280 Nonlinear Programming I: One-Dimensional Minimization Methods
5.11 Cubic Interpolation Method
The cubic interpolation method finds the minimizing step lengthλ∗in four stages [5.5,
5 .11]. It makes use of the derivative of the functionf:
f′(λ)=
df
dλ
=
d
dλ
f (X+λS)=ST∇ f(X+λS)
The first stage normalizes theSvector so that a step sizeλ=1 is acceptable. The
second stage establishes bounds onλ∗, and the third stage finds the value ofλ ̃∗by
approximatingf (λ)by a cubic polynomial h(λ). If theλ ̃∗found in stage 3 does
not satisfy the prescribed convergence criteria, the cubic polynomial is refitted in the
fourth stage.
Stage 1. Calculate =maxi|si| where, |si| s the absolute value of thei ith compo-
nent ofS, and divide each component ofSby. An alternative method of normalization
is to find
=(s^21 +s 22 + · · · +s^2 n)^1 /^2
and divide each component ofSby.
Stage 2. To establish lower and upper bounds on the optimal step sizeλ∗, we need
to find two pointsAandBat which the slopedf/dλhas different signs. We know that
atλ=0,
df
dλ
∣
∣
∣
∣
λ= 0
=ST∇ f(X) < 0
sinceSis presumed to be a direction of descent.†
Hence to start with we can takeA= 0 and try to find a pointλ=Bat which the
slopedf/dλis positive. PointBcan be taken as the first value out oft 0 , 2 t 0 , 4 t 0 , 8 t 0 ,...
at whichf′is nonnegative, wheret 0 is a preassigned initial step size. It then follows
thatλ∗is bounded in the intervalA< λ∗≤ B(Fig. 5.16).
Stage 3. If the cubic equation
h(λ)=a+bλ+cλ^2 +λ^3 (5.45)
f(l)
l
Figure 5.16 Minimum off (λ)lies betweenAandB.
†In this case the angle between the direction of steepest descent andSwill be less than 90◦.