In optimization, a descent direction is a vector
p
∈
R
n
that, in the sense below, moves us closer towards a local minimum
x
∗
of our objective function
f
:
R
n
→
R
.
Suppose we are computing
x
∗
by an iterative method, such as line search. We define a descent direction
p
k
∈
R
n
at the
k
th iterate to be any
p
k
such that
⟨
p
k
,
∇
f
(
x
k
)
⟩
<
0
, where
⟨
,
⟩
denotes the inner product. The motivation for such an approach is that small steps along
p
k
guarantee that
f
is reduced, by Taylor's theorem.
Using this definition, the negative of a non-zero gradient is always a descent direction, as
⟨
−
∇
f
(
x
k
)
,
∇
f
(
x
k
)
⟩
=
−
⟨
∇
f
(
x
k
)
,
∇
f
(
x
k
)
⟩
<
0
.
Numerous methods exist to compute descent directions, all with differing merits. For example, one could use gradient descent or the conjugate gradient method.
More generally, if
P
is a positive definite matrix, then
d
=
−
P
∇
f
(
x
)
is a descent direction at
x
. This generality is used in preconditioned gradient descent methods.