Introduction
Model Predictive Path-Integral Control (MPPI) is one of the major sampling-based control methods used in robotics, autonomous driving, and model-based reinforcement learning [11 ] . While very successful in applications, the theoretical understanding of MPPI is still limited. Introduced originally by information-theoretic arguments connected to stochastic optimal control [16 ] , it has since been recognised that MPPI is essentially a zeroth-order sampling-based optimisation method combined with receding-horizon control. In particular, a rigorous theoretical analysis must start with the optimisation behaviour of MPPI. Most existing works have been focused on local perspectives, e.g., interpreting MPPI as a form of gradient descent [3 ] , or considering restricted classes of objectives like quadratic functions [17 ] . This is a severe limitation, since MPPI is particularly interesting for challenging non-convex and even non-smooth problems arising in control tasks. Motivated by this shortcoming, we investigate MPPI as a sampling-based global optimisation method for non-convex, non-smooth objective functions with a unique global optimum. For this, we build on the recent work [5 ] that connects MPPI with global non-convex optimisation. More precisely, this work finds that the plain MPPI iteration on the population level corresponds exactly to Consensus Hopping (CH), a limiting case of Consensus-Based Optimisation (CBO), which is a class of popular global optimisation methods based on interacting particle systems [13 ] with strong theoretical guarantees [4 , 2 ] . In turn, this connection allows one to prove strong convergence results for CH and hence MPPI. We build on this work and establish semi-global convergence results for MPPI on the population level. While already conceptually insightful, MPPI is a sampling-based method, and hence any result with practical relevance needs to take into account the finite-sample error incurred by the sampling process. To address this important issue, we also prove corresponding high-probability error bounds and establish sample complexity bounds for the actual sampling-based MPPI iteration. Apart from requiring a unique global optimum, our results rely only on rather general assumptions on the objective function, covering in particular non-convex and non-smooth functions. Our work provides a significant step towards a deeper theoretical understanding of MPPI and helps to put it on a solid theoretical foundation, on which reliable sampling-based control approaches can be built.
Related Work MPPI was introduced in the context of autonomous driving [15 ] based on information-theoretic considerations [16 ] that connect to stochastic optimal control and path-integral control. However, the connection is mostly heuristic and the rigorous theoretical foundation is lost [8 ] . By now there is a multitude of different MPPI variants, cf. Kazim et al. [11 ] for an overview, and similar to most analyses, we focus on the plain MPPI iteration formally introduced in § 2 , and investigate MPPI as a zeroth-order sampling-based optimisation method, cf. Jordana et al. [10 ] for more on this perspective. In contrast to previous works like Yi et al. [17 ] or Yoon and Kim [18 ] , we consider general non-convex objective functions. For this, we build on the recent work Fornasier et al. [5 ] that identified MPPI on the population level as Consensus Hopping (CH), which is connected to Consensus-Based Optimisation (CBO) [13 ] . Utilising a quantitative Laplace principle as pioneered by the work Fornasier et al. [4 ] , one can establish semi-global convergence guarantees for CH and hence for population-level MPPI. The concept of semi-global results is common in nonlinear control, see (Grune and Pannek, 2017, Chapter 6) for more details. The use of the Laplace principle in connection with MPPI has already been noticed in Homburger et al. [8 ] ; however, a different scaling is used there, and no non-asymptotic results are provided. Note that it is well-known that MPPI has problems with functions having multiple global optima [12 ] due to averaging effects, which motivates our focus on objectives with a unique global optimum. We also provide finite-sample results, which are particularly relevant since in practice MPPI is always sampling-based. Prior works have investigated this in restricted settings like LQR problems, cf. [19 , 18 ] , whereas we work with very general non-convex objectives. Instead of relying on propagation of chaos techniques as suggested in Fornasier et al. [5 ] , we directly analyse the estimation error in a single step from the perspective of self-normalised importance weighted sampling, cf. [1 ] for an overview and similar results. While this has been used in prior works on MPPI like Yoon et al. [19 ] ; Yoon and Kim [18 ] , our simple but effective estimation error bounds become particularly powerful in connection with our semi-global convergence results. In particular, to the best of our knowledge we achieve the first finite-sample complexity bounds for general nonconvex objective functions with only modest regularity assumptions.
Contributions First , we isolate, generalise, and sharpen the mean-field analysis of the MPPI iteration from Fornasier et al. [5 ] , establishing semi-global convergence to the global optimum. Second , we provide the first semi-global finite-sample accuracy results for MPPI. Third , under additional assumptions, we provide sharpened finite-sample results that include the LQR problem.
Outline In § 2 , we introduce the mean-field and empirical versions of MPPI that we will investigate. § 3 contains results for the mean-field iteration, and § 4 the corresponding finite-sample results. In the main text, we present simplified results that highlight the main features; the supplementary material contains more detailed and quantitative results with full proofs. In § 5 we illustrate some results with the standard LQR problem. Finally, § 6 contains a discussion and outlook.
MPPI and Consensus Hopping
Consider an objective function f : R d → R f : \mathbb{R}^{d} \to\mathbb{R} f : R d → R that is measurable and bounded from below. Let α ∈ R > 0 \alpha\in\mathbb{R}_{>0} α ∈ R > 0 and let Σ ∈ R d × d \Sigma\in\mathbb{R}^{d \times d} Σ ∈ R d × d be a symmetric positive definite matrix. Throughout, σ min ( Σ ) \sigma_{\min}(\Sigma) σ m i n ( Σ ) and σ max ( Σ ) \sigma_{\max}(\Sigma) σ m a x ( Σ ) denote the square roots of the smallest and largest eigenvalue of Σ \Sigma Σ , respectively. The MPPI iteration with inverse temperature α \alpha α and sampling covariance matrix Σ \Sigma Σ is defined as follows. Starting from an initial guess x ^ 0 ∈ R d \hat{x}_0 \in\mathbb{R}^{d} x ^ 0 ∈ R d , in order to compute the next iterate x ^ k + 1 \hat{x}_{k+1} x ^ k + 1 from the current iterate x ^ k \hat{x}_k x ^ k , we sample X 1 , … , X N ∼ i . i . d . N ( x ^ k , Σ ) X_1,\ldots,X_N \mathrel{\overset{\mathrm{i.i.d.}}{\sim}} \mathcal{N}(\hat{x}_k,\Sigma) X 1 , … , X N ∼ i.i.d. N ( x ^ k , Σ ) and then set
x ^ k + 1 = ∑ i = 1 N exp ( − α f ( X i ) ) X i ∑ i = 1 N exp ( − α f ( X i ) ) (1) \hat{x}_{k+1} = \frac{\sum_{i=1}^{N} \exp(-\alpha f(X_i))X_i}{\sum_{i=1}^{N} \exp(-\alpha f(X_i))} \tag*{(1)} x ^ k + 1 = ∑ i = 1 N exp ( − α f ( X i )) ∑ i = 1 N exp ( − α f ( X i )) X i (1)
Note that in each iteration we draw a new sample.
We start the analysis of the MPPI iteration on the mean-field or population-level . For this, define the transition map
T α : R d → R d , T α ( x ) : = E X ∼ N ( x , Σ ) [ exp ( − α f ( X ) ) X ] E X ∼ N ( x , Σ ) [ exp ( − α f ( X ) ) ] = 1 Z α ( x ) E X ∼ N ( x , Σ ) [ exp ( − α f ( X ) ) X ] , (2) \begin{aligned}
\mathcal{T}_{\alpha} : \mathbb{R}^{d} \to\mathbb{R}^{d}, \qquad\mathcal{T}_{\alpha}(x) := \frac{\mathbb{E}_{X \sim\mathcal{N}(x,\Sigma)}[\exp(-\alpha f(X))X]}{\mathbb{E}_{X \sim\mathcal{N}(x,\Sigma)}[\exp(-\alpha f(X))]} \\
&= \frac{1}{Z_{\alpha}(x)}\mathbb{E}_{X \sim\mathcal{N}(x,\Sigma)}[\exp(-\alpha f(X))X], \tag*{(2)}
\end{aligned} T α : R d → R d , T α ( x ) := E X ∼ N ( x , Σ ) [ exp ( − α f ( X ))] E X ∼ N ( x , Σ ) [ exp ( − α f ( X )) X ] = Z α ( x ) 1 E X ∼ N ( x , Σ ) [ exp ( − α f ( X )) X ] , (2)
where Z α ( x ) = E X ∼ N ( x , Σ ) [ exp ( − α f ( X ) ) ] Z_{\alpha}(x) = \mathbb{E}_{X \sim\mathcal{N}(x,\Sigma)}[\exp(-\alpha f(X))] Z α ( x ) = E X ∼ N ( x , Σ ) [ exp ( − α f ( X ))] is a normalisation constant, and for later use we define the free energy by
E α ( x ) = − 1 α log Z α ( x ) = − 1 α log E X ∼ N ( x , Σ ) [ exp ( − α f ( X ) ) ] . (3) \mathcal{E}_{\alpha}(x) = -\frac{1}{\alpha}\log Z_{\alpha}(x) = -\frac{1}{\alpha}\log\mathbb{E}_{X \sim\mathcal{N}(x,\Sigma)}[\exp(-\alpha f(X))]. \tag*{(3)} E α ( x ) = − α 1 log Z α ( x ) = − α 1 log E X ∼ N ( x , Σ ) [ exp ( − α f ( X ))] . (3)
Given some initial state x 0 ∈ R d x_0 \in\mathbb{R}^{d} x 0 ∈ R d , we define the mean-field MPPI iteration by
x k + 1 = T α ( x k ) , k ∈ N 0 . (4) x_{k+1} = \mathcal{T}_{\alpha}(x_k), \qquad k \in\mathbb{N}_0. \tag*{(4)} x k + 1 = T α ( x k ) , k ∈ N 0 . (4)
As is well-known, each step of the MPPI iteration (1) is a self-normalised importance-weighted estimate of the corresponding mean-field step. As pointed out in Fornasier et al. [5 ] , the mean-field MPPI iteration is exactly the Consensus Hopping (CH) iteration from this reference.
Assume now that f f f has a global minimiser x ∗ x_* x ∗ . The core question of the present work is in which sense the MPPI iterates ( x ^ k ) k (\hat{x}_k)_k ( x ^ k ) k and the mean-field MPPI iterates ( x k ) k (x_k)_k ( x k ) k approximate the global minimiser x ∗ x_* x ∗ of f f f . It is clear that without any additional assumptions on f f f , little progress can be made, so our analysis will be based mostly on the following properties of f f f .
Assumption 1 (Objective function) The function f : R d → R f : \mathbb{R}^{d} \to\mathbb{R} f : R d → R satisfies the following.
(F1) f f f is measurable and attains its global minimum f ∗ ∈ R f_* \in\mathbb{R} f ∗ ∈ R at x ∗ ∈ R d x_* \in\mathbb{R}^{d} x ∗ ∈ R d .
(F2) f f f is upper semicontinuous at x ∗ x_* x ∗ .
(F3) For every ϵ > 0 \epsilon> 0 ϵ > 0 there exists S ϵ > 0 S_{\epsilon} > 0 S ϵ > 0 such that
inf ∥ x − x ∗ ∥ 2 ≥ ϵ f ( x ) ≥ f ∗ + S ϵ . (5) \inf_{\lVert x-x_\ast\rVert_2\ge\epsilon} f(x) \ge f_\ast+ S_\epsilon. \tag*{(5)} ∥ x − x ∗ ∥ 2 ≥ ϵ inf f ( x ) ≥ f ∗ + S ϵ . (5)
(F2) is a very weak continuity assumption. (F3) makes the notion precise that the global minimiser x ∗ x_\ast x ∗ is sufficiently separated from any other local minimisers and that the global minimum is sufficiently pronounced. In particular, an elementary computation shows that (F3) implies that x ∗ x_\ast x ∗ is indeed the unique global minimiser of f f f .
To obtain explicit thresholds, we also consider the following simplified assumption, which only requires strong convexity close to the global minimiser and a suboptimality gap away from it.
Assumption 2 (Local strong convexity and energy gap) The function f : R d → R f : \mathbb{R}^d \to\mathbb{R} f : R d → R is continuous and attains its global minimum f ∗ f_\ast f ∗ at x ∗ ∈ R d x_\ast\in\mathbb{R}^d x ∗ ∈ R d , and there are R , m , Δ > 0 R,m,\Delta> 0 R , m , Δ > 0 such that the following hold.
(SF1) The function y ↦ f ( y ) − m 2 ∥ y − x ∗ ∥ 2 2 y \mapsto f(y) - \frac{m}{2}\lVert y-x_\ast\rVert_2^2 y ↦ f ( y ) − 2 m ∥ y − x ∗ ∥ 2 2 is convex on B R ( x ∗ ) B_R(x_\ast) B R ( x ∗ ) .
(SF2) f ( y ) − f ∗ ≥ Δ f(y)-f_\ast\ge\Delta f ( y ) − f ∗ ≥ Δ for all y ∈ R d ∖ B R ( x ∗ ) y \in\mathbb{R}^d \setminus B_R(x_\ast) y ∈ R d ∖ B R ( x ∗ ) .
(SF3) There is M ≥ m M \ge m M ≥ m such that f ( y ) − f ∗ ≤ M 2 ∥ y − x ∗ ∥ 2 2 f(y)-f_\ast\le\frac{M}{2}\lVert y-x_\ast\rVert_2^2 f ( y ) − f ∗ ≤ 2 M ∥ y − x ∗ ∥ 2 2 for all y ∈ B R ( x ∗ ) y \in B_R(x_\ast) y ∈ B R ( x ∗ ) .
(SF1) and (SF2) imply Assumption 1, and (F3) holds with S ϵ = min { m 2 ϵ 2 , Δ } S_\epsilon= \min\{\frac{m}{2}\epsilon^2,\Delta\} S ϵ = min { 2 m ϵ 2 , Δ } , cf. Lemma 19 in the supplementary material. We stress that outside the well B R ( x ∗ ) B_R(x_\ast) B R ( x ∗ ) , the objective can still be non-convex and non-smooth. Furthermore, condition (SF3) is only needed for the finite-sample results, and it holds, for instance, if f f f has an M M M -Lipschitz gradient on B R ( x ∗ ) B_R(x_\ast) B R ( x ∗ ) .
Semi-Global Convergence in Mean-Field
Elementary computations show that
T α ( x ) = x − α Σ ∇ E α ( x ) , (6) T_\alpha(x) = x - \alpha\Sigma\nabla\mathcal{E}_\alpha(x), \tag*{(6)} T α ( x ) = x − α Σ∇ E α ( x ) , (6)
suggesting that mean-field MPPI is a form of preconditioned gradient descent on a smoothed objective function, cf. Fazlyab et al. [3 ] for an in-depth discussion of this connection. The following example illustrates this point of view.
Example 1 Let d = 1 d = 1 d = 1 , and consider the objective
f ( y ) = min { 1 2 y 2 , 1 + 1 2 ( y − 3 ) 2 } , (7) f(y) = \min\left\{\frac{1}{2}y^2,\ 1+\frac{1}{2}(y-3)^2\right\}, \tag*{(7)} f ( y ) = min { 2 1 y 2 , 1 + 2 1 ( y − 3 ) 2 } , (7) which has the global minimiser x ∗ = 0 x_\ast= 0 x ∗ = 0 with f ∗ = 0 f_\ast= 0 f ∗ = 0 , a strict local minimiser at y = 3 y = 3 y = 3 with f ( 3 ) = 1 f(3) = 1 f ( 3 ) = 1 , and a kink at y = 11 / 6 y = 11/6 y = 11/6 , where the two branches meet. Note that f f f fulfils Assumption 1, though it is neither convex nor differentiable. Figure 1 shows E α \mathcal{E}_\alpha E α for Σ ∈ { 0.35 2 , 1 } \Sigma\in\{0.35^2,1\} Σ ∈ { 0.3 5 2 , 1 } and α ∈ { 1 , 3 , 8 } \alpha\in\{1,3,8\} α ∈ { 1 , 3 , 8 } . For Σ = 0.35 2 \Sigma= 0.35^2 Σ = 0.3 5 2 , all three smoothed objectives keep a second well near the local minimiser 3. For Σ = 1 \Sigma= 1 Σ = 1 , the proposal reaches the deeper well from every centre in the displayed interval, and E α \mathcal{E}_\alpha E α has a single minimum near x ∗ x_\ast x ∗ already for α = 1 \alpha= 1 α = 1 .
Figure 1. The free energy E α E_\alpha E α for the objective (7) of Example 1 with Σ = 0.35 2 ≈ 0.12 \Sigma= 0.35^2 \approx0.12 Σ = 0.3 5 2 ≈ 0.12 (left) and Σ = 1 \Sigma= 1 Σ = 1 (right), for α ∈ { 1 , 3 , 8 } \alpha\in\{1,3,8\} α ∈ { 1 , 3 , 8 } (solid, darker for larger α \alpha α ). The dashed curve is the objective f f f .
As expected, Figure 1 shows that E α E_\alpha E α smooths the objective function. However, we observe that more than that happens: for large sampling covariance and inverse temperature, the smoothed objective looks increasingly convex. Interpreting mean-field MPPI as preconditioned gradient descent then suggests that we should be able to get some form of global convergence to the global minimiser x ∗ x_\ast x ∗ , cf. Figure 2 in the supplementary material for an illustration.
Figure 2. Mean-field MPPI iterates for the objective (7) of Example 1 with Σ = 0.35 2 ≈ 0.12 \Sigma= 0.35^{2} \approx0.12 Σ = 0.3 5 2 ≈ 0.12 (left) and Σ = 1 \Sigma= 1 Σ = 1 (right), started at x 0 = − 1 x_{0} = -1 x 0 = − 1 , 2 2 2 and 4 4 4 (top to bottom; the dotted vertical line marks x 0 x_{0} x 0 ). The markers show ( x k , E α ( x k ) ) (x_{k},\mathcal{E}_{\alpha}(x_{k})) ( x k , E α ( x k )) , k = 0 , … , 12 k = 0,\ldots,12 k = 0 , … , 12 , for α = 1 \alpha= 1 α = 1 (squares), α = 3 \alpha= 3 α = 3 (triangles) and α = 8 \alpha= 8 α = 8 (circles), and lie on the solid curve E α \mathcal{E}_{\alpha} E α of the same colour (darker for larger α \alpha α ). The dashed curve is the objective f f f .
We will now make this precise by building directly on the developments in Fornasier et al. [5 ] . The cornerstone of this approach is a quantitative Laplace principle , which makes precise the intuition that for large α > 0 \alpha> 0 α > 0 we should have
E X ∼ N ( x , Σ ) [ exp ( − α f ( X ) ) X ] E X ∼ N ( x , Σ ) [ exp ( − α f ( X ) ) ] ≈ x ∗ (8) \frac{\mathbb{E}_{X \sim\mathcal{N}(x,\Sigma)}[\exp(-\alpha f(X))X]}{\mathbb{E}_{X \sim\mathcal{N}(x,\Sigma)}[\exp(-\alpha f(X))]} \approx x_\ast \tag*{(8)} E X ∼ N ( x , Σ ) [ exp ( − α f ( X ))] E X ∼ N ( x , Σ ) [ exp ( − α f ( X )) X ] ≈ x ∗ (8)
even if x x x is far away from x ∗ x_\ast x ∗ , cf. Section A.2.1 in the supplementary material for details. Such a result can be used to show that T α T_\alpha T α is a contraction on a region around x ∗ x_\ast x ∗ , which in turn leads to our first result: mean-field MPPI converges semi-globally . We would like to stress that this result is a direct generalisation of developments in Fornasier et al. [5 ] , and we do not claim any novelty in terms of proof techniques.
Theorem 1 Let f f f satisfy Assumption 1, and let Σ ∈ R d × d \Sigma\in\mathbb{R}^{d \times d} Σ ∈ R d × d be symmetric positive definite. For all r > 0 r > 0 r > 0 and ρ ∈ ( 0 , 1 ) \rho\in(0,1) ρ ∈ ( 0 , 1 ) there exists α 0 = α 0 ( f , Σ , r , ρ ) ≥ 0 \alpha_0 = \alpha_0(f,\Sigma,r,\rho) \ge0 α 0 = α 0 ( f , Σ , r , ρ ) ≥ 0 such that for all α > 0 \alpha> 0 α > 0 with α ≥ α 0 \alpha\ge\alpha_0 α ≥ α 0 the following hold.
(i) We have T α ( B r ( x ∗ ) ) ⊆ B r ( x ∗ ) T_\alpha(B_r(x_\ast)) \subseteq B_r(x_\ast) T α ( B r ( x ∗ )) ⊆ B r ( x ∗ ) , and
∥ T α ( x ) − T α ( x ′ ) ∥ 2 ≤ ρ ∥ x − x ′ ∥ 2 for all x , x ′ ∈ B r ( x ∗ ) . (9) \lVert T_\alpha(x) - T_\alpha(x') \rVert_2 \le\rho\lVert x - x' \rVert_2 \qquad\text{for all } x,x' \in B_r(x_\ast). \tag*{(9)} ∥ T α ( x ) − T α ( x ′ ) ∥ 2 ≤ ρ ∥ x − x ′ ∥ 2 for all x , x ′ ∈ B r ( x ∗ ) . (9) (ii) T α T_\alpha T α has exactly one fixed point m α m_\alpha m α in B r ( x ∗ ) B_r(x_\ast) B r ( x ∗ ) . For every x 0 ∈ B r ( x ∗ ) x_0 \in B_r(x_\ast) x 0 ∈ B r ( x ∗ ) , the mean-field iterates x k + 1 = T α ( x k ) x_{k+1} = T_\alpha(x_k) x k + 1 = T α ( x k ) satisfy x k ∈ B r ( x ∗ ) x_k \in B_r(x_\ast) x k ∈ B r ( x ∗ ) and
∥ x k − m α ∥ 2 ≤ ρ k ∥ x 0 − m α ∥ 2 for all k ∈ N 0 . (10) \lVert x_k - m_\alpha\rVert_2 \le\rho^k \lVert x_0 - m_\alpha\rVert_2 \qquad\text{for all } k \in\mathbb{N}_0. \tag*{(10)} ∥ x k − m α ∥ 2 ≤ ρ k ∥ x 0 − m α ∥ 2 for all k ∈ N 0 . (10)
Moreover, ∥ m α − x ∗ ∥ 2 → 0 \lVert m_{\alpha}-x_{\ast}\rVert_{2} \to0 ∥ m α − x ∗ ∥ 2 → 0 as α → ∞ \alpha\to\infty α → ∞ .
This result makes the notion of semi-global convergence precise: We cannot guarantee global convergence in the sense that for α ∈ R > 0 \alpha\in\mathbb{R}_{>0} α ∈ R > 0 large enough mean-field MPPI converges for all initial x 0 ∈ R d x_{0} \in\mathbb{R}^{d} x 0 ∈ R d . Rather, for all r ∈ R > 0 r \in\mathbb{R}_{>0} r ∈ R > 0 we can find a lower bound α 0 \alpha_{0} α 0 on the inverse temperature – depending only on r r r , the contraction factor ρ \rho ρ , properties of f f f as described in Assumption 1, and spectral properties of Σ \Sigma Σ – such that for all α ≥ α 0 \alpha\ge\alpha_{0} α ≥ α 0 and all x 0 ∈ B r ( x ∗ ) x_{0} \in B_{r}(x_{\ast}) x 0 ∈ B r ( x ∗ ) we have geometric convergence, uniformly in B r ( x ∗ ) B_{r}(x_{\ast}) B r ( x ∗ ) , to a fixed point m α m_{\alpha} m α . Furthermore, m α m_{\alpha} m α does not depend on r r r in the sense that if 0 < r ≤ r ′ 0 < r \le r' 0 < r ≤ r ′ and α \alpha α exceeds both corresponding thresholds, then the fixed point in B r ( x ∗ ) B_{r}(x_{\ast}) B r ( x ∗ ) coincides with the unique fixed point in B r ′ ( x ∗ ) B_{r'}(x_{\ast}) B r ′ ( x ∗ ) .
Theorem 1 does not yet quantify how close the iterates get to the global minimiser x ∗ x_{\ast} x ∗ . However, for the fixed point from above we have m α → x ∗ m_{\alpha} \to x_{\ast} m α → x ∗ for α → ∞ \alpha\to\infty α → ∞ , and this convergence can be made quantitative. This leads to our second main result: mean-field MPPI converges semi-globally to the global minimiser x ∗ x_{\ast} x ∗ . Similar results are contained in Fornasier et al. [5 ] under stronger assumptions, cf. Section A.2.3 for a discussion.
Theorem 2 Let f f f satisfy Assumption 1, and let Σ ∈ R d × d \Sigma\in\mathbb{R}^{d \times d} Σ ∈ R d × d be symmetric positive definite. For every β > 0 \beta> 0 β > 0 there exists α b = α b ( f , Σ , β ) ≥ 0 \alpha_{\mathrm{b}} = \alpha_{\mathrm{b}}(f,\Sigma,\beta) \ge0 α b = α b ( f , Σ , β ) ≥ 0 such that for all r > 0 r > 0 r > 0 , ρ ∈ ( 0 , 1 ) \rho\in(0,1) ρ ∈ ( 0 , 1 ) , and all α > 0 \alpha> 0 α > 0 with α ≥ max { α 0 ( f , Σ , r , ρ ) , α b ( f , Σ , β ) } \alpha\ge\max\{\alpha_{0}(f,\Sigma,r,\rho), \alpha_{\mathrm{b}}(f,\Sigma,\beta)\} α ≥ max { α 0 ( f , Σ , r , ρ ) , α b ( f , Σ , β )} , where α 0 \alpha_{0} α 0 is from Theorem 1, the following hold.
(i) For all x ∈ B r ( x ∗ ) x \in B_{r}(x_{\ast}) x ∈ B r ( x ∗ ) ,
∥ T α ( x ) − x ∗ ∥ 2 ≤ β + ρ ∥ x − x ∗ ∥ 2 . (11) \lVert T_{\alpha}(x)-x_{\ast}\rVert_{2} \le\beta+\rho\lVert x-x_{\ast}\rVert_{2}. \tag*{(11)} ∥ T α ( x ) − x ∗ ∥ 2 ≤ β + ρ ∥ x − x ∗ ∥ 2 . (11) (ii) For every x 0 ∈ B r ( x ∗ ) x_{0} \in B_{r}(x_{\ast}) x 0 ∈ B r ( x ∗ ) , the mean-field iterates satisfy
∥ x k − x ∗ ∥ 2 ≤ ρ k ∥ x 0 − x ∗ ∥ 2 + β 1 − ρ for all k ∈ N 0 . (12) \lVert x_{k}-x_{\ast}\rVert_{2} \le\rho^{k}\lVert x_{0}-x_{\ast}\rVert_{2}+\frac{\beta}{1-\rho}
\qquad\text{for all } k \in\mathbb{N}_{0}. \tag*{(12)} ∥ x k − x ∗ ∥ 2 ≤ ρ k ∥ x 0 − x ∗ ∥ 2 + 1 − ρ β for all k ∈ N 0 . (12)
Observe that the threshold α 0 ( f , Σ , r , ρ ) \alpha_{0}(f,\Sigma,r,\rho) α 0 ( f , Σ , r , ρ ) depends on the radius r r r , but not on the accuracy parameter β \beta β , while α b ( f , Σ , β ) \alpha_{\mathrm{b}}(f,\Sigma,\beta) α b ( f , Σ , β ) depends on β \beta β , but not on r r r . Interestingly, (11) shows that for large α \alpha α already one mean-field MPPI step can lead to good accuracy, which provides a partial theoretical justification, on the mean-field level, for the single-step variant originally proposed in Williams et al. [16 ] . However, (12) shows that iteration on the mean-field can indeed lead to improvements of the accuracy. This phenomenon has been pointed out already in (Fornasier et al., [5 ] , Remark 4.5), albeit under stronger assumptions. The preceding results imply in particular that for an arbitrary initial guess x 0 x_{0} x 0 an arbitrary accuracy can be achieved after enough iterations for a large enough α > 0 \alpha> 0 α > 0 , as made precise in the next result.
Corollary 3 Let f f f satisfy Assumption 1, and let Σ ∈ R d × d \Sigma\in\mathbb{R}^{d \times d} Σ ∈ R d × d be symmetric positive definite. Let r > 0 r > 0 r > 0 and ϵ > 0 \epsilon> 0 ϵ > 0 , and set K 0 ( r , ϵ ) : = max { 0 , ⌈ log 2 ( 2 r / ϵ ) ⌉ } K_{0}(r,\epsilon) := \max\{0,\lceil\log_{2}(2r/\epsilon)\rceil\} K 0 ( r , ϵ ) := max { 0 , ⌈ log 2 ( 2 r / ϵ )⌉} . Then for all α > 0 \alpha> 0 α > 0 with α ≥ max { α 0 ( f , Σ , r , 1 2 ) , α b ( f , Σ , ϵ 4 ) } \alpha\ge\max\{\alpha_{0}(f,\Sigma,r,\frac{1}{2}), \alpha_{\mathrm{b}}(f,\Sigma,\frac{\epsilon}{4})\} α ≥ max { α 0 ( f , Σ , r , 2 1 ) , α b ( f , Σ , 4 ϵ )} and all x 0 ∈ B r ( x ∗ ) x_{0} \in B_{r}(x_{\ast}) x 0 ∈ B r ( x ∗ ) , the mean-field iterates satisfy
∥ x k − x ∗ ∥ 2 ≤ ϵ for all k ≥ K 0 ( r , ϵ ) . (13) \lVert x_{k}-x_{\ast}\rVert_{2} \le\epsilon\qquad\text{for all } k \ge K_{0}(r,\epsilon). \tag*{(13)} ∥ x k − x ∗ ∥ 2 ≤ ϵ for all k ≥ K 0 ( r , ϵ ) . (13)
Note that the number of iterations K 0 ( r , ϵ ) K_{0}(r,\epsilon) K 0 ( r , ϵ ) depends neither on f f f nor on Σ \Sigma Σ nor on α \alpha α . We would like to stress that for these results it is enough that f f f fulfils Assumption 1. In particular, f f f might be severely non-convex and non-smooth.
Finally, under Assumption 2, the thresholds become explicit. We state them up to logarithmic factors and refer to Corollary 23 and Lemma 27 in the supplementary material for detailed statements.
Corollary 4 Let f f f satisfy (SF1) and (SF2) of Assumption 2, let Σ ∈ R d × d \Sigma\in\mathbb{R}^{d \times d} Σ ∈ R d × d be symmetric positive definite, and let r > 0 r > 0 r > 0 and ϵ ∈ ( 0 , 2 r ] \epsilon\in(0, 2r] ϵ ∈ ( 0 , 2 r ] . There is a threshold α m f S ( r , ϵ ) \alpha_{\mathrm{mf}}^{\mathrm{S}}(r,\epsilon) α mf S ( r , ϵ ) with
α m f S ( r , ϵ ) = O ~ ( d m ϵ 2 + ( r + R ) 2 Δ σ min ( Σ ) 2 + 1 m σ min ( Σ ) 2 + d Δ ) (14) \alpha_{\mathrm{mf}}^{\mathrm{S}}(r,\epsilon) = \widetilde{O}\left(\frac{d}{m\epsilon^{2}}+\frac{(r+R)^{2}}{\Delta\sigma_{\min}(\Sigma)^{2}}+\frac{1}{m\sigma_{\min}(\Sigma)^{2}}+\frac{d}{\Delta}\right) \tag*{(14)} α mf S ( r , ϵ ) = O ( m ϵ 2 d + Δ σ m i n ( Σ ) 2 ( r + R ) 2 + m σ m i n ( Σ ) 2 1 + Δ d ) (14) such that for all α ≥ α m f S ( r , ϵ ) \alpha\ge\alpha_{\mathrm{mf}}^{\mathrm{S}}(r,\epsilon) α ≥ α mf S ( r , ϵ ) and all x 0 ∈ B r ( x ∗ ) x_{0} \in B_{r}(x_{*}) x 0 ∈ B r ( x ∗ ) , the mean-field iterates satisfy ∥ x k − x ∗ ∥ 2 ≤ ϵ \lVert x_{k}-x_{*}\rVert_{2} \le\epsilon ∥ x k − x ∗ ∥ 2 ≤ ϵ for all k ≥ K 0 ( r , ϵ ) k \ge K_{0}(r,\epsilon) k ≥ K 0 ( r , ϵ ) .
Note that the accuracy and the radius enter the inverse temperature additively, where the dependence on the accuracy is of order d / ( m ϵ 2 ) d/(m\epsilon^{2}) d / ( m ϵ 2 ) , and the radius only enters through the term ( r + R ) 2 / ( Δ σ min ( Σ ) 2 ) (r+R)^{2}/(\Delta\sigma_{\min}(\Sigma)^{2}) ( r + R ) 2 / ( Δ σ m i n ( Σ ) 2 ) , which does not depend on ϵ \epsilon ϵ .
Finite-Sample Results
While the mean-field results from the preceding section are strong, it is a priori unclear how applicable they are to the actual finite-sample method, i.e., the MPPI iteration (1) . To investigate this issue, we now turn to finite-sample results. We focus on finite horizons, i.e., on a fixed number of iterations. To simplify the notation, for x ∈ R d x \in\mathbb{R}^{d} x ∈ R d define the empirical transition map
T ^ α , N ( x ) : = ∑ i = 1 N e − α f ( X i ) X i ∑ i = 1 N e − α f ( X i ) (15) \widehat{T}_{\alpha,N}(x) := \frac{\sum_{i=1}^{N}e^{-\alpha f(X_{i})}X_{i}}{\sum_{i=1}^{N}e^{-\alpha f(X_{i})}} \tag*{(15)} T α , N ( x ) := ∑ i = 1 N e − α f ( X i ) ∑ i = 1 N e − α f ( X i ) X i (15)
with X 1 , … , X N X_{1},\ldots,X_{N} X 1 , … , X N independent with X i ∼ N ( x , Σ ) X_{i} \sim\mathcal{N}(x,\Sigma) X i ∼ N ( x , Σ ) as before.
Fornasier et al. [5 ] suggest deriving finite-sample results via propagation of chaos results which are already available for CH (Haug [7 ] ). However, as pointed out above, it is well-known that T ^ α , N ( x ) \widehat{T}_{\alpha,N}(x) T α , N ( x ) is a self-normalised importance-weighted sampling estimate of T α ( x ) T_{\alpha}(x) T α ( x ) . From an abstract standpoint, this means that the MPPI or empirical CH iteration is not a full interacting particle system, but rather a time-homogeneous Markov chain with continuous state space. We can use this connection to directly derive finite-sample results without first establishing propagation of chaos, similar to earlier works like Yoon et al. [19 ] ; Yoon and Kim [18 ] . We start with a non-asymptotic error bound for the self-normalised importance sampling estimate T ^ α , N ( x ) \widehat{T}_{\alpha,N}(x) T α , N ( x ) of T α ( x ) T_{\alpha}(x) T α ( x ) . While we are not aware of this exact result in the literature, such bounds are well-known in the context of self-normalised estimators, see for example (Agapiou et al. [1 ] , Theorem 2.1).
Proposition 5 Let f f f satisfy Assumption 1, let Σ ∈ R d × d \Sigma\in\mathbb{R}^{d \times d} Σ ∈ R d × d be symmetric positive definite, and let α > 0 \alpha> 0 α > 0 . Then for all x ∈ R d x \in\mathbb{R}^{d} x ∈ R d , N ∈ N + N \in\mathbb{N}_{+} N ∈ N + and t > 0 t > 0 t > 0 ,
P ( ∥ T ^ α , N ( x ) − T α ( x ) ∥ 2 > t ) ≤ 4 N 1 ζ α ( x ) ( 1 + V α ( x ) t 2 ) (16) \mathbb{P}\left(\left\lVert\widehat{T}_{\alpha,N}(x)-T_{\alpha}(x)\right\rVert_{2}>t\right)\leq\frac{4}{N}\frac{1}{\zeta_{\alpha}(x)}\left(1+\frac{V_{\alpha}(x)}{t^{2}}\right) \tag*{(16)} P ( T α , N ( x ) − T α ( x ) 2 > t ) ≤ N 4 ζ α ( x ) 1 ( 1 + t 2 V α ( x ) ) (16) where
ζ α ( x ) : = E X ∼ N ( x , Σ ) [ e − α ( f ( X ) − f ( x ∗ ) ) ] ∈ ( 0 , 1 ] , \zeta_{\alpha}(x):=\mathbb{E}_{X\sim\mathcal{N}(x,\Sigma)}\left[e^{-\alpha(f(X)-f(x_{*}))}\right]\in(0,1], ζ α ( x ) := E X ∼ N ( x , Σ ) [ e − α ( f ( X ) − f ( x ∗ )) ] ∈ ( 0 , 1 ] , V α ( x ) : = E X ∼ N ( x , Σ ) [ e − α f ( X ) ∥ X − T α ( x ) ∥ 2 2 ] Z α ( x ) < ∞ . (17) V_{\alpha}(x):=\frac{\mathbb{E}_{X\sim\mathcal{N}(x,\Sigma)}\left[e^{-\alpha f(X)}\lVert X-T_{\alpha}(x)\rVert_{2}^{2}\right]}{Z_{\alpha}(x)}<\infty. \tag*{(17)} V α ( x ) := Z α ( x ) E X ∼ N ( x , Σ ) [ e − α f ( X ) ∥ X − T α ( x ) ∥ 2 2 ] < ∞. (17)
The factor 1 / ζ α ( x ) 1/\zeta_{\alpha}(x) 1/ ζ α ( x ) plays the role of an inverse effective sample size. It is nondecreasing in the inverse temperature α \alpha α and typically grows with the distance of x x x to x ∗ x_{*} x ∗ , since then fewer samples carry a substantial weight. We now combine Proposition 5 with the semi-global mean-field convergence result Theorem 2 . The key observation for establishing the following result is that the sampling error only needs to be small compared with the current distance to x ∗ x_{*} x ∗ . Hence steps far away from x ∗ x_{*} x ∗ only require a coarse accuracy, and the fine accuracy only has to be achieved close to x ∗ x_{*} x ∗ .
