets}
89     sets the number of level set functions to be found during the inversion.
90     \member{w0}, \member{w1} and \member{wc} define the weighting factors
91     $\omega^{(0)}$,
92     $\omega^{(1)}$ and
93     $\omega^{(c)}$, respectively. A value for \member{w0} or \member{w1} or both must be given.
94 caltinay 4285 If more than one level set function is involved \member{wc} must be given.
95 gross 4099 \member{location_of_set_m} sets the characteristic function $q$
96     to define locations where the level set function is set to zero, see equation~(\ref{ref:EQU:REG:7}).
97     \member{scale} and
98     \member{scale_c} set the scales $\alpha_k$ in equation~(\ref{ref:EQU:REG:5}) and
99     $\alpha^{(c)}_{lk}$ in equation~(\ref{ref:EQU:REG:6}), respectively. By default, their values are set to one.
100     Notice that weighting factors are rescaled to meet the scaling conditions. \member{tol} sets the
101     tolerance for the calculation of the Hessian approximation. \member{useDiagonalHessianApproximation}
102     indicates to ignore coupling in the Hessian approximation produced by the
103     cross-gradient term. This can speed-up an individual iteration step in the inversion but typically leads to more
104     inversion steps.
105     \end{classdesc}
106    
107     \section{Gradient Calculation}
108    
109    
110     The cost function kernel\index{cost function!kernel} is given as
111     \begin{equation}\label{ref:EQU:REG:100}
112     K^{reg}(m) = \frac{1}{2}
113     \sum_{k} \mu_k \cdot ( \omega^{(0)}_k \cdot m_k^2 + \omega^{(1)}_{ki}m_{k,i}^2 )
114     + \sum_{l<k} \mu^{(c)}_{lk} \cdot \omega^{(c)}_{lk} \cdot \chi(m_l,m_k)
115 gross 4074 \end{equation}
116 gross 4099 We need to provide the gradient of the cost function $J^{reg}$ with respect to the level set functions $m$.
117     The gradient is represented by two functions $Y$ and $X$ which define the
118     derivative of the cost function kernel with respect to $m$ and to the gradient $m_{,i}$, respectively:
119     \begin{equation}\label{ref:EQU:REG:101}
120     \begin{array}{rcl}
121     Y_k & = & \displaystyle{\frac{\partial K^{reg}}{\partial m_k}} \\
122 gross 4122 X_{ki} & = & \displaystyle{\frac{\partial K^{reg}}{\partial m_{k,i}}}
123 gross 4074 \end{array}
124     \end{equation}
125 gross 4122 For the case of a single valued level set function $m$ we get
126     \begin{equation}\label{ref:EQU:REG:202}
127     Y = \mu \cdot \omega^{(0)} \cdot m
128 gross 4074 \end{equation}
129 gross 4122 and
130     \begin{equation}\label{ref:EQU:REG:203}
131     X_{i} = \mu \cdot \omega^{(1)}_{i} \cdot m_{,i}
132     \end{equation}
133     For a two-valued level set function $(m_0,m_1)$ we have
134     \begin{equation}\label{ref:EQU:REG:302}
135     Y_k = \mu_k \cdot \omega^{(0)}_k \cdot m_k \mbox{ for } k=0,1
136     \end{equation}
137     and for $X$
138     \begin{equation}\label{ref:EQU:REG:303}
139     \begin{array}{rcl}
140     X_{0i} & = & \mu_0 \cdot \omega^{(1)}_{0i} \cdot m_{0,i} + \mu^{(c)}_{01} \cdot \omega^{(c)}_{01} \cdot
141     \left( (m_{1,j}m_{1,j} ) \cdot m_{0,i} - (m_{1,j}m_{0,j} ) \cdot m_{1,i} \right) \\
142     X_{1i} & = & \mu_1 \cdot \omega^{(1)}_{1i} \cdot m_{1,i} + \mu^{(c)}_{01} \cdot \omega^{(c)}_{01} \cdot
143     \left( (m_{0,j}m_{0,j} ) \cdot m_{1,i} - (m_{1,j}m_{0,j} ) \cdot m_{0,i} \right)
144     \\
145     \end{array}
146 gross 4099 \end{equation}
147 gross 4142 We also need to provide an approximation of the inverse of the Hessian operator as discussed in section~\ref{chapter:ref:inversion cost function:gradient}.
148 gross 4122 For the case of a single valued level set function $m$ we get
149     \begin{equation}\label{ref:EQU:REG:601}
150 gross 4074 \begin{array}{rcl}
151 gross 4099 A_{ij} & =& \mu \cdot \omega^{(1)}_i \cdot \delta_{ij} \\
152     D & = & \mu \cdot \omega^{(0)}
153 gross 4074 \end{array}
154 gross 4122 \end{equation}
155     For a two-valued level set function $(m_0,m_1)$ we have
156     \begin{equation}\label{ref:EQU:REG:602}
157     D_{kl} = \mu_k \cdot \omega^{(0)}_k \cdot \delta_{kl}
158 gross 4074 \end{equation}
159 gross 4122 and
160     \begin{equation}\label{ref:EQU:REG:603}
161 gross 4074 \begin{array}{rcl}
162 gross 4122 A_{0i0j} & = & \mu_0 \cdot \omega^{(1)}_{0i} \cdot \delta_{ij} + \mu^{(c)}_{01} \cdot \omega^{(c)}_{01} \cdot
163     \left( (m_{1,j'}m_{1,j'} )\cdot \delta_{ij} - m_{1,i} \cdot m_{1,j} \right) \\
164     A_{0i1j} & = & \mu^{(c)}_{01} \cdot \omega^{(c)}_{01} \cdot \left( 2 \cdot m_{0,i} \cdot m_{1,j}
165     - m_{1,i} \cdot m_{0,j} - ( m_{1,j'} m_{0,j'} ) \cdot \delta_{ij}
166     \right) \\
167     A_{1i0j} & = & \mu^{(c)}_{01} \cdot \omega^{(c)}_{01} \cdot \left( 2 \cdot m_{1,i} \cdot m_{0,j}
168     - m_{0,i} \cdot m_{1,j} - ( m_{1,j'} m_{0,j'} ) \cdot \delta_{ij} \right) \\
169     A_{1i1j} & = & \mu_1 \cdot \omega^{(1)}_{1i} \cdot \delta_{ij} + \mu^{(c)}_{01} \cdot \omega^{(c)}_{01} \cdot
170     \left( (m_{0,j'}m_{0,j'} ) \cdot \delta_{ij} - m_{0,i} \cdot m_{0,j} ) \right)
171 gross 4074 \end{array}
172     \end{equation}
173    
174 gross 4122

  ViewVC Help
Powered by ViewVC 1.1.26