or
Hence,
А|а||ЛГ* l≤M = l-.n.
∂x∙i
ll⅛∣μ1 - ⅛ = 2∣∣Λτ(Ar - 6)∣∣co ≤ 1
(4.14)
As we mentioned before, for very small regularization coefficients λ, zero is close
to the optimal point. Thus, putting x = 0 in Equation 4.14 determines a value
for λ. i.e., if x = 0 is a optimal solution,
Thus, the value for λ corresponding to zero would be
λ° = p‰
Kim et al. [40] suggest determining λ based on Λq∙ They use ʌɪ = 10λo∙ For
the problem shown in Figure 4.5, ʌɪ = lθʌɑ = 5.56 × IO-4. This estimation of
the λ is far from λopt (λopt ɪs shown in Figure 4.5).
Hale et al. [35] use distribution of measurement error to find λ. Assuming
independent normal distribution for measurement noise, they suggest
_ 2 / N
^2 — ~∖ ~2----
£ V ⅛-<*,m
53