Theorem 6. Let f f f satisfy Assumption 1, let Σ ∈ R d × d \Sigma\in\mathbb{R}^{d \times d} Σ ∈ R d × d be symmetric positive definite, and let ( x ^ k ) k ∈ N 0 (\hat{x}_{k})_{k\in\mathbb{N}_{0}} ( x ^ k ) k ∈ N 0 be the MPPI iterates (1). Let r > 0 r > 0 r > 0 , ρ ∈ ( 0 , 1 ) \rho\in(0,1) ρ ∈ ( 0 , 1 ) and β > 0 \beta> 0 β > 0 with 4 β ≤ ( 1 − ρ ) r 4\beta\le(1-\rho)r 4 β ≤ ( 1 − ρ ) r , let γ > 0 \gamma> 0 γ > 0 , and set ρ ′ : = ( 1 + ρ ) / 2 \rho' := (1+\rho)/2 ρ ′ := ( 1 + ρ ) /2 . There exist α 0 ′ = α 0 ′ ( f , Σ , r , ρ ) \alpha'_0=\alpha'_0(f,\Sigma,r,\rho) α 0 ′ = α 0 ′ ( f , Σ , r , ρ ) , α b ′ = α b ′ ( f , Σ , β ) \alpha'_b=\alpha'_b(f,\Sigma,\beta) α b ′ = α b ′ ( f , Σ , β ) and C = C ( f , Σ , r , γ ) C=C(f,\Sigma,r,\gamma) C = C ( f , Σ , r , γ ) such that for all α > 0 \alpha>0 α > 0 with α ≥ max { α 0 ′ , α b ′ } \alpha\ge\max\{\alpha'_0,\alpha'_b\} α ≥ max { α 0 ′ , α b ′ } , all K ∈ N + K\in\mathbb{N}_{+} K ∈ N + , δ ∈ ( 0 , 1 ) \delta\in(0,1) δ ∈ ( 0 , 1 ) , N ∈ N + N\in\mathbb{N}_{+} N ∈ N + with
N ≥ K δ C e γ α , (18) N \ge\frac{K}{\delta} C e^{\gamma\alpha}, \tag*{(18)} N ≥ δ K C e γ α , (18) and all x ^ 0 ∈ B r ( x ∗ ) \hat{x}_{0}\in B_{r}(x_{*}) x ^ 0 ∈ B r ( x ∗ ) , we have
P ( x ^ k ∈ B r ( x ∗ ) and ∥ x ^ k − x ∗ ∥ 2 ≤ ρ ′ k ∥ x ^ 0 − x ∗ ∥ 2 + 4 β 1 − ρ for all k = 0 , … , K ) ≥ 1 − δ . (19) \mathbb{P}\left(\hat{x}_{k}\in B_{r}(x_{*})\ \text{and}\ \|\hat{x}_{k}-x_{*}\|_{2}\le\rho'^{k}\|\hat{x}_{0}-x_{*}\|_{2}+\frac{4\beta}{1-\rho}\ \text{for all } k=0,\ldots,K\right)\ge1-\delta. \tag*{(19)} P ( x ^ k ∈ B r ( x ∗ ) and ∥ x ^ k − x ∗ ∥ 2 ≤ ρ ′ k ∥ x ^ 0 − x ∗ ∥ 2 + 1 − ρ 4 β for all k = 0 , … , K ) ≥ 1 − δ . (19) In particular, let ϵ ∈ ( 0 , 2 r ] \epsilon\in(0,2r] ϵ ∈ ( 0 , 2 r ] and K ϵ : = max { 0 , ⌈ log ( 2 r / ϵ ) / log ( 4 / 3 ) ⌉ } K_{\epsilon}:=\max\{0,\lceil\log(2r/\epsilon)/\log(4/3)\rceil\} K ϵ := max { 0 , ⌈ log ( 2 r / ϵ ) / log ( 4/3 )⌉} . Then for all α > 0 \alpha>0 α > 0 with α ≥ max { α 0 ′ ( f , Σ , r , 1 2 ) , α b ′ ( f , Σ , ϵ 16 ) } \alpha\ge\max\{\alpha'_0(f,\Sigma,r,\frac{1}{2}),\alpha'_b(f,\Sigma,\frac{\epsilon}{16})\} α ≥ max { α 0 ′ ( f , Σ , r , 2 1 ) , α b ′ ( f , Σ , 16 ϵ )} , all K ∈ N + K\in\mathbb{N}_{+} K ∈ N + , δ ∈ ( 0 , 1 ) \delta\in(0,1) δ ∈ ( 0 , 1 ) , N N N as in (18) , and x ^ 0 ∈ B r ( x ∗ ) \hat{x}_{0}\in B_{r}(x_{*}) x ^ 0 ∈ B r ( x ∗ ) ,
P ( ∥ x ^ k − x ∗ ∥ 2 ≤ ϵ for all k with K ϵ ≤ k ≤ K ) ≥ 1 − δ . (20) \mathbb{P}\left(\|\hat{x}_{k}-x_{*}\|_{2}\le\epsilon\ \text{for all } k\text{ with }K_{\epsilon}\le k\le K\right)\ge1-\delta. \tag*{(20)} P ( ∥ x ^ k − x ∗ ∥ 2 ≤ ϵ for all k with K ϵ ≤ k ≤ K ) ≥ 1 − δ . (20)
As in the mean-field case, the inverse temperature thresholds are decoupled: α 0 ′ \alpha'_0 α 0 ′ depends on the radius r r r , but not on the accuracy, and α b ′ \alpha'_b α b ′ depends on the accuracy, but not on r r r . The sample size (18) grows linearly in K / δ K/\delta K / δ and like e γ α e^{\gamma\alpha} e γ α in α \alpha α , where γ > 0 \gamma>0 γ > 0 can be chosen arbitrarily small at the price of a larger constant C C C . Note that since the constant C C C depends on r r r , the sample size still couples the radius with the inverse temperature, and hence with the final accuracy. In particular, also this result is semi-global like Theorem 2 .
Under Assumption 2, the sample size becomes explicit as well. The precise statements are Corollary 26 and Lemma 27 in the supplementary material.
Corollary 7. Let f f f satisfy Assumption 2, let Σ ∈ R d × d \Sigma\in\mathbb{R}^{d \times d} Σ ∈ R d × d be symmetric positive definite, and let r > 0 r>0 r > 0 and ϵ ∈ ( 0 , 2 r ] \epsilon\in(0,2r] ϵ ∈ ( 0 , 2 r ] . There is a threshold α f s ( r , ϵ ) \alpha_{\mathrm{fs}}(r,\epsilon) α fs ( r , ϵ ) of the same order (14) as in Corollary 4 such that for all α ≥ α f s ( r , ϵ ) \alpha\ge\alpha_{\mathrm{fs}}(r,\epsilon) α ≥ α fs ( r , ϵ ) , K ∈ N + K\in\mathbb{N}_{+} K ∈ N + , δ ∈ ( 0 , 1 ) \delta\in(0,1) δ ∈ ( 0 , 1 ) , N ∈ N + N\in\mathbb{N}_{+} N ∈ N + with
N ≥ 12 K δ ( 1 + α M σ max ( Σ ) 2 ) d / 2 exp ( r 2 2 σ min ( Σ ) 2 ) , (21) N \ge\frac{12K}{\delta}\left(1+\alpha M\sigma_{\max}(\Sigma)^{2}\right)^{d/2}\exp\left(\frac{r^{2}}{2\sigma_{\min}(\Sigma)^{2}}\right), \tag*{(21)} N ≥ δ 12 K ( 1 + α M σ m a x ( Σ ) 2 ) d /2 exp ( 2 σ m i n ( Σ ) 2 r 2 ) , (21) and all x ^ 0 ∈ B r ( x ∗ ) \hat{x}_{0}\in B_{r}(x_{*}) x ^ 0 ∈ B r ( x ∗ ) , the MPPI iterates satisfy
P ( ∥ x ^ k − x ∗ ∥ 2 ≤ ϵ for all k with K 0 ( r , ϵ ) ≤ k ≤ K ) ≥ 1 − δ . \mathbb{P}\left(\|\hat{x}_{k}-x_{*}\|_{2}\le\epsilon\ \text{for all } k\text{ with }K_{0}(r,\epsilon)\le k\le K\right)\ge1-\delta. P ( ∥ x ^ k − x ∗ ∥ 2 ≤ ϵ for all k with K 0 ( r , ϵ ) ≤ k ≤ K ) ≥ 1 − δ .
For small ϵ \epsilon ϵ , the threshold is dominated by the term 384 d / ( m ϵ 2 ) 384d/(m\epsilon^{2}) 384 d / ( m ϵ 2 ) , cf. Lemma 27 in the supplementary material, and for α \alpha α of this order the sample size (21) is of order
K δ e r 2 / ( 2 σ min ( Σ ) 2 ) ( 384 d M σ max ( Σ ) 2 m ϵ 2 ) d / 2 , \frac{K}{\delta}e^{r^{2}/(2\sigma_{\min}(\Sigma)^{2})}\left(\frac{384dM\sigma_{\max}(\Sigma)^{2}}{m\epsilon^{2}}\right)^{d/2}, δ K e r 2 / ( 2 σ m i n ( Σ ) 2 ) ( m ϵ 2 384 d M σ m a x ( Σ ) 2 ) d /2 ,
i.e., it grows like e − d e^{-d} e − d in the accuracy.
Applications to Optimal Control
As a concrete illustration of the developments above, let us consider the classic finite-horizon LQR problem. We consider a deterministic LTI system
s + = A s + B a , (22) s_{+}=As+Ba, \tag*{(22)} s + = A s + B a , (22)
where A ∈ R n × n A \in\mathbb{R}^{n \times n} A ∈ R n × n is the state transition matrix, B ∈ R n × p B \in\mathbb{R}^{n \times p} B ∈ R n × p the control input matrix, s ∈ R n s \in\mathbb{R}^{n} s ∈ R n the current state, s + s_{+} s + the next state, and a ∈ R p a \in\mathbb{R}^{p} a ∈ R p the control input. For an initial state s 0 ∈ R n s_{0} \in\mathbb{R}^{n} s 0 ∈ R n , a horizon T ∈ N + T \in\mathbb{N}_{+} T ∈ N + and a control sequence a = ( a 0 , … , a T − 1 ) ∈ ( R p ) T a=(a_{0},\ldots,a_{T-1}) \in(\mathbb{R}^{p})^{T} a = ( a 0 , … , a T − 1 ) ∈ ( R p ) T , let s ( t ; s 0 , a ) s(t;s_{0},a) s ( t ; s 0 , a ) denote the state at time t = 0 , … , T t=0,\ldots,T t = 0 , … , T , i.e., s ( 0 ; s 0 , a ) = s 0 s(0;s_{0},a)=s_{0} s ( 0 ; s 0 , a ) = s 0 and s ( t + 1 ; s 0 , a ) = A s ( t ; s 0 , a ) + B a t s(t+1;s_{0},a)=As(t;s_{0},a)+Ba_{t} s ( t + 1 ; s 0 , a ) = A s ( t ; s 0 , a ) + B a t for t = 0 , … , T − 1 t=0,\ldots,T-1 t = 0 , … , T − 1 . If no confusion can arise, we also write s t = s ( t ; s 0 , a ) s_{t}=s(t;s_{0},a) s t = s ( t ; s 0 , a ) . Note that we use RL-type notation to avoid a notation clash with our preceding analysis.
For horizon T ∈ N + T \in\mathbb{N}_{+} T ∈ N + , consider the total cost functional
J T ( s 0 , a ) = ∑ t = 0 T − 1 s t T Q s t + a t T R a t + 2 s t T S a t (23) J_{T}(s_{0},a)=\sum_{t=0}^{T-1}s_{t}^{\mathsf{T}}Qs_{t}+a_{t}^{\mathsf{T}}Ra_{t}+2s_{t}^{\mathsf{T}}Sa_{t} \tag*{(23)} J T ( s 0 , a ) = t = 0 ∑ T − 1 s t T Q s t + a t T R a t + 2 s t T S a t (23)
for matrices Q ∈ R n × n Q \in\mathbb{R}^{n \times n} Q ∈ R n × n , R ∈ R p × p R \in\mathbb{R}^{p \times p} R ∈ R p × p and S ∈ R n × p S \in\mathbb{R}^{n \times p} S ∈ R n × p . We are interested in the optimisation problem
min a ∈ ( R p ) T J T ( s 0 , a ) . (24) \min_{a \in(\mathbb{R}^{p})^{T}} J_{T}(s_{0},a). \tag*{(24)} a ∈ ( R p ) T min J T ( s 0 , a ) . (24)
Without loss of generality, we assume that Q Q Q and R R R are symmetric, since only the quadratic forms s T Q s s^{\mathsf{T}}Qs s T Q s and a T R a a^{\mathsf{T}}Ra a T R a enter the cost. As is well-known, the objective function J T J_{T} J T can be rewritten as
J T ( s 0 , a ) = a T H T a + 2 a T F T s 0 + s 0 T Y T s 0 , (25) J_{T}(s_{0},a)=a^{\mathsf{T}}H_{T}a+2a^{\mathsf{T}}F_{T}s_{0}+s_{0}^{\mathsf{T}}Y_{T}s_{0}, \tag*{(25)} J T ( s 0 , a ) = a T H T a + 2 a T F T s 0 + s 0 T Y T s 0 , (25)
where the control sequence a a a is identified with the stacked vector ( a 0 T , … , a T − 1 T ) T ∈ R p T (a_{0}^{\mathsf{T}},\ldots,a_{T-1}^{\mathsf{T}})^{\mathsf{T}} \in\mathbb{R}^{pT} ( a 0 T , … , a T − 1 T ) T ∈ R pT , and H T ∈ R p T × p T H_{T} \in\mathbb{R}^{pT \times pT} H T ∈ R pT × pT (symmetric), F T ∈ R p T × n F_{T} \in\mathbb{R}^{pT \times n} F T ∈ R pT × n and Y T ∈ R n × n Y_{T} \in\mathbb{R}^{n \times n} Y T ∈ R n × n depend only on A A A , B B B , Q Q Q , R R R , S S S and T T T , cf. Section A.5 in the supplementary material. For simplicity, assume that ( Q S S T R ) \begin{pmatrix} Q & S \\ S^{\mathsf{T}} & R \end{pmatrix} ( Q S T S R ) is positive definite (which implies that R R R is positive definite), so that H T H_{T} H T is also positive definite, cf. Lemma 29 in the supplementary material, and hence (24) has the unique solution a ∗ ( s 0 ) = − H T − 1 F T s 0 a_{*}(s_{0})=-H_{T}^{-1}F_{T}s_{0} a ∗ ( s 0 ) = − H T − 1 F T s 0 . Finally, let Σ c ∈ R p × p \Sigma_{c} \in\mathbb{R}^{p \times p} Σ c ∈ R p × p be symmetric positive definite, so that also Σ = diag ( Σ c , … , Σ c ) \Sigma=\operatorname{diag}(\Sigma_{c},\ldots,\Sigma_{c}) Σ = diag ( Σ c , … , Σ c ) is symmetric positive definite. For fixed s 0 ∈ R n s_{0} \in\mathbb{R}^{n} s 0 ∈ R n we can then set d = p T d=pT d = pT and apply the MPPI iteration to the objective function f = J T ( s 0 , ⋅ ) f=J_{T}(s_{0},\cdot) f = J T ( s 0 , ⋅ ) with Σ \Sigma Σ from above and inverse temperature α \alpha α . Let us write a ∗ : = a ∗ ( s 0 ) a_{*}:=a_{*}(s_{0}) a ∗ := a ∗ ( s 0 ) , a ˉ k + 1 : = T α ( a ˉ k ) \bar{a}_{k+1}:=T_{\alpha}(\bar{a}_{k}) a ˉ k + 1 := T α ( a ˉ k ) for the mean-field iterates, and ( a ^ k ) k ∈ N 0 (\hat{a}_{k})_{k \in\mathbb{N}_{0}} ( a ^ k ) k ∈ N 0 for the MPPI iterates (1) with this objective f f f .
Corollary 8 (Explicit results for the LQR problem). Let r > 0 r>0 r > 0 and ϵ ∈ ( 0 , 2 r ] \epsilon\in(0,2r] ϵ ∈ ( 0 , 2 r ] .
(i) For all α > 16 p T λ min ( H T ) ϵ 2 + 6 λ min ( H T ) σ min ( Σ c ) 2 \alpha> \frac{16pT}{\lambda_{\min}(H_{T})\epsilon^{2}}+\frac{6}{\lambda_{\min}(H_{T})\sigma_{\min}(\Sigma_{c})^{2}} α > λ m i n ( H T ) ϵ 2 16 pT + λ m i n ( H T ) σ m i n ( Σ c ) 2 6 and a ˉ 0 ∈ B r ( a ∗ ) \bar{a}_{0} \in B_{r}(a_{*}) a ˉ 0 ∈ B r ( a ∗ ) , the mean-field iterates satisfy
∥ a ˉ k − a ∗ ∥ 2 ≤ ϵ for all k ≥ K 0 ( r , ϵ ) . \lVert\bar{a}_{k}-a_{*}\rVert_{2}\leq\epsilon\text{ for all } k\geq K_{0}(r,\epsilon). ∥ a ˉ k − a ∗ ∥ 2 ≤ ϵ for all k ≥ K 0 ( r , ϵ ) . (ii) For all α > 192 p T λ min ( H T ) ϵ 2 + 10 λ min ( H T ) σ min ( Σ c ) 2 \alpha> \frac{192pT}{\lambda_{\min}(H_{T})\epsilon^{2}}+\frac{10}{\lambda_{\min}(H_{T})\sigma_{\min}(\Sigma_{c})^{2}} α > λ m i n ( H T ) ϵ 2 192 pT + λ m i n ( H T ) σ m i n ( Σ c ) 2 10 , K ∈ N + K \in\mathbb{N}_{+} K ∈ N + , δ ∈ ( 0 , 1 ) \delta\in(0,1) δ ∈ ( 0 , 1 ) , N ∈ N + N \in\mathbb{N}_{+} N ∈ N + with
N ≥ 12 K δ ( 1 + 2 α λ max ( H T ) σ max ( Σ c ) 2 ) p T / 2 exp ( r 2 2 σ min ( Σ c ) 2 ) , (26) N\geq\frac{12K}{\delta}\left(1+2\alpha\lambda_{\max}(H_{T})\sigma_{\max}(\Sigma_{c})^{2}\right)^{pT/2}\exp\left(\frac{r^{2}}{2\sigma_{\min}(\Sigma_{c})^{2}}\right), \tag*{(26)} N ≥ δ 12 K ( 1 + 2 α λ m a x ( H T ) σ m a x ( Σ c ) 2 ) pT /2 exp ( 2 σ m i n ( Σ c ) 2 r 2 ) , (26) and a ^ 0 ∈ B r ( a ∗ ) \hat{a}_{0} \in B_{r}(a_{*}) a ^ 0 ∈ B r ( a ∗ ) , the MPPI iterates satisfy
P ( ∥ a ^ k − a ∗ ∥ 2 ≤ ϵ for all k with K 0 ( r , ϵ ) ≤ k ≤ K ) ≥ 1 − δ . \mathbb{P}\left(\lVert\hat{a}_{k}-a_{*}\rVert_{2}\leq\epsilon\text{ for all } k \text{ with } K_{0}(r,\epsilon)\leq k\leq K\right)\geq1-\delta. P ( ∥ a ^ k − a ∗ ∥ 2 ≤ ϵ for all k with K 0 ( r , ϵ ) ≤ k ≤ K ) ≥ 1 − δ .
For a quadratic objective, the radius r r r does not enter the inverse temperature, since the well of Assumption 2 can be taken arbitrarily large. For small ϵ \epsilon ϵ and α \alpha α at the threshold in (ii), the sample size (26) is of order K δ e r 2 / ( 2 σ min ( Σ c ) 2 ) ( 384 p T κ σ max ( Σ c ) 2 ϵ 2 ) p T / 2 \frac{K}{\delta}e^{r^2/(2\sigma_{\min}(\Sigma_c)^2)}\left(384\,pT\kappa\frac{\sigma_{\max}(\Sigma_c)^2}{\epsilon^2}\right)^{pT/2} δ K e r 2 / ( 2 σ m i n ( Σ c ) 2 ) ( 384 pT κ ϵ 2 σ m a x ( Σ c ) 2 ) pT /2 with κ : = λ max ( H T ) / λ min ( H T ) \kappa:= \lambda_{\max}(H_T)/\lambda_{\min}(H_T) κ := λ m a x ( H T ) / λ m i n ( H T ) , i.e., the exponent of 1 / ϵ 1/\epsilon 1/ ϵ grows linearly in the horizon T T T . Since λ min ( H T ) ≥ λ min ( ( Q S S ⊤ R ) ) \lambda_{\min}(H_T) \ge\lambda_{\min}\left(\begin{pmatrix} Q & S \\ S^{\top} & R \end{pmatrix}\right) λ m i n ( H T ) ≥ λ m i n ( ( Q S ⊤ S R ) ) by Lemma 29 , the thresholds on α \alpha α remain valid with this lower bound, and then depend on the dynamics only through the dimension p T pT pT . By the same lemma, κ \kappa κ is bounded uniformly in T T T if A A A is Schur stable. Finally, we would like to stress that for a quadratic objective function as arising in the LQR problem, one can conduct a more tailored analysis, as has been done in related work like Yi et al. [17 ] or Yoon and Kim [18 ] . Since we use the LQR problem as a mere illustration, we do not pursue this here.
Conclusion
We investigated MPPI as a sampling-based global optimisation method for non-convex and non-smooth objectives. For objective functions with a unique global minimum, under weak additional assumptions, we established semi-global convergence to the minimiser of the mean-field MPPI iteration, as well as semi-global finite-sample results for the MPPI iteration, both in the low-temperature setting. To the best of our knowledge, these are the first such results for general non-smooth and non-convex objective functions. We would like to stress that we considered only the low-temperature (large inverse temperature) setting. We expect that for general objective functions as considered here, no reliable semi-global convergence can be shown at low inverse temperatures, but for restricted objective classes such results should be possible. Furthermore, we point out again that our finite-sample results are based on effective, but simple error bounds for the self-normalised importance weighted estimator, and we expect that with more sophisticated techniques and for more restricted function classes better results can be achieved. Regarding the practical applicability of our results, we stress that our results are quantitative and require only reasonable assumptions, but we do not claim that the various constants necessary to apply the results can be easily found. Rather, our work contributes to a better conceptual understanding of MPPI, and similarly to no-regret results in the bandit literature, our results show that there are algorithmic parameters (in particular, inverse temperature thresholds and sample bounds) that are benign and that ensure the algorithmic goal (semi-global convergence) is achieved. Let us close by pointing out two important directions for future work. First, we focused on the popular plain MPPI iteration (1) in the low-temperature setting. However, there is by now a multitude of MPPI variants (Kazin et al., 2024) and also other scaling limits as in Homburger et al. [8 ] . Investigating more variants with our techniques is therefore an interesting and relevant problem. Second, we focused on the optimisation behaviour of MPPI. Since its main use is in control applications, it is important to investigate control-theoretic properties that make our results applicable to the optimal control objective function, similar to our illustration in § 5 for LQR. Furthermore, the suboptimality needs to be accounted for in the control scheme. We expect that our results will be an important step towards such results.
Generative AI Use
The present project and the high-level approach have been conceived by the authors without AI use. The mean-field analysis is based on initial work by the authors without AI use, then refined using Claude Opus 5.5 (in particular, the reduction to Assumption 1 was found by this AI model), with final editing and rewriting by the authors. The finite-sample results as well as the results based on the simplified Assumption 2 have been developed interactively with ChatGPT-6, and the write-up of technical intermediate results as well as their proofs was supported by Claude Opus 5.5. The LQR example was developed by the authors with the writeup supported by Claude Opus 5.5. All illustrations have been conceptualised by the authors without AI, and then typeset and programmed with Claude Opus 5.5. The main text has been written by the authors; several parts (in particular Assumption 2 and its discussion, Corollaries 4 , 7 and 8 , parts of § 5 , and the figure captions) were drafted with Claude Opus 5.5 and subsequently revised by the authors.
Acknowledgments
We would like to thank Jona Klemenc for very helpful discussions, especially related to Fornasier et al. [5 ] . C.F. acknowledges funding from DFG Project FO 767/10-2 (eBer-24-32734) “Implicit Bias in Adversarial Training”. G.B. acknowledges support from the Munich Center for Machine Learning.
Figure 3.
Funded by the European Union. Views and opinions expressed are however those of the author(s) only and do not necessarily reflect those of the European Union or the European Research Council Executive Agency. Neither the European Union nor the granting authority can be held responsible for them. This project has received funding from the European Research Council (ERC) under the European Union’s Horizon Europe research and innovation programme (grant agreement No. 101198055, project acronym NEITALG).
References
Supplementary Material
Notation. We denote the Euclidean norm of vectors by ∥ ⋅ ∥ 2 \lVert\cdot\rVert_{2} ∥ ⋅ ∥ 2 and the spectral norm of matrices by ∥ ⋅ ∥ \lVert\cdot\rVert ∥ ⋅ ∥ . For x ∈ R d x \in\mathbb{R}^{d} x ∈ R d and r ∈ R > 0 r \in\mathbb{R}_{>0} r ∈ R > 0 , we define B r ( x ) = { x ′ ∈ R d ∣ ∥ x − x ′ ∥ 2 ≤ r } B_{r}(x) = \{x^{\prime} \in\mathbb{R}^{d} \mid\lVert x - x^{\prime} \rVert_{2} \le r\} B r ( x ) = { x ′ ∈ R d ∣ ∥ x − x ′ ∥ 2 ≤ r } . We denote the set of Borel probability measures on R d \mathbb{R}^{d} R d by P ( R d ) \mathcal{P}(\mathbb{R}^{d}) P ( R d ) , and with measurable we always mean Borel measurable. We denote the Gaussian measure on R d \mathbb{R}^{d} R d with mean x x x and (symmetric positive definite) covariance matrix Σ \Sigma Σ by N ( x , Σ ) \mathcal{N}(x, \Sigma) N ( x , Σ ) , and the corresponding Gaussian density evaluated at y ∈ R d y \in\mathbb{R}^{d} y ∈ R d by N ( y ∣ x , Σ ) \mathcal{N}(y \mid x, \Sigma) N ( y ∣ x , Σ ) . For a symmetric matrix A ∈ R n × n A \in\mathbb{R}^{n \times n} A ∈ R n × n we write λ min ( A ) \lambda_{\min}(A) λ m i n ( A ) and λ max ( A ) \lambda_{\max}(A) λ m a x ( A ) for the smallest and largest eigenvalue, respectively (note that eigenvalues of a symmetric matrix are necessarily real). For a symmetric positive definite matrix Σ \Sigma Σ , we write σ min ( Σ ) : = λ min ( Σ ) 1 / 2 \sigma_{\min}(\Sigma) := \lambda_{\min}(\Sigma)^{1/2} σ m i n ( Σ ) := λ m i n ( Σ ) 1/2 and σ max ( Σ ) : = λ max ( Σ ) 1 / 2 \sigma_{\max}(\Sigma) := \lambda_{\max}(\Sigma)^{1/2} σ m a x ( Σ ) := λ m a x ( Σ ) 1/2 for the square roots of its smallest and largest eigenvalue, and κ Σ : = σ max ( Σ ) 2 / σ min ( Σ ) 2 \kappa_{\Sigma} := \sigma_{\max}(\Sigma)^{2}/\sigma_{\min}(\Sigma)^{2} κ Σ := σ m a x ( Σ ) 2 / σ m i n ( Σ ) 2 . In particular, ∥ Σ ∥ = σ max ( Σ ) 2 \lVert\Sigma\rVert= \sigma_{\max}(\Sigma)^{2} ∥ Σ ∥ = σ m a x ( Σ ) 2 and ∥ Σ − 1 ∥ = σ min ( Σ ) − 2 \lVert\Sigma^{-1} \rVert= \sigma_{\min}(\Sigma)^{-2} ∥ Σ − 1 ∥ = σ m i n ( Σ ) − 2 . For ν ∈ P ( R d ) \nu\in\mathcal{P}(\mathbb{R}^{d}) ν ∈ P ( R d ) and α > 0 \alpha> 0 α > 0 , we write Z α ( ν ) : = ∫ R d e − α f ( y ) ν ( d y ) Z_{\alpha}(\nu) := \int_{\mathbb{R}^{d}} e^{-\alpha f(y)} \nu(\mathrm{d}y) Z α ( ν ) := ∫ R d e − α f ( y ) ν ( d y ) , so that Z α ( x ) = Z α ( N ( x , Σ ) ) Z_{\alpha}(x) = Z_{\alpha}(\mathcal{N}(x, \Sigma)) Z α ( x ) = Z α ( N ( x , Σ )) in the notation of the main text; in Lemma 9, where ν \nu ν is fixed, we simply write Z α Z_{\alpha} Z α . Finally, as usual the Landau symbol O ~ ( ⋅ ) \widetilde{O}(\cdot) O ( ⋅ ) suppresses poly-logarithmic terms.
Details for Example 1
Let us verify that the function f ( y ) = min { 1 2 y 2 , 1 + 1 2 ( y − 3 ) 2 } f(y) = \min\left\{\frac{1}{2}y^{2}, 1 + \frac{1}{2}(y - 3)^{2}\right\} f ( y ) = min { 2 1 y 2 , 1 + 2 1 ( y − 3 ) 2 } fulfils Assumption 1. Clearly, f f f is continuous and has a unique global minimum, so (F1) and (F2) hold. Moreover, if ∣ y ∣ ≥ ϵ |y| \ge\epsilon ∣ y ∣ ≥ ϵ , then either y ≤ 11 / 6 y \le11/6 y ≤ 11/6 and f ( y ) = 1 2 y 2 ≥ 1 2 ϵ 2 f(y) = \frac{1}{2}y^{2} \ge\frac{1}{2}\epsilon^{2} f ( y ) = 2 1 y 2 ≥ 2 1 ϵ 2 , or y > 11 / 6 y > 11/6 y > 11/6 and f ( y ) ≥ 1 f(y) \ge1 f ( y ) ≥ 1 . Hence (F3) holds with S ϵ = min { ϵ 2 / 2 , 1 } S_{\epsilon} = \min\{\epsilon^{2}/2, 1\} S ϵ = min { ϵ 2 /2 , 1 } , and f f f satisfies Assumption 1, although it is neither convex nor differentiable.
Figure 2 shows the mean-field MPPI iterates x k + 1 = T α ( x k ) x_{k+1} = T_{\alpha}(x_{k}) x k + 1 = T α ( x k ) , k = 0 , … , 11 k = 0, \ldots, 11 k = 0 , … , 11 , started at x 0 ∈ { − 1 , 2 , 4 } x_{0} \in\{-1, 2, 4\} x 0 ∈ { − 1 , 2 , 4 } . For Σ = 0.35 2 \Sigma= 0.35^{2} Σ = 0.3 5 2 , the iterates started at x 0 = 4 x_{0} = 4 x 0 = 4 converge to the fixed point near the local minimiser 3 for all three values of α \alpha α , in line with the second well of E α E_{\alpha} E α in Figure 1 . Started at x 0 = 2 x_{0} = 2 x 0 = 2 , the iterates reach a neighbourhood of x ∗ x_{*} x ∗ for α ∈ { 3 , 8 } \alpha\in\{3, 8\} α ∈ { 3 , 8 } , whereas for α = 1 \alpha= 1 α = 1 they have barely moved after twelve steps (x 12 ≈ 1.97 x_{12} \approx1.97 x 12 ≈ 1.97 ). For α = 1 \alpha= 1 α = 1 , the small step size α Σ ≈ 0.12 \alpha\Sigma\approx0.12 α Σ ≈ 0.12 in (6) also slows down the iterates started at x 0 = − 1 x_{0} = -1 x 0 = − 1 . For Σ = 1 \Sigma= 1 Σ = 1 , all iterates approach x ∗ x_{*} x ∗ . For α = 1 \alpha= 1 α = 1 , the iterates started at − 1 -1 − 1 and 2 settle at about 0.04, while the one started at 4 is still at about 0.2 after twelve steps; for α = 3 \alpha= 3 α = 3 , all iterates settle at about 0.004. This remaining offset reflects that the fixed point m α m_{\alpha} m α differs from x ∗ x_{*} x ∗ for finite α \alpha α .
General Mean-Field Analysis
Laplace Principle
Our analysis is based on the following version of the quantitative Laplace principle, which is a variation of [5 Lemma 3.3] and [9 Theorem 4.3] . Its proof is based directly on arguments from these references, but since we use slightly different assumptions, we provide a complete and self-contained proof below.
Lemma 9 (Quantitative Laplace principle under separation). Let f : R d → R f : \mathbb{R}^{d} \to\mathbb{R} f : R d → R be a Borel measurable function with global minimum f ∗ ∈ R f_{*} \in\mathbb{R} f ∗ ∈ R , attained at x ∗ ∈ R d x_{*} \in\mathbb{R}^{d} x ∗ ∈ R d , i.e. f ∗ = f ( x ∗ ) ≤ f ( x ) f_{*} = f(x_{*}) \le f(x) f ∗ = f ( x ∗ ) ≤ f ( x ) for all x ∈ R d x \in\mathbb{R}^{d} x ∈ R d . Furthermore, assume that f f f is upper semicontinuous at x ∗ x_{*} x ∗ , and that for every ϵ > 0 \epsilon> 0 ϵ > 0 there exists S ϵ > 0 S_{\epsilon} > 0 S ϵ > 0 such that
inf ∥ x − x ∗ ∥ 2 ≥ ϵ f ( x ) ≥ f ∗ + S ϵ , (27) \inf_{\lVert x-x_{*}\rVert_{2}\ge\epsilon} f(x) \ge f_{*} + S_{\epsilon}, \tag*{(27)} ∥ x − x ∗ ∥ 2 ≥ ϵ inf f ( x ) ≥ f ∗ + S ϵ , (27)
which implies that x ∗ x_* x ∗ is the unique global minimiser of f f f .
Let ν \nu ν be a Borel probability measure on R d \mathbb{R}^d R d such that
ν ( B s ( x ∗ ) ) ≥ C s > 0 for all s > 0. (28) \nu\left(B_s(x_*)\right) \ge C_s > 0 \qquad\text{for all } s > 0. \tag*{(28)} ν ( B s ( x ∗ ) ) ≥ C s > 0 for all s > 0. (28)
Then for all α > 0 \alpha> 0 α > 0 ,
η ν α ( A ) : = 1 Z α ∫ A e − α f ( x ) ν ( d x ) , Z α : = ∫ R d e − α f ( x ) ν ( d x ) , A ∈ B ( R d ) , (29) \eta_\nu^\alpha(A) := \frac{1}{Z_\alpha}\int_A e^{-\alpha f(x)}\,\nu(\mathrm{d}x), \qquad Z_\alpha:= \int_{\mathbb{R}^d} e^{-\alpha f(x)}\,\nu(\mathrm{d}x), \qquad A \in\mathcal{B}(\mathbb{R}^d), \tag*{(29)} η ν α ( A ) := Z α 1 ∫ A e − α f ( x ) ν ( d x ) , Z α := ∫ R d e − α f ( x ) ν ( d x ) , A ∈ B ( R d ) , (29)
is a well-defined Borel probability measure, and for every ϵ > 0 \epsilon> 0 ϵ > 0 , there exists s f ( ϵ ) > 0 s_f(\epsilon) > 0 s f ( ϵ ) > 0 with
sup ∥ x − x ∗ ∥ 2 ≤ s f ( ϵ ) f ( x ) ≤ f ∗ + 1 2 S ϵ , (30) \sup_{\lVert x-x_*\rVert_2 \le s_f(\epsilon)} f(x) \le f_* + \frac{1}{2}S_\epsilon, \tag*{(30)} ∥ x − x ∗ ∥ 2 ≤ s f ( ϵ ) sup f ( x ) ≤ f ∗ + 2 1 S ϵ , (30)
such that the following hold.
(i) For all ϵ > 0 \epsilon> 0 ϵ > 0 and α > 0 \alpha> 0 α > 0 ,
η ν α ( { x ∈ R d : ∥ x − x ∗ ∥ 2 ≥ ϵ } ) ≤ e − α S ϵ / 2 C s f ( ϵ ) . (31) \eta_\nu^\alpha\left(\left\{x \in\mathbb{R}^d : \lVert x-x_*\rVert_2 \ge\epsilon\right\}\right) \le\frac{e^{-\alpha S_\epsilon/2}}{C_{s_f(\epsilon)}}. \tag*{(31)} η ν α ( { x ∈ R d : ∥ x − x ∗ ∥ 2 ≥ ϵ } ) ≤ C s f ( ϵ ) e − α S ϵ /2 . (31)
In particular, for every δ > 0 \delta> 0 δ > 0 and
α 0 = α 0 ( δ , S ϵ , C s f ( ϵ ) ) : = 2 S ϵ max { 0 , log 1 δ C s f ( ϵ ) } , (32) \alpha_0 = \alpha_0\left(\delta,S_\epsilon,C_{s_f(\epsilon)}\right) := \frac{2}{S_\epsilon}\max\left\{0,\log\frac{1}{\delta C_{s_f(\epsilon)}}\right\}, \tag*{(32)} α 0 = α 0 ( δ , S ϵ , C s f ( ϵ ) ) := S ϵ 2 max { 0 , log δ C s f ( ϵ ) 1 } , (32)
we have
e − α S ϵ / 2 C s f ( ϵ ) ≤ δ for all α ≥ α 0 , (33) \frac{e^{-\alpha S_\epsilon/2}}{C_{s_f(\epsilon)}} \le\delta\qquad\text{for all } \alpha\ge\alpha_0, \tag*{(33)} C s f ( ϵ ) e − α S ϵ /2 ≤ δ for all α ≥ α 0 , (33)
so that the left-hand side of (31) is at most δ \delta δ for all α ≥ α 0 \alpha\ge\alpha_0 α ≥ α 0 . Moreover,
η ν α ( { x ∈ R d : ∥ x − x ∗ ∥ 2 ≥ ϵ } ) → 0 for α → ∞ . (34) \eta_\nu^\alpha\left(\left\{x \in\mathbb{R}^d : \lVert x-x_*\rVert_2 \ge\epsilon\right\}\right) \to0 \qquad\text{for } \alpha\to\infty. \tag*{(34)} η ν α ( { x ∈ R d : ∥ x − x ∗ ∥ 2 ≥ ϵ } ) → 0 for α → ∞. (34)
(ii) Moments. Let p ∈ [ 1 , ∞ ) p \in[1,\infty) p ∈ [ 1 , ∞ ) and assume that
E X ∼ ν [ ∥ X − x ∗ ∥ 2 p ] ≤ M p < ∞ . (35) \mathbb{E}_{X\sim\nu}\left[\lVert X-x_*\rVert_2^p\right] \le M_p < \infty. \tag*{(35)} E X ∼ ν [ ∥ X − x ∗ ∥ 2 p ] ≤ M p < ∞. (35)
Then E X ∼ η ν α [ ∥ X − x ∗ ∥ 2 p ] < ∞ \mathbb{E}_{X\sim\eta_\nu^\alpha}\left[\lVert X-x_*\rVert_2^p\right] < \infty E X ∼ η ν α [ ∥ X − x ∗ ∥ 2 p ] < ∞ and, for all ϵ > 0 \epsilon> 0 ϵ > 0 and α > 0 \alpha> 0 α > 0 ,
E X ∼ η ν α [ ∥ X − x ∗ ∥ 2 p ] ≤ ϵ p + e − α S ϵ / 2 C s f ( ϵ ) M p . (36) \mathbb{E}_{X\sim\eta_\nu^\alpha}\left[\lVert X-x_*\rVert_2^p\right] \le\epsilon^p + \frac{e^{-\alpha S_\epsilon/2}}{C_{s_f(\epsilon)}}M_p. \tag*{(36)} E X ∼ η ν α [ ∥ X − x ∗ ∥ 2 p ] ≤ ϵ p + C s f ( ϵ ) e − α S ϵ /2 M p . (36)
Consequently, E X ∼ η ν α [ ∥ X − x ∗ ∥ 2 p ] → 0 \mathbb{E}_{X\sim\eta_\nu^\alpha}\left[\lVert X-x_*\rVert_2^p\right] \to0 E X ∼ η ν α [ ∥ X − x ∗ ∥ 2 p ] → 0 as α → ∞ \alpha\to\infty α → ∞ . In particular, if (35) holds for p = 1 p=1 p = 1 , then the mean E X ∼ η ν α [ X ] \mathbb{E}_{X\sim\eta_\nu^\alpha}[X] E X ∼ η ν α [ X ] exists, and for all ϵ > 0 \epsilon> 0 ϵ > 0 and α > 0 \alpha> 0 α > 0 we have
∥ E X ∼ η ν α [ X ] − x ∗ ∥ 2 ≤ ϵ + e − α S ϵ / 2 C s f ( ϵ ) M 1 , (37) \left\lVert\mathbb{E}_{X\sim\eta_\nu^\alpha}[X]-x_*\right\rVert_2 \le\epsilon+ \frac{e^{-\alpha S_\epsilon/2}}{C_{s_f(\epsilon)}}M_1, \tag*{(37)} E X ∼ η ν α [ X ] − x ∗ 2 ≤ ϵ + C s f ( ϵ ) e − α S ϵ /2 M 1 , (37)
and E X ∼ η ν α [ X ] → x ∗ \mathbb{E}_{X\sim\eta_\nu^\alpha}[X] \to x_* E X ∼ η ν α [ X ] → x ∗ as α → ∞ \alpha\to\infty α → ∞ .
