GP 506-Inverse Theory

Error Analysis
Inverse Problem => Gm=d B)
A)
15 15
Þ mest = G-gd
Þ dpre = Gmest
diobs
10 10
ei
If,dpre = dobs; Exact dipre

d
predication 5 5
If, dpre ≠ dobs ; lead to error!

Definition of error: 0
0 5 10
0
0 5 zi 10
z z

E = dobs – dpre;
 mtrue Quantitative Model
dpre

due to
≠ ≠
due to error
observational
propagation
error

mest Quantitative Model

dobs
 For the simple (two parameter) case of a line fit, N = M means two points and the answer is
simple enough to avoid matrix algebra:

y1 = mx1 + b m = (y2-y1)/(x2-x1)
(x2,y2)
y2 = mx2 + b b = y1 - mx1
(x1,y1)

But what should we do for the overdetermined problem, N > M?

Here, in the presence of errors e, no single line will pass through all points so
we want to minimize the effects of error in the model. An arbitrary choice of
model m predicts data:
M

d pred = Gm d ipred = ∑G m ij j
j=1

We can define a residual vector e as

M

e = d − €d pred = d − Gm ei = d i − ∑G m ij j

€ j=1
 15
Recall that the measurement error is
L1
ε = d − Gm t 10 L2
so would like to find the m vector that minimizes e in some sense,

d
on the assumption that that yields e » e.
5
L∞
The “sense” in which we minimize e is in terms of some
measure€of “length” of the residual vector, called a “norm”:
outlier
0
N 0 2 4 6 8 10

L1 norm: e 1= ∑e z
i
Straight line problem solved under three different
norms. This is a worst-case example, since the data
i=1
N
have a horrible outlier (be sure to point it out).
L2 norm: e = ∑e 2
i = eTe
2
i=1
€
The definition of norm can be generalized to any power n : L∞ norm:

N
€
Ln norm: e =n ∑e n
i
n
i=1 but in practice we generally only use L1 & L2…
 Most commonly we use L2, for two reasons:

(1) It most accurately predicts m » mt if errors e are zero-mean & Gaussian (or normally distributed);
(2) It turns out to be the most computationally simple.

When we minimize L2, we call it the method of “least squares”.

Note that outlier measurements are emphasized more in model fits when n is larger in the Ln norm being
minimized.

For L2 norm minimization (ordinary least squares), we seek to minimize

⎛
N M ⎞2
∑
E = e T e = (d − Gm ) T (d − Gm ) = ⎜ d i −
⎜ ∑ Gij m j ⎟
⎟
i=1 ⎝ j=1 ⎠

€
• Impossible to match data exactly
• In theory, possible to exactly
resolve all model parameters for a
model that minimizes misfit to
error

• Can always find model to match

data exactly, but many models are
possible (m has no unique solution)

The real world:

• Impossible to satisfy data exactly
• Some combinations of model
parameters are not independently
sampled and cannot be resolved
We will back later to this!