Self-interacting CBO: existence, uniqueness, and long-time convergence (Q7029930)

The authors consider a complete filtered probability space \((\Omega , \mathfrak{F},\mathfrak{F}_{t},\mathbb{P})\), a standard \(d\)-dimensional Brownian motion \((B_{t})_{t\geq 0}\), and a McKean-Vlasov process \(X_{\cdot }\) on \(\mathbb{R}^{d}\) whose dynamics is governed by the equation: \( dX_{t}=-\lambda (X_{t}-m_{\alpha }(\mathcal{L}_{t}[X]))dt+\sigma D(X_{t}-m_{\alpha }(\mathcal{L}_{t}[X])))dB_{t}\), where \(\mathcal{L} [X]=Law(X_{t})\), \(\sigma >0\) is a real constant, \(D(X)=diag(\left\vert x_{1}\right\vert ,\ldots ,\left\vert x_{d}\right\vert )\) for any \( X=(x_{1},\ldots ,x_{d})\in \mathbb{R}^{d}\), and \(m_{\alpha }(\mathcal{L} _{t}[X])\) the current global consensus point defined by \(m_{\alpha }( \mathcal{L}_{t}[X])=\int_{\mathbb{R}^{d}}x\omega _{\alpha }^{f}(x)\mathcal{L} _{t}[X](dx)/\int_{\mathbb{R}^{d}}\omega _{\alpha }^{f}(x)\mathcal{L} _{t}[X](dx)\), \(\omega _{\alpha }^{f}\) being the weight function chosen as \( \omega _{\alpha }^{f}(x)=exp(-\alpha f(x))\). They introduce the rescaled consensus-based optimization model: \(dX_{t}=-\lambda (X_{t}-\kappa m_{\alpha }(\mathcal{L}_{t}[X]))dt+\sigma (\frac{1}{\alpha }\mathbb{I} _{d}+D(X_{t}-\kappa m_{\alpha }(\mathcal{L}_{t}[X])))dB_{t}\), where \( 0<\kappa <1\), then the self-interacting consensus-based optimization model: \( dY_{t}=-\lambda (Y_{t}-\kappa _{\alpha }(\mathcal{E}_{t}[Y]))dt+\sigma ( \frac{1}{\alpha }\mathbb{I}_{d}+D(Y_{t}-\kappa m_{\alpha }(\mathcal{E} _{t}[Y])))dB_{t}\), complemented with the deterministic initial condition \( Y_{0}=x\in \mathbb{R}^{d}\), and where \(\mathcal{E}_{t}\) is the occupation measure \(\mathcal{E}_{t}[Y]=\frac{1}{t}\int_{0}^{t}\delta _{Y_{\tau }}d\tau \) , and finally the mean-field interacting model: \(dX_{t}=-\lambda (X_{t}-\kappa _{\alpha }(\mathcal{L}_{t}[X]))dt+\sigma (\frac{1}{\alpha } \mathbb{I}_{d}+D(X_{t}-\kappa m_{\alpha }(\mathcal{L}_{t}[X])))dB_{t}\). The optimization problem is: Find \(x^{\ast }\in \arg \min_{x\in \mathbb{R} ^{d}}f(x)\), where \(f:\mathbb{R}^{d}\rightarrow \mathbb{R}\) is a non-convex and non-smooth objective function which is bounded from below by \( \underline{f}\), which satisfies: \(\left\vert f(x)-f(y)\right\vert \leq L_{f}(1+\left\vert x\right\vert +\left\vert y\right\vert )^{s}\left\vert x-y\right\vert \), \(\forall x,y\in \mathbb{R}^{d}\), and \(c_{1}(\left\vert x\right\vert ^{\ell }-1)\leq f-\underline{f}\leq c_{2}(\left\vert x\right\vert ^{\ell }+1)\), \(\forall x\in \mathbb{R}^{d}\). Using Leray-Schauder fixed point theorem, the authors prove that the two last problems have strong solutions. The main result of the paper proves some equivalence between the mean-field interacting and self-interacting models, in the sense that they both evolve eventually towards the same equilibrium state. For the proof, the authors establish some estimates and the existence of an invariant measure for the mean-field interacting model. The paper ends with a quite similar result in the case of multi-self-interacting particles.