Smoothing Technique for Nonsmooth Optimization I
Solving a nonsmooth optimization problem via solving a sequence of approximating smooth problem is a useful and widely used technique in optimization. Consider the first order methods in optimization, if the objective is differentiable, then the normal gradient descent can achieve \(\epsilon\) accuracy in function value in \(\mathcal{O}(1/\epsilon)\) iterations and those accelerated gradient methods can achieve \(\epsilon\) accuracy in \(\mathcal{O}(1/\sqrt{\epsilon})\) iterations. However, if the objective is nondifferentiable, then the subgradient descent and bundle methods need \(\mathcal{O}(1/\epsilon^2)\) iterations to achieve \(\epsilon\) accuracy. This illustrates why we can benefit from proper smoothing technique.
In this post, I will discuss the deterministic smoothing techniques for nonsmooth optimization and the randomized ones will be introduced in the next post. Most of the theoretical results here can be found in ^{1}, ^{2} and ^{3}.
Smooth Approximation
Throughout this post, we consider \(f:\mathbb{R}^n \to \mathbb{R}\) to be a convex but not always differentiable function. We begin by defining the concept of a smooth approximation.
Definition Let \(f_\mu : \mathbb{R}^n \to \mathbb{R}\) be a convex funtion. Then we say \(f_\mu\) is a \(\mu\)smooth approximation of \(f\) with parameter \((\alpha, \beta)\) if
 \(f_\mu\) is \(\frac{\alpha}{\mu}\)smooth
 \(f_\mu \leq f \leq f_\mu + \beta\mu\).
Here we can view \(\mu\) as a tradeoff between approximation accuracy and smothness.
Example (absolute value) Consider \(f(x)=x\) and define the smooth approximation \(f_\mu\) as
Note that \(f_\mu\) is known to be the Huber function. It is then easy to see that \(f_\mu\) is \(\frac{1}{\mu}\)smooth and satisfies
\[f_\mu(x) \leq f(x) \leq f_\mu(x) + \frac{\mu}{2}.\]Therefore, \(f_\mu\) is a \(\mu\)smooth approximation of \(f\) with parameter \((1,\frac{1}{2})\). \(\square\)
Example (max) Consider \(f(x) = \max\{x_1, \dots, x_n\}\) and define the smooth approximation \(f_\mu\) as
\[f_\mu(x) = \mu\log\left(\sum_{i = 1}^n \exp(\frac{x_i}{\mu})\right)  \mu\log(n).\]Note that \(f_\mu\) is known to be the softmax function. It is not hard to check that \(f_\mu\) is \(\frac{1}{\mu}\)smooth and satisfies
\[f_\mu(x) \leq f(x) \leq f_\mu(x) + \mu\log(n).\]Therefore, \(f_\mu\) is a \(\mu\)smooth approximation of \(f\) with parameter \((1, \log(n))\). \(\square\)
Next, I am going to introduce two widely used smoothing techniques.
Moreau proximal smoothing
Moreau proximal smoothing is a widely used smoothing technique in the Euclidean setting. Given a convex function \(f:\mathbb{R}^n \to \mathbb{R}\). The Moreau proximal approximation for \(f\) is given by
\[f_\mu(x) = \inf_{u \in \mathbb{R}^n} \left\{f(u) + \frac{1}{2\mu}\x  u\^2\right\}.\]It has been shown by Moreau that \(f_\mu\) is \(\frac{1}{\mu}\)smooth. Let \(u_x\) denote the unique point that achieves the above infimum, which is also known as the proximal point, then the gradient of \(f_\mu\) at \(x\) is given by
\[\nabla f_\mu(x) = \frac{1}{\mu}(x  u_x).\]Moreover, when \(f\) is \(L\)Lipschitz, then
\[f_\mu(x) \leq f(x) \leq f_\mu(x) + \frac{L^2}{2}\mu.\]In other words, when \(f\) is \(L\)Lipschitz, then \(f_\mu\) is a \(\mu\)smooth approximation of \(f\) with parameter \((1,\frac{L^2}{2})\). It is also worth noting that \(f\) and \(f_\mu\) have the same set of minimizers, and therefore minimizing \(f\) and \(f_\mu\) are equivalent.
Nesterov smoothing
Nesterov’s smoothing is a smoothing technique for a class of nonsmooth functions:
\[f(x) = \max\{\langle u, Ax\rangle  \phi(u) \mid u \in \mathcal{C}\},\]where \(\phi:\mathbb{R}^m\to\mathbb{R}\) is a continuous convex function, \(A:\mathbb{R}^n \to \mathbb{R}^m\) is a linear operator and \(\mathcal{C}\) is a convex compact set in \(\mathbb{R}^m\). The smooth approximation is given by
\[f_\mu(x) = \max\{\langle u, Ax\rangle  \phi(u)  \mu d(u): u \in \mathcal{C}\},\]where \(d\) is a \(\sigma\)strongly convex function on \(\mathcal{C}\). It has been shown by Nestrov that \(f_\mu\) is \(\frac{\A\^2}{\sigma\mu}\)smooth with gradient
\[\nabla f_\mu(x) = A^*u_x,\]where \(u_x\) is the unique maximizer in the definition of \(f_\mu(x)\).
Infimal convolution smoothing technique
Beck and Teboulle propose a smoothing technique based on infimal convolution, which can be seen as a generalized framework for the above two discussed smoothing techniques. Let \(f:\mathbb{R}^n \to \mathbb{R}\) be a closed and convex function. Then \(f\) can be expressed as
\[f(x) = \sup_{u \in \mathbb{R}^n}\enspace \langle u, x\rangle  f^*(u),\]where \(f^*\) is the conjugate function of \(f\). Then we can build a smooth approximation of \(f\) by adding a strongly convex component to its conjugate, namely,
\[f_\mu(x) = (f^* + \mu d)^*(x) = \sup_{u \in \mathbb{R}^n}\enspace \langle u, x\rangle  f^*(u)  \mu d(u),\]where \(d:\mathbb{R}^n\to\mathbb{R}\) is a nonnegative, continuous and 1strongly convex function. It follows that \(f^* + \mu d\) is \(\mu\)strongly convex, and therefore \(f_\mu\) is \(1/\mu\)smooth. It can also be shown that
\[f_\mu(x) \leq f(x) \leq f_\mu(x) + \mu D,\]where \(D = \sup_x d(x)\). Therefore, \(f_\mu\) is a \(1/\mu\)smooth approximation of \(f\) with parameters \((1, D)\).
Complexity Improvement
Consider the minimization problem:
\[\min_{x \in \mathbb{R}^n} \enspace f(x),\]where \(f\) is a nonsmooth convex funtion. Let \(x^*\) denote the optimizer for \(f\). Let \(f_\mu\) be a \(1/\mu\)smooth approximation of \(f\) with parameters \((\alpha, \beta)\). Now we consider solving the following problem:
\[\min_{x \in \mathbb{R}^n} \enspace f_\mu(x).\]Let \(x_\mu^*\) denote the optimizer for \(f_\mu\). It is widely known that to attain \(\epsilon\)accuracy for minimizing a \(L\)smooth convex function by accelerated gradient descent, one needs \(\mathcal{O}\left(\sqrt{\frac{L}{\epsilon}}\right)\) iterations. Now we know that \(f_\mu\) is \(\frac{\alpha}{\mu}\)smooth, in order to attain \(\frac{\epsilon}{2}\)accuracy, we need \(\mathcal{O}\left(\sqrt{\frac{\alpha}{\mu\epsilon}}\right)\) iteration. Let \(x^t\) denote the output. Finally, we set \(\mu\) such that \(\beta\mu = \frac{\epsilon}{2}\). Then we have
The final iteration complexity is given by
\[\mathcal{O}\left(\sqrt{\frac{\alpha}{\mu\epsilon}}\right) = \mathcal{O}\left(\frac{\sqrt{\alpha\beta}}{\epsilon}\right).\]References

Nesterov, Yu. Smooth minimization of nonsmooth functions. Mathematical programming 103.1, 2005. ↩

Beck, Amir, and Marc Teboulle. Smoothing and first order methods: A unified framework. SIAM Journal on Optimization 22.2, 2012. ↩

Chen, Yuxin. Smoothing for nonsmooth optimization. Lecture Notes for ELE 522: LargeScale Optimization for Data Science, Princeton University, Fall 2019. ↩