📂Optimization

Subgradient

Definition ^{1 2}

For a function $f : \mathbb{R}^{n} \to \mathbb{R}$, the $\mathbf{g} \in \mathbb{R}^{n}$ that satisfies the following is called a subgradient of $f$ at point $\mathbf{x}$.

$$f(\mathbf{y}) \ge f(\mathbf{x}) + \mathbf{g}^{T}(\mathbf{y} - \mathbf{x}) \qquad \forall \mathbf{y} \in \mathbb{R}^{n}$$

Explanation

If the convex function $f$ is differentiable at $\mathbf{x}$, then $\mathbf{g} =$ $\nabla f(\mathbf{x})$ is unique. Conversely, if $\partial f(\mathbf{x}) = \left\{ \mathbf{g} \right\}$, then $f$ is differentiable at $\mathbf{x}$ and $\nabla f (\mathbf{x}) = \mathbf{g}$.

The set of subgradients of $f$ at $\mathbf{x}$ is called the subdifferential and is denoted as follows:

$$\partial f(\mathbf{x}) := \left\{ \mathbf{g} : \mathbf{g} \text{ is subgradient of } f \text{ at } \mathbf{x} \right\}$$

If $\partial f(\mathbf{x}) \ne \varnothing$, then $f$ is said to be subdifferentiable at $\mathbf{x}$.

Properties

• Scaling: For $a \gt 0$, $\partial (af) = a \partial f$
• Addition: $\partial(f + g) = \partial f + \partial g$
• Affine composition: If $g(\mathbf{x}) = f(A \mathbf{x} + \mathbf{b})$, $$\partial g(\mathbf{x}) = A^{T}\partial f(A \mathbf{x} + \mathbf{b})$$
• Subgradient optimality condition: For any function $f$, $$f(\mathbf{x}_{\ast}) = \min\limits_{\mathbf{x}} f(\mathbf{x}) \iff 0 \in \partial f(\mathbf{x}_{\ast})$$

See Also

• Subgradient method