= Projected gradient descent
Let $C$ be a nonempty closed <convex set> in a real finite-dimensional <Euclidean space>, and let $F$ be a differentiable <convex function>. A <projected gradient descent> step is
$$
x_{k+1}=\Pi_C(x_k-\eta\nabla F(x_k)),\qquad\eta>0.
$$
The <Euclidean projection onto a convex set> is single-valued under these assumptions and restores feasibility after the unconstrained <gradient descent> step. Closedness matters: the nearest point to zero in the convex set $(0,1)$ does not exist.
This is the <proximal gradient method> with nonsmooth term equal to the <indicator functional of a constraint set>, whose <proximal operator> is $\Pi_C$. If $\nabla F$ is $L$-<Lipschitz continuous>, a standard choice $0<\eta\leq1/L$ gives the usual convex objective-error bound relative to an existing minimizer.<Proximal Algorithms by Neal Parikh and Stephen Boyd>[sections1.2 and4.2, projected-gradient specialization, printed pp125–126 and149]{ref}
For $C=[0,1]$, $\Pi_C(z)=\max(0,\min(z,1))$. Minimizing $F(x)=(x-2)^2/2$ from $x_0=0$ with $\eta=1$ gives $x_1=\Pi_C(2)=1$, the constrained optimum.
Back to article page