Luis M. Briceño-Arias,Cristóbal Vivar-Vargas
Abstract:In this paper we provide an efficient computation of the projection onto the cone generated by the epigraph of the perspective of any convex lower semicontinuous function. Our formula requires solving only two scalar equations involving the proximity operator of the function. This enables the computation of projections, for instance, onto exponential and power cones, and extends to previously unexplored conic projections, such as the projection onto the hyperbolic cone. We compare numerically the efficiency of the proposed approach in the case of exponential cones with an open source available method in the literature, illustrating its efficiency.
What problem does this paper attempt to address?
The problem that this paper attempts to solve is to provide a method for efficiently computing the projection of cones generated by the perspective of any convex lower - semicontinuous function. Specifically, the author proposes a formula that only requires solving two scalar equations involving the function proximity operator, which enables the computation of projections of specific cones such as the exponential cone and the power cone, and extends to the projection of cones that have not been explored before, such as the projection of hyperbolic cones. In addition, the author demonstrates the effectiveness of the proposed method by comparing the numerical efficiency with existing open - source methods in the case of the exponential cone.
### Background and Motivation of the Paper
In optimization theory, many problems can be simplified by transforming the constraint conditions into the form of closed convex cones. For a specific type of function \(f\), the epigraph of its perspective \(\tilde{f}\) is a closed convex cone, and computing the projections of these cones is crucial for solving certain optimization problems. For example, when dealing with entropy functions, softmax and softplus activation functions in neural networks, and generalized polynomials in geometric programming, it is necessary to compute the projection of the exponential cone. However, for some complex functions, existing methods may not be able to compute these projections efficiently.
### Main Contributions of the Paper
1. **General Formula**: The paper provides a general formula for computing the projection of cones generated by the perspective of any convex lower - semicontinuous function. This formula only needs to solve two scalar equations involving the function proximity operator, greatly simplifying the calculation process.
2. **Application Examples**: The author shows how to use this formula to compute the projections of the exponential cone and the hyperbolic cone. In particular, for the projection of the hyperbolic cone, existing methods cannot handle it, while the method in this paper can effectively solve this problem.
3. **Numerical Experiments**: By comparing the numerical efficiency with existing open - source methods in the case of the exponential cone, the effectiveness and flexibility of the proposed method are demonstrated. The experimental results show that this method has higher precision and faster calculation speed in some cases.
### Mathematical Background
- **Perspective Function**: Given a convex lower - semicontinuous function \(f\) defined on a real Hilbert space \(H\), its perspective \(\tilde{f}\) is defined as:
\[
\tilde{f}: H\times\mathbb{R}\to[-\infty,+\infty]:(x,\eta)\mapsto
\begin{cases}
\eta f\left(\frac{x}{\eta}\right), & \text{if }\eta > 0;\\
(\text{rec}f)(x), & \text{if }\eta = 0;\\
+\infty, & \text{if }\eta < 0.
\end{cases}
\]
- **Proximity Operator**: The proximity operator \(\text{prox}_f\) of the function \(f\) is defined as:
\[
\text{prox}_f: H\to H: x\mapsto\arg\min_{y\in H}\left(f(y)+\frac{1}{2}\|x - y\|^2\right).
\]
- **Projection Formula**: The core result of the paper is to provide a formula for computing the projection of the cone \(\text{epi}\tilde{f}\):
\[
P_{\text{epi}\tilde{f}}(x,\eta,\delta)=
\begin{cases}
(P_{\text{dom}\tilde{f}}(x,\eta),\delta), & \text{if }\tilde{f}(P_{\text{dom}\tilde{f}}(x,\eta))\leq\delta;\\
(\text{prox}_{\mu\tilde{f}}(x,\eta),\delta+\mu), & \text{if }\