Online learning with exponential weights in metric spaces

Let (M,d) be a complete geodesic space and f:M\to\mathbb R be a given function.

1. Suppose that f is geodesically \beta-expconcave for some \beta>0. Then, the function f is geodesically convex.

2. Suppose that f is geodesically \alpha-convex and L-Lipchitz for some \alpha,L>0. Then, it is geodesically \beta-expconcave for all 0< \beta\le \frac{\alpha}{L^2}.

A geodesic space (M,d) satisfies {\rm curv}(M)\ge0 iff, for all p,x,y\in M, any geodesic \gamma:[0,1]\to M connecting x to y in M, and any t\in [0,1],

Similarly, {\rm curv}(M)\le0 if the same holds with opposite inequality in (2.5).

Let (M,d) be a geodesic space satisfying {\rm curv}(M)\le 0 and with finite diameter. Then, for all p\in M, the function d^2(p,.):M\to \mathbb R_+ is \beta-concave for all 0<\beta\le 1/( 2\,\mathrm{diam}(M)^2).

Let (M,d) be a complete geodesic space with {\rm curv}(M)\le 0. Let m\in \mathcal P_2(M) and let x^* be its unique barycenter. Let f:M\to \mathbb R be convex and lower semi-continuous. Then we have

provided f is either positive or in L^1(m).

Let (M,d) be a complete and separable geodesic space with {\rm curv}(M)\ge \kappa, for some \kappa\in \mathbb R. Let m\in \mathcal P_2(M) and suppose that it admits at least one barycenter x^*. Let f:M\to \mathbb R be \alpha-convex, for some \alpha\in \mathbb R, and Lipschitz in a neighborhood of x^*. Then we have

provided f is either positive or in L^1(m).

Assume that ?, ? and ? hold. Suppose that there exists \beta>0 such that every \ell\in\mathcal L is geodesically \beta-expconcave. Then taking \beta_t=\beta for all t\ge 1 and selecting any

the regret of the ewb forecaster satisfies, for all n\ge 2,

Assume that ? and ? hold. Assume that there exists constants p,r_0>0 and a function c:M\to\mathbb R_+ such that

Assume finally that there exist \beta,L>0 such that every \ell\in\mathcal L is geodesically \beta-expconcave and L-Lipschitz. Then taking \beta_t=\beta for all t\ge 1, the regret of the ewb forecaster satisfies, for all n\ge 1/r_0,

for x^*_n\in M as in (3.2).

Suppose that (M,d,m) satisfies the measure contraction property mcp(\kappa,p) for some \kappa\in \mathbb R and some p>1. Then, there exists p,r_0>0 and c:M\to\mathbb R_+ such that property (3.3) holds.

Assume that ? and ? hold. Suppose that every \ell\in\mathcal L is geodesically convex and [a,b]-valued. Then setting, for all t\ge 1,

the regret of the ewb forecaster satisfies the following.

1. If Assumption ? holds, then for all n\ge 2,

   \frac{R_n}{b-a}\le 1+c_1\left(1+\ln\frac{1}{c_{\kappa,p}(x^{\star}_n)}\right)\sqrt{pn\ln n},

   where c_{\kappa,p} and x^*_n are as in Theorem 3.2.

2. If more generally property (3.3) holds and all losses \ell\in\mathcal L are also L-Lipschitz, then for all n\ge 1/r_0,

   \frac{R_n}{b-a}\le \frac{L}{b-a}+c_1\left(1+\ln\frac{1}{c(x^*_n)}\right)\sqrt{pn\ln n},

   for x^*_n\in M as in (3.2).

Suppose that the metric space (M,d) satisfies Assumptions ? and ?. Suppose that (?) holds uniformly over the outcome sequence (z_1,\dots,z_n)\in\mathcal Z^n and that, for all z\in\mathcal Z, the function \ell(.,z):M\to \mathbb R is geodesically convex. Then, for all n\ge 1,

Suppose that (M,d) satisfies Assumptions ? and ?. Let m be a prior distribution over M such that (M,d,m) satisfies Assumption ? and condition (?) if \kappa<0. Suppose that there exists \beta>0 such that, for all z\in\mathcal Z, the function \ell(.,z):M\to \mathbb R is geodesically \beta-expconcave. Let \hat\theta_n be as in (?) where \vartheta_1 is a barycenter of m and, for 2\le t\le n+1, \vartheta_t(Z_1,\dots,Z_{t-1}) is a barycenter of the (random) probability measure m_{t} defined by

where m_1:=m. Then, for all n\ge 2,

Let P\in\mathcal P_2(\mathcal P_2(\mathbb R^D)) and suppose \Delta:=\mathrm{diam}(\mathrm{supp}(P))<+\infty. Let \mu^* be a barycenter of P and suppose that, for P-almost all \nu, \varphi_{\mu^*\to\nu} is \alpha(\nu)-strongly convex for a measurable \alpha:\mathcal P_2(\mathbb R^D)\to \mathbb R_+. Then \mu^* is unique and, for all \mu in the support of P, and all

we have

where C_{\rm var}:=\int \alpha(\nu){\rm d}P(\nu).

Provided 0<\beta_{t+1}\le\beta_t, for all t\ge 1, we have for all n\ge 1,

where L_n:=\sum_{t=1}^n\ell_t.

Assume that ? and ? hold. Suppose that there exists \beta>0 such that every \ell\in\mathcal L is geodesically \beta-expconcave. Then the cumulative loss of the ewb forecaster, with \beta_t=\beta for all t\ge 1, satisfies

where L_n:=\sum_{t=1}^n\ell_t and where the inf runs over all probability measures \mu on M.

1. For all 0\le r\le \pi and all \varepsilon\in(0,1),

   \frac{\sin(\varepsilon r)}{\sin(r)}\ge \varepsilon,

   with the convention 0/0=1.

2. For all r\ge 0 and all \varepsilon\in(0,1/2],

   \frac{\sinh(\varepsilon r)}{\sinh(r)}\ge \varepsilon \psi(r) ,

   with \psi(r):=(r \coth r)\exp(-r \coth r) and the convention 0/0=1.

Let P\in\mathcal P_2(\mathcal P_2(\mathbb R^d)) and \mu^* be a barycenter of P. Suppose that, for P-almost all \nu, \varphi_{\mu^*\to\nu} is \alpha(\nu)-strongly convex for a measurable \alpha:\mathcal P_2(\mathbb R^d)\to \mathbb R_+. Then \mu^* is unique and, for all \mu\in\mathcal P_2(\mathbb R^d),

for C_{\rm var}:=\int \alpha(\nu)\,\mathrm{d}P(\nu).

Under the assumptions of Lemma 4.4, for all \mu\in\mathcal P_2(\mathbb R^D), and all \beta>0 satisfying

we have

Assume that ? and ? hold. Suppose that every \ell\in\mathcal L is geodesically convex. Then the cumulative loss of the ewb forecaster, with 0<\beta_{t+1}<\beta_t for all t\ge 1, satisfies

where \bar\ell_t:=\int_M\ell_t\mathrm{d}m_t, where L_n=\sum_{t=1}^n\ell_t and where the inf runs over all probability measures \mu on M.