Scaling-invariant functions versus positively homogeneous functions (Q2664909)

A function \(f:\mathbb{R}^n\to\mathbb{R}\) is scaling-invariant (SI) with respect to a reference point \(x^\star\in\mathbb{R}^n\) if for all \(x,y\in\mathbb{R}^n\) and for all \(\rho>0\): \[ f(x^\star+x)\leq f(x^\star+y)\Leftrightarrow f(x^\star+\rho x)\leq f(x^\star+\rho y). \] Moreover, a function \(p:\mathbb{R}^n\to\mathbb{R}\) is positively homogeneous (PH) with degree \(\alpha > 0\) (\(PH_\alpha\)) if for all \(x\in\mathbb{R}^n\) and \(\rho> 0\): \[ p(\rho x) = \rho^\alpha p(x). \] Composites of strictly monotonic functions with positively homogeneous functions are scaling-invariant with respect to zero. The authors prove in this paper that also the reverse is true for large classes of scaling-invariant functions. More specifically, the authors prove the following fundamental result. Let \(f\) be a continuous SI function. Then for all \(\alpha>0\), there exists a continuous \(PH_\alpha\) function \(p\) and a strictly increasing and continuous bijection (thus a homeomorphism) \(\varphi\) such that \(f =\varphi\circ p\). For a nonzero \(f\) and \(\alpha > 0\), the choice of \((\varphi, p)\) is unique up to a left composition of \(p\) with a piece-wise linear function. If \(f\) admits a global optimum, then \(0\) is also a global optimum. For any \(x_0\in\mathbb{R}^n\) such that \(f (x_0)\neq0\), in the case where \(p\) is a \(PH_1\) function, the homeomorphism \(\varphi\) can be chosen as \(f_{x_0}\) and is at least as smooth as \(f\). If \(f\) does not admit a global optimum, then there exist \(x_1, x_{-1}\in\mathbb{R}^n\) such that \(f(x_1) > 0\) and \(f (x_{-1}) < 0\). For any such \(x_1\) and \(x_{-1}\), in the case where \(p\) is a \(PH_1\) function, the homeomorphism \(\varphi\) can be chosen as the function equal to \(f_{x_1}\) on \(\mathbb{R}_+\) and equal to \(t\mapsto f_{x_{-1}}(-t)\) on \(\mathbb{R}_-\), where \[ f_x:t\in[0,\infty)\mapsto f(tx) \] is the surjective restriction of \(f\) to half-lines along the vector \(x\in\mathbb{R}^n\). Relaxing continuity, they also prove some necessary and sufficient condition under which an SI function is the composite of a PH function with a strictly monotonic function. In particular, let \(f\) be an SI function. There exist a \(PH_1\) function \(p\) and a strictly increasing function \(\varphi\) such that \(f =\varphi\circ p\) if and only if for all \(x\in\mathbb{R}^n\), the function \(f_x\) is either constant or strictly monotonic and the strictly increasing \(f_x\) share the same image (i.e., if \(\lambda\in\mathbb{R}\) is reached for one of these functions, then it is reached for all of them) and the strictly decreasing ones too. For a nonzero \(f\), up to a left composition of \(p\) with a piece-wise linear function, the choice of \((\varphi, p)\) is unique. The authors also study sublevel sets of scaling-invariant functions generalizing well-known properties of positively homogeneous functions.