Proof Uniqueness of global minimiser. Consider some x ≠ x ∗ x \ne x_* x = x ∗ . Using (27) with ϵ = ∥ x − x ∗ ∥ 2 > 0 \epsilon= \lVert x-x_*\rVert_2 > 0 ϵ = ∥ x − x ∗ ∥ 2 > 0 results in f ( x ) ≥ f ∗ + S ϵ > f ∗ f(x) \ge f_* + S\epsilon> f_* f ( x ) ≥ f ∗ + S ϵ > f ∗ , so x ∗ x_* x ∗ is indeed the unique global minimiser.
Existence of s f ( ϵ ) s_f(\epsilon) s f ( ϵ ) . Let ϵ > 0 \epsilon> 0 ϵ > 0 . By definition of upper semicontinuity at x ∗ x_* x ∗ there is r > 0 r > 0 r > 0 with f ( x ) < f ∗ + 1 2 S ϵ f(x) < f_* + \frac{1}{2}S\epsilon f ( x ) < f ∗ + 2 1 S ϵ for all x x x with ∥ x − x ∗ ∥ 2 < r \lVert x-x_*\rVert_2 < r ∥ x − x ∗ ∥ 2 < r , and any s f ( ϵ ) ∈ ( 0 , r ) s_f(\epsilon) \in(0,r) s f ( ϵ ) ∈ ( 0 , r ) satisfies (30) .
For the remainder of the proof, let ϵ > 0 \epsilon> 0 ϵ > 0 be arbitrary, but fixed, and define
A ϵ : = { x ∈ R d : ∥ x − x ∗ ∥ 2 ≥ ϵ } . (38) A_\epsilon:= \{x \in\mathbb{R}^d : \lVert x-x_*\rVert_2 \ge\epsilon\}. \tag*{(38)} A ϵ := { x ∈ R d : ∥ x − x ∗ ∥ 2 ≥ ϵ } . (38)
Furthermore, to improve readability, we will write s f = s f ( ϵ ) s_f=s_f(\epsilon) s f = s f ( ϵ ) .
η ν α \eta_\nu^\alpha η ν α is well defined. We have to show that Z α Z_\alpha Z α is finite and positive. Since f ≥ f ∗ f \ge f_* f ≥ f ∗ and f f f is measurable, the integrand e − α f e^{-\alpha f} e − α f is measurable and bounded by e − α f ∗ e^{-\alpha f_*} e − α f ∗ , so Z α ≤ e − α f ∗ < ∞ Z_\alpha\le e^{-\alpha f_*} < \infty Z α ≤ e − α f ∗ < ∞ . To show positivity, restrict the integral to B s f ( x ∗ ) B_{s_f}(x_*) B s f ( x ∗ ) and use (30) and (28) ,
Z α ≥ ∫ B s f ( x ∗ ) e − α f ( x ) ν ( d x ) ≥ e − α ( f ∗ + 1 2 S ϵ ) ν ( B s f ( x ∗ ) ) ≥ e − α ( f ∗ + 1 2 S ϵ ) C s f > 0. (39) Z_\alpha\ge\int_{B_{s_f}(x_*)} e^{-\alpha f(x)} \,\nu(dx) \ge e^{-\alpha(f_*+\frac{1}{2}S\epsilon)}\nu(B_{s_f}(x_*)) \ge e^{-\alpha(f_*+\frac{1}{2}S\epsilon)}C_{s_f} > 0. \tag*{(39)} Z α ≥ ∫ B s f ( x ∗ ) e − α f ( x ) ν ( d x ) ≥ e − α ( f ∗ + 2 1 S ϵ ) ν ( B s f ( x ∗ )) ≥ e − α ( f ∗ + 2 1 S ϵ ) C s f > 0. (39)
In particular, η ν α \eta_\nu^\alpha η ν α in (29) is a well-defined Borel probability measure.
A density bound away from the minimiser. Let g : R d → [ 0 , ∞ ] g:\mathbb{R}^d \to[0,\infty] g : R d → [ 0 , ∞ ] be measurable. By (27) , e − α f ( x ) ≤ e − α ( f ∗ + S ϵ ) e^{-\alpha f(x)} \le e^{-\alpha(f_*+S\epsilon)} e − α f ( x ) ≤ e − α ( f ∗ + S ϵ ) for all x ∈ A ϵ x \in A_\epsilon x ∈ A ϵ , so
∫ A ϵ g ( x ) e − α f ( x ) ν ( d x ) ≤ e − α ( f ∗ + S ϵ ) ∫ A ϵ g ( x ) ν ( d x ) . (40) \int_{A_\epsilon} g(x)e^{-\alpha f(x)}\,\nu(dx) \le e^{-\alpha(f_*+S\epsilon)}\int_{A_\epsilon}g(x)\,\nu(dx). \tag*{(40)} ∫ A ϵ g ( x ) e − α f ( x ) ν ( d x ) ≤ e − α ( f ∗ + S ϵ ) ∫ A ϵ g ( x ) ν ( d x ) . (40)
Combining this with the lower bound (39) for Z α Z_\alpha Z α results in
∫ A ϵ g d η ν α = 1 Z α ∫ A ϵ g ( x ) e − α f ( x ) ν ( d x ) ≤ e − α S ϵ / 2 C s f ∫ A ϵ g d ν . (41) \int_{A_\epsilon}g\,d\eta_\nu^\alpha= \frac{1}{Z_\alpha}\int_{A_\epsilon}g(x)e^{-\alpha f(x)}\,\nu(dx) \le\frac{e^{-\alpha S\epsilon/2}}{C_{s_f}}\int_{A_\epsilon}g\,d\nu. \tag*{(41)} ∫ A ϵ g d η ν α = Z α 1 ∫ A ϵ g ( x ) e − α f ( x ) ν ( d x ) ≤ C s f e − α S ϵ /2 ∫ A ϵ g d ν . (41)
Part (i). Choosing g ≡ 1 g \equiv1 g ≡ 1 in (41) and using ν ( A ϵ ) ≤ 1 \nu(A_\epsilon) \le1 ν ( A ϵ ) ≤ 1 yields
η ν α ( { x ∈ R d : ∥ x − x ∗ ∥ 2 ≥ ϵ } ) = ∫ A ϵ d η ν α ≤ e − α S ϵ / 2 C s f ∫ A ϵ 1 d ν ≤ e − α S ϵ / 2 C s f , \eta_\nu^\alpha(\{x \in\mathbb{R}^d : \lVert x-x_*\rVert_2 \ge\epsilon\}) = \int_{A_\epsilon}d\eta_\nu^\alpha\le\frac{e^{-\alpha S\epsilon/2}}{C_{s_f}}\int_{A_\epsilon}1\,d\nu\le\frac{e^{-\alpha S\epsilon/2}}{C_{s_f}}, η ν α ({ x ∈ R d : ∥ x − x ∗ ∥ 2 ≥ ϵ }) = ∫ A ϵ d η ν α ≤ C s f e − α S ϵ /2 ∫ A ϵ 1 d ν ≤ C s f e − α S ϵ /2 ,
so we established (31) .
Let δ > 0 \delta> 0 δ > 0 . Since C s f > 0 C_{s_f} > 0 C s f > 0 , α 0 \alpha_0 α 0 in (32) is well defined and nonnegative. Let now α ≥ α 0 \alpha\ge\alpha_0 α ≥ α 0 . If δ C s f < 1 \delta C_{s_f} < 1 δ C s f < 1 , then α 0 = 2 S ϵ log 1 δ C s f \alpha_0 = \frac{2}{S\epsilon}\log\frac{1}{\delta C_{s_f}} α 0 = S ϵ 2 log δ C s f 1 , and we get
e − α S ϵ / 2 C s f ≤ e − α 0 S ϵ / 2 C s f = δ C s f C s f = δ . (42) \frac{e^{-\alpha S\epsilon/2}}{C_{s_f}} \le\frac{e^{-\alpha_0 S\epsilon/2}}{C_{s_f}} = \frac{\delta C_{s_f}}{C_{s_f}} = \delta. \tag*{(42)} C s f e − α S ϵ /2 ≤ C s f e − α 0 S ϵ /2 = C s f δ C s f = δ . (42)
If δ C s f ≥ 1 \delta C_{s_f} \ge1 δ C s f ≥ 1 , then e − α S ϵ / 2 / C s f ≤ 1 / C s f ≤ δ e^{-\alpha S\epsilon/2}/C_{s_f} \le1/C_{s_f} \le\delta e − α S ϵ /2 / C s f ≤ 1/ C s f ≤ δ . This shows (33) , and together with (31) that for α ≥ α 0 \alpha\ge\alpha_0 α ≥ α 0 the left-hand side of (31) is bounded by δ \delta δ . Furthermore, since δ > 0 \delta> 0 δ > 0 was arbitrary, this also shows η ν α ( A ϵ ) → 0 \eta_\nu^\alpha(A_\epsilon) \to0 η ν α ( A ϵ ) → 0 as α → ∞ \alpha\to\infty α → ∞ .
Part (ii). Let p ∈ [ 1 , ∞ ) p \in[1,\infty) p ∈ [ 1 , ∞ ) be such that (35) holds. Choosing g ( x ) = ∥ x − x ∗ ∥ 2 p g(x)=\lVert x-x_*\rVert_2^p g ( x ) = ∥ x − x ∗ ∥ 2 p in (41) and using (35) gives
∫ A ϵ ∥ x − x ∗ ∥ 2 p d η ν α ( x ) ≤ e − α S ϵ / 2 C s f ∫ A ϵ ∥ x − x ∗ ∥ 2 p ν ( d x ) ≤ e − α S ϵ / 2 C s f M p . (43) \int_{A_\epsilon}\lVert x-x_*\rVert_2^p\,d\eta_\nu^\alpha(x) \le\frac{e^{-\alpha S\epsilon/2}}{C_{s_f}}\int_{A_\epsilon}\lVert x-x_*\rVert_2^p\,\nu(dx) \le\frac{e^{-\alpha S\epsilon/2}}{C_{s_f}}M_p. \tag*{(43)} ∫ A ϵ ∥ x − x ∗ ∥ 2 p d η ν α ( x ) ≤ C s f e − α S ϵ /2 ∫ A ϵ ∥ x − x ∗ ∥ 2 p ν ( d x ) ≤ C s f e − α S ϵ /2 M p . (43)
On the complement A ϵ c A_{\epsilon}^{c} A ϵ c we have ∥ x − x ∗ ∥ 2 p ≤ ϵ p \lVert x-x_{*}\rVert_{2}^{p}\leq\epsilon^{p} ∥ x − x ∗ ∥ 2 p ≤ ϵ p , and since η ν α \eta_{\nu}^{\alpha} η ν α is a probability measure we also have η ν α ( A ϵ c ) ≤ 1 \eta_{\nu}^{\alpha}(A_{\epsilon}^{c})\leq1 η ν α ( A ϵ c ) ≤ 1 . This implies that the integral over A ϵ c A_{\epsilon}^{c} A ϵ c is at most ϵ p \epsilon^{p} ϵ p . Together with (43) we get
E X ∼ η ν α [ ∥ X − x ∗ ∥ 2 p ] = ∫ A ϵ c ∥ x − x ∗ ∥ 2 p η ν α ( d x ) + ∫ A ϵ ∥ x − x ∗ ∥ 2 p η ν α ( d x ) ≤ ϵ p + e − α S ϵ / 2 C s f M p < ∞ , (44) \mathbb{E}_{X\sim\eta_{\nu}^{\alpha}}\left[\lVert X-x_{*}\rVert_{2}^{p}\right]=\int_{A_{\epsilon}^{c}}\lVert x-x_{*}\rVert_{2}^{p}\eta_{\nu}^{\alpha}(\mathrm{d}x)+\int_{A_{\epsilon}}\lVert x-x_{*}\rVert_{2}^{p}\eta_{\nu}^{\alpha}(\mathrm{d}x)\leq\epsilon^{p}+\frac{\mathrm{e}^{-\alpha S_{\epsilon}/2}}{C_{s f}}M_{p}<\infty, \tag*{(44)} E X ∼ η ν α [ ∥ X − x ∗ ∥ 2 p ] = ∫ A ϵ c ∥ x − x ∗ ∥ 2 p η ν α ( d x ) + ∫ A ϵ ∥ x − x ∗ ∥ 2 p η ν α ( d x ) ≤ ϵ p + C s f e − α S ϵ /2 M p < ∞ , (44)
which is (36) . In particular, since ∥ x ∥ 2 p ≤ 2 p − 1 ( ∥ x ∗ ∥ 2 p + ∥ x − x ∗ ∥ 2 p ) \lVert x\rVert_{2}^{p}\leq2^{p-1}\left(\lVert x_{*}\rVert_{2}^{p}+\lVert x-x_{*}\rVert_{2}^{p}\right) ∥ x ∥ 2 p ≤ 2 p − 1 ( ∥ x ∗ ∥ 2 p + ∥ x − x ∗ ∥ 2 p ) by convexity of t ↦ t p t\mapsto t^{p} t ↦ t p on [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) , we have E X ∼ η ν α [ ∥ X ∥ 2 p ] < ∞ \mathbb{E}_{X\sim\eta_{\nu}^{\alpha}}[\lVert X\rVert_{2}^{p}]<\infty E X ∼ η ν α [∥ X ∥ 2 p ] < ∞ .
It remains to show convergence. Let γ > 0 \gamma>0 γ > 0 and set ϵ : = ( γ / 2 ) 1 / p \epsilon:=(\gamma/2)^{1/p} ϵ := ( γ /2 ) 1/ p . If M p = 0 M_{p}=0 M p = 0 , then (36) gives E X ∼ η ν α [ ∥ X − x ∗ ∥ 2 p ] ≤ γ / 2 \mathbb{E}_{X\sim\eta_{\nu}^{\alpha}}[\lVert X-x_{*}\rVert_{2}^{p}]\leq\gamma/2 E X ∼ η ν α [∥ X − x ∗ ∥ 2 p ] ≤ γ /2 for all α > 0 \alpha>0 α > 0 . Otherwise, (33) with δ = γ / ( 2 M p ) \delta=\gamma/(2M_{p}) δ = γ / ( 2 M p ) shows that
e − α S ϵ / 2 C s f ( ϵ ) M p ≤ γ 2 for all α ≥ α 0 ( γ 2 M p , S ϵ , C s f ( ϵ ) ) , \frac{\mathrm{e}^{-\alpha S_{\epsilon}/2}}{C_{s f}(\epsilon)}M_{p}\leq\frac{\gamma}{2}\qquad\text{for all }\alpha\geq\alpha_{0}\left(\frac{\gamma}{2M_{p}},S_{\epsilon},C_{s f}(\epsilon)\right), C s f ( ϵ ) e − α S ϵ /2 M p ≤ 2 γ for all α ≥ α 0 ( 2 M p γ , S ϵ , C s f ( ϵ ) ) ,
and (36) gives E X ∼ η ν α [ ∥ X − x ∗ ∥ 2 p ] ≤ ϵ p + γ / 2 = γ \mathbb{E}_{X\sim\eta_{\nu}^{\alpha}}[\lVert X-x_{*}\rVert_{2}^{p}]\leq\epsilon^{p}+\gamma/2=\gamma E X ∼ η ν α [∥ X − x ∗ ∥ 2 p ] ≤ ϵ p + γ /2 = γ for all these α \alpha α . Since γ > 0 \gamma>0 γ > 0 was arbitrary, E X ∼ η ν α [ ∥ X − x ∗ ∥ 2 p ] → 0 \mathbb{E}_{X\sim\eta_{\nu}^{\alpha}}[\lVert X-x_{*}\rVert_{2}^{p}]\to0 E X ∼ η ν α [∥ X − x ∗ ∥ 2 p ] → 0 as α → ∞ \alpha\to\infty α → ∞ .
Finally, let (35) hold for p = 1 p=1 p = 1 . Then E X ∼ η ν α [ ∥ X ∥ 2 2 ] < ∞ \mathbb{E}_{X\sim\eta_{\nu}^{\alpha}}[\lVert X\rVert_{2}^{2}]<\infty E X ∼ η ν α [∥ X ∥ 2 2 ] < ∞ , so the mean E X ∼ η ν α [ X ] \mathbb{E}_{X\sim\eta_{\nu}^{\alpha}}[X] E X ∼ η ν α [ X ] exists. Jensen’s inequality and (36) with p = 1 p=1 p = 1 give
∥ E X ∼ η ν α [ X ] − x ∗ ∥ 2 = ∥ E X ∼ η ν α [ X − x ∗ ] ∥ 2 ≤ E X ∼ η ν α [ ∥ X − x ∗ ∥ 2 ] ≤ ϵ + e − α S ϵ / 2 C s f M 1 , (45) \left\lVert\mathbb{E}_{X\sim\eta_{\nu}^{\alpha}}[X]-x_{*}\right\rVert_{2}=\left\lVert\mathbb{E}_{X\sim\eta_{\nu}^{\alpha}}[X-x_{*}]\right\rVert_{2}\leq\mathbb{E}_{X\sim\eta_{\nu}^{\alpha}}\left[\lVert X-x_{*}\rVert_{2}\right]\leq\epsilon+\frac{\mathrm{e}^{-\alpha S_{\epsilon}/2}}{C_{s f}}M_{1}, \tag*{(45)} E X ∼ η ν α [ X ] − x ∗ 2 = E X ∼ η ν α [ X − x ∗ ] 2 ≤ E X ∼ η ν α [ ∥ X − x ∗ ∥ 2 ] ≤ ϵ + C s f e − α S ϵ /2 M 1 , (45)
which is (37) . Since the middle term in (45) tends to 0 as α → ∞ \alpha\to\infty α → ∞ , so does the left-hand side.
Let us briefly discuss the differences to Fornasier et al., [5 ] and Huang et al., [9 ] . Assumption (28) is also present in these references and used in the same way to get a lower bound on the normalisation constant Z α Z_{\alpha} Z α . Assumption (27) is implied by the setting of these references, cf. Huang et al., [9 ] . The splitting technique in our proof is essentially the approach from the proof of Huang et al., [9 ] .
We now show that N ( x , Σ ) \mathcal{N}(x,\Sigma) N ( x , Σ ) for any x ∈ R d x\in\mathbb{R}^{d} x ∈ R d is a suitable probability measure and that we can apply a quantitative Laplace principle, cf. Lemma 9 . This result is essentially the first step of the proof of Fornasier et al., [5 ] , generalised from a scaled identity matrix to a general positive definite covariance matrix.
Lemma 10 Let x ∗ ∈ R d x_{*}\in\mathbb{R}^{d} x ∗ ∈ R d , r > 0 r>0 r > 0 , and let Σ ∈ R d × d \Sigma\in\mathbb{R}^{d\times d} Σ ∈ R d × d be symmetric positive definite. For all x ∈ B r ( x ∗ ) x\in B_{r}(x_{*}) x ∈ B r ( x ∗ ) and all s ∈ R > 0 s\in\mathbb{R}_{>0} s ∈ R > 0 it holds
N ( x , Σ ) ( B s ( x ∗ ) ) ≥ π d / 2 s d Γ ( d 2 + 1 ) ( 2 π σ max ( Σ ) 2 ) − d / 2 exp ( − ( r + s ) 2 2 σ min ( Σ ) 2 ) = : C s > 0. (46) \mathcal{N}(x,\Sigma)(B_{s}(x_{*}))\geq\frac{\pi^{d/2}s^{d}}{\Gamma\left(\frac{d}{2}+1\right)}(2\pi\sigma_{\max}(\Sigma)^{2})^{-d/2}\exp\left(-\frac{(r+s)^{2}}{2\sigma_{\min}(\Sigma)^{2}}\right)=:C_{s}>0. \tag*{(46)} N ( x , Σ ) ( B s ( x ∗ )) ≥ Γ ( 2 d + 1 ) π d /2 s d ( 2 π σ m a x ( Σ ) 2 ) − d /2 exp ( − 2 σ m i n ( Σ ) 2 ( r + s ) 2 ) =: C s > 0. (46) Furthermore, for X ∼ N ( x , Σ ) X\sim\mathcal{N}(x,\Sigma) X ∼ N ( x , Σ ) we also have
E [ ∥ X − x ∗ ∥ 2 ] ≤ r + σ max ( Σ ) d = : M 1 (47) \mathbb{E}\left[\lVert X-x_{*}\rVert_{2}\right]\leq r+\sigma_{\max}(\Sigma)\sqrt{d}=:M_{1} \tag*{(47)} E [ ∥ X − x ∗ ∥ 2 ] ≤ r + σ m a x ( Σ ) d =: M 1 (47) E [ ∥ X − x ∗ ∥ 2 2 ] ≤ r 2 + d σ max ( Σ ) 2 = : M 2 . (48) \mathbb{E}\left[\lVert X-x_{*}\rVert_{2}^{2}\right]\leq r^{2}+d\sigma_{\max}(\Sigma)^{2}=:M_{2}. \tag*{(48)} E [ ∥ X − x ∗ ∥ 2 2 ] ≤ r 2 + d σ m a x ( Σ ) 2 =: M 2 . (48)
Proof. Let x ∈ B r ( x ∗ ) x\in B_{r}(x_{*}) x ∈ B r ( x ∗ ) be arbitrary. Since Σ \Sigma Σ is symmetric positive definite, we have det ( Σ ) ≤ λ max ( Σ ) d = σ max ( Σ ) 2 d \det(\Sigma)\leq\lambda_{\max}(\Sigma)^{d}=\sigma_{\max}(\Sigma)^{2d} det ( Σ ) ≤ λ m a x ( Σ ) d = σ m a x ( Σ ) 2 d , and for all v ∈ R d v\in\mathbb{R}^{d} v ∈ R d we have v T Σ − 1 v ≤ λ min ( Σ ) − 1 ∥ v ∥ 2 2 = σ min ( Σ ) − 2 ∥ v ∥ 2 2 v^{\mathsf{T}}\Sigma^{-1}v\leq\lambda_{\min}(\Sigma)^{-1}\lVert v\rVert_{2}^{2}=\sigma_{\min}(\Sigma)^{-2}\lVert v\rVert_{2}^{2} v T Σ − 1 v ≤ λ m i n ( Σ ) − 1 ∥ v ∥ 2 2 = σ m i n ( Σ ) − 2 ∥ v ∥ 2 2 . Let now s ∈ R > 0 s \in\mathbb{R}_{>0} s ∈ R > 0 and y ∈ B s ( x ∗ ) y \in B_s(x_*) y ∈ B s ( x ∗ ) be arbitrary. Since ∥ y − x ∥ 2 ≤ ∥ y − x ∗ ∥ 2 + ∥ x ∗ − x ∥ 2 ≤ s + r \lVert y-x\rVert_2 \le\lVert y-x_*\rVert_2+\lVert x_*-x\rVert_2 \le s+r ∥ y − x ∥ 2 ≤ ∥ y − x ∗ ∥ 2 + ∥ x ∗ − x ∥ 2 ≤ s + r , we have
N ( y ∣ x , Σ ) = 1 ( 2 π ) d det ( Σ ) exp ( − 1 2 ( y − x ) T Σ − 1 ( y − x ) ) ≥ ( 2 π σ max ( Σ ) 2 ) − d / 2 exp ( − ( r + s ) 2 2 σ min ( Σ ) 2 ) . \begin{aligned}
\mathcal{N}(y\mid x,\Sigma)=\frac{1}{\sqrt{(2\pi)^d\det(\Sigma)}}\exp\left(-\frac{1}{2}(y-x)^{\mathsf{T}}\Sigma^{-1}(y-x)\right) \\
&\ge\left(2\pi\sigma_{\max}(\Sigma)^2\right)^{-d/2}\exp\left(-\frac{(r+s)^2}{2\sigma_{\min}(\Sigma)^2}\right).
\end{aligned} N ( y ∣ x , Σ ) = ( 2 π ) d det ( Σ ) 1 exp ( − 2 1 ( y − x ) T Σ − 1 ( y − x ) ) ≥ ( 2 π σ m a x ( Σ ) 2 ) − d /2 exp ( − 2 σ m i n ( Σ ) 2 ( r + s ) 2 ) .
Integrating over B s ( x ∗ ) B_s(x_*) B s ( x ∗ ) and using Vol ( B s ( x ∗ ) ) = π d / 2 s d / Γ ( d 2 + 1 ) \operatorname{Vol}(B_s(x_*))=\pi^{d/2}s^d/\Gamma\left(\frac{d}{2}+1\right) Vol ( B s ( x ∗ )) = π d /2 s d /Γ ( 2 d + 1 ) leads to
N ( x , Σ ) ( B s ( x ∗ ) ) = ∫ B s ( x ∗ ) 1 ( 2 π ) d det ( Σ ) exp ( − 1 2 ( y − x ) T Σ − 1 ( y − x ) ) d y ≥ ( 2 π σ max ( Σ ) 2 ) − d / 2 exp ( − ( r + s ) 2 2 σ min ( Σ ) 2 ) Vol ( B s ( x ∗ ) ) = π d / 2 s d Γ ( d 2 + 1 ) ( 2 π σ max ( Σ ) 2 ) − d / 2 exp ( − ( r + s ) 2 2 σ min ( Σ ) 2 ) > 0. \begin{aligned}
\mathcal{N}(x,\Sigma)(B_s(x_*))=\int_{B_s(x_*)}\frac{1}{\sqrt{(2\pi)^d\det(\Sigma)}}\exp\left(-\frac{1}{2}(y-x)^{\mathsf{T}}\Sigma^{-1}(y-x)\right)\,\mathrm{d}y \\
&\ge\left(2\pi\sigma_{\max}(\Sigma)^2\right)^{-d/2}\exp\left(-\frac{(r+s)^2}{2\sigma_{\min}(\Sigma)^2}\right)\operatorname{Vol}(B_s(x_*)) \\
&=\frac{\pi^{d/2}s^d}{\Gamma\left(\frac{d}{2}+1\right)}\left(2\pi\sigma_{\max}(\Sigma)^2\right)^{-d/2}\exp\left(-\frac{(r+s)^2}{2\sigma_{\min}(\Sigma)^2}\right)>0.
\end{aligned} N ( x , Σ ) ( B s ( x ∗ )) = ∫ B s ( x ∗ ) ( 2 π ) d det ( Σ ) 1 exp ( − 2 1 ( y − x ) T Σ − 1 ( y − x ) ) d y ≥ ( 2 π σ m a x ( Σ ) 2 ) − d /2 exp ( − 2 σ m i n ( Σ ) 2 ( r + s ) 2 ) Vol ( B s ( x ∗ )) = Γ ( 2 d + 1 ) π d /2 s d ( 2 π σ m a x ( Σ ) 2 ) − d /2 exp ( − 2 σ m i n ( Σ ) 2 ( r + s ) 2 ) > 0.
For the moment bounds, note that with X = x + Z X=x+Z X = x + Z with Z ∼ N ( 0 , Σ ) Z\sim\mathcal{N}(0,\Sigma) Z ∼ N ( 0 , Σ ) we have X ∼ N ( x , Σ ) X\sim\mathcal{N}(x,\Sigma) X ∼ N ( x , Σ ) . By the triangle inequality, Jensen’s inequality, E [ ∥ Z ∥ 2 2 ] = tr ( Σ ) \mathbb{E}[\lVert Z\rVert_2^2]=\operatorname{tr}(\Sigma) E [∥ Z ∥ 2 2 ] = tr ( Σ ) and tr ( Σ ) ≤ d λ max ( Σ ) \operatorname{tr}(\Sigma)\le d\lambda_{\max}(\Sigma) tr ( Σ ) ≤ d λ m a x ( Σ ) , we get
E [ ∥ X − x ∗ ∥ 2 ] ≤ ∥ x − x ∗ ∥ 2 + E [ ∥ Z ∥ 2 ] ≤ r + ( E [ ∥ Z ∥ 2 2 ] ) 1 / 2 = r + ( tr ( Σ ) ) 1 / 2 ≤ r + σ max ( Σ ) d = M 1 . \begin{aligned}
\mathbb{E}[\lVert X-x_*\rVert_2]\le\lVert x-x_*\rVert_2+\mathbb{E}[\lVert Z\rVert_2] \\
&\le r+\left(\mathbb{E}[\lVert Z\rVert_2^2]\right)^{1/2} \\
&=r+\left(\operatorname{tr}(\Sigma)\right)^{1/2} \\
&\le r+\sigma_{\max}(\Sigma)\sqrt{d}=M_1.
\end{aligned} E [∥ X − x ∗ ∥ 2 ] ≤ ∥ x − x ∗ ∥ 2 + E [∥ Z ∥ 2 ] ≤ r + ( E [∥ Z ∥ 2 2 ] ) 1/2 = r + ( tr ( Σ ) ) 1/2 ≤ r + σ m a x ( Σ ) d = M 1 .
For the second moment, expanding the square and using E [ Z ] = 0 \mathbb{E}[Z]=0 E [ Z ] = 0 gives
E [ ∥ X − x ∗ ∥ 2 2 ] = ∥ x − x ∗ ∥ 2 2 + 2 ( x − x ∗ ) T E [ Z ] + E [ ∥ Z ∥ 2 2 ] = ∥ x − x ∗ ∥ 2 2 + tr ( Σ ) ≤ r 2 + d σ max ( Σ ) 2 = M 2 . \begin{aligned}
\mathbb{E}[\lVert X-x_*\rVert_2^2]=\lVert x-x_*\rVert_2^2+2(x-x_*)^{\mathsf{T}}\mathbb{E}[Z]+\mathbb{E}[\lVert Z\rVert_2^2] \\
&=\lVert x-x_*\rVert_2^2+\operatorname{tr}(\Sigma)\le r^2+d\sigma_{\max}(\Sigma)^2=M_2.
\end{aligned} E [∥ X − x ∗ ∥ 2 2 ] = ∥ x − x ∗ ∥ 2 2 + 2 ( x − x ∗ ) T E [ Z ] + E [∥ Z ∥ 2 2 ] = ∥ x − x ∗ ∥ 2 2 + tr ( Σ ) ≤ r 2 + d σ m a x ( Σ ) 2 = M 2 .
Mean-Field Transition Map as a Contraction
Our next goal is to show that the transition map T α \mathcal{T}_\alpha T α is a contraction for α \alpha α large enough. We follow the strategy of the proof of Fornasier et al. [5 ] and bound the norm of the gradient of T α \mathcal{T}_\alpha T α . For this, we need the following technical result, which is contained, under slightly stronger assumptions, in Step 2 of the aforementioned proof, and can be proved using exactly the same arguments.
Lemma 11 (Differentiability of the transition map). Let f : R d → R f:\mathbb{R}^d\to\mathbb{R} f : R d → R be Borel measurable and bounded below, f ‾ : = inf y ∈ R d f ( y ) > − ∞ \underline{f}:=\inf_{y\in\mathbb{R}^d}f(y)>-\infty f := inf y ∈ R d f ( y ) > − ∞ . Let Σ ∈ R d × d \Sigma\in\mathbb{R}^{d\times d} Σ ∈ R d × d be symmetric positive definite and α > 0 \alpha>0 α > 0 .
:::`
(i) For every x ∈ R d x \in\mathbb{R}^{d} x ∈ R d we have 0 < Z α ( N ( x , Σ ) ) ≤ e − α f ∗ 0 < Z_{\alpha}(\mathcal{N}(x,\Sigma)) \le e^{-\alpha f_*} 0 < Z α ( N ( x , Σ )) ≤ e − α f ∗ . Moreover, the Gibbs tilt η N ( x , Σ ) α \eta^{\alpha}_{\mathcal{N}(x,\Sigma)} η N ( x , Σ ) α , defined as in (29) with ν = N ( x , Σ ) \nu= \mathcal{N}(x,\Sigma) ν = N ( x , Σ ) , is a Borel probability measure with finite moments of all orders, and for every Borel measurable h : R d → [ 0 , ∞ ) h : \mathbb{R}^{d} \to[0,\infty) h : R d → [ 0 , ∞ ) ,
E Y ∼ η N ( x , Σ ) α [ h ( Y ) ] = 1 Z α ( N ( x , Σ ) ) E X ∼ N ( x , Σ ) [ h ( X ) e − α f ( X ) ] . (49) \mathbb{E}_{Y \sim\eta^{\alpha}_{\mathcal{N}(x,\Sigma)}}[h(Y)] = \frac{1}{Z_{\alpha}(\mathcal{N}(x,\Sigma))} \mathbb{E}_{X \sim\mathcal{N}(x,\Sigma)}[h(X)e^{-\alpha f(X)}]. \tag*{(49)} E Y ∼ η N ( x , Σ ) α [ h ( Y )] = Z α ( N ( x , Σ )) 1 E X ∼ N ( x , Σ ) [ h ( X ) e − α f ( X ) ] . (49)
In particular, T α ( x ) T_{\alpha}(x) T α ( x ) is well defined and coincides with the mean E Y ∼ η N ( x , Σ ) α [ Y ] \mathbb{E}_{Y \sim\eta^{\alpha}_{\mathcal{N}(x,\Sigma)}}[Y] E Y ∼ η N ( x , Σ ) α [ Y ] of the Gibbs tilt.
(ii) T α ∈ C 1 ( R d ; R d ) T_{\alpha} \in C^{1}(\mathbb{R}^{d};\mathbb{R}^{d}) T α ∈ C 1 ( R d ; R d ) , and its Jacobian ∇ T α ( x ) = ( ∂ x j T α , i ( x ) ) i , j \nabla T_{\alpha}(x) = (\partial_{x_j}T_{\alpha,i}(x))_{i,j} ∇ T α ( x ) = ( ∂ x j T α , i ( x ) ) i , j is
∇ T α ( x ) = Cov ( η N ( x , Σ ) α ) Σ − 1 for all x ∈ R d , (50) \nabla T_{\alpha}(x) = \operatorname{Cov}(\eta^{\alpha}_{\mathcal{N}(x,\Sigma)})\Sigma^{-1} \qquad\text{for all } x \in\mathbb{R}^{d}, \tag*{(50)} ∇ T α ( x ) = Cov ( η N ( x , Σ ) α ) Σ − 1 for all x ∈ R d , (50)
where Cov ( η N ( x , Σ ) α ) : = E Y ∼ η N ( x , Σ ) α [ ( Y − T α ( x ) ) ( Y − T α ( x ) ) ⊤ ] \operatorname{Cov}(\eta^{\alpha}_{\mathcal{N}(x,\Sigma)}) := \mathbb{E}_{Y \sim\eta^{\alpha}_{\mathcal{N}(x,\Sigma)}}[(Y-T_{\alpha}(x))(Y-T_{\alpha}(x))^{\top}] Cov ( η N ( x , Σ ) α ) := E Y ∼ η N ( x , Σ ) α [( Y − T α ( x )) ( Y − T α ( x ) ) ⊤ ] .
We can now bound the norm of the gradient of T α T_{\alpha} T α .
Lemma 12 (Jacobian bound). Let f f f satisfy Assumption 1, let Σ ∈ R d × d \Sigma\in\mathbb{R}^{d \times d} Σ ∈ R d × d be symmetric positive definite, and let r > 0 r > 0 r > 0 . Let C s C_s C s and M 2 M_2 M 2 be the constants of Lemma 10 for r r r and Σ \Sigma Σ , and let s f ( ϵ ) s_f(\epsilon) s f ( ϵ ) be as in (30) . Then, with T α T_{\alpha} T α as in Lemma 11 , for all ϵ > 0 \epsilon> 0 ϵ > 0 , α > 0 \alpha> 0 α > 0 and x ∈ B r ( x ∗ ) x \in B_r(x_*) x ∈ B r ( x ∗ ) ,
∥ ∇ T α ( x ) ∥ ≤ L α ( ϵ ) : = 1 σ min ( Σ ) 2 ( ϵ 2 + e − α S ϵ / 2 C s f ( ϵ ) M 2 ) . (51) \lVert\nabla T_{\alpha}(x) \rVert\le L_{\alpha}(\epsilon) := \frac{1}{\sigma_{\min}(\Sigma)^2}\left(\epsilon^2 + \frac{e^{-\alpha S_{\epsilon}/2}}{C_{s_f(\epsilon)}}M_2\right). \tag*{(51)} ∥ ∇ T α ( x )∥ ≤ L α ( ϵ ) := σ m i n ( Σ ) 2 1 ( ϵ 2 + C s f ( ϵ ) e − α S ϵ /2 M 2 ) . (51)
Proof. By (F1), f f f is measurable and bounded below by f ∗ f_* f ∗ , so Lemma 11 applies. Let x ∈ B r ( x ∗ ) x \in B_r(x_*) x ∈ B r ( x ∗ ) , ϵ > 0 \epsilon> 0 ϵ > 0 and α > 0 \alpha> 0 α > 0 be arbitrary. Using (50) , we can start with
∥ ∇ T α ( x ) ∥ ≤ ∥ Cov ( η N ( x , Σ ) α ) ∥ ∥ Σ − 1 ∥ ≤ tr Cov ( η N ( x , Σ ) α ) σ min ( Σ ) 2 . \lVert\nabla T_{\alpha}(x) \rVert\le\lVert\operatorname{Cov}(\eta^{\alpha}_{\mathcal{N}(x,\Sigma)}) \rVert\lVert\Sigma^{-1} \rVert\le\frac{\operatorname{tr}\operatorname{Cov}(\eta^{\alpha}_{\mathcal{N}(x,\Sigma)})}{\sigma_{\min}(\Sigma)^2}. ∥ ∇ T α ( x )∥ ≤ ∥ Cov ( η N ( x , Σ ) α )∥ ∥ Σ − 1 ∥ ≤ σ m i n ( Σ ) 2 tr Cov ( η N ( x , Σ ) α ) .
Let now Y ∼ η N ( x , Σ ) α Y \sim\eta^{\alpha}_{\mathcal{N}(x,\Sigma)} Y ∼ η N ( x , Σ ) α . Then we have
tr Cov ( η N ( x , Σ ) α ) = E [ ∥ Y − T α ( x ) ∥ 2 2 ] ≤ E [ ∥ Y − x ∗ ∥ 2 2 ] , \operatorname{tr}\operatorname{Cov}(\eta^{\alpha}_{\mathcal{N}(x,\Sigma)}) = \mathbb{E}[\lVert Y-T_{\alpha}(x)\rVert_2^2] \le\mathbb{E}[\lVert Y-x_*\rVert_2^2], tr Cov ( η N ( x , Σ ) α ) = E [∥ Y − T α ( x ) ∥ 2 2 ] ≤ E [∥ Y − x ∗ ∥ 2 2 ] ,
where we used the definition of the covariance matrix and of the trace together with the fact E [ Y ] = T α ( x ) \mathbb{E}[Y]=T_{\alpha}(x) E [ Y ] = T α ( x ) in the first step, and in the second step we used the well-known identity E ∥ Y − T α ( x ) ∥ 2 2 = E ∥ Y − z ∥ 2 2 − ∥ T α ( x ) − z ∥ 2 2 \mathbb{E}\lVert Y-T_{\alpha}(x)\rVert_2^2=\mathbb{E}\lVert Y-z\rVert_2^2-\lVert T_{\alpha}(x)-z\rVert_2^2 E ∥ Y − T α ( x ) ∥ 2 2 = E ∥ Y − z ∥ 2 2 − ∥ T α ( x ) − z ∥ 2 2 , which holds for any z ∈ R d z \in\mathbb{R}^{d} z ∈ R d .
We know from Lemma 10 that N ( x , Σ ) \mathcal{N}(x,\Sigma) N ( x , Σ ) satisfies (28) with the constants C s C_s C s and (35) for p = 2 p=2 p = 2 with M 2 M_2 M 2 , so Lemma 9 (ii) gives
E [ ∥ Y − x ∗ ∥ 2 2 ] ≤ ϵ 2 + e − α S ϵ / 2 C s f ( ϵ ) M 2 . \mathbb{E}[\lVert Y-x_*\rVert_2^2] \le\epsilon^2+\frac{e^{-\alpha S_{\epsilon}/2}}{C_{s_f(\epsilon)}}M_2. E [∥ Y − x ∗ ∥ 2 2 ] ≤ ϵ 2 + C s f ( ϵ ) e − α S ϵ /2 M 2 .
Altogether, we have
∥ ∇ T α ( x ) ∥ ≤ tr Cov ( η N ( x , Σ ) α ) σ min ( Σ ) 2 ≤ 1 σ min ( Σ ) 2 ( ϵ 2 + e − α S ϵ / 2 C s f ( ϵ ) M 2 ) . \lVert\nabla T_{\alpha}(x) \rVert\le\frac{\operatorname{tr}\operatorname{Cov}(\eta^{\alpha}_{\mathcal{N}(x,\Sigma)})}{\sigma_{\min}(\Sigma)^2} \le\frac{1}{\sigma_{\min}(\Sigma)^2}\left(\epsilon^2+\frac{e^{-\alpha S_{\epsilon}/2}}{C_{s_f(\epsilon)}}M_2\right). ∥ ∇ T α ( x )∥ ≤ σ m i n ( Σ ) 2 tr Cov ( η N ( x , Σ ) α ) ≤ σ m i n ( Σ ) 2 1 ( ϵ 2 + C s f ( ϵ ) e − α S ϵ /2 M 2 ) .
As an immediate corollary of the previous bound, we find that T α T_{\alpha} T α is a contraction for all α \alpha α large enough.
Corollary 13 (Contraction of the transition map) Let the assumptions of Lemma 12 hold, let L ∈ ( 0 , 1 ) L \in(0,1) L ∈ ( 0 , 1 ) , and set ϵ : = σ min ( Σ ) L / 2 \epsilon:= \sigma_{\min}(\Sigma)\sqrt{L/2} ϵ := σ m i n ( Σ ) L /2 . With α 0 \alpha_0 α 0 from (32) , we have
α 0 ( L σ min ( Σ ) 2 2 M 2 , S ϵ , C s f ( ϵ ) ) = 2 S ϵ max { 0 , log 2 M 2 L σ min ( Σ ) 2 C s f ( ϵ ) } . (52) \alpha_0\left(\frac{L\sigma_{\min}(\Sigma)^2}{2M_2},S_\epsilon,C_{s_f}(\epsilon)\right)=\frac{2}{S_\epsilon}\max\left\{0,\log\frac{2M_2}{L\sigma_{\min}(\Sigma)^2C_{s_f}(\epsilon)}\right\}. \tag*{(52)} α 0 ( 2 M 2 L σ m i n ( Σ ) 2 , S ϵ , C s f ( ϵ ) ) = S ϵ 2 max { 0 , log L σ m i n ( Σ ) 2 C s f ( ϵ ) 2 M 2 } . (52) Then, for all α ≥ α 0 ( L σ min ( Σ ) 2 2 M 2 , S ϵ , C s f ( ϵ ) ) \alpha\geq\alpha_0\left(\frac{L\sigma_{\min}(\Sigma)^2}{2M_2},S_\epsilon,C_{s_f}(\epsilon)\right) α ≥ α 0 ( 2 M 2 L σ m i n ( Σ ) 2 , S ϵ , C s f ( ϵ ) ) ,
sup x ∈ B r ( x ∗ ) ∥ ∇ T α ( x ) ∥ ≤ L , (53) \sup_{x\in B_r(x_*)}\left\|\nabla T_\alpha(x)\right\|\leq L, \tag*{(53)} x ∈ B r ( x ∗ ) sup ∥ ∇ T α ( x ) ∥ ≤ L , (53) and T α T_\alpha T α is Lipschitz continuous on B r ( x ∗ ) B_r(x_*) B r ( x ∗ ) with constant L L L , i.e.
∥ T α ( x ) − T α ( x ′ ) ∥ 2 ≤ L ∥ x − x ′ ∥ 2 for all x , x ′ ∈ B r ( x ∗ ) . (54) \left\|T_\alpha(x)-T_\alpha(x')\right\|_2\leq L\left\|x-x'\right\|_2\qquad\text{for all }x,x'\in B_r(x_*). \tag*{(54)} ∥ T α ( x ) − T α ( x ′ ) ∥ 2 ≤ L ∥ x − x ′ ∥ 2 for all x , x ′ ∈ B r ( x ∗ ) . (54) In particular, sup x ∈ B r ( x ∗ ) ∥ ∇ T α ( x ) ∥ → 0 \sup_{x\in B_r(x_*)}\left\|\nabla T_\alpha(x)\right\|\to0 sup x ∈ B r ( x ∗ ) ∥ ∇ T α ( x ) ∥ → 0 as α → ∞ \alpha\to\infty α → ∞ .
Proof. With ϵ = σ min ( Σ ) L / 2 \epsilon=\sigma_{\min}(\Sigma)\sqrt{L/2} ϵ = σ m i n ( Σ ) L /2 , the first term of L α ( ϵ ) L_\alpha(\epsilon) L α ( ϵ ) in (51) equals ϵ 2 / σ min ( Σ ) 2 = L / 2 \epsilon^2/\sigma_{\min}(\Sigma)^2=L/2 ϵ 2 / σ m i n ( Σ ) 2 = L /2 . For α ≥ α 0 ( L σ min ( Σ ) 2 2 M 2 , S ϵ , C s f ( ϵ ) ) \alpha\geq\alpha_0\left(\frac{L\sigma_{\min}(\Sigma)^2}{2M_2},S_\epsilon,C_{s_f}(\epsilon)\right) α ≥ α 0 ( 2 M 2 L σ m i n ( Σ ) 2 , S ϵ , C s f ( ϵ ) ) , (33) with δ = L σ min ( Σ ) 2 / ( 2 M 2 ) \delta=L\sigma_{\min}(\Sigma)^2/(2M_2) δ = L σ m i n ( Σ ) 2 / ( 2 M 2 ) gives
1 σ min ( Σ ) 2 e − α S ϵ / 2 C s f ( ϵ ) M 2 ≤ 1 σ min ( Σ ) 2 L σ min ( Σ ) 2 2 M 2 M 2 = L 2 . (55) \frac{1}{\sigma_{\min}(\Sigma)^2}\frac{e^{-\alpha S_\epsilon/2}}{C_{s_f}(\epsilon)}M_2\leq\frac{1}{\sigma_{\min}(\Sigma)^2}\frac{L\sigma_{\min}(\Sigma)^2}{2M_2}M_2=\frac{L}{2}. \tag*{(55)} σ m i n ( Σ ) 2 1 C s f ( ϵ ) e − α S ϵ /2 M 2 ≤ σ m i n ( Σ ) 2 1 2 M 2 L σ m i n ( Σ ) 2 M 2 = 2 L . (55)
Hence L α ( ϵ ) ≤ L L_\alpha(\epsilon)\leq L L α ( ϵ ) ≤ L , and Lemma 12 yields (53) .
Let x , x ′ ∈ B r ( x ∗ ) x,x'\in B_r(x_*) x , x ′ ∈ B r ( x ∗ ) . Since B r ( x ∗ ) B_r(x_*) B r ( x ∗ ) is convex, x ′ + t ( x − x ′ ) ∈ B r ( x ∗ ) x'+t(x-x')\in B_r(x_*) x ′ + t ( x − x ′ ) ∈ B r ( x ∗ ) for all t ∈ [ 0 , 1 ] t\in[0,1] t ∈ [ 0 , 1 ] . Since T α ∈ C 1 T_\alpha\in C^1 T α ∈ C 1 by Lemma 11 , the fundamental theorem of calculus and (53) give
∥ T α ( x ) − T α ( x ′ ) ∥ 2 = ∥ ∫ 0 1 ∇ T α ( x ′ + t ( x − x ′ ) ) ( x − x ′ ) d t ∥ 2 ≤ L ∥ x − x ′ ∥ 2 . (56) \left\|T_\alpha(x)-T_\alpha(x')\right\|_2=\left\|\int_0^1\nabla T_\alpha\left(x'+t(x-x')\right)(x-x')\,\mathrm{d}t\right\|_2\leq L\left\|x-x'\right\|_2. \tag*{(56)} ∥ T α ( x ) − T α ( x ′ ) ∥ 2 = ∫ 0 1 ∇ T α ( x ′ + t ( x − x ′ ) ) ( x − x ′ ) d t 2 ≤ L ∥ x − x ′ ∥ 2 . (56)
Since L ∈ ( 0 , 1 ) L\in(0,1) L ∈ ( 0 , 1 ) was arbitrary, the final claim follows.
We also need to show that a region near the global minimiser is invariant under the transition map for α \alpha α large enough.
Lemma 14 (Invariance of balls around the minimiser) Let f f f satisfy Assumption 1, let Σ ∈ R d × d \Sigma\in\mathbb{R}^{d\times d} Σ ∈ R d × d be symmetric positive definite, and let r > 0 r>0 r > 0 . Let C s C_s C s and M 1 M_1 M 1 be the constants of Lemma 10 for r r r and Σ \Sigma Σ , and let s f ( ϵ ) s_f(\epsilon) s f ( ϵ ) be as in (30) . Then the following hold.
(i) For all ϵ > 0 \epsilon>0 ϵ > 0 , α > 0 \alpha>0 α > 0 and x ∈ B r ( x ∗ ) x\in B_r(x_*) x ∈ B r ( x ∗ ) ,
∥ T α ( x ) − x ∗ ∥ 2 ≤ ϵ + e − α S ϵ / 2 C s f ( ϵ ) M 1 . (57) \left\|T_\alpha(x)-x_*\right\|_2\leq\epsilon+\frac{e^{-\alpha S_\epsilon/2}}{C_{s_f}(\epsilon)}M_1. \tag*{(57)} ∥ T α ( x ) − x ∗ ∥ 2 ≤ ϵ + C s f ( ϵ ) e − α S ϵ /2 M 1 . (57) (ii) Let ρ > 0 \rho>0 ρ > 0 . With α 0 \alpha_0 α 0 from (32) , we have
α 0 ( ρ 4 M 1 , S ρ / 2 , C s f ( ρ / 2 ) ) = 2 S ρ / 2 max { 0 , log 4 M 1 ρ C s f ( ρ / 2 ) } , (58) \alpha_0\left(\frac{\rho}{4M_1},S_{\rho/2},C_{s_f}(\rho/2)\right)=\frac{2}{S_{\rho/2}}\max\left\{0,\log\frac{4M_1}{\rho C_{s_f}(\rho/2)}\right\}, \tag*{(58)} α 0 ( 4 M 1 ρ , S ρ /2 , C s f ( ρ /2 ) ) = S ρ /2 2 max { 0 , log ρ C s f ( ρ /2 ) 4 M 1 } , (58) and for all α ≥ α 0 ( ρ 4 M 1 , S ρ / 2 , C s f ( ρ / 2 ) ) \alpha\geq\alpha_0\left(\frac{\rho}{4M_1},S_{\rho/2},C_{s_f}(\rho/2)\right) α ≥ α 0 ( 4 M 1 ρ , S ρ /2 , C s f ( ρ /2 ) ) ,
sup x ∈ B r ( x ∗ ) ∥ T α ( x ) − x ∗ ∥ 2 ≤ 3 4 ρ , in particular T α ( B r ( x ∗ ) ) ⊆ B ρ ( x ∗ ) . (59) \sup_{x\in B_r(x_*)}\left\|T_\alpha(x)-x_*\right\|_2\leq\frac{3}{4}\rho,\qquad\text{in particular}\qquad T_\alpha\left(B_r(x_*)\right)\subseteq B_\rho(x_*). \tag*{(59)} x ∈ B r ( x ∗ ) sup ∥ T α ( x ) − x ∗ ∥ 2 ≤ 4 3 ρ , in particular T α ( B r ( x ∗ ) ) ⊆ B ρ ( x ∗ ) . (59)
(iii) In particular, for ρ ∈ ( 0 , r ] \rho\in(0,r] ρ ∈ ( 0 , r ] and all α ≥ α 0 ( ρ 4 M 1 , S ρ / 2 , C s f ( ρ / 2 ) ) \alpha\ge\alpha_0\left(\frac{\rho}{4M_1},S_{\rho/2},C_{sf}(\rho/2)\right) α ≥ α 0 ( 4 M 1 ρ , S ρ /2 , C s f ( ρ /2 ) ) , the ball B r ( x ∗ ) B_r(x_*) B r ( x ∗ ) is invariant under T α T_\alpha T α :
T α ( B r ( x ∗ ) ) ⊆ B ρ ( x ∗ ) ⊆ B r ( x ∗ ) . (60) T_\alpha(B_r(x_*)) \subseteq B_\rho(x_*) \subseteq B_r(x_*). \tag*{(60)} T α ( B r ( x ∗ )) ⊆ B ρ ( x ∗ ) ⊆ B r ( x ∗ ) . (60)
Moreover, sup x ∈ B r ( x ∗ ) ∥ T α ( x ) − x ∗ ∥ 2 → 0 \sup_{x\in B_r(x_*)}\|T_\alpha(x)-x_*\|_2 \to0 sup x ∈ B r ( x ∗ ) ∥ T α ( x ) − x ∗ ∥ 2 → 0 as α → ∞ \alpha\to\infty α → ∞ .
Proof Part (i). Let x ∈ B r ( x ∗ ) x \in B_r(x_*) x ∈ B r ( x ∗ ) and write ν : = N ( x , Σ ) \nu:= \mathcal{N}(x,\Sigma) ν := N ( x , Σ ) for brevity. By Lemma 10 , ν \nu ν satisfies (28) with the constants C s C_s C s , and (35) for p = 1 p=1 p = 1 with M 1 M_1 M 1 . Both constants depend only on r r r , Σ \Sigma Σ and d d d , not on x x x . By Lemma 9 (ii) with p = 1 p=1 p = 1 , more precisely (37) , for all ϵ > 0 \epsilon> 0 ϵ > 0 and α > 0 \alpha> 0 α > 0 ,
∥ T α ( x ) − x ∗ ∥ 2 = ∥ E X ∼ η α x [ X ] − x ∗ ∥ 2 ≤ ϵ + e − α S ϵ / 2 C s f ( ϵ ) M 1 , (61) \|T_\alpha(x)-x_*\|_2=\|\mathbb{E}_{X\sim\eta_\alpha^x}[X]-x_*\|_2\leq\epsilon+\frac{e^{-\alpha S_{\epsilon/2}}}{C_{sf}(\epsilon)}M_1, \tag*{(61)} ∥ T α ( x ) − x ∗ ∥ 2 = ∥ E X ∼ η α x [ X ] − x ∗ ∥ 2 ≤ ϵ + C s f ( ϵ ) e − α S ϵ /2 M 1 , (61)
which is (57) .
Part (ii). Let ρ > 0 \rho>0 ρ > 0 and ϵ : = ρ / 2 \epsilon:=\rho/2 ϵ := ρ /2 . For α ≥ α 0 ( ρ 4 M 1 , S ρ / 2 , C s f ( ρ / 2 ) ) \alpha\geq\alpha_0\left(\frac{\rho}{4M_1},S_{\rho/2},C_{sf}(\rho/2)\right) α ≥ α 0 ( 4 M 1 ρ , S ρ /2 , C s f ( ρ /2 ) ) , (33) with δ = ρ / ( 4 M 1 ) \delta=\rho/(4M_1) δ = ρ / ( 4 M 1 ) gives
e − α S ρ / 2 / 2 C s f ( ρ / 2 ) M 1 ≤ ρ 4 M 1 M 1 = ρ 4 . (62) \frac{e^{-\alpha S_{\rho/2}/2}}{C_{sf}(\rho/2)}M_1\leq\frac{\rho}{4M_1}M_1=\frac{\rho}{4}. \tag*{(62)} C s f ( ρ /2 ) e − α S ρ /2 /2 M 1 ≤ 4 M 1 ρ M 1 = 4 ρ . (62)
Inserting (62) into (57) with ϵ = ρ / 2 \epsilon=\rho/2 ϵ = ρ /2 yields ∥ T α ( x ) − x ∗ ∥ 2 ≤ ρ 2 + ρ 4 = 3 4 ρ < ρ \|T_\alpha(x)-x_*\|_2\leq\frac{\rho}{2}+\frac{\rho}{4}=\frac{3}{4}\rho<\rho ∥ T α ( x ) − x ∗ ∥ 2 ≤ 2 ρ + 4 ρ = 4 3 ρ < ρ for all x ∈ B r ( x ∗ ) x\in B_r(x_*) x ∈ B r ( x ∗ ) . This proves (59) .
Part (iii). For ρ ∈ ( 0 , r ] \rho\in(0,r] ρ ∈ ( 0 , r ] , (60) follows from (59) and B ρ ( x ∗ ) ⊆ B r ( x ∗ ) B_\rho(x_*)\subseteq B_r(x_*) B ρ ( x ∗ ) ⊆ B r ( x ∗ ) . Since ρ > 0 \rho>0 ρ > 0 was arbitrary in (ii), sup x ∈ B r ( x ∗ ) ∥ T α ( x ) − x ∗ ∥ 2 ≤ 3 4 ρ \sup_{x\in B_r(x_*)}\|T_\alpha(x)-x_*\|_2\leq\frac{3}{4}\rho sup x ∈ B r ( x ∗ ) ∥ T α ( x ) − x ∗ ∥ 2 ≤ 4 3 ρ for all α ≥ α 0 ( ρ 4 M 1 , S ρ / 2 , C s f ( ρ / 2 ) ) \alpha\geq\alpha_0\left(\frac{\rho}{4M_1},S_{\rho/2},C_{sf}(\rho/2)\right) α ≥ α 0 ( 4 M 1 ρ , S ρ /2 , C s f ( ρ /2 ) ) , and hence the supremum tends to 0 0 0 as α → ∞ \alpha\to\infty α → ∞ .
Contraction and invariance alone do not quantify the distance of the iterates to x ∗ x_* x ∗ independently of the radius. For this, we evaluate the bias of the transition map at the minimiser itself, where we can use the constants of Lemma 10 with radius 0 0 0 ,
C s ∘ : = π d / 2 s d Γ ( d 2 + 1 ) ( 2 π σ max ( Σ ) 2 ) − d / 2 exp ( − s 2 2 σ min ( Σ ) 2 ) , M 1 ∘ : = σ max ( Σ ) d , (63) C_s^\circ:=\frac{\pi^{d/2}s^d}{\Gamma\left(\frac{d}{2}+1\right)}(2\pi\sigma_{\max}(\Sigma)^2)^{-d/2}\exp\left(-\frac{s^2}{2\sigma_{\min}(\Sigma)^2}\right),\qquad M_1^\circ:=\sigma_{\max}(\Sigma)\sqrt{d}, \tag*{(63)} C s ∘ := Γ ( 2 d + 1 ) π d /2 s d ( 2 π σ m a x ( Σ ) 2 ) − d /2 exp ( − 2 σ m i n ( Σ ) 2 s 2 ) , M 1 ∘ := σ m a x ( Σ ) d , (63)
which depend only on Σ \Sigma Σ , d d d and s > 0 s>0 s > 0 .
Lemma 15 (Bias at the minimiser) Let f f f satisfy Assumption 1, and let Σ ∈ R d × d \Sigma\in\mathbb{R}^{d\times d} Σ ∈ R d × d be symmetric positive definite. For β > 0 \beta>0 β > 0 , define
α b ( f , Σ , β ) : = α 0 ( β 2 M 1 ∘ , S β / 2 , C s f ∘ ( β / 2 ) ) (64) \alpha_b(f,\Sigma,\beta):=\alpha_0\left(\frac{\beta}{2M_1^\circ},S_{\beta/2},C_{sf}^\circ(\beta/2)\right) \tag*{(64)} α b ( f , Σ , β ) := α 0 ( 2 M 1 ∘ β , S β /2 , C s f ∘ ( β /2 ) ) (64) with α 0 \alpha_0 α 0 from (32) . Then ∥ T α ( x ∗ ) − x ∗ ∥ 2 ≤ β \|T_\alpha(x_*)-x_*\|_2\leq\beta ∥ T α ( x ∗ ) − x ∗ ∥ 2 ≤ β for all α > 0 \alpha>0 α > 0 with α ≥ α b ( f , Σ , β ) \alpha\geq\alpha_b(f,\Sigma,\beta) α ≥ α b ( f , Σ , β ) .
Proof Let ν ∘ : = N ( x ∗ , Σ ) \nu^\circ:=\mathcal{N}(x_*,\Sigma) ν ∘ := N ( x ∗ , Σ ) . For every r ′ > 0 r'>0 r ′ > 0 , Lemma 10 with radius r ′ r' r ′ and x = x ∗ ∈ B r ′ ( x ∗ ) x=x_*\in B_{r'}(x_*) x = x ∗ ∈ B r ′ ( x ∗ ) gives lower bounds on ν ∘ ( B s ( x ∗ ) ) \nu^\circ(B_s(x_*)) ν ∘ ( B s ( x ∗ )) and upper bounds on E X ∼ ν ∘ [ ∥ X − x ∗ ∥ 2 ] \mathbb{E}_{X\sim\nu^\circ}[\|X-x_*\|_2] E X ∼ ν ∘ [ ∥ X − x ∗ ∥ 2 ] , which depend continuously on r ′ r' r ′ , while the left-hand sides do not depend on r ′ r' r ′ . Letting r ′ → 0 r'\to0 r ′ → 0 yields ν ∘ ( B s ( x ∗ ) ) ≥ C s ∘ > 0 \nu^\circ(B_s(x_*))\geq C_s^\circ>0 ν ∘ ( B s ( x ∗ )) ≥ C s ∘ > 0 for all s > 0 s>0 s > 0 and E X ∼ ν ∘ [ ∥ X − x ∗ ∥ 2 ] ≤ M 1 ∘ \mathbb{E}_{X\sim\nu^\circ}[\|X-x_*\|_2]\leq M_1^\circ E X ∼ ν ∘ [ ∥ X − x ∗ ∥ 2 ] ≤ M 1 ∘ . Hence Lemma 9 (ii) with p = 1 p=1 p = 1 applies to ν ∘ \nu^\circ ν ∘ with the constants C s ∘ C_s^\circ C s ∘ and M 1 ∘ M_1^\circ M 1 ∘ . By Lemma 11 (i), T α ( x ∗ ) T_\alpha(x_*) T α ( x ∗ ) is the mean of η ν ∘ α \eta_{\nu^\circ}^\alpha η ν ∘ α , so (37) with ϵ = β / 2 \epsilon=\beta/2 ϵ = β /2 gives
∥ T α ( x ∗ ) − x ∗ ∥ 2 ≤ β 2 + e − α S β / 2 / 2 C s f ∘ ( β / 2 ) M 1 ∘ . \|T_\alpha(x_*)-x_*\|_2\leq\frac{\beta}{2}+\frac{e^{-\alpha S_{\beta/2}/2}}{C_{sf}^\circ(\beta/2)}M_1^\circ. ∥ T α ( x ∗ ) − x ∗ ∥ 2 ≤ 2 β + C s f ∘ ( β /2 ) e − α S β /2 /2 M 1 ∘ . For α ≥ α b ( f , Σ , β ) \alpha\geq\alpha_b(f,\Sigma,\beta) α ≥ α b ( f , Σ , β ) , (33) with δ = β / ( 2 M 1 ∘ ) \delta=\beta/(2M_1^\circ) δ = β / ( 2 M 1 ∘ ) bounds the second term by β / 2 \beta/2 β /2 .
Main Results: Proofs and Discussion
We now prove the results of § 3 . First, we establish an invariance and contraction property of the mean-field transition map. This result follows from the results above using essentially the same arguments as in Fornasier et al., [5 ] .
Proof [Proof of Theorem 1] Let r > 0 r > 0 r > 0 and ρ ∈ ( 0 , 1 ) \rho\in(0,1) ρ ∈ ( 0 , 1 ) , let C s C_s C s and M 2 M_2 M 2 be the constants of Lemma 10 for r r r and Σ \Sigma Σ , let ϵ ρ : = σ min ( Σ ) ρ / 2 \epsilon_\rho:= \sigma_{\min}(\Sigma)\sqrt{\rho/2} ϵ ρ := σ m i n ( Σ ) ρ /2 , and define
α 0 ( f , Σ , r , ρ ) : = max { α 0 ( ρ σ min ( Σ ) 2 2 M 2 , S ϵ ρ , C s f ( ϵ ρ ) ) , α b ( f , Σ , ( 1 − ρ ) r ) } , (65) \alpha_0(f,\Sigma,r,\rho) := \max\left\{\alpha_0\left(\frac{\rho\sigma_{\min}(\Sigma)^2}{2M_2}, S_{\epsilon_\rho}, C_{s_f(\epsilon_\rho)}\right), \alpha_b(f,\Sigma,(1-\rho)r)\right\}, \tag*{(65)} α 0 ( f , Σ , r , ρ ) := max { α 0 ( 2 M 2 ρ σ m i n ( Σ ) 2 , S ϵ ρ , C s f ( ϵ ρ ) ) , α b ( f , Σ , ( 1 − ρ ) r ) } , (65)
with α 0 ( ⋅ , ⋅ , ⋅ ) \alpha_0(\cdot,\cdot,\cdot) α 0 ( ⋅ , ⋅ , ⋅ ) from (32) and α b \alpha_b α b from (64) . It depends only on f f f (through S ϵ S_\epsilon S ϵ and s f ( ϵ ) s_f(\epsilon) s f ( ϵ ) ), on Σ \Sigma Σ , r r r , ρ \rho ρ and the dimension d d d . Let α > 0 \alpha> 0 α > 0 with α ≥ α 0 ( f , Σ , r , ρ ) \alpha\ge\alpha_0(f,\Sigma,r,\rho) α ≥ α 0 ( f , Σ , r , ρ ) .
Ad (i). The first threshold in (65) is (52) with L = ρ L=\rho L = ρ , so Corollary 13 yields (9) . By Lemma 15, ∥ T α ( x ∗ ) − x ∗ ∥ 2 ≤ ( 1 − ρ ) r \lVert T_\alpha(x_\ast)-x_\ast\rVert_2 \le(1-\rho)r ∥ T α ( x ∗ ) − x ∗ ∥ 2 ≤ ( 1 − ρ ) r . Hence, for x ∈ B r ( x ∗ ) x \in B_r(x_\ast) x ∈ B r ( x ∗ ) ,
∥ T α ( x ) − x ∗ ∥ 2 ≤ ∥ T α ( x ) − T α ( x ∗ ) ∥ 2 + ∥ T α ( x ∗ ) − x ∗ ∥ 2 ≤ ρ r + ( 1 − ρ ) r = r , (66) \lVert T_\alpha(x)-x_\ast\rVert_2 \le\lVert T_\alpha(x)-T_\alpha(x_\ast)\rVert_2+\lVert T_\alpha(x_\ast)-x_\ast\rVert_2 \le\rho r+(1-\rho)r=r, \tag*{(66)} ∥ T α ( x ) − x ∗ ∥ 2 ≤ ∥ T α ( x ) − T α ( x ∗ ) ∥ 2 + ∥ T α ( x ∗ ) − x ∗ ∥ 2 ≤ ρ r + ( 1 − ρ ) r = r , (66)
i.e. T α ( B r ( x ∗ ) ) ⊆ B r ( x ∗ ) T_\alpha(B_r(x_\ast))\subseteq B_r(x_\ast) T α ( B r ( x ∗ )) ⊆ B r ( x ∗ ) .
Ad (ii). The closed ball B r ( x ∗ ) B_r(x_\ast) B r ( x ∗ ) is compact, T α T_\alpha T α maps it into itself by (i), and it is a contraction on it by (9) . Banach’s fixed-point theorem yields a unique fixed point m α ∈ B r ( x ∗ ) m_\alpha\in B_r(x_\ast) m α ∈ B r ( x ∗ ) . Let x 0 ∈ B r ( x ∗ ) x_0 \in B_r(x_\ast) x 0 ∈ B r ( x ∗ ) . By (i) and induction, x k ∈ B r ( x ∗ ) x_k \in B_r(x_\ast) x k ∈ B r ( x ∗ ) for all k ∈ N 0 k \in\mathbb{N}_0 k ∈ N 0 . Since m α = T α ( m α ) m_\alpha=T_\alpha(m_\alpha) m α = T α ( m α ) , (9) gives ∥ x k + 1 − m α ∥ 2 ≤ ρ ∥ x k − m α ∥ 2 \lVert x_{k+1}-m_\alpha\rVert_2\le\rho\lVert x_k-m_\alpha\rVert_2 ∥ x k + 1 − m α ∥ 2 ≤ ρ ∥ x k − m α ∥ 2 , and induction yields (10) . Finally, since m α ∈ B r ( x ∗ ) m_\alpha\in B_r(x_\ast) m α ∈ B r ( x ∗ ) and m α = T α ( m α ) m_\alpha=T_\alpha(m_\alpha) m α = T α ( m α ) , we have
∥ m α − x ∗ ∥ 2 ≤ sup x ∈ B r ( x ∗ ) ∥ T α ( x ) − x ∗ ∥ 2 , \lVert m_\alpha-x_\ast\rVert_2\le\sup_{x\in B_r(x_\ast)}\lVert T_\alpha(x)-x_\ast\rVert_2, ∥ m α − x ∗ ∥ 2 ≤ x ∈ B r ( x ∗ ) sup ∥ T α ( x ) − x ∗ ∥ 2 ,
and the right-hand side tends to 0 0 0 as α → ∞ \alpha\to\infty α → ∞ by Lemma 14 (iii). ■ \blacksquare ■
We can now turn to Theorem 2 . The first part is similar to Fornasier et al., [5 ] , though we show it under weaker assumptions with a more direct argument. The second part has been anticipated in Fornasier et al., [5 ] , but not formally stated and proven yet.
Proof [Proof of Theorem 2 ] Let β > 0 \beta>0 β > 0 , let α b ( f , Σ , β ) \alpha_b(f,\Sigma,\beta) α b ( f , Σ , β ) be as in (64) , which only depends on f f f , Σ \Sigma Σ and β \beta β , and let r > 0 r>0 r > 0 , ρ ∈ ( 0 , 1 ) \rho\in(0,1) ρ ∈ ( 0 , 1 ) and α ≥ max { α 0 ( f , Σ , r , ρ ) , α b ( f , Σ , β ) } \alpha\ge\max\{\alpha_0(f,\Sigma,r,\rho),\alpha_b(f,\Sigma,\beta)\} α ≥ max { α 0 ( f , Σ , r , ρ ) , α b ( f , Σ , β )} . Set β r : = min { β , ( 1 − ρ ) r } \beta_r:=\min\{\beta,(1-\rho)r\} β r := min { β , ( 1 − ρ ) r } . Since α b ( f , Σ , β r ) \alpha_b(f,\Sigma,\beta_r) α b ( f , Σ , β r ) equals either α b ( f , Σ , β ) \alpha_b(f,\Sigma,\beta) α b ( f , Σ , β ) or α b ( f , Σ , ( 1 − ρ ) r ) ≤ α 0 ( f , Σ , r , ρ ) \alpha_b(f,\Sigma,(1-\rho)r)\le\alpha_0(f,\Sigma,r,\rho) α b ( f , Σ , ( 1 − ρ ) r ) ≤ α 0 ( f , Σ , r , ρ ) , Lemma 15 gives ∥ T α ( x ∗ ) − x ∗ ∥ 2 ≤ β r ≤ β \lVert T_\alpha(x_\ast)-x_\ast\rVert_2\le\beta_r\le\beta ∥ T α ( x ∗ ) − x ∗ ∥ 2 ≤ β r ≤ β .
Ad (i). Let x ∈ B r ( x ∗ ) x\in B_r(x_\ast) x ∈ B r ( x ∗ ) . Since also x ∗ ∈ B r ( x ∗ ) x_\ast\in B_r(x_\ast) x ∗ ∈ B r ( x ∗ ) , (9) gives
∥ T α ( x ) − x ∗ ∥ 2 ≤ ∥ T α ( x ) − T α ( x ∗ ) ∥ 2 + ∥ T α ( x ∗ ) − x ∗ ∥ 2 ≤ ρ ∥ x − x ∗ ∥ 2 + β . \lVert T_\alpha(x)-x_\ast\rVert_2\le\lVert T_\alpha(x)-T_\alpha(x_\ast)\rVert_2+\lVert T_\alpha(x_\ast)-x_\ast\rVert_2\le\rho\lVert x-x_\ast\rVert_2+\beta. ∥ T α ( x ) − x ∗ ∥ 2 ≤ ∥ T α ( x ) − T α ( x ∗ ) ∥ 2 + ∥ T α ( x ∗ ) − x ∗ ∥ 2 ≤ ρ ∥ x − x ∗ ∥ 2 + β .
Ad (ii). By Theorem 1(ii), x k ∈ B r ( x ∗ ) x_k\in B_r(x_\ast) x k ∈ B r ( x ∗ ) for all k ∈ N 0 k\in\mathbb{N}_0 k ∈ N 0 . Writing e k : = ∥ x k − x ∗ ∥ 2 e_k:=\lVert x_k-x_\ast\rVert_2 e k := ∥ x k − x ∗ ∥ 2 , (i) gives e k + 1 ≤ β + ρ e k e_{k+1}\le\beta+\rho e_k e k + 1 ≤ β + ρ e k , and induction yields
e k ≤ ρ k e 0 + β ∑ j = 0 k − 1 ρ j ≤ ρ k e 0 + β 1 − ρ . e_k\le\rho^k e_0+\beta\sum_{j=0}^{k-1}\rho^j\le\rho^k e_0+\frac{\beta}{1-\rho}. e k ≤ ρ k e 0 + β j = 0 ∑ k − 1 ρ j ≤ ρ k e 0 + 1 − ρ β .
■ \blacksquare ■
As an immediate consequence, we get Corollary 3.
Proof [Proof of Corollary 3] Apply Theorem 2 with ρ = 1 2 \rho= \frac{1}{2} ρ = 2 1 and β = ϵ 4 \beta= \frac{\epsilon}{4} β = 4 ϵ . For x 0 ∈ B r ( x ∗ ) x_0 \in B_r(x_*) x 0 ∈ B r ( x ∗ ) and k ≥ K 0 ( r , ϵ ) k \ge K_0(r,\epsilon) k ≥ K 0 ( r , ϵ ) , (12) gives ∥ x k − x ∗ ∥ 2 ≤ 2 − k r + ϵ 2 ≤ ϵ \|x_k-x_*\|_2 \le2^{-k}r+\frac{\epsilon}{2} \le\epsilon ∥ x k − x ∗ ∥ 2 ≤ 2 − k r + 2 ϵ ≤ ϵ , since 2 k ≥ 2 r / ϵ 2^k \ge2r/\epsilon 2 k ≥ 2 r / ϵ if K 0 ( r , ϵ ) > 0 K_0(r,\epsilon)>0 K 0 ( r , ϵ ) > 0 , and r ≤ ϵ / 2 r \le\epsilon/2 r ≤ ϵ /2 otherwise.
Finally, let us discuss in detail how the preceding results show that iterating in mean-field MPPI can indeed improve over just a single step of MPPI. This phenomenon has already been described in [5 Remark 4.5] , but using two different thresholds for the inverse temperature allows us to discuss this quantitatively.
Remark 16 (Iterations help) To reach accuracy ϵ \epsilon ϵ in a single step from every point of B r ( x ∗ ) B_r(x_*) B r ( x ∗ ) , Lemma 14 (ii) requires
α ≥ α 0 ( ϵ 4 M 1 , S ϵ / 2 , C S f ( ϵ / 2 ) ) ≥ 2 S ϵ / 2 log 4 M 1 ϵ C S f ( ϵ / 2 ) . \alpha\ge\alpha_0\left(\frac{\epsilon}{4M_1},S_{\epsilon/2},C_{S_f}(\epsilon/2)\right) \ge\frac{2}{S_{\epsilon/2}}\log\frac{4M_1}{\epsilon C_{S_f}(\epsilon/2)}. α ≥ α 0 ( 4 M 1 ϵ , S ϵ /2 , C S f ( ϵ /2 ) ) ≥ S ϵ /2 2 log ϵ C S f ( ϵ /2 ) 4 M 1 .
and by Lemma 10 ,
log 1 C s = ( r + s ) 2 2 σ min ( Σ ) 2 − log π d / 2 s d Γ ( d 2 + 1 ) + d 2 log ( 2 π σ max ( Σ ) 2 ) . \log\frac{1}{C_s}=\frac{(r+s)^2}{2\sigma_{\min}(\Sigma)^2}-\log\frac{\pi^{d/2}s^d}{\Gamma\left(\frac{d}{2}+1\right)}+\frac{d}{2}\log\left(2\pi\sigma_{\max}(\Sigma)^2\right). log C s 1 = 2 σ m i n ( Σ ) 2 ( r + s ) 2 − log Γ ( 2 d + 1 ) π d /2 s d + 2 d log ( 2 π σ m a x ( Σ ) 2 ) .
Hence this threshold contains the product of 1 / S ϵ / 2 1/S_{\epsilon/2} 1/ S ϵ /2 , which is large for small ϵ \epsilon ϵ , and r 2 / ( 2 σ min ( Σ ) 2 ) r^2/(2\sigma_{\min}(\Sigma)^2) r 2 / ( 2 σ m i n ( Σ ) 2 ) , which is large for large r r r . In Corollary 3, the radius only enters the threshold α 0 ( f , Σ , r , 1 2 ) \alpha_0(f,\Sigma,r,\frac{1}{2}) α 0 ( f , Σ , r , 2 1 ) . There, r 2 / ( 2 σ min ( Σ ) 2 ) r^2/(2\sigma_{\min}(\Sigma)^2) r 2 / ( 2 σ m i n ( Σ ) 2 ) is divided by quantities that do not depend on ϵ \epsilon ϵ . The accuracy only enters the threshold α b ( f , Σ , ϵ 4 ) \alpha_b(f,\Sigma,\frac{\epsilon}{4}) α b ( f , Σ , 4 ϵ ) , which does not depend on r r r . If, for instance, S ϵ = ( c ϵ ) q S_\epsilon=(c\epsilon)^q S ϵ = ( cϵ ) q for small ϵ \epsilon ϵ with constants c , q > 0 c,q>0 c , q > 0 , then up to logarithmic factors the single-step threshold grows like r 2 ϵ − q r^2\epsilon^{-q} r 2 ϵ − q , whereas the threshold of Corollary 3 grows like max { r 2 , ϵ − q } \max\{r^2,\epsilon^{-q}\} max { r 2 , ϵ − q } . The price is K 0 ( r , ϵ ) = O ( log ( r / ϵ ) ) K_0(r,\epsilon)=O(\log(r/\epsilon)) K 0 ( r , ϵ ) = O ( log ( r / ϵ )) iterations.
Finite-Sample Analysis
We start with a finite-sample bound for the estimation error of the empirical transition map interpreted as a self-normalised importance-weighted estimator of the mean-field transition map. For clarity, we state and prove a more general result that makes clear which properties of the Gibbs tilt and the self-normalised importance weighted estimator are actually used.
Lemma 17 (One-step concentration of the empirical transition map) Let f : R d → R f:\mathbb{R}^d\to\mathbb{R} f : R d → R be Borel measurable and bounded below, f ‾ : = inf y ∈ R d f ( y ) > − ∞ \underline{f}:=\inf_{y\in\mathbb{R}^d}f(y)>-\infty f := inf y ∈ R d f ( y ) > − ∞ , let α > 0 \alpha>0 α > 0 , let Σ ∈ R d × d \Sigma\in\mathbb{R}^{d\times d} Σ ∈ R d × d be symmetric positive definite, let N ∈ N + N\in\mathbb{N}_+ N ∈ N + and x ∈ R d x\in\mathbb{R}^d x ∈ R d , and let X 1 , … , X N X_1,\ldots,X_N X 1 , … , X N be independent with X i ∼ N ( x , Σ ) X_i\sim\mathcal{N}(x,\Sigma) X i ∼ N ( x , Σ ) .
(i) For all t > 0 t>0 t > 0 ,
P ( ∥ T ~ α , N ( x ) − T α ( x ) ∥ 2 > t ) ≤ 4 N e − α f ‾ Z α ( N ( x , Σ ) ) ( 1 + V α ( x ) t 2 ) , (67) \mathbb{P}\left(\left\|\widetilde{T}_{\alpha,N}(x)-T_\alpha(x)\right\|_2>t\right)\le\frac{4}{N}\frac{e^{-\alpha\underline{f}}}{Z_\alpha(\mathcal{N}(x,\Sigma))}\left(1+\frac{V_\alpha(x)}{t^2}\right), \tag*{(67)} P ( T α , N ( x ) − T α ( x ) 2 > t ) ≤ N 4 Z α ( N ( x , Σ )) e − α f ( 1 + t 2 V α ( x ) ) , (67) where
V α ( x ) : = E Y ∼ η N ( x , Σ ) α [ ∥ Y − T α ( x ) ∥ 2 2 ] < ∞ (68) V_\alpha(x):=\mathbb{E}_{Y\sim\eta^\alpha_{\mathcal{N}(x,\Sigma)}}\left[\left\|Y-T_\alpha(x)\right\|_2^2\right]<\infty \tag*{(68)} V α ( x ) := E Y ∼ η N ( x , Σ ) α [ ∥ Y − T α ( x ) ∥ 2 2 ] < ∞ (68) is the total variance of the Gibbs tilt η N ( x , Σ ) α \eta^\alpha_{\mathcal{N}(x,\Sigma)} η N ( x , Σ ) α .
(ii) Let f f f in addition satisfy Assumption 1, let r > 0 r>0 r > 0 , let C s C_s C s and M 2 M_2 M 2 be the constants of Lemma 10 for r r r and Σ \Sigma Σ , and let s f ( ϵ ) s_f(\epsilon) s f ( ϵ ) be as in (30) . Then for all x ∈ B r ( x ∗ ) x\in B_r(x_*) x ∈ B r ( x ∗ ) , ϵ , ϵ ′ > 0 \epsilon,\epsilon'>0 ϵ , ϵ ′ > 0 and t > 0 t>0 t > 0 ,
P ( ∥ T ^ α , N ( x ) − T α ( x ) ∥ 2 > t ) ≤ 4 N e α S ϵ / 2 C s f ( ϵ ) ( 1 + 1 t 2 ( ϵ ′ 2 + e − α S ϵ ′ / 2 C s f ( ϵ ′ ) M 2 ) ) . (69) \mathbb{P}\left(\left\|\widehat{T}_{\alpha,N}(x)-T_\alpha(x)\right\|_2>t\right)
\leq\frac{4}{N}\frac{e^{\alpha S\epsilon/2}}{C s_f(\epsilon)}
\left(1+\frac{1}{t^2}\left(\epsilon'^2+\frac{e^{-\alpha S\epsilon'/2}}{C s_f(\epsilon')}M_2\right)\right). \tag*{(69)} P ( T α , N ( x ) − T α ( x ) 2 > t ) ≤ N 4 C s f ( ϵ ) e α S ϵ /2 ( 1 + t 2 1 ( ϵ ′2 + C s f ( ϵ ′ ) e − α S ϵ ′ /2 M 2 ) ) . (69)
Proof Ad (i). Write ν : = N ( x , Σ ) \nu:=\mathcal{N}(x,\Sigma) ν := N ( x , Σ ) , w ( y ) : = e − α f ( y ) w(y):=e^{-\alpha f(y)} w ( y ) := e − α f ( y ) , W : = e − α f ∗ W:=e^{-\alpha f_*} W := e − α f ∗ , Z : = Z α ( ν ) Z:=Z_\alpha(\nu) Z := Z α ( ν ) and c : = T α ( x ) c:=T_\alpha(x) c := T α ( x ) , and let X ∼ ν X\sim\nu X ∼ ν . Then 0 < w ≤ W 0<w\leq W 0 < w ≤ W , and by Lemma 11(i) we have 0 < Z ≤ W 0<Z\leq W 0 < Z ≤ W . Moreover, by Lemma 11(i), c c c is well defined and is the mean of η ν α \eta^\alpha_\nu η ν α , the definition (2) gives E [ w ( X ) X ] = Z c \mathbb{E}[w(X)X]=Zc E [ w ( X ) X ] = Z c , and since η ν α \eta^\alpha_\nu η ν α has finite second moments, V α ( x ) ≤ 2 E Y ∼ η ν α [ ∥ Y ∥ 2 2 ] + 2 ∥ c ∥ 2 2 < ∞ V_\alpha(x)\leq2\mathbb{E}_{Y\sim\eta^\alpha_\nu}[\|Y\|_2^2]+2\|c\|_2^2<\infty V α ( x ) ≤ 2 E Y ∼ η ν α [ ∥ Y ∥ 2 2 ] + 2∥ c ∥ 2 2 < ∞ .
We first collect two moment bounds. Since 0 < w ≤ W 0<w\leq W 0 < w ≤ W ,
Var ( w ( X ) ) ≤ E [ w ( X ) 2 ] ≤ W E [ w ( X ) ] = W Z . (70) \operatorname{Var}(w(X))\leq\mathbb{E}[w(X)^2]\leq W\mathbb{E}[w(X)]=WZ. \tag*{(70)} Var ( w ( X )) ≤ E [ w ( X ) 2 ] ≤ W E [ w ( X )] = W Z . (70)
Moreover, (49) with h ( y ) = ∥ y − c ∥ 2 2 h(y)=\|y-c\|_2^2 h ( y ) = ∥ y − c ∥ 2 2 gives
E [ w ( X ) 2 ∥ X − c ∥ 2 2 ] ≤ W E [ w ( X ) ∥ X − c ∥ 2 2 ] = W Z V α ( x ) . (71) \mathbb{E}[w(X)^2\|X-c\|_2^2]\leq W\mathbb{E}[w(X)\|X-c\|_2^2]=WZV_\alpha(x). \tag*{(71)} E [ w ( X ) 2 ∥ X − c ∥ 2 2 ] ≤ W E [ w ( X ) ∥ X − c ∥ 2 2 ] = W Z V α ( x ) . (71)
Now define
Y i : = w ( X i ) ( X i − c ) , A N : = 1 N ∑ i = 1 N Y i , B N : = 1 N ∑ i = 1 N w ( X i ) . Y_i:=w(X_i)(X_i-c),\qquad A_N:=\frac{1}{N}\sum_{i=1}^{N}Y_i,\qquad B_N:=\frac{1}{N}\sum_{i=1}^{N}w(X_i). Y i := w ( X i ) ( X i − c ) , A N := N 1 i = 1 ∑ N Y i , B N := N 1 i = 1 ∑ N w ( X i ) .
The Y i Y_i Y i are independent and identically distributed, square-integrable by (71) , and centred, since E [ Y i ] = E [ w ( X ) X ] − c E [ w ( X ) ] = Z c − c Z = 0 \mathbb{E}[Y_i]=\mathbb{E}[w(X)X]-c\mathbb{E}[w(X)]=Zc-cZ=0 E [ Y i ] = E [ w ( X ) X ] − c E [ w ( X )] = Z c − c Z = 0 . Since B N > 0 B_N>0 B N > 0 , we have
T ^ α , N ( x ) − c = A N B N . \widehat{T}_{\alpha,N}(x)-c=\frac{A_N}{B_N}. T α , N ( x ) − c = B N A N .
Let E N : = { B N ≥ Z / 2 } E_N:=\{B_N\geq Z/2\} E N := { B N ≥ Z /2 } . Then
P ( ∥ T ^ α , N ( x ) − c ∥ 2 > t ) ≤ P ( E N c ) + P ( ∥ T ^ α , N ( x ) − c ∥ 2 > t , E N ) . \mathbb{P}\left(\left\|\widehat{T}_{\alpha,N}(x)-c\right\|_2>t\right)
\leq\mathbb{P}(E_N^c)+\mathbb{P}\left(\left\|\widehat{T}_{\alpha,N}(x)-c\right\|_2>t,E_N\right). P ( T α , N ( x ) − c 2 > t ) ≤ P ( E N c ) + P ( T α , N ( x ) − c 2 > t , E N ) .
Since E [ B N ] = Z \mathbb{E}[B_N]=Z E [ B N ] = Z and Var ( B N ) = Var ( w ( X ) ) / N \operatorname{Var}(B_N)=\operatorname{Var}(w(X))/N Var ( B N ) = Var ( w ( X )) / N , Chebyshev’s inequality and (70) yield
P ( E N c ) ≤ P ( ∣ B N − Z ∣ ≥ Z / 2 ) ≤ 4 Var ( w ( X ) ) N Z 2 ≤ 4 N W Z . \mathbb{P}(E_N^c)\leq\mathbb{P}(|B_N-Z|\geq Z/2)\leq\frac{4\operatorname{Var}(w(X))}{NZ^2}\leq\frac{4}{N}\frac{W}{Z}. P ( E N c ) ≤ P ( ∣ B N − Z ∣ ≥ Z /2 ) ≤ N Z 2 4 Var ( w ( X )) ≤ N 4 Z W .
On E N E_N E N , we have ∥ T ^ α , N ( x ) − c ∥ 2 2 = ∥ A N ∥ 2 2 / B N ≤ 2 ∥ A N ∥ 2 2 / Z \|\widehat{T}_{\alpha,N}(x)-c\|_2^2=\|A_N\|_2^2/B_N\leq2\|A_N\|_2^2/Z ∥ T α , N ( x ) − c ∥ 2 2 = ∥ A N ∥ 2 2 / B N ≤ 2∥ A N ∥ 2 2 / Z . Since the Y i Y_i Y i are independent and centred, the cross terms vanish, and with (71) we get
E [ ∥ A N ∥ 2 2 ] = 1 N 2 ∑ i , j = 1 N E [ ⟨ Y i , Y j ⟩ ] = 1 N E [ ∥ Y 1 ∥ 2 2 ] ≤ W Z V α ( x ) N . \mathbb{E}[\|A_N\|_2^2]=\frac{1}{N^2}\sum_{i,j=1}^{N}\mathbb{E}[\langle Y_i,Y_j\rangle]=\frac{1}{N}\mathbb{E}[\|Y_1\|_2^2]\leq\frac{WZV_\alpha(x)}{N}. E [ ∥ A N ∥ 2 2 ] = N 2 1 i , j = 1 ∑ N E [⟨ Y i , Y j ⟩] = N 1 E [ ∥ Y 1 ∥ 2 2 ] ≤ N W Z V α ( x ) .
Markov’s inequality applied to ∥ A N ∥ 2 2 \|A_N\|_2^2 ∥ A N ∥ 2 2 therefore gives
P ( ∥ T ^ α , N ( x ) − c ∥ 2 > t , E N ) ≤ P ( ∥ A N ∥ 2 > t Z / 2 ) ≤ 4 E [ ∥ A N ∥ 2 2 ] t 2 Z 2 ≤ 4 N W Z V α ( x ) t 2 . \mathbb{P}\left(\left\|\widehat{T}_{\alpha,N}(x)-c\right\|_2>t,E_N\right)
\leq\mathbb{P}(\|A_N\|_2>tZ/2)
\leq\frac{4\mathbb{E}[\|A_N\|_2^2]}{t^2Z^2}
\leq\frac{4}{N}\frac{W}{Z}\frac{V_\alpha(x)}{t^2}. P ( T α , N ( x ) − c 2 > t , E N ) ≤ P ( ∥ A N ∥ 2 > tZ /2 ) ≤ t 2 Z 2 4 E [ ∥ A N ∥ 2 2 ] ≤ N 4 Z W t 2 V α ( x ) .
Adding the two bounds proves (67) .
Ad (ii). By (F1), f = f ∗ f=f_\ast f = f ∗ . Let x ∈ B r ( x ∗ ) x\in B_r(x_\ast) x ∈ B r ( x ∗ ) , ν : = N ( x , Σ ) \nu:=\mathcal{N}(x,\Sigma) ν := N ( x , Σ ) , and ϵ , ϵ ′ > 0 \epsilon,\epsilon'>0 ϵ , ϵ ′ > 0 . By Lemma 10 , ν \nu ν satisfies (28) with the constants C s C_s C s and (35) for p = 2 p=2 p = 2 with M 2 M_2 M 2 . Restricting the integral to B s f ( ϵ ) ( x ∗ ) B_{s_f(\epsilon)}(x_\ast) B s f ( ϵ ) ( x ∗ ) and using (30) and (28) , we obtain, as in (39) ,
Z α ( ν ) ≥ ∫ B s f ( ϵ ) ( x ∗ ) e − α f ( y ) ν ( d y ) ≥ e − α ( f ∗ + S ϵ / 2 ) ν ( B s f ( ϵ ) ( x ∗ ) ) ≥ e − α ( f ∗ + S ϵ / 2 ) C s f ( ϵ ) , Z_\alpha(\nu)\ge\int_{B_{s_f(\epsilon)}(x_\ast)}e^{-\alpha f(y)}\,\nu(\mathrm{d}y)\ge e^{-\alpha(f_\ast+S_\epsilon/2)}\nu\left(B_{s_f(\epsilon)}(x_\ast)\right)\ge e^{-\alpha(f_\ast+S_\epsilon/2)}C_{s_f(\epsilon)}, Z α ( ν ) ≥ ∫ B s f ( ϵ ) ( x ∗ ) e − α f ( y ) ν ( d y ) ≥ e − α ( f ∗ + S ϵ /2 ) ν ( B s f ( ϵ ) ( x ∗ ) ) ≥ e − α ( f ∗ + S ϵ /2 ) C s f ( ϵ ) ,
and hence e − α f ∗ / Z α ( ν ) ≤ e α S ϵ / 2 / C s f ( ϵ ) e^{-\alpha f_\ast}/Z_\alpha(\nu)\le e^{\alpha S_\epsilon/2}/C_{s_f(\epsilon)} e − α f ∗ / Z α ( ν ) ≤ e α S ϵ /2 / C s f ( ϵ ) . Furthermore, since the mean minimises the expected squared distance, Lemma 9 (ii) with p = 2 p=2 p = 2 and ϵ ′ \epsilon' ϵ ′ in place of ϵ \epsilon ϵ yields
V α ( x ) ≤ E Y ∼ η α ν [ ∥ Y − x ∗ ∥ 2 2 ] ≤ ϵ ′ 2 + e − α S ϵ ′ / 2 C s f ( ϵ ′ ) M 2 . V_\alpha(x)\le\mathbb{E}_{Y\sim\eta_\alpha^\nu}\left[\lVert Y-x_\ast\rVert_2^2\right]\le\epsilon'^2+\frac{e^{-\alpha S_{\epsilon'}/2}}{C_{s_f(\epsilon')}}M_2. V α ( x ) ≤ E Y ∼ η α ν [ ∥ Y − x ∗ ∥ 2 2 ] ≤ ϵ ′2 + C s f ( ϵ ′ ) e − α S ϵ ′ /2 M 2 .
Since the right-hand side of (67) is nondecreasing in e − α f ∗ / Z α ( ν ) e^{-\alpha f_\ast}/Z_\alpha(\nu) e − α f ∗ / Z α ( ν ) and in V α ( x ) V_\alpha(x) V α ( x ) , inserting both bounds proves (69) .
We can now establish Proposition 5 . Note that this result only uses (F1) of Assumption 1, namely that f f f is measurable and bounded from below with inf y ∈ R d f ( y ) = f ( x ∗ ) \inf_{y\in\mathbb{R}^d}f(y)=f(x_\ast) inf y ∈ R d f ( y ) = f ( x ∗ ) . More generally, by Lemma 17 (i), the bound (16) holds for every measurable f f f that is bounded from below, if f ( x ∗ ) f(x_\ast) f ( x ∗ ) in the definition (17) of ζ α ( x ) \zeta_\alpha(x) ζ α ( x ) is replaced by f ‾ = inf y ∈ R d f ( y ) \underline{f}=\inf_{y\in\mathbb{R}^d}f(y) f = inf y ∈ R d f ( y ) .
Proof [Proof of Proposition 5] By (F1), f f f is measurable and bounded from below with f ‾ : = inf y ∈ R d f ( y ) = f ( x ∗ ) \underline{f}:=\inf_{y\in\mathbb{R}^d}f(y)=f(x_\ast) f := inf y ∈ R d f ( y ) = f ( x ∗ ) , so Lemmas 11 and 17 apply. Let x ∈ R d x\in\mathbb{R}^d x ∈ R d and ν : = N ( x , Σ ) \nu:=\mathcal{N}(x,\Sigma) ν := N ( x , Σ ) . Since 0 < e − α ( f ( y ) − f ( x ∗ ) ) ≤ 1 0<e^{-\alpha(f(y)-f(x_\ast))}\le1 0 < e − α ( f ( y ) − f ( x ∗ )) ≤ 1 for all y ∈ R d y\in\mathbb{R}^d y ∈ R d , we have ζ α ( x ) ∈ ( 0 , 1 ] \zeta_\alpha(x)\in(0,1] ζ α ( x ) ∈ ( 0 , 1 ] . Moreover, Z α ( x ) = Z α ( ν ) Z_\alpha(x)=Z_\alpha(\nu) Z α ( x ) = Z α ( ν ) , and hence ζ α ( x ) = e α f ( x ∗ ) Z α ( ν ) \zeta_\alpha(x)=e^{\alpha f(x_\ast)}Z_\alpha(\nu) ζ α ( x ) = e α f ( x ∗ ) Z α ( ν ) , i.e. 1 / ζ α ( x ) = e − α f ∗ / Z α ( ν ) 1/\zeta_\alpha(x)=e^{-\alpha f_\ast}/Z_\alpha(\nu) 1/ ζ α ( x ) = e − α f ∗ / Z α ( ν ) . Furthermore, (49) with h ( y ) = ∥ y − T α ( x ) ∥ 2 2 h(y)=\lVert y-T_\alpha(x)\rVert_2^2 h ( y ) = ∥ y − T α ( x ) ∥ 2 2 shows that V α ( x ) V_\alpha(x) V α ( x ) in (17) coincides with the total variance (68) of the Gibbs tilt η ν α \eta_\nu^\alpha η ν α , which is finite by Lemma 17 (i). Hence (16) is (67) .
We can now turn to sample complexity results. To illustrate the approach, we first consider only a single step in the MPPI iteration.
Corollary 18 (Sample complexity, single step). Let f f f satisfy Assumption 1, let Σ ∈ R d × d \Sigma\in\mathbb{R}^{d\times d} Σ ∈ R d × d be symmetric positive definite, and let r > 0 r>0 r > 0 . Let C s C_s C s , M 1 M_1 M 1 and M 2 M_2 M 2 be the constants of Lemma 10 for r r r and Σ \Sigma Σ , and let s f ( ϵ ) s_f(\epsilon) s f ( ϵ ) be as in (30) . For ϵ > 0 \epsilon>0 ϵ > 0 and δ ∈ ( 0 , 1 ) \delta\in(0,1) δ ∈ ( 0 , 1 ) , set
C ‾ ϵ : = C s f ( ϵ / 4 ) , S ‾ ϵ : = S ϵ / 4 , Λ ϵ : = max { 8 M 1 ϵ , 16 M 2 ϵ 2 } , (72) \overline{C}_\epsilon:=C_{s_f(\epsilon/4)},\qquad\overline{S}_\epsilon:=S_{\epsilon/4},\qquad\Lambda_\epsilon:=\max\left\{\frac{8M_1}{\epsilon},\frac{16M_2}{\epsilon^2}\right\}, \tag*{(72)} C ϵ := C s f ( ϵ /4 ) , S ϵ := S ϵ /4 , Λ ϵ := max { ϵ 8 M 1 , ϵ 2 16 M 2 } , (72) and
α min ( ϵ ) : = 2 S ‾ ϵ max { 0 , log Λ ϵ C ‾ ϵ } , N min ( α , ϵ , δ ) : = 6 e α S ‾ ϵ / 2 δ C ‾ ϵ . (73) \alpha_{\min}(\epsilon):=\frac{2}{\overline{S}_\epsilon}\max\left\{0,\log\frac{\Lambda_\epsilon}{\overline{C}_\epsilon}\right\},\qquad N_{\min}(\alpha,\epsilon,\delta):=\frac{6e^{\alpha\overline{S}_\epsilon/2}}{\delta\overline{C}_\epsilon}. \tag*{(73)} α m i n ( ϵ ) := S ϵ 2 max { 0 , log C ϵ Λ ϵ } , N m i n ( α , ϵ , δ ) := δ C ϵ 6 e α S ϵ /2 . (73) Then for all α > 0 \alpha>0 α > 0 with α ≥ α min ( ϵ ) \alpha\ge\alpha_{\min}(\epsilon) α ≥ α m i n ( ϵ ) , all N ∈ N + N\in\mathbb{N}_+ N ∈ N + with N ≥ N min ( α , ϵ , δ ) N\ge N_{\min}(\alpha,\epsilon,\delta) N ≥ N m i n ( α , ϵ , δ ) , and all x ∈ B r ( x ∗ ) x\in B_r(x_\ast) x ∈ B r ( x ∗ ) ,
P ( ∥ T ^ α , N ( x ) − x ∗ ∥ 2 < ϵ ) ≥ 1 − δ . (74) \mathbb{P}\left(\left\lVert\widehat{T}_{\alpha,N}(x)-x_\ast\right\rVert_2<\epsilon\right)\ge1-\delta. \tag*{(74)} P ( T α , N ( x ) − x ∗ 2 < ϵ ) ≥ 1 − δ . (74) In particular, if α min ( ϵ ) > 0 \alpha_{\min}(\epsilon)>0 α m i n ( ϵ ) > 0 , then for α = α min ( ϵ ) \alpha=\alpha_{\min}(\epsilon) α = α m i n ( ϵ ) the sample size condition reads
N ≥ N min ( α min ( ϵ ) , ϵ , δ ) = 6 δ C ‾ ϵ max { 1 , Λ ϵ C ‾ ϵ } . (75) N\ge N_{\min}(\alpha_{\min}(\epsilon),\epsilon,\delta)=\frac{6}{\delta\overline{C}_\epsilon}\max\left\{1,\frac{\Lambda_\epsilon}{\overline{C}_\epsilon}\right\}. \tag*{(75)} N ≥ N m i n ( α m i n ( ϵ ) , ϵ , δ ) = δ C ϵ 6 max { 1 , C ϵ Λ ϵ } . (75)
Proof. Fix α \alpha α , N N N and x x x as in the statement. Recall that α 0 ( δ ′ , S , C ) = 2 S max { 0 , log 1 δ ′ C } \alpha_0(\delta',S,C)=\frac{2}{S}\max\{0,\log\frac{1}{\delta'C}\} α 0 ( δ ′ , S , C ) = S 2 max { 0 , log δ ′ C 1 } by (32) , which is nondecreasing in 1 / δ ′ 1/\delta' 1/ δ ′ .
Bias. Lemma 14 (ii) with ρ = ϵ / 2 \rho=\epsilon/2 ρ = ϵ /2 applies for α ≥ α 0 ( ϵ 8 M 1 , S ˉ ϵ , C ˉ ϵ ) = 2 S ˉ ϵ max { 0 , log 8 M 1 ϵ C ˉ ϵ } \alpha\geq\alpha_0\left(\frac{\epsilon}{8M_1},\bar{S}_{\epsilon},\bar{C}_{\epsilon}\right)=\frac{2}{\bar{S}_{\epsilon}}\max\left\{0,\log\frac{8M_1}{\epsilon\bar{C}_{\epsilon}}\right\} α ≥ α 0 ( 8 M 1 ϵ , S ˉ ϵ , C ˉ ϵ ) = S ˉ ϵ 2 max { 0 , log ϵ C ˉ ϵ 8 M 1 } , and this holds since 8 M 1 / ϵ ≤ Λ ϵ 8M_1/\epsilon\leq\Lambda_{\epsilon} 8 M 1 / ϵ ≤ Λ ϵ and α ≥ α min ( ϵ ) \alpha\geq\alpha_{\min}(\epsilon) α ≥ α m i n ( ϵ ) . Hence
∥ T α ( x ) − x ∗ ∥ 2 ≤ 3 4 ⋅ ϵ 2 = 3 ϵ 8 . (76) \left\|\mathcal{T}_{\alpha}(x)-x_*\right\|_2\leq\frac{3}{4}\cdot\frac{\epsilon}{2}=\frac{3\epsilon}{8}. \tag*{(76)} ∥ T α ( x ) − x ∗ ∥ 2 ≤ 4 3 ⋅ 2 ϵ = 8 3 ϵ . (76)
Variance of the Gibbs tilt. Similarly, 16 M 2 / ϵ 2 ≤ Λ ϵ 16M_2/\epsilon^2\leq\Lambda_{\epsilon} 16 M 2 / ϵ 2 ≤ Λ ϵ gives α ≥ α 0 ( ϵ 2 16 M 2 , S ˉ ϵ , C ˉ ϵ ) \alpha\geq\alpha_0\left(\frac{\epsilon^2}{16M_2},\bar{S}_{\epsilon},\bar{C}_{\epsilon}\right) α ≥ α 0 ( 16 M 2 ϵ 2 , S ˉ ϵ , C ˉ ϵ ) , so (33) yields
e − α S ˉ ϵ / 2 C ˉ ϵ M 2 ≤ ϵ 2 16 . \frac{e^{-\alpha\bar{S}_{\epsilon}/2}}{\bar{C}_{\epsilon}}M_2\leq\frac{\epsilon^2}{16}. C ˉ ϵ e − α S ˉ ϵ /2 M 2 ≤ 16 ϵ 2 .
Sampling error. Lemma 17 (ii), applied with ϵ / 4 \epsilon/4 ϵ /4 in place of both ϵ \epsilon ϵ and ϵ ′ \epsilon' ϵ ′ and with t = ϵ / 2 t=\epsilon/2 t = ϵ /2 , together with the previous estimate gives
P ( ∥ T ^ α , N ( x ) − T α ( x ) ∥ 2 > ϵ 2 ) ≤ 4 N e α S ˉ ϵ / 2 C ˉ ϵ ( 1 + 4 ϵ 2 ( ϵ 2 16 + ϵ 2 16 ) ) = 6 N e α S ˉ ϵ / 2 C ˉ ϵ ≤ δ , \mathbb{P}\left(\left\|\widehat{\mathcal{T}}_{\alpha,N}(x)-\mathcal{T}_{\alpha}(x)\right\|_2>\frac{\epsilon}{2}\right)\leq\frac{4}{N}\frac{e^{\alpha\bar{S}_{\epsilon}/2}}{\bar{C}_{\epsilon}}\left(1+\frac{4}{\epsilon^2}\left(\frac{\epsilon^2}{16}+\frac{\epsilon^2}{16}\right)\right)=\frac{6}{N}\frac{e^{\alpha\bar{S}_{\epsilon}/2}}{\bar{C}_{\epsilon}}\leq\delta, P ( T α , N ( x ) − T α ( x ) 2 > 2 ϵ ) ≤ N 4 C ˉ ϵ e α S ˉ ϵ /2 ( 1 + ϵ 2 4 ( 16 ϵ 2 + 16 ϵ 2 ) ) = N 6 C ˉ ϵ e α S ˉ ϵ /2 ≤ δ ,
where the last step uses N ≥ N min ( α , ϵ , δ ) N\geq N_{\min}(\alpha,\epsilon,\delta) N ≥ N m i n ( α , ϵ , δ ) .
Conclusion. On the complement of the event in the last display, (76) and the triangle inequality give ∥ T ^ α , N ( x ) − x ∗ ∥ 2 ≤ ϵ 2 + 3 ϵ 8 < ϵ \left\|\widehat{\mathcal{T}}_{\alpha,N}(x)-x_*\right\|_2\leq\frac{\epsilon}{2}+\frac{3\epsilon}{8}<\epsilon T α , N ( x ) − x ∗ 2 ≤ 2 ϵ + 8 3 ϵ < ϵ , which proves (74) . Finally, if α = α min ( ϵ ) > 0 \alpha=\alpha_{\min}(\epsilon)>0 α = α m i n ( ϵ ) > 0 , then e α min ( ϵ ) S ˉ ϵ / 2 = Λ ϵ / C ˉ ϵ = max { 1 , Λ ϵ / C ˉ ϵ } e^{\alpha_{\min}(\epsilon)\bar{S}_{\epsilon}/2}=\Lambda_{\epsilon}/\bar{C}_{\epsilon}=\max\{1,\Lambda_{\epsilon}/\bar{C}_{\epsilon}\} e α m i n ( ϵ ) S ˉ ϵ /2 = Λ ϵ / C ˉ ϵ = max { 1 , Λ ϵ / C ˉ ϵ } , which gives (75) .
Let us now prove Theorem 6 , which considers a finite trajectory of the MPPI iteration.
Proof [Proof of Theorem 6 ]. Throughout, for R > 0 R>0 R > 0 we write C s ( R ) C_s(R) C s ( R ) and M 2 ( R ) M_2(R) M 2 ( R ) for the constants of Lemma 10 with radius R R R . Let r r r , ρ \rho ρ , β \beta β , γ \gamma γ and ρ ′ \rho' ρ ′ be as in the statement, and set
θ : = 1 − ρ 2 , R 0 : = σ min ( Σ ) , \theta:=\frac{1-\rho}{2},\qquad R_0:=\sigma_{\min}(\Sigma), θ := 2 1 − ρ , R 0 := σ m i n ( Σ ) ,
so that ρ ′ = ρ + θ \rho'=\rho+\theta ρ ′ = ρ + θ and θ + ρ ′ = 1 \theta+\rho'=1 θ + ρ ′ = 1 . By (F2), there is s γ > 0 s_{\gamma}>0 s γ > 0 , depending only on f f f and γ \gamma γ , such that f ≤ f ∗ + γ f\leq f_*+\gamma f ≤ f ∗ + γ on B s γ ( x ∗ ) B_{s_{\gamma}}(x_*) B s γ ( x ∗ ) . With α 0 ( ⋅ , ⋅ , ⋅ ) \alpha_0(\cdot,\cdot,\cdot) α 0 ( ⋅ , ⋅ , ⋅ ) from (32) , α 0 ( f , Σ , r , ρ ) \alpha_0(f,\Sigma,r,\rho) α 0 ( f , Σ , r , ρ ) from (65) and α b \alpha_b α b from (64) , define
α 0 ′ ( f , Σ , r , ρ ) : = max { α 0 ( f , Σ , r , ρ ) , α 0 ( θ 2 R 0 2 4 M 2 ( r ) , S θ R 0 / 2 , C s f ( θ R 0 / 2 ) ( r ) ) } , \alpha'_0(f,\Sigma,r,\rho):=\max\left\{\alpha_0(f,\Sigma,r,\rho),\alpha_0\left(\frac{\theta^2R_0^2}{4M_2(r)},S_{\theta R_0/2},C_{s f}(\theta R_0/2)(r)\right)\right\}, α 0 ′ ( f , Σ , r , ρ ) := max { α 0 ( f , Σ , r , ρ ) , α 0 ( 4 M 2 ( r ) θ 2 R 0 2 , S θ R 0 /2 , C s f ( θ R 0 /2 ) ( r ) ) } ,
α b ′ ( f , Σ , β ) : = max { α b ( f , Σ , β ) , α 0 ( β 2 4 M 2 ( R 0 ) , S β / 2 , C s f ( β / 2 ) ( R 0 ) ) } , \alpha'_b(f,\Sigma,\beta):=\max\left\{\alpha_b(f,\Sigma,\beta),\alpha_0\left(\frac{\beta^2}{4M_2(R_0)},S_{\beta/2},C_{s f(\beta/2)}(R_0)\right)\right\}, α b ′ ( f , Σ , β ) := max { α b ( f , Σ , β ) , α 0 ( 4 M 2 ( R 0 ) β 2 , S β /2 , C s f ( β /2 ) ( R 0 ) ) } ,
C ( f , Σ , r , γ ) : = 6 C s γ ( r ) . (77) C(f,\Sigma,r,\gamma):=\frac{6}{C_{s_\gamma}(r)}. \tag*{(77)} C ( f , Σ , r , γ ) := C s γ ( r ) 6 . (77)
Then α 0 ′ \alpha'_0 α 0 ′ depends only on f f f , Σ \Sigma Σ , r r r and ρ \rho ρ , α b ′ \alpha'_b α b ′ only on f f f , Σ \Sigma Σ and β \beta β , and C C C only on f f f , Σ \Sigma Σ , r r r and γ \gamma γ . Fix α \alpha α , K K K , δ \delta δ , N N N and x ^ 0 \hat{x}_0 x ^ 0 as in the statement.
Step 0: Realisation of the iteration. Let ( G i ( k ) ) k ∈ N 0 , i = 1 , … , N (G_i^{(k)})_{k\in\mathbb{N}_0,\ i=1,\ldots,N} ( G i ( k ) ) k ∈ N 0 , i = 1 , … , N be independent N ( 0 , I d ) \mathcal{N}(0,I_d) N ( 0 , I d ) -distributed random vectors, let Σ 1 / 2 \Sigma^{1/2} Σ 1/2 be the symmetric positive definite square root of Σ \Sigma Σ , and write G ( k ) : = ( G 1 ( k ) , … , G N ( k ) ) G^{(k)}:=(G_1^{(k)},\ldots,G_N^{(k)}) G ( k ) := ( G 1 ( k ) , … , G N ( k ) ) . For x ∈ R d x\in\mathbb{R}^d x ∈ R d and g ∈ ( R d ) N g\in(\mathbb{R}^d)^N g ∈ ( R d ) N , let T ^ ( x , g ) \widehat{T}(x,g) T ( x , g ) be the right-hand side of (15) with X i = x + Σ 1 / 2 g i X_i=x+\Sigma^{1/2}g_i X i = x + Σ 1/2 g i . Since f f f is measurable and the denominator is strictly positive, T ^ \widehat{T} T is Borel measurable, and the MPPI iterates are given by x ^ k + 1 = T ^ ( x ^ k , G ( k ) ) \hat{x}_{k+1}=\widehat{T}(\hat{x}_k,G^{(k)}) x ^ k + 1 = T ( x ^ k , G ( k ) ) . By induction, x ^ k \hat{x}_k x ^ k is a Borel measurable function of G ( 0 ) , … , G ( k − 1 ) G^{(0)},\ldots,G^{(k-1)} G ( 0 ) , … , G ( k − 1 ) , and hence independent of G ( k ) G^{(k)} G ( k ) . We write e k : = ∥ x ^ k − x ∗ ∥ 2 e_k:=\|\hat{x}_k-x_*\|_2 e k := ∥ x ^ k − x ∗ ∥ 2 .
Step 1: Normaliser and variance. Let x ∈ B r ( x ∗ ) x \in B_r(x_\ast) x ∈ B r ( x ∗ ) , e : = ∥ x − x ∗ ∥ 2 e := \lVert x-x_\ast\rVert_2 e := ∥ x − x ∗ ∥ 2 and ν : = N ( x , Σ ) \nu:= \mathcal{N}(x,\Sigma) ν := N ( x , Σ ) . By (F1), f = f ∗ f=f_\ast f = f ∗ . Restricting the integral to B s γ ( x ∗ ) B_{s_\gamma}(x_\ast) B s γ ( x ∗ ) and using Lemma 10 with radius r r r gives Z α ( x ) ≥ e − α ( f ∗ + γ ) C s γ ( r ) Z_\alpha(x) \ge e^{-\alpha(f_\ast+\gamma)}C_{s_\gamma}(r) Z α ( x ) ≥ e − α ( f ∗ + γ ) C s γ ( r ) , i.e.
1 ζ α ( x ) ≤ e α γ C s γ ( r ) . (78) \frac{1}{\zeta_\alpha(x)} \le\frac{e^{\alpha\gamma}}{C_{s_\gamma}(r)}. \tag*{(78)} ζ α ( x ) 1 ≤ C s γ ( r ) e α γ . (78)
Next, we show
V α ( x ) ≤ 1 2 ( θ e + β ) 2 . (79) V_\alpha(x) \le\frac{1}{2}(\theta e+\beta)^2. \tag*{(79)} V α ( x ) ≤ 2 1 ( θ e + β ) 2 . (79)
By Proposition 5 and its proof, V α ( x ) V_\alpha(x) V α ( x ) is the total variance of η ν α \eta^\alpha_\nu η ν α , and since the mean minimises the expected squared distance, V α ( x ) ≤ E Y ∼ η ν α [ ∥ Y − x ∗ ∥ 2 2 ] V_\alpha(x) \le\mathbb{E}_{Y\sim\eta^\alpha_\nu}[\lVert Y-x_\ast\rVert_2^2] V α ( x ) ≤ E Y ∼ η ν α [∥ Y − x ∗ ∥ 2 2 ] . We bound the right-hand side with Lemma 9 (ii) for p = 2 p=2 p = 2 in two regimes. If e ≤ R 0 e\le R_0 e ≤ R 0 , then x ∈ B R 0 ( x ∗ ) x\in B_{R_0}(x_\ast) x ∈ B R 0 ( x ∗ ) , so ν \nu ν satisfies (28) and (35) for p = 2 p=2 p = 2 with C s ( R 0 ) C_s(R_0) C s ( R 0 ) and M 2 ( R 0 ) M_2(R_0) M 2 ( R 0 ) . With ϵ = β / 2 \epsilon=\beta/2 ϵ = β /2 , (36) , α ≥ α b ′ ( f , Σ , β ) \alpha\ge\alpha'_b(f,\Sigma,\beta) α ≥ α b ′ ( f , Σ , β ) and (33) give
E Y ∼ η ν α [ ∥ Y − x ∗ ∥ 2 2 ] ≤ β 2 4 + β 2 4 ≤ 1 2 ( θ e + β ) 2 . \mathbb{E}_{Y\sim\eta^\alpha_\nu}[\lVert Y-x_\ast\rVert_2^2]
\le\frac{\beta^2}{4}+\frac{\beta^2}{4}\le\frac{1}{2}(\theta e+\beta)^2. E Y ∼ η ν α [∥ Y − x ∗ ∥ 2 2 ] ≤ 4 β 2 + 4 β 2 ≤ 2 1 ( θ e + β ) 2 .
If R 0 < e ≤ r R_0<e\le r R 0 < e ≤ r , we use the constants C s ( r ) C_s(r) C s ( r ) and M 2 ( r ) M_2(r) M 2 ( r ) . With ϵ = θ R 0 / 2 \epsilon=\theta R_0/2 ϵ = θ R 0 /2 and α ≥ α 0 ′ ( f , Σ , r , ρ ) \alpha\ge\alpha'_0(f,\Sigma,r,\rho) α ≥ α 0 ′ ( f , Σ , r , ρ ) , we obtain
E Y ∼ η ν α [ ∥ Y − x ∗ ∥ 2 2 ] ≤ θ 2 R 0 2 4 + θ 2 R 0 2 4 ≤ 1 2 θ 2 e 2 ≤ 1 2 ( θ e + β ) 2 . \mathbb{E}_{Y\sim\eta^\alpha_\nu}[\lVert Y-x_\ast\rVert_2^2]
\le\frac{\theta^2R_0^2}{4}+\frac{\theta^2R_0^2}{4}
\le\frac{1}{2}\theta^2e^2\le\frac{1}{2}(\theta e+\beta)^2. E Y ∼ η ν α [∥ Y − x ∗ ∥ 2 2 ] ≤ 4 θ 2 R 0 2 + 4 θ 2 R 0 2 ≤ 2 1 θ 2 e 2 ≤ 2 1 ( θ e + β ) 2 .
Step 2: One sampled step. For x ∈ R d x\in\mathbb{R}^d x ∈ R d , let
ψ ( x ) : = P ( ∥ T ^ ( x , G ( 0 ) ) − T α ( x ) ∥ 2 > θ ∥ x − x ∗ ∥ 2 + β ) . \psi(x):=\mathbb{P}\left(\lVert\widehat{T}(x,G^{(0)})-T_\alpha(x)\rVert_2>\theta\lVert x-x_\ast\rVert_2+\beta\right). ψ ( x ) := P ( ∥ T ( x , G ( 0 ) ) − T α ( x ) ∥ 2 > θ ∥ x − x ∗ ∥ 2 + β ) .
Let x ∈ B r ( x ∗ ) x\in B_r(x_\ast) x ∈ B r ( x ∗ ) . We apply Proposition 5 with t = θ ∥ x − x ∗ ∥ 2 + β t=\theta\lVert x-x_\ast\rVert_2+\beta t = θ ∥ x − x ∗ ∥ 2 + β and use (78) , (79) and (18) . This yields
ψ ( x ) ≤ 4 N 1 ζ α ( x ) ( 1 + V α ( x ) ( θ ∥ x − x ∗ ∥ 2 + β ) 2 ) ≤ 4 N e α γ C s γ ( r ) ⋅ 3 2 = C ( f , Σ , r , γ ) e γ α N ≤ δ K . (80) \psi(x)\le\frac{4}{N}\frac{1}{\zeta_\alpha(x)}\left(1+\frac{V_\alpha(x)}{(\theta\lVert x-x_\ast\rVert_2+\beta)^2}\right)
\le\frac{4}{N}\frac{e^{\alpha\gamma}}{C_{s_\gamma}(r)}\cdot\frac{3}{2}
=\frac{C(f,\Sigma,r,\gamma)e^{\gamma\alpha}}{N}\le\frac{\delta}{K}. \tag*{(80)} ψ ( x ) ≤ N 4 ζ α ( x ) 1 ( 1 + ( θ ∥ x − x ∗ ∥ 2 + β ) 2 V α ( x ) ) ≤ N 4 C s γ ( r ) e α γ ⋅ 2 3 = N C ( f , Σ , r , γ ) e γ α ≤ K δ . (80)
Step 3: Deterministic progress on good events. For k ∈ N 0 k\in\mathbb{N}_0 k ∈ N 0 , let A k : = { ∥ x ^ k + 1 − T α ( x ^ k ) ∥ 2 ≤ θ e k + β } A_k:=\{\lVert\widehat{x}_{k+1}-T_\alpha(\widehat{x}_k)\rVert_2\le\theta e_k+\beta\} A k := {∥ x k + 1 − T α ( x k ) ∥ 2 ≤ θ e k + β } . Suppose that x ^ k ∈ B r ( x ∗ ) \widehat{x}_k\in B_r(x_\ast) x k ∈ B r ( x ∗ ) and that A k A_k A k occurs. Since α ≥ α 0 ′ ≥ α 0 ( f , Σ , r , ρ ) \alpha\ge\alpha'_0\ge\alpha_0(f,\Sigma,r,\rho) α ≥ α 0 ′ ≥ α 0 ( f , Σ , r , ρ ) and α ≥ α b ′ ≥ α b ( f , Σ , β ) \alpha\ge\alpha'_b\ge\alpha_b(f,\Sigma,\beta) α ≥ α b ′ ≥ α b ( f , Σ , β ) , Theorem 2 (i) gives ∥ T α ( x ^ k ) − x ∗ ∥ 2 ≤ β + ρ e k \lVert T_\alpha(\widehat{x}_k)-x_\ast\rVert_2\le\beta+\rho e_k ∥ T α ( x k ) − x ∗ ∥ 2 ≤ β + ρ e k , and hence
e k + 1 ≤ ∥ T α ( x ^ k ) − x ∗ ∥ 2 + θ e k + β ≤ 2 β + ρ ′ e k . (81) e_{k+1}\le\lVert T_\alpha(\widehat{x}_k)-x_\ast\rVert_2+\theta e_k+\beta\le2\beta+\rho'e_k. \tag*{(81)} e k + 1 ≤ ∥ T α ( x k ) − x ∗ ∥ 2 + θ e k + β ≤ 2 β + ρ ′ e k . (81)
In particular, e k + 1 ≤ 2 β + ρ ′ r ≤ θ r + ρ ′ r = r e_{k+1}\le2\beta+\rho'r\le\theta r+\rho'r=r e k + 1 ≤ 2 β + ρ ′ r ≤ θ r + ρ ′ r = r , where we used 2 β ≤ θ r 2\beta\le\theta r 2 β ≤ θ r , i.e. 4 β ≤ ( 1 − ρ ) r 4\beta\le(1-\rho)r 4 β ≤ ( 1 − ρ ) r . By induction, on ⋂ j = 0 k − 1 A j \bigcap_{j=0}^{k-1}A_j ⋂ j = 0 k − 1 A j we have x ^ j ∈ B r ( x ∗ ) \widehat{x}_j\in B_r(x_\ast) x j ∈ B r ( x ∗ ) for j = 0 , … , k j=0,\ldots,k j = 0 , … , k , and (81) yields
e k ≤ ρ ′ k e 0 + 2 β ∑ j = 0 k − 1 ρ ′ j ≤ ρ ′ k e 0 + 2 β 1 − ρ ′ = ρ ′ k e 0 + 4 β 1 − ρ . e_k\le\rho'^k e_0+2\beta\sum_{j=0}^{k-1}\rho'^j\le\rho'^k e_0+\frac{2\beta}{1-\rho'}=\rho'^k e_0+\frac{4\beta}{1-\rho}. e k ≤ ρ ′ k e 0 + 2 β j = 0 ∑ k − 1 ρ ′ j ≤ ρ ′ k e 0 + 1 − ρ ′ 2 β = ρ ′ k e 0 + 1 − ρ 4 β .
Hence the event in (19) contains ⋂ k = 0 K − 1 A k \bigcap_{k=0}^{K-1}A_k ⋂ k = 0 K − 1 A k .
Step 4: Probability of the good events. Since x ^ k \widehat{x}_k x k is independent of G ( k ) G^{(k)} G ( k ) , T ^ \widehat{T} T is Borel measurable and T α T_\alpha T α is continuous by Lemma 11 , Fubini’s theorem and (80) give P ( A k c ∩ { x ^ k ∈ B r ( x ∗ ) } ) = E [ 1 B r ( x ∗ ) ( x ^ k ) ψ ( x ^ k ) ] ≤ δ / K \mathbb{P}(A_k^c\cap\{\widehat{x}_k\in B_r(x_\ast)\})=\mathbb{E}[\mathbf{1}_{B_r(x_\ast)}(\widehat{x}_k)\psi(\widehat{x}_k)]\le\delta/K P ( A k c ∩ { x k ∈ B r ( x ∗ )}) = E [ 1 B r ( x ∗ ) ( x k ) ψ ( x k )] ≤ δ / K . Decomposing according to the first failure time and using Step 3,
P ( ⋃ k = 0 K − 1 A k c ) = ∑ k = 0 K − 1 P ( A k c ∩ ⋂ j = 0 k − 1 A j ) ≤ ∑ k = 0 K − 1 P ( A k c ∩ { x ^ k ∈ B r ( x ∗ ) } ) ≤ δ . \mathbb{P}\left(\bigcup_{k=0}^{K-1}A_k^c\right)
=\sum_{k=0}^{K-1}\mathbb{P}\left(A_k^c\cap\bigcap_{j=0}^{k-1}A_j\right)
\le\sum_{k=0}^{K-1}\mathbb{P}\left(A_k^c\cap\{\widehat{x}_k\in B_r(x_\ast)\}\right)
\le\delta. P ( k = 0 ⋃ K − 1 A k c ) = k = 0 ∑ K − 1 P ( A k c ∩ j = 0 ⋂ k − 1 A j ) ≤ k = 0 ∑ K − 1 P ( A k c ∩ { x k ∈ B r ( x ∗ )} ) ≤ δ .
which proves (19) .
Step 5: Accuracy ϵ \epsilon ϵ . Let ϵ ∈ ( 0 , 2 r ] \epsilon\in(0,2r] ϵ ∈ ( 0 , 2 r ] , and apply (19) with ρ = 1 2 \rho=\frac{1}{2} ρ = 2 1 and β = ϵ 16 \beta=\frac{\epsilon}{16} β = 16 ϵ . Then 4 β = ϵ 4 ≤ r 2 = ( 1 − ρ ) r 4\beta=\frac{\epsilon}{4}\le\frac{r}{2}=(1-\rho)r 4 β = 4 ϵ ≤ 2 r = ( 1 − ρ ) r , ρ ′ = 3 4 \rho'=\frac{3}{4} ρ ′ = 4 3 and 4 β 1 − ρ = ϵ 2 \frac{4\beta}{1-\rho}=\frac{\epsilon}{2} 1 − ρ 4 β = 2 ϵ . For k ≥ K ϵ k\ge K_\epsilon k ≥ K ϵ we have ( 3 4 ) k ≤ ϵ 2 r \left(\frac{3}{4}\right)^k\le\frac{\epsilon}{2r} ( 4 3 ) k ≤ 2 r ϵ , and hence ρ ′ k ∥ x ^ 0 − x ∗ ∥ 2 + ϵ 2 ≤ ϵ \rho'^k\lVert\hat{x}_0-x_*\rVert_2+\frac{\epsilon}{2}\le\epsilon ρ ′ k ∥ x ^ 0 − x ∗ ∥ 2 + 2 ϵ ≤ ϵ on the event in (19) . This proves (20) .
Analysis under a Simplified Assumption
In this section, we prove the results stated under Assumption 2. Under this assumption, the covariance of the Gibbs tilt inside the well can be controlled by a log-concavity argument, and all thresholds become explicit. Throughout, Σ ∈ R d × d \Sigma\in\mathbb{R}^{d\times d} Σ ∈ R d × d is symmetric positive definite, and κ Σ = σ max ( Σ ) 2 / σ min ( Σ ) 2 \kappa_\Sigma=\sigma_{\max}(\Sigma)^2/\sigma_{\min}(\Sigma)^2 κ Σ = σ m a x ( Σ ) 2 / σ m i n ( Σ ) 2 as in the notation paragraph.
Assumption 2 implies Assumption 1
For s > 0 s>0 s > 0 , we use the abbreviation
D s : = Γ ( d 2 + 1 ) ( 2 σ max ( Σ ) 2 s 2 ) d / 2 , (82) D_s:=\Gamma\left(\frac{d}{2}+1\right)\left(\frac{2\sigma_{\max}(\Sigma)^2}{s^2}\right)^{d/2}, \tag*{(82)} D s := Γ ( 2 d + 1 ) ( s 2 2 σ m a x ( Σ ) 2 ) d /2 , (82)
so that the constants of Lemma 10 and (63) read
1 C s = D s exp ( ( r + s ) 2 2 σ min ( Σ ) 2 ) , 1 C s ∞ = D s exp ( s 2 2 σ min ( Σ ) 2 ) . (83) \frac{1}{C_s}=D_s\exp\left(\frac{(r+s)^2}{2\sigma_{\min}(\Sigma)^2}\right),\qquad
\frac{1}{C_s^\infty}=D_s\exp\left(\frac{s^2}{2\sigma_{\min}(\Sigma)^2}\right). \tag*{(83)} C s 1 = D s exp ( 2 σ m i n ( Σ ) 2 ( r + s ) 2 ) , C s ∞ 1 = D s exp ( 2 σ m i n ( Σ ) 2 s 2 ) . (83)
Lemma 19 (Consequences of the simplified assumption). Let f f f satisfy (SF1) and (SF2) of Assumption 2. Then the following hold.
(i) Quadratic growth in the well. f ( y ) − f ∗ ≥ m 2 ∥ y − x ∗ ∥ 2 2 f(y)-f_*\ge\frac{m}{2}\lVert y-x_*\rVert_2^2 f ( y ) − f ∗ ≥ 2 m ∥ y − x ∗ ∥ 2 2 for all y ∈ B R ( x ∗ ) y\in B_R(x_*) y ∈ B R ( x ∗ ) .
(ii) Radial monotonicity. The function g ( y ) : = f ( y ) − f ∗ − m 2 ∥ y − x ∗ ∥ 2 2 g(y):=f(y)-f_*-\frac{m}{2}\lVert y-x_*\rVert_2^2 g ( y ) := f ( y ) − f ∗ − 2 m ∥ y − x ∗ ∥ 2 2 is convex and nonnegative on B R ( x ∗ ) B_R(x_*) B R ( x ∗ ) with g ( x ∗ ) = 0 g(x_*)=0 g ( x ∗ ) = 0 , and for every unit vector u u u , the map t ↦ g ( x ∗ + t u ) t\mapsto g(x_*+tu) t ↦ g ( x ∗ + t u ) is nondecreasing on [ 0 , R ] [0,R] [ 0 , R ] .
(iii) Assumption 1. f f f satisfies Assumption 1, x ∗ x_* x ∗ is its unique global minimiser, and (5) holds with
S ϵ : = min { m 2 ϵ 2 , Δ } for all ϵ > 0. (84) S_\epsilon:=\min\left\{\frac{m}{2}\epsilon^2,\Delta\right\}\qquad\text{for all }\epsilon>0. \tag*{(84)} S ϵ := min { 2 m ϵ 2 , Δ } for all ϵ > 0. (84)
(iv) Small ball below half the gap. There is s ∈ ( 0 , R ] s\in(0,R] s ∈ ( 0 , R ] with sup y ∈ B s ( x ∗ ) ( f ( y ) − f ∗ ) ≤ Δ 2 \sup_{y\in B_s(x_*)}(f(y)-f_*)\le\frac{\Delta}{2} sup y ∈ B s ( x ∗ ) ( f ( y ) − f ∗ ) ≤ 2 Δ . If f f f also satisfies (SF3), then s : = min { R , Δ / M } s:=\min\{R,\sqrt{\Delta/M}\} s := min { R , Δ/ M } is admissible.
Proof Without loss of generality, x ∗ = 0 x_*=0 x ∗ = 0 and f ∗ = 0 f_*=0 f ∗ = 0 .
Ad (i). Let y ∈ B R ( 0 ) y\in B_R(0) y ∈ B R ( 0 ) and u ∈ ( 0 , 1 ) u\in(0,1) u ∈ ( 0 , 1 ) . The point u y uy u y lies in B R ( 0 ) B_R(0) B R ( 0 ) , and convexity of g g g from (SF1) on the segment from 0 to y y y gives g ( u y ) ≤ ( 1 − u ) g ( 0 ) + u g ( y ) g(uy)\le(1-u)g(0)+ug(y) g ( u y ) ≤ ( 1 − u ) g ( 0 ) + ug ( y ) , i.e.
f ( u y ) − m 2 u 2 ∥ y ∥ 2 2 ≤ u ( f ( y ) − m 2 ∥ y ∥ 2 2 ) . f(uy)-\frac{m}{2}u^2\lVert y\rVert_2^2\le u\left(f(y)-\frac{m}{2}\lVert y\rVert_2^2\right). f ( u y ) − 2 m u 2 ∥ y ∥ 2 2 ≤ u ( f ( y ) − 2 m ∥ y ∥ 2 2 ) .
Since f ( u y ) ≥ 0 f(uy)\ge0 f ( u y ) ≥ 0 , this yields 0 ≤ u f ( y ) − m 2 u ( 1 − u ) ∥ y ∥ 2 2 0\le uf(y)-\frac{m}{2}u(1-u)\lVert y\rVert_2^2 0 ≤ u f ( y ) − 2 m u ( 1 − u ) ∥ y ∥ 2 2 . Dividing by u u u and letting u ↓ 0 u\downarrow0 u ↓ 0 gives f ( y ) ≥ m 2 ∥ y ∥ 2 2 f(y)\ge\frac{m}{2}\lVert y\rVert_2^2 f ( y ) ≥ 2 m ∥ y ∥ 2 2 .
Ad (ii). Convexity is (SF1), nonnegativity is (i), and g ( 0 ) = f ( 0 ) = 0 g(0)=f(0)=0 g ( 0 ) = f ( 0 ) = 0 . Let u u u be a unit vector and 0 ≤ t 1 < t 2 ≤ R 0\le t_1<t_2\le R 0 ≤ t 1 < t 2 ≤ R . Writing t 1 = ( 1 − λ ) ⋅ 0 + λ t 2 t_1=(1-\lambda)\cdot0+\lambda t_2 t 1 = ( 1 − λ ) ⋅ 0 + λ t 2 with λ = t 1 / t 2 ∈ [ 0 , 1 ) \lambda=t_1/t_2\in[0,1) λ = t 1 / t 2 ∈ [ 0 , 1 ) , convexity gives g ( t 1 u ) ≤ ( 1 − λ ) g ( 0 ) + λ g ( t 2 u ) = λ g ( t 2 u ) ≤ g ( t 2 u ) g(t_1u)\le(1-\lambda)g(0)+\lambda g(t_2u)=\lambda g(t_2u)\le g(t_2u) g ( t 1 u ) ≤ ( 1 − λ ) g ( 0 ) + λ g ( t 2 u ) = λ g ( t 2 u ) ≤ g ( t 2 u ) , where we used g ( 0 ) = 0 g(0)=0 g ( 0 ) = 0 and g ( t 2 u ) ≥ 0 g(t_2u)\ge0 g ( t 2 u ) ≥ 0 .
Ad (iii). f f f is continuous, hence Borel measurable and upper semicontinuous at 0 0 0 , and f ≥ 0 = f ( 0 ) f\ge0=f(0) f ≥ 0 = f ( 0 ) by assumption, so (F1) and (F2) hold. Let ϵ > 0 \epsilon>0 ϵ > 0 and ∥ y ∥ 2 ≥ ϵ \lVert y\rVert_2\ge\epsilon ∥ y ∥ 2 ≥ ϵ . If y ∈ B R ( 0 ) y\in B_R(0) y ∈ B R ( 0 ) , then (i) gives f ( y ) ≥ m 2 ∥ y ∥ 2 2 ≥ m 2 ϵ 2 f(y)\ge\frac{m}{2}\lVert y\rVert_2^2\ge\frac{m}{2}\epsilon^2 f ( y ) ≥ 2 m ∥ y ∥ 2 2 ≥ 2 m ϵ 2 ; otherwise, (SF2) gives f ( y ) ≥ Δ f(y)\ge\Delta f ( y ) ≥ Δ . Hence f ( y ) ≥ S ϵ > 0 f(y)\ge S_\epsilon>0 f ( y ) ≥ S ϵ > 0 , which is (F3) with (84), and uniqueness of the global minimiser follows.
Ad (iv). Since f f f is continuous with f ( 0 ) = 0 f(0)=0 f ( 0 ) = 0 , there is s ∈ ( 0 , R ] s\in(0,R] s ∈ ( 0 , R ] with f ≤ Δ 2 f\le\frac{\Delta}{2} f ≤ 2 Δ on B s ( 0 ) B_s(0) B s ( 0 ) . Under (SF3), f ( y ) ≤ M 2 ∥ y ∥ 2 2 ≤ M 2 s 2 ≤ Δ 2 f(y)\le\frac{M}{2}\lVert y\rVert_2^2\le\frac{M}{2}s^2\le\frac{\Delta}{2} f ( y ) ≤ 2 M ∥ y ∥ 2 2 ≤ 2 M s 2 ≤ 2 Δ for ∥ y ∥ 2 ≤ s = min { R , Δ / M } \lVert y\rVert_2\le s=\min\{R,\sqrt{\Delta/M}\} ∥ y ∥ 2 ≤ s = min { R , Δ/ M } .
In the remainder of this subsection, s ∈ ( 0 , R ] s\in(0,R] s ∈ ( 0 , R ] is fixed as in Lemma 19(iv). For α > 0 \alpha>0 α > 0 , we abbreviate
τ α : = 1 1 + α m σ max ( Σ ) 2 . (85) \tau_\alpha:=\frac{1}{1+\alpha m\sigma_{\max}(\Sigma)^2}. \tag*{(85)} τ α := 1 + α m σ m a x ( Σ ) 2 1 . (85)
and for x ∈ R d x\in\mathbb{R}^d x ∈ R d , we write P x : = η N ( x , Σ ) α P_x:=\eta^\alpha_{\mathcal{N}(x,\Sigma)} P x := η N ( x , Σ ) α for the Gibbs tilt, so that T α ( x ) T_\alpha(x) T α ( x ) is the mean of P x P_x P x .
Mean-field Analysis under the Simplified Assumption
The simplified Assumption 2 allows us to make the mean-field analysis more concrete and quantitative. We start with investigating how the various components of the Gibbs tilt behave.
Lemma 20 (The Gibbs tilt inside and outside the well). Let f f f satisfy (SF1) and (SF2) of Assumption 2, let α > 0 \alpha>0 α > 0 and x ∈ R d x\in\mathbb{R}^d x ∈ R d , and let e : = ∥ x − x ∗ ∥ 2 e:=\lVert x-x_*\rVert_2 e := ∥ x − x ∗ ∥ 2 .
(i) Outside the well. On R d ∖ B R ( x ∗ ) \mathbb{R}^d\setminus B_R(x_*) R d ∖ B R ( x ∗ ) , the density of P x P_x P x with respect to N ( x , Σ ) \mathcal{N}(x,\Sigma) N ( x , Σ ) is bounded by
q α ( e ) : = D s exp ( ( e + s ) 2 2 σ min ( Σ ) 2 − α Δ 2 ) . (86) q_\alpha(e):=D_s\exp\left(\frac{(e+s)^2}{2\sigma_{\min}(\Sigma)^2}-\frac{\alpha\Delta}{2}\right). \tag*{(86)} q α ( e ) := D s exp ( 2 σ m i n ( Σ ) 2 ( e + s ) 2 − 2 α Δ ) . (86) (ii) Inside the well. P x ( B R ( x ∗ ) ) > 0 P_x(B_R(x_*))>0 P x ( B R ( x ∗ )) > 0 , and the conditional law P x i n : = P x ( ⋅ ∣ B R ( x ∗ ) ) P_x^{\mathrm{in}}:=P_x(\,\cdot\mid B_R(x_*)) P x in := P x ( ⋅ ∣ B R ( x ∗ )) satisfies
Cov ( P x i n ) ⪯ ( Σ − 1 + α m I d ) − 1 ⪯ σ max ( Σ ) 2 τ α I d . (87) \operatorname{Cov}(P_x^{\mathrm{in}})\preceq(\Sigma^{-1}+\alpha m I_d)^{-1}\preceq\sigma_{\max}(\Sigma)^2\tau_\alpha I_d. \tag*{(87)} Cov ( P x in ) ⪯ ( Σ − 1 + α m I d ) − 1 ⪯ σ m a x ( Σ ) 2 τ α I d . (87) (iii) Inside the well, centred at the minimiser. For x = x ∗ x=x_* x = x ∗ ,
E Y ∼ P x ∗ i n [ ∥ Y − x ∗ ∥ 2 2 ] ≤ d λ min ( Σ − 1 + α m I d ) = d σ max ( Σ ) 2 τ α . (88) \mathbb{E}_{Y\sim P_{x_*}^{\mathrm{in}}}\left[\lVert Y-x_*\rVert_2^2\right]\le\frac{d}{\lambda_{\min}(\Sigma^{-1}+\alpha m I_d)}=d\sigma_{\max}(\Sigma)^2\tau_\alpha. \tag*{(88)} E Y ∼ P x ∗ in [ ∥ Y − x ∗ ∥ 2 2 ] ≤ λ m i n ( Σ − 1 + α m I d ) d = d σ m a x ( Σ ) 2 τ α . (88)
Proof Without loss of generality, x ∗ = 0 x_*=0 x ∗ = 0 and f ∗ = 0 f_*=0 f ∗ = 0 . Write ν : = N ( x , Σ ) \nu:=\mathcal{N}(x,\Sigma) ν := N ( x , Σ ) , P : = P x P:=P_x P := P x , B : = B R ( 0 ) B:=B_R(0) B := B R ( 0 ) , and A α : = Σ − 1 + α m I d A_\alpha:=\Sigma^{-1}+\alpha mI_d A α := Σ − 1 + α m I d .
Ad (i). By Lemma 19(iv), e − α f ≥ e − α Δ / 2 e^{-\alpha f}\ge e^{-\alpha\Delta/2} e − α f ≥ e − α Δ/2 on B s ( 0 ) B_s(0) B s ( 0 ) , and Lemma 10 with radius e e e (for e = 0 e=0 e = 0 by the limiting argument in the proof of Lemma 15) together with (83) gives
Z α ( ν ) ≥ e − α Δ / 2 ν ( B s ( 0 ) ) ≥ e − α Δ / 2 D s − 1 exp ( − ( e + s ) 2 2 σ min ( Σ ) 2 ) . Z_\alpha(\nu)\ge e^{-\alpha\Delta/2}\nu(B_s(0))\ge e^{-\alpha\Delta/2}D_s^{-1}\exp\left(-\frac{(e+s)^2}{2\sigma_{\min}(\Sigma)^2}\right). Z α ( ν ) ≥ e − α Δ/2 ν ( B s ( 0 )) ≥ e − α Δ/2 D s − 1 exp ( − 2 σ m i n ( Σ ) 2 ( e + s ) 2 ) .
On R d ∖ B \mathbb{R}^d\setminus B R d ∖ B , (SF2) gives e − α f ≤ e − α Δ e^{-\alpha f}\le e^{-\alpha\Delta} e − α f ≤ e − α Δ , so the density e − α f / Z α ( ν ) e^{-\alpha f}/Z_\alpha(\nu) e − α f / Z α ( ν ) of P P P with respect to ν \nu ν is at most q α ( e ) q_\alpha(e) q α ( e ) there. Ad (ii). Since e − α f > 0 e^{-\alpha f}>0 e − α f > 0 and the Gaussian density is positive, P ( B ) > 0 P(B)>0 P ( B ) > 0 . Let g g g be as in Lemma 19(ii), and let w ( y ) : = 1 B ( y ) e − α g ( y ) w(y):=\mathbf{1}_{B}(y)e^{-\alpha g(y)} w ( y ) := 1 B ( y ) e − α g ( y ) , which is log-concave on R d \mathbb{R}^{d} R d , as g g g is convex on the convex set B B B . With Z i n ( x ) : = ∫ B e − α f ( y ) N ( y ∣ x , Σ ) d y > 0 Z_{\mathrm{in}}(x):=\int_{B}e^{-\alpha f(y)}\mathcal{N}(y\mid x,\Sigma)\,\mathrm{d}y>0 Z in ( x ) := ∫ B e − α f ( y ) N ( y ∣ x , Σ ) d y > 0 , completing the square gives
α m 2 ∥ y ∥ 2 2 + 1 2 ( y − x ) T Σ − 1 ( y − x ) = 1 2 ( y − A α − 1 Σ − 1 x ) T A α ( y − A α − 1 Σ − 1 x ) + 1 2 x T E x , \frac{\alpha m}{2}\lVert y\rVert_{2}^{2}+\frac{1}{2}(y-x)^{\mathsf{T}}\Sigma^{-1}(y-x)=\frac{1}{2}(y-A_{\alpha}^{-1}\Sigma^{-1}x)^{\mathsf{T}}A_{\alpha}(y-A_{\alpha}^{-1}\Sigma^{-1}x)+\frac{1}{2}x^{\mathsf{T}}Ex, 2 α m ∥ y ∥ 2 2 + 2 1 ( y − x ) T Σ − 1 ( y − x ) = 2 1 ( y − A α − 1 Σ − 1 x ) T A α ( y − A α − 1 Σ − 1 x ) + 2 1 x T E x ,
with the symmetric positive semidefinite matrix E : = Σ − 1 − Σ − 1 A α − 1 Σ − 1 E:=\Sigma^{-1}-\Sigma^{-1}A_{\alpha}^{-1}\Sigma^{-1} E := Σ − 1 − Σ − 1 A α − 1 Σ − 1 . Hence
log Z i n ( x ) + 1 2 x T E x = c + log ∫ R d w ( y ) exp ( − 1 2 ( y − A α − 1 Σ − 1 x ) T A α ( y − A α − 1 Σ − 1 x ) ) d y (89) \log Z_{\mathrm{in}}(x)+\frac{1}{2}x^{\mathsf{T}}Ex=c+\log\int_{\mathbb{R}^{d}}w(y)\exp\left(-\frac{1}{2}(y-A_{\alpha}^{-1}\Sigma^{-1}x)^{\mathsf{T}}A_{\alpha}(y-A_{\alpha}^{-1}\Sigma^{-1}x)\right)\,\mathrm{d}y \tag*{(89)} log Z in ( x ) + 2 1 x T E x = c + log ∫ R d w ( y ) exp ( − 2 1 ( y − A α − 1 Σ − 1 x ) T A α ( y − A α − 1 Σ − 1 x ) ) d y (89)
with a constant c c c that does not depend on x x x . The integrand is a log-concave function of ( x , y ) ∈ R d × R d (x,y)\in\mathbb{R}^{d}\times\mathbb{R}^{d} ( x , y ) ∈ R d × R d , as the logarithm of the Gaussian factor is a concave quadratic function of ( x , y ) (x,y) ( x , y ) . By Prekopa’s theorem Prekopa [14 ] , the right-hand side of (89) is a concave function of x x x .
On the other hand, the proof of Lemma 11 applies verbatim with all integrands multiplied by 1 B \mathbf{1}_{B} 1 B , and shows that x ↦ Z i n ( x ) x\mapsto Z_{\mathrm{in}}(x) x ↦ Z in ( x ) and the mean T i n ( x ) T^{\mathrm{in}}(x) T in ( x ) of P x i n P_{x}^{\mathrm{in}} P x in are continuously differentiable, with ∇ log Z i n ( x ) = Σ − 1 ( T i n ( x ) − x ) \nabla\log Z_{\mathrm{in}}(x)=\Sigma^{-1}(T^{\mathrm{in}}(x)-x) ∇ log Z in ( x ) = Σ − 1 ( T in ( x ) − x ) and ∇ T i n ( x ) = Cov ( P x i n ) Σ − 1 \nabla T^{\mathrm{in}}(x)=\operatorname{Cov}(P_{x}^{\mathrm{in}})\Sigma^{-1} ∇ T in ( x ) = Cov ( P x in ) Σ − 1 . Hence log Z i n \log Z_{\mathrm{in}} log Z in is twice continuously differentiable with
∇ 2 log Z i n ( x ) = Σ − 1 Cov ( P x i n ) Σ − 1 − Σ − 1 . \nabla^{2}\log Z_{\mathrm{in}}(x)=\Sigma^{-1}\operatorname{Cov}(P_{x}^{\mathrm{in}})\Sigma^{-1}-\Sigma^{-1}. ∇ 2 log Z in ( x ) = Σ − 1 Cov ( P x in ) Σ − 1 − Σ − 1 .
Concavity of the left-hand side of (89) gives ∇ 2 log Z i n ( x ) + E ⪯ 0 \nabla^{2}\log Z_{\mathrm{in}}(x)+E\preceq0 ∇ 2 log Z in ( x ) + E ⪯ 0 , i.e.
Σ − 1 Cov ( P x i n ) Σ − 1 ⪯ Σ − 1 − E = Σ − 1 A α − 1 Σ − 1 . \Sigma^{-1}\operatorname{Cov}(P_{x}^{\mathrm{in}})\Sigma^{-1}\preceq\Sigma^{-1}-E=\Sigma^{-1}A_{\alpha}^{-1}\Sigma^{-1}. Σ − 1 Cov ( P x in ) Σ − 1 ⪯ Σ − 1 − E = Σ − 1 A α − 1 Σ − 1 .
Multiplying from both sides by Σ \Sigma Σ gives the first inequality in (87) . The second follows from A α ⪰ ( σ max ( Σ ) − 2 + α m ) I d A_{\alpha}\succeq(\sigma_{\max}(\Sigma)^{-2}+\alpha m)I_{d} A α ⪰ ( σ m a x ( Σ ) − 2 + α m ) I d and ( σ max ( Σ ) − 2 + α m ) − 1 = σ max ( Σ ) 2 τ (\sigma_{\max}(\Sigma)^{-2}+\alpha m)^{-1}=\sigma_{\max}(\Sigma)^{2}\tau ( σ m a x ( Σ ) − 2 + α m ) − 1 = σ m a x ( Σ ) 2 τ .
Ad (iii). For x = 0 x=0 x = 0 , completing the square shows that P 0 i n P_{0}^{\mathrm{in}} P 0 in has the Lebesgue density proportional to w ( y ) exp ( − 1 2 y T A α y ) w(y)\exp(-\frac{1}{2}y^{\mathsf{T}}A_{\alpha}y) w ( y ) exp ( − 2 1 y T A α y ) . We use polar coordinates y = t u y=tu y = t u with t > 0 t>0 t > 0 and u ∈ S d − 1 u\in S^{d-1} u ∈ S d − 1 , and write a ( u ) : = u T A α u ≥ λ min ( A α ) a(u):=u^{\mathsf{T}}A_{\alpha}u\ge\lambda_{\min}(A_{\alpha}) a ( u ) := u T A α u ≥ λ m i n ( A α ) . For fixed u u u , let π u \pi_{u} π u be the probability measure on ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) with density proportional to t d − 1 e − a ( u ) t 2 / 2 t^{d-1}e^{-a(u)t^{2}/2} t d − 1 e − a ( u ) t 2 /2 , let c u : = ∫ 0 ∞ t d − 1 e − a ( u ) t 2 / 2 d t c_{u}:=\int_{0}^{\infty}t^{d-1}e^{-a(u)t^{2}/2}\,\mathrm{d}t c u := ∫ 0 ∞ t d − 1 e − a ( u ) t 2 /2 d t , and let W u ( t ) : = w ( t u ) W_{u}(t):=w(tu) W u ( t ) := w ( t u ) . By Lemma 19(ii), W u W_{u} W u is nonincreasing on ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) , as it equals e − α g ( t u ) e^{-\alpha g(tu)} e − α g ( t u ) for t ≤ R t\le R t ≤ R and vanishes for t > R t>R t > R . Since t ↦ t 2 t\mapsto t^{2} t ↦ t 2 is increasing, Chebyshev’s integral inequality gives
∫ 0 ∞ t 2 W u ( t ) π u ( d t ) ≤ ∫ 0 ∞ t 2 π u ( d t ) ∫ 0 ∞ W u ( t ) π u ( d t ) = d a ( u ) ∫ 0 ∞ W u ( t ) π u ( d t ) . \int_{0}^{\infty}t^{2}W_{u}(t)\,\pi_{u}(\mathrm{d}t)\le\int_{0}^{\infty}t^{2}\,\pi_{u}(\mathrm{d}t)\int_{0}^{\infty}W_{u}(t)\,\pi_{u}(\mathrm{d}t)=\frac{d}{a(u)}\int_{0}^{\infty}W_{u}(t)\,\pi_{u}(\mathrm{d}t). ∫ 0 ∞ t 2 W u ( t ) π u ( d t ) ≤ ∫ 0 ∞ t 2 π u ( d t ) ∫ 0 ∞ W u ( t ) π u ( d t ) = a ( u ) d ∫ 0 ∞ W u ( t ) π u ( d t ) .
where we used ∫ 0 ∞ t 2 π u ( d t ) = d / a ( u ) \int_{0}^{\infty}t^{2}\,\pi_{u}(\mathrm{d}t)=d/a(u) ∫ 0 ∞ t 2 π u ( d t ) = d / a ( u ) . Integrating over u u u with respect to the surface measure, with weights c u c_{u} c u ,
E Y ∼ P 0 i n ∥ Y ∥ 2 2 = ∫ S d − 1 c u ∫ 0 ∞ t 2 W u d π u d u ∫ S d − 1 c u ∫ 0 ∞ W u d π u d u ≤ sup u ∈ S d − 1 d a ( u ) = d λ min ( A α ) , \mathbb{E}_{Y\sim P_{0}^{\mathrm{in}}}\lVert Y\rVert_{2}^{2}=\frac{\int_{S^{d-1}}c_{u}\int_{0}^{\infty}t^{2}W_{u}\,\mathrm{d}\pi_{u}\,\mathrm{d}u}{\int_{S^{d-1}}c_{u}\int_{0}^{\infty}W_{u}\,\mathrm{d}\pi_{u}\,\mathrm{d}u}\le\sup_{u\in S^{d-1}}\frac{d}{a(u)}=\frac{d}{\lambda_{\min}(A_{\alpha})}, E Y ∼ P 0 in ∥ Y ∥ 2 2 = ∫ S d − 1 c u ∫ 0 ∞ W u d π u d u ∫ S d − 1 c u ∫ 0 ∞ t 2 W u d π u d u ≤ u ∈ S d − 1 sup a ( u ) d = λ m i n ( A α ) d ,
and λ min ( A α ) ≥ σ max ( Σ ) − 2 + α m \lambda_{\min}(A_{\alpha})\ge\sigma_{\max}(\Sigma)^{-2}+\alpha m λ m i n ( A α ) ≥ σ m a x ( Σ ) − 2 + α m gives (88) .
Recall that to prove contractivity of the transition map on the mean-field, we followed Fornasier et al. [5 ] and bounded the gradient. Under the Assumption 2, this simplifies again.
Lemma 21 (Explicit Jacobian and bias bounds). Let f f f satisfy (SF1) and (SF2) of Assumption 2, and let α > 0 \alpha> 0 α > 0 .
(i) For all r > 0 r > 0 r > 0 and x ∈ B r ( x ∗ ) x \in B_r(x_\ast) x ∈ B r ( x ∗ ) ,
∥ ∇ T α ( x ) ∥ ≤ Q α ( r ) : = κ Σ τ α + κ Σ ( 1 + ( r + R ) 2 σ max ( Σ ) 2 ) D s exp ( ( r + s ) 2 2 σ min ( Σ ) 2 − α Δ 2 ) . (90) \lVert\nabla T_\alpha(x) \rVert\le Q_\alpha(r) := \kappa_\Sigma\tau_\alpha+ \kappa_\Sigma\left(1 + \frac{(r+R)^2}{\sigma_{\max}(\Sigma)^2}\right)D_s \exp\left(\frac{(r+s)^2}{2\sigma_{\min}(\Sigma)^2} - \frac{\alpha\Delta}{2}\right). \tag*{(90)} ∥ ∇ T α ( x )∥ ≤ Q α ( r ) := κ Σ τ α + κ Σ ( 1 + σ m a x ( Σ ) 2 ( r + R ) 2 ) D s exp ( 2 σ m i n ( Σ ) 2 ( r + s ) 2 − 2 α Δ ) . (90) In particular, T α T_\alpha T α is Lipschitz continuous on B r ( x ∗ ) B_r(x_\ast) B r ( x ∗ ) with constant Q α ( r ) Q_\alpha(r) Q α ( r ) .
(ii) At the minimiser,
∥ T α ( x ∗ ) − x ∗ ∥ 2 2 ≤ d σ max ( Σ ) 2 ( τ α + D s exp ( s 2 2 σ min ( Σ ) 2 − α Δ 2 ) ) . (91) \lVert T_\alpha(x_\ast)-x_\ast\rVert_2^2 \le d\sigma_{\max}(\Sigma)^2\left(\tau_\alpha+D_s\exp\left(\frac{s^2}{2\sigma_{\min}(\Sigma)^2}-\frac{\alpha\Delta}{2}\right)\right). \tag*{(91)} ∥ T α ( x ∗ ) − x ∗ ∥ 2 2 ≤ d σ m a x ( Σ ) 2 ( τ α + D s exp ( 2 σ m i n ( Σ ) 2 s 2 − 2 α Δ ) ) . (91)
Proof. Without loss of generality, x ∗ = 0 x_\ast=0 x ∗ = 0 and f ∗ = 0 f_\ast=0 f ∗ = 0 , and write B : = B R ( 0 ) B:=B_R(0) B := B R ( 0 ) .
Ad (i). Let x ∈ B r ( 0 ) x\in B_r(0) x ∈ B r ( 0 ) , P : = P x P:=P_x P := P x , ν : = N ( x , Σ ) \nu:=\mathcal{N}(x,\Sigma) ν := N ( x , Σ ) , and let h h h be the mean of P x i n P_x^{\mathrm{in}} P x in . Since P x i n P_x^{\mathrm{in}} P x in is concentrated on the closed convex set B B B , h ∈ B h\in B h ∈ B , and hence ∥ x − h ∥ 2 ≤ r + R \lVert x-h\rVert_2\le r+R ∥ x − h ∥ 2 ≤ r + R . Let v v v be a unit vector. Since the mean minimises the expected squared distance,
v T Cov ( P ) v ≤ E Y ∼ P [ ( v T ( Y − h ) ) 2 ] = P ( B ) v T Cov ( P x i n ) v + ∫ R d ∖ B ( v T ( y − h ) ) 2 P ( d y ) . v^\mathsf{T}\operatorname{Cov}(P)v \le\mathbb{E}_{Y\sim P}\left[(v^\mathsf{T}(Y-h))^2\right] = P(B)v^\mathsf{T}\operatorname{Cov}(P_x^{\mathrm{in}})v+\int_{\mathbb{R}^d\setminus B}(v^\mathsf{T}(y-h))^2P(\mathrm{d}y). v T Cov ( P ) v ≤ E Y ∼ P [ ( v T ( Y − h ) ) 2 ] = P ( B ) v T Cov ( P x in ) v + ∫ R d ∖ B ( v T ( y − h ) ) 2 P ( d y ) .
By Lemma 20(ii), the first term is at most σ max ( Σ ) 2 τ α \sigma_{\max}(\Sigma)^2\tau_\alpha σ m a x ( Σ ) 2 τ α . By Lemma 20(i) with ϵ ≤ r \epsilon\le r ϵ ≤ r , the second term is at most
q α ( r ) E X ∼ ν [ ( v T ( X − h ) ) 2 ] = q α ( r ) ( v T Σ v + ( v T ( x − h ) ) 2 ) ≤ q α ( r ) ( σ max ( Σ ) 2 + ( r + R ) 2 ) . q_\alpha(r)\mathbb{E}_{X\sim\nu}\left[(v^\mathsf{T}(X-h))^2\right]=q_\alpha(r)\left(v^\mathsf{T}\Sigma v+(v^\mathsf{T}(x-h))^2\right)\le q_\alpha(r)\left(\sigma_{\max}(\Sigma)^2+(r+R)^2\right). q α ( r ) E X ∼ ν [ ( v T ( X − h ) ) 2 ] = q α ( r ) ( v T Σ v + ( v T ( x − h ) ) 2 ) ≤ q α ( r ) ( σ m a x ( Σ ) 2 + ( r + R ) 2 ) .
Hence ∥ Cov ( P ) ∥ ≤ σ max ( Σ ) 2 τ α + q α ( r ) ( σ max ( Σ ) 2 + ( r + R ) 2 ) \lVert\operatorname{Cov}(P)\rVert\le\sigma_{\max}(\Sigma)^2\tau_\alpha+q_\alpha(r)(\sigma_{\max}(\Sigma)^2+(r+R)^2) ∥ Cov ( P )∥ ≤ σ m a x ( Σ ) 2 τ α + q α ( r ) ( σ m a x ( Σ ) 2 + ( r + R ) 2 ) . By Lemma 11 , ∥ ∇ T α ( x ) ∥ ≤ ∥ Cov ( P ) ∥ ∥ Σ − 1 ∥ \lVert\nabla T_\alpha(x)\rVert\le\lVert\operatorname{Cov}(P)\rVert\lVert\Sigma^{-1}\rVert ∥ ∇ T α ( x )∥ ≤ ∥ Cov ( P )∥ ∥ Σ − 1 ∥ , and ∥ Σ − 1 ∥ = σ min ( Σ ) − 2 \lVert\Sigma^{-1}\rVert=\sigma_{\min}(\Sigma)^{-2} ∥ Σ − 1 ∥ = σ m i n ( Σ ) − 2 gives (90) . The Lipschitz bound follows as in the proof of Corollary 13 , since B r ( 0 ) B_r(0) B r ( 0 ) is convex.
Ad (ii). Let P : = P 0 P:=P_0 P := P 0 . By Jensen’s inequality,
∥ T α ( 0 ) ∥ 2 2 ≤ E Y ∼ P [ ∥ Y ∥ 2 2 ] = P ( B ) E Y ∼ P 0 i n [ ∥ Y ∥ 2 2 ] + ∫ R d ∖ B ∥ y ∥ 2 2 P ( d y ) . \lVert T_\alpha(0)\rVert_2^2\le\mathbb{E}_{Y\sim P}\left[\lVert Y\rVert_2^2\right]=P(B)\mathbb{E}_{Y\sim P_0^{\mathrm{in}}}\left[\lVert Y\rVert_2^2\right]+\int_{\mathbb{R}^d\setminus B}\lVert y\rVert_2^2P(\mathrm{d}y). ∥ T α ( 0 ) ∥ 2 2 ≤ E Y ∼ P [ ∥ Y ∥ 2 2 ] = P ( B ) E Y ∼ P 0 in [ ∥ Y ∥ 2 2 ] + ∫ R d ∖ B ∥ y ∥ 2 2 P ( d y ) .
By Lemma 20(iii), the first term is at most d σ max ( Σ ) 2 τ α d\sigma_{\max}(\Sigma)^2\tau_\alpha d σ m a x ( Σ ) 2 τ α . By Lemma 20(i) with ϵ = 0 \epsilon=0 ϵ = 0 , the second term is at most q α ( 0 ) E X ∼ N ( 0 , Σ ) [ ∥ X ∥ 2 2 ] = q α ( 0 ) tr ( Σ ) ≤ q α ( 0 ) d σ max ( Σ ) 2 q_\alpha(0)\mathbb{E}_{X\sim\mathcal{N}(0,\Sigma)}[\lVert X\rVert_2^2]=q_\alpha(0)\operatorname{tr}(\Sigma)\le q_\alpha(0)d\sigma_{\max}(\Sigma)^2 q α ( 0 ) E X ∼ N ( 0 , Σ ) [∥ X ∥ 2 2 ] = q α ( 0 ) tr ( Σ ) ≤ q α ( 0 ) d σ m a x ( Σ ) 2 .
We are now ready to investigate the convergence behaviour in the mean-field. First, let us define explicit thresholds. For r > 0 r>0 r > 0 , ρ ∈ ( 0 , 1 ) \rho\in(0,1) ρ ∈ ( 0 , 1 ) and β > 0 \beta>0 β > 0 , let
α c S ( r , ρ ) : = max { 2 ρ m σ min ( Σ ) 2 , 2 Δ [ ( r + s ) 2 2 σ min ( Σ ) 2 + log ( 2 κ Σ D s ρ ( 1 + ( r + R ) 2 σ max ( Σ ) 2 ) ) ] } , \alpha_c^{\mathrm{S}}(r,\rho):=\max\left\{\frac{2}{\rho m\sigma_{\min}(\Sigma)^2},\frac{2}{\Delta}\left[\frac{(r+s)^2}{2\sigma_{\min}(\Sigma)^2}+\log\left(\frac{2\kappa_\Sigma D_s}{\rho}\left(1+\frac{(r+R)^2}{\sigma_{\max}(\Sigma)^2}\right)\right)\right]\right\}, α c S ( r , ρ ) := max { ρ m σ m i n ( Σ ) 2 2 , Δ 2 [ 2 σ m i n ( Σ ) 2 ( r + s ) 2 + log ( ρ 2 κ Σ D s ( 1 + σ m a x ( Σ ) 2 ( r + R ) 2 ) ) ] } ,
α b S ( β ) : = max { 2 d m β 2 , 2 Δ [ s 2 2 σ min ( Σ ) 2 + log 2 d σ max ( Σ ) 2 D s β 2 ] } , (92) \alpha_b^{\mathrm{S}}(\beta):=\max\left\{\frac{2d}{m\beta^2},\frac{2}{\Delta}\left[\frac{s^2}{2\sigma_{\min}(\Sigma)^2}+\log\frac{2d\sigma_{\max}(\Sigma)^2D_s}{\beta^2}\right]\right\}, \tag*{(92)} α b S ( β ) := max { m β 2 2 d , Δ 2 [ 2 σ m i n ( Σ ) 2 s 2 + log β 2 2 d σ m a x ( Σ ) 2 D s ] } , (92)
α 0 S ( r , ρ ) : = max { α c S ( r , ρ ) , α b S ( ( 1 − ρ ) r ) } . \alpha_0^{\mathrm{S}}(r,\rho):=\max\left\{\alpha_c^{\mathrm{S}}(r,\rho),\alpha_b^{\mathrm{S}}((1-\rho)r)\right\}. α 0 S ( r , ρ ) := max { α c S ( r , ρ ) , α b S (( 1 − ρ ) r ) } .
Note that α c S ( r , ρ ) \alpha_c^{\mathrm{S}}(r,\rho) α c S ( r , ρ ) and α 0 S ( r , ρ ) \alpha_0^{\mathrm{S}}(r,\rho) α 0 S ( r , ρ ) depend on r r r and ρ \rho ρ but not on β \beta β , and that α b S ( β ) \alpha_b^{\mathrm{S}}(\beta) α b S ( β ) does not depend on r r r .
The next result corresponds to Theorems 1 and 2 under Assumption 2.
Proposition 22 (Explicit mean-field convergence). Let f f f satisfy (SF1) and (SF2) of Assumption 2, and let r > 0 r > 0 r > 0 and ρ ∈ ( 0 , 1 ) \rho\in(0,1) ρ ∈ ( 0 , 1 ) .
(i) Contraction. For all α ≥ α c S ( r , ρ ) \alpha\ge\alpha_{\mathrm{c}}^{\mathrm{S}}(r,\rho) α ≥ α c S ( r , ρ ) , T α \mathcal{T}_{\alpha} T α is Lipschitz continuous on B r ( x ∗ ) B_r(x_{\ast}) B r ( x ∗ ) with constant ρ \rho ρ .
(ii) Bias at the minimiser. For all β > 0 \beta> 0 β > 0 and α ≥ α b S ( β ) \alpha\ge\alpha_{\mathrm{b}}^{\mathrm{S}}(\beta) α ≥ α b S ( β ) , ∥ T α ( x ∗ ) − x ∗ ∥ 2 ≤ β \lVert\mathcal{T}_{\alpha}(x_{\ast}) - x_{\ast} \rVert_2 \le\beta ∥ T α ( x ∗ ) − x ∗ ∥ 2 ≤ β .
(iii) Invariance and convergence. For all α ≥ α 0 S ( r , ρ ) \alpha\ge\alpha_0^{\mathrm{S}}(r,\rho) α ≥ α 0 S ( r , ρ ) , we have T α ( B r ( x ∗ ) ) ⊂ B r ( x ∗ ) \mathcal{T}_{\alpha}(B_r(x_{\ast})) \subset B_r(x_{\ast}) T α ( B r ( x ∗ )) ⊂ B r ( x ∗ ) , T α \mathcal{T}_{\alpha} T α has exactly one fixed point m α m_{\alpha} m α in B r ( x ∗ ) B_r(x_{\ast}) B r ( x ∗ ) , and for every x 0 ∈ B r ( x ∗ ) x_0 \in B_r(x_{\ast}) x 0 ∈ B r ( x ∗ ) , the mean-field iterates x k + 1 = T α ( x k ) x_{k+1} = \mathcal{T}_{\alpha}(x_k) x k + 1 = T α ( x k ) satisfy x k ∈ B r ( x ∗ ) x_k \in B_r(x_{\ast}) x k ∈ B r ( x ∗ ) and ∥ x k − m α ∥ 2 ≤ ρ k ∥ x 0 − m α ∥ 2 \lVert x_k - m_{\alpha} \rVert_2 \le\rho^k \lVert x_0 - m_{\alpha} \rVert_2 ∥ x k − m α ∥ 2 ≤ ρ k ∥ x 0 − m α ∥ 2 for all k ∈ N 0 k \in\mathbb{N}_0 k ∈ N 0 .
(iv) One-step improvement. For all β > 0 \beta> 0 β > 0 and α ≥ max { α 0 S ( r , ρ ) , α b S ( β ) } \alpha\ge\max\{\alpha_0^{\mathrm{S}}(r,\rho),\alpha_{\mathrm{b}}^{\mathrm{S}}(\beta)\} α ≥ max { α 0 S ( r , ρ ) , α b S ( β )} , and all x ∈ B r ( x ∗ ) x \in B_r(x_{\ast}) x ∈ B r ( x ∗ ) ,
∥ T α ( x ) − x ∗ ∥ 2 ≤ β + ρ ∥ x − x ∗ ∥ 2 . (93) \lVert\mathcal{T}_{\alpha}(x) - x_{\ast} \rVert_2 \le\beta+ \rho\lVert x - x_{\ast} \rVert_2. \tag*{(93)} ∥ T α ( x ) − x ∗ ∥ 2 ≤ β + ρ ∥ x − x ∗ ∥ 2 . (93) Consequently, the iterates in (iii) satisfy ∥ x k − x ∗ ∥ 2 ≤ ρ k ∥ x 0 − x ∗ ∥ 2 + β 1 − ρ \lVert x_k - x_{\ast} \rVert_2 \le\rho^k \lVert x_0 - x_{\ast} \rVert_2 + \frac{\beta}{1-\rho} ∥ x k − x ∗ ∥ 2 ≤ ρ k ∥ x 0 − x ∗ ∥ 2 + 1 − ρ β for all k ∈ N 0 k \in\mathbb{N}_0 k ∈ N 0 , and ∥ m α − x ∗ ∥ 2 ≤ β 1 − ρ \lVert m_{\alpha} - x_{\ast} \rVert_2 \le\frac{\beta}{1-\rho} ∥ m α − x ∗ ∥ 2 ≤ 1 − ρ β .
Proof. Ad (i). Let α ≥ α c S ( r , ρ ) \alpha\ge\alpha_{\mathrm{c}}^{\mathrm{S}}(r,\rho) α ≥ α c S ( r , ρ ) . Since τ α ≤ ( α m σ max ( Σ ) 2 ) − 1 \tau_{\alpha} \le(\alpha m \sigma_{\max}(\Sigma)^2)^{-1} τ α ≤ ( α m σ m a x ( Σ ) 2 ) − 1 and κ Σ / σ max ( Σ ) 2 = σ min ( Σ ) − 2 \kappa_{\Sigma}/\sigma_{\max}(\Sigma)^2 = \sigma_{\min}(\Sigma)^{-2} κ Σ / σ m a x ( Σ ) 2 = σ m i n ( Σ ) − 2 , the first threshold gives κ Σ τ α ≤ ( α m σ min ( Σ ) 2 ) − 1 ≤ ρ / 2 \kappa_{\Sigma}\tau_{\alpha} \le(\alpha m \sigma_{\min}(\Sigma)^2)^{-1} \le\rho/2 κ Σ τ α ≤ ( α m σ m i n ( Σ ) 2 ) − 1 ≤ ρ /2 . Taking logarithms, the second threshold is equivalent to
κ Σ ( 1 + ( r + R ) 2 σ max ( Σ ) 2 ) D s exp ( ( r + s ) 2 2 σ min ( Σ ) 2 − α Δ 2 ) ≤ ρ 2 . \kappa_{\Sigma}\left(1 + \frac{(r+R)^2}{\sigma_{\max}(\Sigma)^2}\right)D_s \exp\left(\frac{(r+s)^2}{2\sigma_{\min}(\Sigma)^2} - \frac{\alpha\Delta}{2}\right) \le\frac{\rho}{2}. κ Σ ( 1 + σ m a x ( Σ ) 2 ( r + R ) 2 ) D s exp ( 2 σ m i n ( Σ ) 2 ( r + s ) 2 − 2 α Δ ) ≤ 2 ρ .
Hence Q α ( r ) ≤ ρ Q_{\alpha}(r) \le\rho Q α ( r ) ≤ ρ , and Lemma 21(i) gives the claim.
Ad (ii). Let α ≥ α b S ( β ) \alpha\ge\alpha_{\mathrm{b}}^{\mathrm{S}}(\beta) α ≥ α b S ( β ) . As in (i), the first threshold gives τ α ≤ ( α m σ max ( Σ ) 2 ) − 1 ≤ β 2 2 d σ max ( Σ ) 2 \tau_{\alpha} \le(\alpha m \sigma_{\max}(\Sigma)^2)^{-1} \le\frac{\beta^2}{2d\sigma_{\max}(\Sigma)^2} τ α ≤ ( α m σ m a x ( Σ ) 2 ) − 1 ≤ 2 d σ m a x ( Σ ) 2 β 2 , and the second gives D s exp ( s 2 / ( 2 σ min ( Σ ) 2 ) − α Δ / 2 ) ≤ β 2 2 d σ max ( Σ ) 2 D_s \exp(s^2/(2\sigma_{\min}(\Sigma)^2) - \alpha\Delta/2) \le\frac{\beta^2}{2d\sigma_{\max}(\Sigma)^2} D s exp ( s 2 / ( 2 σ m i n ( Σ ) 2 ) − α Δ/2 ) ≤ 2 d σ m a x ( Σ ) 2 β 2 . Inserting both into (91) gives ∥ T α ( x ∗ ) − x ∗ ∥ 2 2 ≤ β 2 \lVert\mathcal{T}_{\alpha}(x_{\ast}) - x_{\ast} \rVert_2^2 \le\beta^2 ∥ T α ( x ∗ ) − x ∗ ∥ 2 2 ≤ β 2 .
Ad (iii). Let α ≥ α 0 S ( r , ρ ) \alpha\ge\alpha_0^{\mathrm{S}}(r,\rho) α ≥ α 0 S ( r , ρ ) . By (i) and (ii) with β = ( 1 − ρ ) r \beta= (1-\rho)r β = ( 1 − ρ ) r , for x ∈ B r ( x ∗ ) x \in B_r(x_{\ast}) x ∈ B r ( x ∗ ) ,
∥ T α ( x ) − x ∗ ∥ 2 ≤ ∥ T α ( x ) − T α ( x ∗ ) ∥ 2 + ∥ T α ( x ∗ ) − x ∗ ∥ 2 ≤ ρ r + ( 1 − ρ ) r = r . \lVert\mathcal{T}_{\alpha}(x) - x_{\ast} \rVert_2 \le\lVert\mathcal{T}_{\alpha}(x) - \mathcal{T}_{\alpha}(x_{\ast}) \rVert_2 + \lVert\mathcal{T}_{\alpha}(x_{\ast}) - x_{\ast} \rVert_2 \le\rho r + (1-\rho)r = r. ∥ T α ( x ) − x ∗ ∥ 2 ≤ ∥ T α ( x ) − T α ( x ∗ ) ∥ 2 + ∥ T α ( x ∗ ) − x ∗ ∥ 2 ≤ ρ r + ( 1 − ρ ) r = r .
The remaining claims follow from Banach’s fixed-point theorem on the compact set B r ( x ∗ ) B_r(x_{\ast}) B r ( x ∗ ) and induction, exactly as in the proof of Theorem 1.
Ad (iv). Let β r : = min { β , ( 1 − ρ ) r } \beta_r := \min\{\beta,(1-\rho)r\} β r := min { β , ( 1 − ρ ) r } . Since α b S ( β r ) \alpha_{\mathrm{b}}^{\mathrm{S}}(\beta_r) α b S ( β r ) equals α b S ( β ) \alpha_{\mathrm{b}}^{\mathrm{S}}(\beta) α b S ( β ) or α b S ( ( 1 − ρ ) r ) ≤ α 0 S ( r , ρ ) \alpha_{\mathrm{b}}^{\mathrm{S}}((1-\rho)r) \le\alpha_0^{\mathrm{S}}(r,\rho) α b S (( 1 − ρ ) r ) ≤ α 0 S ( r , ρ ) , (ii) gives ∥ T α ( x ∗ ) − x ∗ ∥ 2 ≤ β r ≤ β \lVert\mathcal{T}_{\alpha}(x_{\ast}) - x_{\ast} \rVert_2 \le\beta_r \le\beta ∥ T α ( x ∗ ) − x ∗ ∥ 2 ≤ β r ≤ β . Together with (i), the triangle inequality as in (iii) yields (93) . With e k : = ∥ x k − x ∗ ∥ 2 e_k := \lVert x_k - x_{\ast} \rVert_2 e k := ∥ x k − x ∗ ∥ 2 , we get e k + 1 ≤ β + ρ e k e_{k+1} \le\beta+ \rho e_k e k + 1 ≤ β + ρ e k , since x k ∈ B r ( x ∗ ) x_k \in B_r(x_{\ast}) x k ∈ B r ( x ∗ ) by (iii), and induction gives e k ≤ ρ k e 0 + β ∑ j < k ρ j ≤ ρ k e 0 + β 1 − ρ e_k \le\rho^k e_0 + \beta\sum_{j<k}\rho^j \le\rho^k e_0 + \frac{\beta}{1-\rho} e k ≤ ρ k e 0 + β ∑ j < k ρ j ≤ ρ k e 0 + 1 − ρ β . Applying this with x 0 = m α x_0 = m_{\alpha} x 0 = m α , for which x k = m α x_k = m_{\alpha} x k = m α for all k k k , and letting k → ∞ k \to\infty k → ∞ gives the bound on m α m_{\alpha} m α .
The next result summarises the semi-global convergence behaviour under Assumption 2 and is a more detailed version of Corollary 4 from the main text.
Corollary 23 (Explicit accuracy of mean-field MPPI). Let f f f satisfy (SF1) and (SF2) of Assumption 2, let r > 0 r > 0 r > 0 and ϵ > 0 \epsilon> 0 ϵ > 0 , and set K 0 ( r , ϵ ) : = max { 0 , ⌈ log 2 ( 2 r / ϵ ) ⌉ } K_0(r,\epsilon) := \max\{0,\lceil\log_2(2r/\epsilon)\rceil\} K 0 ( r , ϵ ) := max { 0 , ⌈ log 2 ( 2 r / ϵ )⌉} . Then for all
α ≥ α m f S ( r , ϵ ) : = max { α c S ( r , 1 2 ) , α b S ( r 2 ) , α b S ( ϵ 4 ) } (94) \alpha\ge\alpha_{\mathrm{mf}}^{\mathrm{S}}(r,\epsilon) := \max\left\{\alpha_{\mathrm{c}}^{\mathrm{S}}\left(r,\frac{1}{2}\right),\alpha_{\mathrm{b}}^{\mathrm{S}}\left(\frac{r}{2}\right),\alpha_{\mathrm{b}}^{\mathrm{S}}\left(\frac{\epsilon}{4}\right)\right\} \tag*{(94)} α ≥ α mf S ( r , ϵ ) := max { α c S ( r , 2 1 ) , α b S ( 2 r ) , α b S ( 4 ϵ ) } (94) and all x 0 ∈ B r ( x ∗ ) x_0 \in B_r(x_{\ast}) x 0 ∈ B r ( x ∗ ) , the mean-field iterates satisfy ∥ x k − x ∗ ∥ 2 ≤ ϵ \lVert x_k - x_{\ast} \rVert_2 \le\epsilon ∥ x k − x ∗ ∥ 2 ≤ ϵ for all k ≥ K 0 ( r , ϵ ) k \ge K_0(r,\epsilon) k ≥ K 0 ( r , ϵ ) .
Proof. The right-hand side of (94) equals max { α 0 S ( r , 1 2 ) , α b S ( ϵ 4 ) } \max\{\alpha_0^{\mathrm{S}}(r,\frac{1}{2}),\alpha_b^{\mathrm{S}}(\frac{\epsilon}{4})\} max { α 0 S ( r , 2 1 ) , α b S ( 4 ϵ )} . Proposition 22(iv) with ρ = 1 2 \rho=\frac{1}{2} ρ = 2 1 and β = ϵ 4 \beta=\frac{\epsilon}{4} β = 4 ϵ gives ∥ x k − x ∗ ∥ 2 ≤ 2 − k r + ϵ 2 ≤ ϵ \lVert x_k-x_*\rVert_2\le2^{-k}r+\frac{\epsilon}{2}\le\epsilon ∥ x k − x ∗ ∥ 2 ≤ 2 − k r + 2 ϵ ≤ ϵ for k ≥ K 0 ( r , ϵ ) k\ge K_0(r,\epsilon) k ≥ K 0 ( r , ϵ ) , since 2 k ≥ 2 r / ϵ 2^k\ge2r/\epsilon 2 k ≥ 2 r / ϵ if K 0 ( r , ϵ ) > 0 K_0(r,\epsilon)>0 K 0 ( r , ϵ ) > 0 , and r ≤ ϵ / 2 r\le\epsilon/2 r ≤ ϵ /2 otherwise. ▪
Finite-sample Results
We now investigate the behaviour of MPPI in a finite-sample setting under Assumption 2. Recall that the starting point of our finite-sample analysis is Lemma 17 , and that the main quantities are the total variance of the Gibbs tilt and the expected normalised weight. The next result bounds these quantities in our simplified setting.
Lemma 24 (Explicit variance and effective sample size) Let f f f satisfy (SF1) and (SF2) of Assumption 2, let α > 0 \alpha>0 α > 0 , r > 0 r>0 r > 0 and x ∈ B r ( x ∗ ) x\in B_r(x_*) x ∈ B r ( x ∗ ) .
(i) The total variance V α ( x ) V_\alpha(x) V α ( x ) from (68) satisfies
V α ( x ) ≤ d σ max ( Σ ) 2 τ α + ( d σ max ( Σ ) 2 + ( r + R ) 2 ) D s exp ( ( r + s ) 2 2 σ min ( Σ ) 2 − α Δ 2 ) . (95) V_\alpha(x)\le d\,\sigma_{\max}(\Sigma)^2\tau_\alpha+\left(d\,\sigma_{\max}(\Sigma)^2+(r+R)^2\right)D_s\exp\left(\frac{(r+s)^2}{2\sigma_{\min}(\Sigma)^2}-\frac{\alpha\Delta}{2}\right). \tag*{(95)} V α ( x ) ≤ d σ m a x ( Σ ) 2 τ α + ( d σ m a x ( Σ ) 2 + ( r + R ) 2 ) D s exp ( 2 σ m i n ( Σ ) 2 ( r + s ) 2 − 2 α Δ ) . (95) (ii) If f f f also satisfies (SF3) and
α ≥ α ζ S ( r ) : = 1 M max { ( 4 d + 8 ) ln 2 R 2 , 2 r R σ min ( Σ ) 2 } , (96) \alpha\ge\alpha_\zeta^{\mathrm{S}}(r):=\frac{1}{M}\max\left\{\frac{(4d+8)\ln2}{R^2},\frac{2r}{R\sigma_{\min}(\Sigma)^2}\right\}, \tag*{(96)} α ≥ α ζ S ( r ) := M 1 max { R 2 ( 4 d + 8 ) ln 2 , R σ m i n ( Σ ) 2 2 r } , (96) then the expected normalised weight ζ α ( x ) \zeta_\alpha(x) ζ α ( x ) from (17) satisfies
1 ζ α ( x ) ≤ 2 det ( I d + α M Σ ) 1 / 2 exp ( 1 2 ( x − x ∗ ) T ( Σ + ( α M ) − 1 I d ) − 1 ( x − x ∗ ) ) \frac{1}{\zeta_\alpha(x)}\le2\det\left(I_d+\alpha M\Sigma\right)^{1/2}\exp\left(\frac{1}{2}(x-x_*)^{\mathsf{T}}\left(\Sigma+(\alpha M)^{-1}I_d\right)^{-1}(x-x_*)\right) ζ α ( x ) 1 ≤ 2 det ( I d + α M Σ ) 1/2 exp ( 2 1 ( x − x ∗ ) T ( Σ + ( α M ) − 1 I d ) − 1 ( x − x ∗ ) ) ≤ 2 ( 1 + α M σ max ( Σ ) 2 ) d / 2 exp ( r 2 2 σ min ( Σ ) 2 ) . (97) \le2\left(1+\alpha M\sigma_{\max}(\Sigma)^2\right)^{d/2}\exp\left(\frac{r^2}{2\sigma_{\min}(\Sigma)^2}\right). \tag*{(97)} ≤ 2 ( 1 + α M σ m a x ( Σ ) 2 ) d /2 exp ( 2 σ m i n ( Σ ) 2 r 2 ) . (97)
Proof. Without loss of generality, x ∗ = 0 x_*=0 x ∗ = 0 and f ∗ = 0 f_*=0 f ∗ = 0 , and write B : = B R ( 0 ) B:=B_R(0) B := B R ( 0 ) , P : = P x P:=P_x P := P x and ν : = N ( x , Σ ) \nu:=\mathcal{N}(x,\Sigma) ν := N ( x , Σ ) .
Ad (i). Let h h h be the mean of P x i n P_x^{\mathrm{in}} P x in , so that ∥ x − h ∥ 2 ≤ r + R \lVert x-h\rVert_2\le r+R ∥ x − h ∥ 2 ≤ r + R as in the proof of Lemma 21 . Since the mean minimises the expected squared distance,
V α ( x ) ≤ E Y ∼ P [ ∥ Y − h ∥ 2 2 ] = P ( B ) tr Cov ( P x i n ) + ∫ R d ∖ B ∥ y − h ∥ 2 2 P ( d y ) . V_\alpha(x)\le\mathbb{E}_{Y\sim P}\left[\lVert Y-h\rVert_2^2\right]=P(B)\operatorname{tr}\operatorname{Cov}(P_x^{\mathrm{in}})+\int_{\mathbb{R}^d\setminus B}\lVert y-h\rVert_2^2P(\mathrm{d}y). V α ( x ) ≤ E Y ∼ P [ ∥ Y − h ∥ 2 2 ] = P ( B ) tr Cov ( P x in ) + ∫ R d ∖ B ∥ y − h ∥ 2 2 P ( d y ) .
By Lemma 20(ii), the first term is at most d σ max ( Σ ) 2 τ α d\sigma_{\max}(\Sigma)^2\tau_\alpha d σ m a x ( Σ ) 2 τ α . By Lemma 20(i), the second term is at most q α ( r ) E X ∼ ν [ ∥ X − h ∥ 2 2 ] = q α ( r ) ( tr ( Σ ) + ∥ x − h ∥ 2 2 ) ≤ q α ( r ) ( d σ max ( Σ ) 2 + ( r + R ) 2 ) q_\alpha(r)\mathbb{E}_{X\sim\nu}[\lVert X-h\rVert_2^2]=q_\alpha(r)\left(\operatorname{tr}(\Sigma)+\lVert x-h\rVert_2^2\right)\le q_\alpha(r)\left(d\sigma_{\max}(\Sigma)^2+(r+R)^2\right) q α ( r ) E X ∼ ν [∥ X − h ∥ 2 2 ] = q α ( r ) ( tr ( Σ ) + ∥ x − h ∥ 2 2 ) ≤ q α ( r ) ( d σ m a x ( Σ ) 2 + ( r + R ) 2 ) .
Ad (ii). Write a : = α M a:=\alpha M a := α M , and for b > 0 b>0 b > 0 let G b : = E X ∼ ν [ e − b 2 ∥ X ∥ 2 2 ] G_b:=\mathbb{E}_{X\sim\nu}\left[e^{-\frac{b}{2}\lVert X\rVert_2^2}\right] G b := E X ∼ ν [ e − 2 b ∥ X ∥ 2 2 ] . Completing the square gives the Gaussian integral
G b = det ( I d + b Σ ) − 1 / 2 exp ( − 1 2 x T ( Σ + b − 1 I d ) − 1 x ) . (98) G_b=\det\left(I_d+b\Sigma\right)^{-1/2}\exp\left(-\frac{1}{2}x^{\mathsf{T}}\left(\Sigma+b^{-1}I_d\right)^{-1}x\right). \tag*{(98)} G b = det ( I d + b Σ ) − 1/2 exp ( − 2 1 x T ( Σ + b − 1 I d ) − 1 x ) . (98)
By (SF3), e − α f ( y ) ≥ e − a 2 ∥ y ∥ 2 2 e^{-\alpha f(y)} \ge e^{-\frac{a}{2}\lVert y\rVert_{2}^{2}} e − α f ( y ) ≥ e − 2 a ∥ y ∥ 2 2 for y ∈ B y \in B y ∈ B , and e − α f ≥ 0 e^{-\alpha f} \ge0 e − α f ≥ 0 on R d ∖ B \mathbb{R}^{d} \setminus B R d ∖ B . Since f ∗ = 0 f_{\ast}=0 f ∗ = 0 , this gives
ζ α ( x ) ≥ E X ∼ ν [ e − a 2 ∥ X ∥ 2 2 1 B ( X ) ] = G a − E X ∼ ν [ e − a 2 ∥ X ∥ 2 2 1 R d ∖ B ( X ) ] . \zeta_{\alpha}(x) \ge\mathbb{E}_{X\sim\nu}\left[e^{-\frac{a}{2}\lVert X\rVert_{2}^{2}}\mathbf{1}_{B}(X)\right] = G_{a}-\mathbb{E}_{X\sim\nu}\left[e^{-\frac{a}{2}\lVert X\rVert_{2}^{2}}\mathbf{1}_{\mathbb{R}^{d}\setminus B}(X)\right]. ζ α ( x ) ≥ E X ∼ ν [ e − 2 a ∥ X ∥ 2 2 1 B ( X ) ] = G a − E X ∼ ν [ e − 2 a ∥ X ∥ 2 2 1 R d ∖ B ( X ) ] .
If ∥ y ∥ 2 > R \lVert y\rVert_{2}>R ∥ y ∥ 2 > R , then e − a 2 ∥ y ∥ 2 2 ≤ e − a R 2 / 4 e − a 4 ∥ y ∥ 2 2 e^{-\frac{a}{2}\lVert y\rVert_{2}^{2}}\le e^{-aR^{2}/4}e^{-\frac{a}{4}\lVert y\rVert_{2}^{2}} e − 2 a ∥ y ∥ 2 2 ≤ e − a R 2 /4 e − 4 a ∥ y ∥ 2 2 , so the last expectation is at most e − a R 2 / 4 G a / 2 e^{-aR^{2}/4}G_{a/2} e − a R 2 /4 G a /2 . We compare G a / 2 G_{a/2} G a /2 with G a G_{a} G a by means of (98) . Let σ 1 2 , … , σ d 2 \sigma_{1}^{2},\ldots,\sigma_{d}^{2} σ 1 2 , … , σ d 2 be the eigenvalues of Σ \Sigma Σ . Then det ( I d + a Σ ) / det ( I d + a 2 Σ ) = ∏ i ( 1 + a σ i 2 ) / ( 1 + a 2 σ i 2 ) ≤ 2 d \det(I_{d}+a\Sigma)/\det(I_{d}+\frac{a}{2}\Sigma)=\prod_{i}(1+a\sigma_{i}^{2})/(1+\frac{a}{2}\sigma_{i}^{2})\le2^{d} det ( I d + a Σ ) / det ( I d + 2 a Σ ) = ∏ i ( 1 + a σ i 2 ) / ( 1 + 2 a σ i 2 ) ≤ 2 d . Moreover, ( Σ + a − 1 I d ) − 1 − ( Σ + 2 a − 1 I d ) − 1 (\Sigma+a^{-1}I_{d})^{-1}-(\Sigma+2a^{-1}I_{d})^{-1} ( Σ + a − 1 I d ) − 1 − ( Σ + 2 a − 1 I d ) − 1 has the eigenvalues
1 σ i 2 + a − 1 − 1 σ i 2 + 2 a − 1 = a − 1 ( σ i 2 + a − 1 ) ( σ i 2 + 2 a − 1 ) ≤ 1 a σ min ( Σ ) 4 . \frac{1}{\sigma_{i}^{2}+a^{-1}}-\frac{1}{\sigma_{i}^{2}+2a^{-1}}=\frac{a^{-1}}{(\sigma_{i}^{2}+a^{-1})(\sigma_{i}^{2}+2a^{-1})}\le\frac{1}{a\sigma_{\min}(\Sigma)^{4}}. σ i 2 + a − 1 1 − σ i 2 + 2 a − 1 1 = ( σ i 2 + a − 1 ) ( σ i 2 + 2 a − 1 ) a − 1 ≤ a σ m i n ( Σ ) 4 1 .
and ∥ x ∥ 2 ≤ r \lVert x\rVert_{2}\le r ∥ x ∥ 2 ≤ r . Hence G a / 2 ≤ 2 d / 2 exp ( r 2 / ( 2 a σ min ( Σ ) 4 ) ) G a G_{a/2}\le2^{d/2}\exp\left(r^{2}/(2a\sigma_{\min}(\Sigma)^{4})\right)G_{a} G a /2 ≤ 2 d /2 exp ( r 2 / ( 2 a σ m i n ( Σ ) 4 ) ) G a , and
ζ α ( x ) ≥ ( 1 − 2 d / 2 exp ( − a R 2 4 + r 2 2 a σ min ( Σ ) 4 ) ) G a . \zeta_{\alpha}(x)\ge\left(1-2^{d/2}\exp\left(-\frac{aR^{2}}{4}+\frac{r^{2}}{2a\sigma_{\min}(\Sigma)^{4}}\right)\right)G_{a}. ζ α ( x ) ≥ ( 1 − 2 d /2 exp ( − 4 a R 2 + 2 a σ m i n ( Σ ) 4 r 2 ) ) G a .
By (96) , a R 2 / 8 ≥ ( d 2 + 1 ) ln 2 aR^{2}/8\ge(\frac{d}{2}+1)\ln2 a R 2 /8 ≥ ( 2 d + 1 ) ln 2 and a 2 ≥ 4 r 2 / ( R 2 σ min ( Σ ) 4 ) a^{2}\ge4r^{2}/(R^{2}\sigma_{\min}(\Sigma)^{4}) a 2 ≥ 4 r 2 / ( R 2 σ m i n ( Σ ) 4 ) , i.e. a R 2 / 8 ≥ r 2 / ( 2 a σ min ( Σ ) 4 ) aR^{2}/8\ge r^{2}/(2a\sigma_{\min}(\Sigma)^{4}) a R 2 /8 ≥ r 2 / ( 2 a σ m i n ( Σ ) 4 ) . Writing − a R 2 4 = − a R 2 8 − a R 2 8 -\frac{aR^{2}}{4}=-\frac{aR^{2}}{8}-\frac{aR^{2}}{8} − 4 a R 2 = − 8 a R 2 − 8 a R 2 , the subtracted term is therefore at most 2 d / 2 ⋅ 2 − d / 2 − 1 = 1 2 2^{d/2}\cdot2^{-d/2-1}=\frac{1}{2} 2 d /2 ⋅ 2 − d /2 − 1 = 2 1 , and ζ α ( x ) ≥ 1 2 G a \zeta_{\alpha}(x)\ge\frac{1}{2}G_{a} ζ α ( x ) ≥ 2 1 G a , which together with (98) is the first inequality in (97) . The second one follows from det ( I d + a Σ ) ≤ ( 1 + a σ max ( Σ ) 2 ) d \det(I_{d}+a\Sigma)\le(1+a\sigma_{\max}(\Sigma)^{2})^{d} det ( I d + a Σ ) ≤ ( 1 + a σ m a x ( Σ ) 2 ) d and ( Σ + a − 1 I d ) − 1 ⪯ ( σ min ( Σ ) 2 + a − 1 ) − 1 I d ⪯ σ min ( Σ ) − 2 I d (\Sigma+a^{-1}I_{d})^{-1}\preceq(\sigma_{\min}(\Sigma)^{2}+a^{-1})^{-1}I_{d}\preceq\sigma_{\min}(\Sigma)^{-2}I_{d} ( Σ + a − 1 I d ) − 1 ⪯ ( σ m i n ( Σ ) 2 + a − 1 ) − 1 I d ⪯ σ m i n ( Σ ) − 2 I d .
For r > 0 r>0 r > 0 and β > 0 \beta>0 β > 0 , define
α ν S ( r , β ) : = max { 4 d m β 2 , 2 Δ [ ( r + s ) 2 2 σ min ( Σ ) 2 + log 4 ( d σ max ( Σ ) 2 + ( r + R ) 2 ) D s β 2 ] } . (99) \alpha_{\nu}^{\mathrm{S}}(r,\beta):=\max\left\{\frac{4d}{m\beta^{2}},\frac{2}{\Delta}\left[\frac{(r+s)^{2}}{2\sigma_{\min}(\Sigma)^{2}}+\log\frac{4(d\sigma_{\max}(\Sigma)^{2}+(r+R)^{2})D_{s}}{\beta^{2}}\right]\right\}. \tag*{(99)} α ν S ( r , β ) := max { m β 2 4 d , Δ 2 [ 2 σ m i n ( Σ ) 2 ( r + s ) 2 + log β 2 4 ( d σ m a x ( Σ ) 2 + ( r + R ) 2 ) D s ] } . (99)
which ensures V α ( x ) ≤ β 2 / 2 V_{\alpha}(x)\le\beta^{2}/2 V α ( x ) ≤ β 2 /2 for all x ∈ B r ( x ∗ ) x\in B_{r}(x_{\ast}) x ∈ B r ( x ∗ ) , cf. the proof below. We can now formulate a result on the finite-sample behaviour for finite horizons under our simplified assumption.
Proposition 25 (Explicit finite-sample convergence) Let f f f satisfy Assumption 2 , including (SF3), let r > 0 r>0 r > 0 , ρ ∈ ( 0 , 1 ) \rho\in(0,1) ρ ∈ ( 0 , 1 ) and β > 0 \beta>0 β > 0 with 2 β ≤ ( 1 − ρ ) r 2\beta\le(1-\rho)r 2 β ≤ ( 1 − ρ ) r , and let
α ≥ max { α 0 S ( r , ρ ) , α ξ S ( β ) , α ξ S ( r , β ) , α ζ S ( r ) } . (100) \alpha\ge\max\left\{\alpha_{0}^{\mathrm{S}}(r,\rho),\alpha_{\xi}^{\mathrm{S}}(\beta),\alpha_{\xi}^{\mathrm{S}}(r,\beta),\alpha_{\zeta}^{\mathrm{S}}(r)\right\}. \tag*{(100)} α ≥ max { α 0 S ( r , ρ ) , α ξ S ( β ) , α ξ S ( r , β ) , α ζ S ( r ) } . (100) Then for all K ∈ N + K\in\mathbb{N}_{+} K ∈ N + , δ ∈ ( 0 , 1 ) \delta\in(0,1) δ ∈ ( 0 , 1 ) , N ∈ N + N\in\mathbb{N}_{+} N ∈ N + with
N ≥ 12 K δ ( 1 + α M σ max ( Σ ) 2 ) d / 2 exp ( r 2 2 σ min ( Σ ) 2 ) , (101) N\ge\frac{12K}{\delta}(1+\alpha M\sigma_{\max}(\Sigma)^{2})^{d/2}\exp\left(\frac{r^{2}}{2\sigma_{\min}(\Sigma)^{2}}\right), \tag*{(101)} N ≥ δ 12 K ( 1 + α M σ m a x ( Σ ) 2 ) d /2 exp ( 2 σ m i n ( Σ ) 2 r 2 ) , (101) and all x ^ 0 ∈ B r ( x ∗ ) \hat{x}_{0}\in B_{r}(x_{\ast}) x ^ 0 ∈ B r ( x ∗ ) , the MPPI iterates (1) satisfy
P ( x ^ k ∈ B r ( x ∗ ) and ∥ x ^ k − x ∗ ∥ 2 ≤ ρ k ∥ x ^ 0 − x ∗ ∥ 2 + 2 β 1 − ρ for all k = 0 , … , K ) ≥ 1 − δ . (102) \mathbb{P}\left(\hat{x}_{k}\in B_{r}(x_{\ast})\ \text{and}\ \lVert\hat{x}_{k}-x_{\ast}\rVert_{2}\le\rho^{k}\lVert\hat{x}_{0}-x_{\ast}\rVert_{2}+\frac{2\beta}{1-\rho}\ \text{for all }k=0,\ldots,K\right)\ge1-\delta. \tag*{(102)} P ( x ^ k ∈ B r ( x ∗ ) and ∥ x ^ k − x ∗ ∥ 2 ≤ ρ k ∥ x ^ 0 − x ∗ ∥ 2 + 1 − ρ 2 β for all k = 0 , … , K ) ≥ 1 − δ . (102)
Proof. Fix α \alpha α , K K K , δ \delta δ , N N N and x ^ 0 \hat{x}_0 x ^ 0 as in the statement, and write e k : = ∥ x ^ k − x ∗ ∥ 2 e_k := \lVert\hat{x}_k-x_*\rVert_2 e k := ∥ x ^ k − x ∗ ∥ 2 .
Step 1: Variance and effective sample size. Let x ∈ B r ( x ∗ ) x \in B_r(x_*) x ∈ B r ( x ∗ ) . The first threshold in (99) gives d σ max ( Σ ) 2 τ α ≤ d / ( α m ) ≤ β 2 / 4 d\sigma_{\max}(\Sigma)^2\tau_\alpha\le d/(\alpha m) \le\beta^2/4 d σ m a x ( Σ ) 2 τ α ≤ d / ( α m ) ≤ β 2 /4 , and the second one bounds the second term in (95) by β 2 / 4 \beta^2/4 β 2 /4 . Hence Lemma 24(i) gives V α ( x ) ≤ β 2 / 2 V_\alpha(x) \le\beta^2/2 V α ( x ) ≤ β 2 /2 . Moreover, Lemma 24(ii) and (101) give 6 N ζ α ( x ) ≤ δ K \frac{6}{N\zeta_\alpha(x)} \le\frac{\delta}{K} N ζ α ( x ) 6 ≤ K δ .
Step 2: One sampled step. For x ∈ R d x \in\mathbb{R}^d x ∈ R d , let ψ ( x ) \psi(x) ψ ( x ) be the probability that the empirical transition map at x x x , computed from N N N independent samples of N ( x , Σ ) \mathcal{N}(x,\Sigma) N ( x , Σ ) , deviates from T α ( x ) T_\alpha(x) T α ( x ) by more than β \beta β . For x ∈ B r ( x ∗ ) x \in B_r(x_*) x ∈ B r ( x ∗ ) , Lemma 17(i) with t = β t=\beta t = β and Step 1 give
ψ ( x ) ≤ 4 N ζ α ( x ) ( 1 + V α ( x ) β 2 ) ≤ 6 N ζ α ( x ) ≤ δ K . \psi(x) \le\frac{4}{N\zeta_\alpha(x)}\left(1+\frac{V_\alpha(x)}{\beta^2}\right) \le\frac{6}{N\zeta_\alpha(x)} \le\frac{\delta}{K}. ψ ( x ) ≤ N ζ α ( x ) 4 ( 1 + β 2 V α ( x ) ) ≤ N ζ α ( x ) 6 ≤ K δ .
Step 3: Deterministic progress on good events. For k ∈ N 0 k \in\mathbb{N}_0 k ∈ N 0 , let A k : = { ∥ x ^ k + 1 − T α ( x ^ k ) ∥ 2 ≤ β } A_k := \{\lVert\hat{x}_{k+1}-T_\alpha(\hat{x}_k)\rVert_2 \le\beta\} A k := {∥ x ^ k + 1 − T α ( x ^ k ) ∥ 2 ≤ β } . Suppose that x ^ k ∈ B r ( x ∗ ) \hat{x}_k \in B_r(x_*) x ^ k ∈ B r ( x ∗ ) and that A k A_k A k occurs. Since α ≥ max { α δ S ( r , ρ ) , α b S ( β ) } \alpha\ge\max\{\alpha_\delta^{\mathrm{S}}(r,\rho),\alpha_b^{\mathrm{S}}(\beta)\} α ≥ max { α δ S ( r , ρ ) , α b S ( β )} , Proposition 22(iv) gives
e k + 1 ≤ ∥ T α ( x ^ k ) − x ∗ ∥ 2 + β ≤ 2 β + ρ e k ≤ 2 β + ρ r ≤ r , e_{k+1} \le\lVert T_\alpha(\hat{x}_k)-x_*\rVert_2+\beta\le2\beta+\rho e_k \le2\beta+\rho r \le r, e k + 1 ≤ ∥ T α ( x ^ k ) − x ∗ ∥ 2 + β ≤ 2 β + ρ e k ≤ 2 β + ρ r ≤ r ,
where the last step uses 2 β ≤ ( 1 − ρ ) r 2\beta\le(1-\rho)r 2 β ≤ ( 1 − ρ ) r . By induction, on ⋂ j = 0 k − 1 A j \bigcap_{j=0}^{k-1}A_j ⋂ j = 0 k − 1 A j we have x ^ j ∈ B r ( x ∗ ) \hat{x}_j \in B_r(x_*) x ^ j ∈ B r ( x ∗ ) for j = 0 , … , k j=0,\ldots,k j = 0 , … , k , and e k ≤ ρ k e 0 + 2 β ∑ j < k ρ j ≤ ρ k e 0 + 2 β 1 − ρ e_k \le\rho^k e_0+2\beta\sum_{j<k}\rho^j \le\rho^k e_0+\frac{2\beta}{1-\rho} e k ≤ ρ k e 0 + 2 β ∑ j < k ρ j ≤ ρ k e 0 + 1 − ρ 2 β . Hence the event in (102) contains ⋂ k = 0 K − 1 A k \bigcap_{k=0}^{K-1}A_k ⋂ k = 0 K − 1 A k .
Step 4: Probability of the good events. As in Steps 0 and 4 of the proof of Theorem 6, x ^ k \hat{x}_k x ^ k is independent of the samples used in step k k k , so Fubini’s theorem and Step 2 give P ( A k c ∩ { x ^ k ∈ B r ( x ∗ ) } ) = E [ 1 B r ( x ∗ ) ( x ^ k ) ψ ( x ^ k ) ] ≤ δ / K \mathbb{P}(A_k^c\cap\{\hat{x}_k\in B_r(x_*)\})=\mathbb{E}[\mathbf{1}_{B_r(x_*)}(\hat{x}_k)\psi(\hat{x}_k)]\le\delta/K P ( A k c ∩ { x ^ k ∈ B r ( x ∗ )}) = E [ 1 B r ( x ∗ ) ( x ^ k ) ψ ( x ^ k )] ≤ δ / K . Decomposing according to the first failure time and using Step 3,
P ( ⋃ k = 0 K − 1 A k c ) ≤ ∑ k = 0 K − 1 P ( A k c ∩ { x ^ k ∈ B r ( x ∗ ) } ) ≤ δ , \mathbb{P}\left(\bigcup_{k=0}^{K-1}A_k^c\right)\le\sum_{k=0}^{K-1}\mathbb{P}\left(A_k^c\cap\{\hat{x}_k\in B_r(x_*)\}\right)\le\delta, P ( k = 0 ⋃ K − 1 A k c ) ≤ k = 0 ∑ K − 1 P ( A k c ∩ { x ^ k ∈ B r ( x ∗ )} ) ≤ δ ,
which proves (102) .
Finally, we can use the preceding result to get an explicit bound on the sample complexity to achieve a given accuracy with high probability.
Corollary 26 (Explicit sample complexity) Let f f f satisfy Assumption 2, including (SF3), let r > 0 r>0 r > 0 and ϵ ∈ ( 0 , 2 r ] \epsilon\in(0,2r] ϵ ∈ ( 0 , 2 r ] , and set K 0 ( r , ϵ ) : = max { 0 , ⌈ log 2 ( 2 r / ϵ ) ⌉ } K_0(r,\epsilon):=\max\{0,\lceil\log_2(2r/\epsilon)\rceil\} K 0 ( r , ϵ ) := max { 0 , ⌈ log 2 ( 2 r / ϵ )⌉} . Let
α ≥ α f s S ( r , ϵ ) : = max { α c S ( r , 1 2 ) , α b S ( r 2 ) , α b S ( ϵ 8 ) , α v S ( r , ϵ 8 ) , α ζ S ( r ) } , (103) \alpha\ge\alpha_{\mathrm{fs}}^{\mathrm{S}}(r,\epsilon):=\max\left\{\alpha_c^{\mathrm{S}}\left(r,\frac{1}{2}\right),\alpha_b^{\mathrm{S}}\left(\frac{r}{2}\right),\alpha_b^{\mathrm{S}}\left(\frac{\epsilon}{8}\right),\alpha_v^{\mathrm{S}}\left(r,\frac{\epsilon}{8}\right),\alpha_\zeta^{\mathrm{S}}(r)\right\}, \tag*{(103)} α ≥ α fs S ( r , ϵ ) := max { α c S ( r , 2 1 ) , α b S ( 2 r ) , α b S ( 8 ϵ ) , α v S ( r , 8 ϵ ) , α ζ S ( r ) } , (103) and let K ∈ N + K\in\mathbb{N}_+ K ∈ N + , δ ∈ ( 0 , 1 ) \delta\in(0,1) δ ∈ ( 0 , 1 ) and N N N as in (101) . Then for all x ^ 0 ∈ B r ( x ∗ ) \hat{x}_0\in B_r(x_*) x ^ 0 ∈ B r ( x ∗ ) ,
P ( ∥ x ^ k − x ∗ ∥ 2 ≤ ϵ for all k with K 0 ( r , ϵ ) ≤ k ≤ K ) ≥ 1 − δ . \mathbb{P}\left(\lVert\hat{x}_k-x_*\rVert_2\le\epsilon\ \text{for all }k\text{ with }K_0(r,\epsilon)\le k\le K\right)\ge1-\delta. P ( ∥ x ^ k − x ∗ ∥ 2 ≤ ϵ for all k with K 0 ( r , ϵ ) ≤ k ≤ K ) ≥ 1 − δ .
Proof. Apply Proposition 25 with ρ = 1 2 \rho=\frac{1}{2} ρ = 2 1 and β = ϵ 8 \beta=\frac{\epsilon}{8} β = 8 ϵ . Then 2 β = ϵ 4 ≤ r 2 = ( 1 − ρ ) r 2\beta=\frac{\epsilon}{4}\le\frac{r}{2}=(1-\rho)r 2 β = 4 ϵ ≤ 2 r = ( 1 − ρ ) r , the threshold on α \alpha α is (100) , and 2 β 1 − ρ = ϵ 2 \frac{2\beta}{1-\rho}=\frac{\epsilon}{2} 1 − ρ 2 β = 2 ϵ . For k ≥ K 0 ( r , ϵ ) k\ge K_0(r,\epsilon) k ≥ K 0 ( r , ϵ ) , we have 2 − k r ≤ ϵ 2 2^{-k}r\le\frac{\epsilon}{2} 2 − k r ≤ 2 ϵ , and hence ∥ x ^ k − x ∗ ∥ 2 ≤ ϵ \lVert\hat{x}_k-x_*\rVert_2\le\epsilon ∥ x ^ k − x ∗ ∥ 2 ≤ ϵ on the event in (102) .
Explicit Sample Complexity Bounds
Finally, a significant advantage of working with the simplified Assumption 2 is the possibility to derive explicit sample complexity bounds. The following result describes the necessary estimates for this.
Lemma 27 (Order of the explicit thresholds). Let f f f satisfy (SF1) and (SF2) of Assumption 2, and let r > 0 r>0 r > 0 and ϵ ∈ ( 0 , 2 r ] \epsilon\in(0,2r] ϵ ∈ ( 0 , 2 r ] . With log + : = max { 0 , log } \log_{+}:=\max\{0,\log\} log + := max { 0 , log } ,
α m f S ( r , ϵ ) ≤ 32 d m ϵ 2 + 4 m σ min ( Σ ) 2 + 4 ( r + R ) 2 Δ σ min ( Σ ) 2 + 2 Δ [ d log + ( d + 2 ) σ max ( Σ ) 2 s 2 + log ( 4 κ Σ ) + log + 32 d σ max ( Σ ) 2 ϵ 2 ] . (104) \alpha_{\mathrm{mf}}^{\mathrm{S}}(r,\epsilon)\leq\frac{32d}{m\epsilon^{2}}+\frac{4}{m\sigma_{\min}(\Sigma)^{2}}+\frac{4(r+R)^{2}}{\Delta\sigma_{\min}(\Sigma)^{2}}+\frac{2}{\Delta}\left[d\log_{+}\frac{(d+2)\sigma_{\max}(\Sigma)^{2}}{s^{2}}+\log(4\kappa_{\Sigma})+\log_{+}\frac{32d\sigma_{\max}(\Sigma)^{2}}{\epsilon^{2}}\right]. \tag*{(104)} α mf S ( r , ϵ ) ≤ m ϵ 2 32 d + m σ m i n ( Σ ) 2 4 + Δ σ m i n ( Σ ) 2 4 ( r + R ) 2 + Δ 2 [ d log + s 2 ( d + 2 ) σ m a x ( Σ ) 2 + log ( 4 κ Σ ) + log + ϵ 2 32 d σ m a x ( Σ ) 2 ] . (104) If f f f also satisfies (SF3), then
α f s S ( r , ϵ ) ≤ 384 d m ϵ 2 + 4 m σ min ( Σ ) 2 + 8 ( r + R ) 2 Δ σ min ( Σ ) 2 + 1 Δ [ 3 d log + ( d + 2 ) σ max ( Σ ) 2 s 2 + 2 log ( 4 κ Σ ) + 4 log + 256 d σ max ( Σ ) 2 ϵ 2 + 5 d ] . (105) \alpha_{\mathrm{fs}}^{\mathrm{S}}(r,\epsilon)\leq\frac{384d}{m\epsilon^{2}}+\frac{4}{m\sigma_{\min}(\Sigma)^{2}}+\frac{8(r+R)^{2}}{\Delta\sigma_{\min}(\Sigma)^{2}}+\frac{1}{\Delta}\left[3d\log_{+}\frac{(d+2)\sigma_{\max}(\Sigma)^{2}}{s^{2}}+2\log(4\kappa_{\Sigma})+4\log_{+}\frac{256d\sigma_{\max}(\Sigma)^{2}}{\epsilon^{2}}+5d\right]. \tag*{(105)} α fs S ( r , ϵ ) ≤ m ϵ 2 384 d + m σ m i n ( Σ ) 2 4 + Δ σ m i n ( Σ ) 2 8 ( r + R ) 2 + Δ 1 [ 3 d log + s 2 ( d + 2 ) σ m a x ( Σ ) 2 + 2 log ( 4 κ Σ ) + 4 log + ϵ 2 256 d σ m a x ( Σ ) 2 + 5 d ] . (105) In particular, both thresholds are of the order (14) .
Proof. Preliminaries. For t ≥ 0 t\geq0 t ≥ 0 , we have Γ ( t + 1 ) ≤ ( t + 1 ) t \Gamma(t+1)\leq(t+1)^t Γ ( t + 1 ) ≤ ( t + 1 ) t : for t ∈ [ 0 , 1 ] t\in[0,1] t ∈ [ 0 , 1 ] , Γ ( t + 1 ) ≤ 1 ≤ ( t + 1 ) t \Gamma(t+1)\leq1\leq(t+1)^t Γ ( t + 1 ) ≤ 1 ≤ ( t + 1 ) t , and for t > 1 t>1 t > 1 with n : = ⌈ t ⌉ − 1 ≥ 1 n:=\lceil t\rceil-1\geq1 n := ⌈ t ⌉ − 1 ≥ 1 , Γ ( t + 1 ) = t ( t − 1 ) ⋯ ( t − n + 1 ) Γ ( t − n + 1 ) ≤ t n ≤ ( t + 1 ) t \Gamma(t+1)=t(t-1)\cdots(t-n+1)\Gamma(t-n+1)\leq t^n\leq(t+1)^t Γ ( t + 1 ) = t ( t − 1 ) ⋯ ( t − n + 1 ) Γ ( t − n + 1 ) ≤ t n ≤ ( t + 1 ) t , since t − n + 1 ∈ ( 1 , 2 ] t-n+1\in(1,2] t − n + 1 ∈ ( 1 , 2 ] . With t = d 2 t=\frac{d}{2} t = 2 d , this gives D s ≤ ( ( d + 2 ) σ max ( Σ ) 2 / s 2 ) d / 2 D_s\leq\left((d+2)\sigma_{\max}(\Sigma)^2/s^2\right)^{d/2} D s ≤ ( ( d + 2 ) σ m a x ( Σ ) 2 / s 2 ) d /2 , i.e. log + D s ≤ d 2 log + ( d + 2 ) σ max ( Σ ) 2 s 2 \log_{+}D_s\leq\frac{d}{2}\log_{+}\frac{(d+2)\sigma_{\max}(\Sigma)^2}{s^2} log + D s ≤ 2 d log + s 2 ( d + 2 ) σ m a x ( Σ ) 2 . Moreover, both entries of α b S ( β ) \alpha_{\mathrm{b}}^{\mathrm{S}}(\beta) α b S ( β ) are nonincreasing in β \beta β , and log ( 1 + y ) ≤ y \log(1+y)\leq y log ( 1 + y ) ≤ y for y ≥ 0 y\geq0 y ≥ 0 . Finally, max of finitely many numbers is at most the sum of upper bounds for their positive parts.
Mean-field threshold. Since ϵ ≤ 2 r \epsilon\leq2r ϵ ≤ 2 r , we have r 2 ≥ ϵ 4 \frac{r}{2}\geq\frac{\epsilon}{4} 2 r ≥ 4 ϵ , so α b S ( r 2 ) ≤ α b S ( ϵ 4 ) \alpha_{\mathrm{b}}^{\mathrm{S}}\left(\frac{r}{2}\right)\leq\alpha_{\mathrm{b}}^{\mathrm{S}}\left(\frac{\epsilon}{4}\right) α b S ( 2 r ) ≤ α b S ( 4 ϵ ) and α m f S ( r , ϵ ) = max { α c S ( r , 1 2 ) , α b S ( ϵ 4 ) } \alpha_{\mathrm{mf}}^{\mathrm{S}}(r,\epsilon)=\max\left\{\alpha_{\mathrm{c}}^{\mathrm{S}}\left(r,\frac{1}{2}\right),\alpha_{\mathrm{b}}^{\mathrm{S}}\left(\frac{\epsilon}{4}\right)\right\} α mf S ( r , ϵ ) = max { α c S ( r , 2 1 ) , α b S ( 4 ϵ ) } . The entries of α c S ( r , 1 2 ) \alpha_{\mathrm{c}}^{\mathrm{S}}\left(r,\frac{1}{2}\right) α c S ( r , 2 1 ) are 4 m σ min ( Σ ) 2 \frac{4}{m\sigma_{\min}(\Sigma)^2} m σ m i n ( Σ ) 2 4 and
( r + s ) 2 Δ σ min ( Σ ) 2 + 2 Δ [ log ( 4 κ Σ ) + log D s + log ( 1 + ( r + R ) 2 σ max ( Σ ) 2 ) ] ≤ 3 ( r + R ) 2 Δ σ min ( Σ ) 2 + 2 Δ [ log ( 4 κ Σ ) + log + D s ] , \begin{aligned}
\frac{(r+s)^2}{\Delta\sigma_{\min}(\Sigma)^2}+\frac{2}{\Delta}\left[\log(4\kappa_{\Sigma})+\log D_s+\log\left(1+\frac{(r+R)^2}{\sigma_{\max}(\Sigma)^2}\right)\right] \\
&\leq\frac{3(r+R)^2}{\Delta\sigma_{\min}(\Sigma)^2}+\frac{2}{\Delta}\left[\log(4\kappa_{\Sigma})+\log_{+}D_s\right],
\end{aligned} Δ σ m i n ( Σ ) 2 ( r + s ) 2 + Δ 2 [ log ( 4 κ Σ ) + log D s + log ( 1 + σ m a x ( Σ ) 2 ( r + R ) 2 ) ] ≤ Δ σ m i n ( Σ ) 2 3 ( r + R ) 2 + Δ 2 [ log ( 4 κ Σ ) + log + D s ] ,
where we used s ≤ R s\leq R s ≤ R and σ min ( Σ ) ≤ σ max ( Σ ) \sigma_{\min}(\Sigma)\leq\sigma_{\max}(\Sigma) σ m i n ( Σ ) ≤ σ m a x ( Σ ) . The entries of α b S ( ϵ 4 ) \alpha_{\mathrm{b}}^{\mathrm{S}}\left(\frac{\epsilon}{4}\right) α b S ( 4 ϵ ) are 32 d m ϵ 2 \frac{32d}{m\epsilon^2} m ϵ 2 32 d and
s 2 Δ σ min ( Σ ) 2 + 2 Δ [ log D s + log 32 d σ max ( Σ ) 2 ϵ 2 ] ≤ ( r + R ) 2 Δ σ min ( Σ ) 2 + 2 Δ [ log + D s + log + 32 d σ max ( Σ ) 2 ϵ 2 ] . \begin{aligned}
\frac{s^2}{\Delta\sigma_{\min}(\Sigma)^2}+\frac{2}{\Delta}\left[\log D_s+\log\frac{32d\sigma_{\max}(\Sigma)^2}{\epsilon^2}\right] \\
&\leq\frac{(r+R)^2}{\Delta\sigma_{\min}(\Sigma)^2}+\frac{2}{\Delta}\left[\log_{+}D_s+\log_{+}\frac{32d\sigma_{\max}(\Sigma)^2}{\epsilon^2}\right].
\end{aligned} Δ σ m i n ( Σ ) 2 s 2 + Δ 2 [ log D s + log ϵ 2 32 d σ m a x ( Σ ) 2 ] ≤ Δ σ m i n ( Σ ) 2 ( r + R ) 2 + Δ 2 [ log + D s + log + ϵ 2 32 d σ m a x ( Σ ) 2 ] .
Adding these bounds and using the bound on log + D s \log_{+}D_s log + D s gives (104) .
Finite-sample threshold. As above, α b S ( r 2 ) ≤ α b S ( ϵ 8 ) \alpha_{\mathrm{b}}^{\mathrm{S}}\left(\frac{r}{2}\right)\leq\alpha_{\mathrm{b}}^{\mathrm{S}}\left(\frac{\epsilon}{8}\right) α b S ( 2 r ) ≤ α b S ( 8 ϵ ) . The entries of α c S ( r , 1 2 ) \alpha_{\mathrm{c}}^{\mathrm{S}}\left(r,\frac{1}{2}\right) α c S ( r , 2 1 ) are bounded as above. The entries of α b S ( ϵ 8 ) \alpha_{\mathrm{b}}^{\mathrm{S}}\left(\frac{\epsilon}{8}\right) α b S ( 8 ϵ ) are 128 d m ϵ 2 \frac{128d}{m\epsilon^2} m ϵ 2 128 d and at most ( r + R ) 2 Δ σ min ( Σ ) 2 + 2 Δ [ log + D s + log + 128 d σ max ( Σ ) 2 ϵ 2 ] \frac{(r+R)^2}{\Delta\sigma_{\min}(\Sigma)^2}+\frac{2}{\Delta}\left[\log_{+}D_s+\log_{+}\frac{128d\sigma_{\max}(\Sigma)^2}{\epsilon^2}\right] Δ σ m i n ( Σ ) 2 ( r + R ) 2 + Δ 2 [ log + D s + log + ϵ 2 128 d σ m a x ( Σ ) 2 ] . The entries of α ν S ( r , ϵ ξ ) \alpha_{\nu}^{\mathrm{S}}(r,\frac{\epsilon}{\xi}) α ν S ( r , ξ ϵ ) are 256 d m ϵ 2 \frac{256d}{m\epsilon^{2}} m ϵ 2 256 d and, writing d σ max ( Σ ) 2 + ( r + R ) 2 = d σ max ( Σ ) 2 ( 1 + ( r + R ) 2 d σ max ( Σ ) 2 ) d\sigma_{\max}(\Sigma)^{2}+(r+R)^{2}=d\sigma_{\max}(\Sigma)^{2}\left(1+\frac{(r+R)^{2}}{d\sigma_{\max}(\Sigma)^{2}}\right) d σ m a x ( Σ ) 2 + ( r + R ) 2 = d σ m a x ( Σ ) 2 ( 1 + d σ m a x ( Σ ) 2 ( r + R ) 2 ) ,
( r + s ) 2 Δ σ min ( Σ ) 2 + 2 Δ [ log D s + log 256 ( d σ max ( Σ ) 2 + ( r + R ) 2 ) ϵ 2 ] ≤ 3 ( r + R ) 2 Δ σ min ( Σ ) 2 + 2 Δ [ log + D s + log + 256 d σ max ( Σ ) 2 ϵ 2 ] . \frac{(r+s)^{2}}{\Delta\sigma_{\min}(\Sigma)^{2}}+\frac{2}{\Delta}\left[\log D_s+\log\frac{256(d\sigma_{\max}(\Sigma)^{2}+(r+R)^{2})}{\epsilon^{2}}\right]
\leq\frac{3(r+R)^{2}}{\Delta\sigma_{\min}(\Sigma)^{2}}+\frac{2}{\Delta}\left[\log_{+}D_s+\log_{+}\frac{256d\sigma_{\max}(\Sigma)^{2}}{\epsilon^{2}}\right]. Δ σ m i n ( Σ ) 2 ( r + s ) 2 + Δ 2 [ log D s + log ϵ 2 256 ( d σ m a x ( Σ ) 2 + ( r + R ) 2 ) ] ≤ Δ σ m i n ( Σ ) 2 3 ( r + R ) 2 + Δ 2 [ log + D s + log + ϵ 2 256 d σ m a x ( Σ ) 2 ] .
For α ζ S ( r ) \alpha_{\zeta}^{\mathrm{S}}(r) α ζ S ( r ) , note that Δ ≤ M 2 R 2 \Delta\leq\frac{M}{2}R^{2} Δ ≤ 2 M R 2 : for ∥ y − x ∗ ∥ 2 = R \lVert y-x_{*}\rVert_{2}=R ∥ y − x ∗ ∥ 2 = R , the points x ∗ + ( 1 + 1 n ) ( y − x ∗ ) x_{*}+\left(1+\frac{1}{n}\right)(y-x_{*}) x ∗ + ( 1 + n 1 ) ( y − x ∗ ) lie outside B R ( x ∗ ) B_{R}(x_{*}) B R ( x ∗ ) , so (SF2) and continuity give f ( y ) − f ∗ ≥ Δ f(y)-f_{*}\geq\Delta f ( y ) − f ∗ ≥ Δ , while (SF3) gives f ( y ) − f ∗ ≤ M 2 R 2 f(y)-f_{*}\leq\frac{M}{2}R^{2} f ( y ) − f ∗ ≤ 2 M R 2 . Hence
( 4 d + 8 ) ln 2 M R 2 ≤ ( 2 d + 4 ) ln 2 M R 2 ≤ 5 d Δ \frac{(4d+8)\ln2}{MR^{2}}\leq\frac{(2d+4)\ln2}{MR^{2}}\leq\frac{5d}{\Delta} M R 2 ( 4 d + 8 ) ln 2 ≤ M R 2 ( 2 d + 4 ) ln 2 ≤ Δ 5 d
and
2 r M R σ min ( Σ ) 2 ≤ r R Δ σ min ( Σ ) 2 ≤ ( r + R ) 2 4 Δ σ min ( Σ ) 2 . \frac{2r}{MR\sigma_{\min}(\Sigma)^{2}}\leq\frac{rR}{\Delta\sigma_{\min}(\Sigma)^{2}}\leq\frac{(r+R)^{2}}{4\Delta\sigma_{\min}(\Sigma)^{2}}. M R σ m i n ( Σ ) 2 2 r ≤ Δ σ m i n ( Σ ) 2 r R ≤ 4Δ σ m i n ( Σ ) 2 ( r + R ) 2 .
Adding all bounds, using log + 128 d σ max ( Σ ) 2 ϵ 2 ≤ log + 256 d σ max ( Σ ) 2 ϵ 2 \log_{+}\frac{128d\sigma_{\max}(\Sigma)^{2}}{\epsilon^{2}}\leq\log_{+}\frac{256d\sigma_{\max}(\Sigma)^{2}}{\epsilon^{2}} log + ϵ 2 128 d σ m a x ( Σ ) 2 ≤ log + ϵ 2 256 d σ m a x ( Σ ) 2 and the bound on log + D s \log_{+}D_s log + D s , gives (105) .
Order. In (104) and (105) , the terms in square brackets are at most d d d times a factor that is polylogarithmic in d d d , κ Σ \kappa_{\Sigma} κ Σ , σ max ( Σ ) / ϵ \sigma_{\max}(\Sigma)/\epsilon σ m a x ( Σ ) / ϵ and σ max ( Σ ) / s \sigma_{\max}(\Sigma)/s σ m a x ( Σ ) / s , since d ≥ 1 d\geq1 d ≥ 1 . This gives (14) .
We can finally provide the missing proofs of the simplified, concrete results from the main text.
Proof [Proof of Corollary 4] With α m f S ( r , ϵ ) \alpha_{\mathrm{mf}}^{\mathrm{S}}(r,\epsilon) α mf S ( r , ϵ ) from (94) , the claim on the iterates is Corollary 23, and the order (14) is Lemma 27. The radius s s s is the one of Lemma 19(iv), which can be any radius with f ≤ f ∗ + Δ 2 f\leq f_{*}+\frac{\Delta}{2} f ≤ f ∗ + 2 Δ on B s ( x ∗ ) B_s(x_{*}) B s ( x ∗ ) .
Proof [Proof of Corollary 7] With α f s S ( r , ϵ ) \alpha_{\mathrm{fs}}^{\mathrm{S}}(r,\epsilon) α fs S ( r , ϵ ) from (103) , the claim is Corollary 26, since (21) is (101) , and the order is Lemma 27.
Quadratic objectives
Finally, we consider a particularly simple class of functions fulfilling Assumption 2, the quadratic functions.
Lemma 28 (Quadratic objectives) Let f ( x ) = f ∗ + 1 2 ( x − x ∗ ) T H ( x − x ∗ ) f(x)=f_{*}+\frac{1}{2}(x-x_{*})^{\mathsf{T}}H(x-x_{*}) f ( x ) = f ∗ + 2 1 ( x − x ∗ ) T H ( x − x ∗ ) with f ∗ ∈ R f_{*}\in\mathbb{R} f ∗ ∈ R , x ∗ ∈ R d x_{*}\in\mathbb{R}^{d} x ∗ ∈ R d and H ∈ R d × d H\in\mathbb{R}^{d\times d} H ∈ R d × d symmetric positive definite, and let λ − : = λ min ( H ) \lambda_{-}:=\lambda_{\min}(H) λ − := λ m i n ( H ) and λ + : = λ max ( H ) \lambda_{+}:=\lambda_{\max}(H) λ + := λ m a x ( H ) .
(i) For every R > 0 R>0 R > 0 , f f f satisfies Assumption 2 with m = λ − m=\lambda_{-} m = λ − , M = λ + M=\lambda_{+} M = λ + and Δ = λ − 2 R 2 \Delta=\frac{\lambda_{-}}{2}R^{2} Δ = 2 λ − R 2 , and s = R λ − / ( 2 λ + ) s=R\sqrt{\lambda_{-}/(2\lambda_{+})} s = R λ − / ( 2 λ + ) is admissible in Lemma 19(iv).
(ii) Let r > 0 r>0 r > 0 and ϵ ∈ ( 0 , 2 r ] \epsilon\in(0,2r] ϵ ∈ ( 0 , 2 r ] . For all α > 32 d λ − ϵ 2 + 12 λ − σ min ( Σ ) 2 \alpha>\frac{32d}{\lambda_{-}\epsilon^{2}}+\frac{12}{\lambda_{-}\sigma_{\min}(\Sigma)^{2}} α > λ − ϵ 2 32 d + λ − σ m i n ( Σ ) 2 12 and x 0 ∈ B r ( x ∗ ) x_{0}\in B_{r}(x_{*}) x 0 ∈ B r ( x ∗ ) , the mean-field iterates satisfy ∥ x k − x ