that determine λ. When there is no noise in measurements, i.e., Axr = b, the
regularization coefficient λ should be set infinity. As measurement noise increases,
we should relax t2^norm constraint or equivalently decrease λ. In addition, sparse
vectors imply small λ. When it is known that vector x is strongly sparse, one
should relax t⅛-norm constraint (decrease λ) to obtain a very sparse solution for
the problem.
Figure 4.5 shows estimation error for different regularization coefficients, λ.
As explained, for very small λ and very large λ estimation error is high. There is
an optimal regularization coefficient λopt in which the variation estimation error is
minimum. Optimizing Equation 4.13 for λ = λopt leads to the minimum variation
estimation error, λopt is a function of the measurement matrix, measurement
noise, and the true variations xr∙, thus, it is not possible to find λopt exactly.
Applying first-order necessity condition for regularization problem in Equa-
tion 4.13 determines minimum value for λ. Let
j(æ)= ∣∣τ∣∣ι + λ∣∣Ar - b∣∣∣.
The first-order necessity condition for optimal solution implies = 0, i =
1... n. Thus,
g∣∣x∣∣1 _ ∂λ∣∣Ar-⅛∣∣j
∂Xi ∂Xi
52
More intriguing information
1. Workforce or Workfare?2. The name is absent
3. The name is absent
4. A parametric approach to the estimation of cointegration vectors in panel data
5. Elicited bid functions in (a)symmetric first-price auctions
6. The name is absent
7. The urban sprawl dynamics: does a neural network understand the spatial logic better than a cellular automata?
8. The name is absent
9. From Aurora Borealis to Carpathians. Searching the Road to Regional and Rural Development
10. The name is absent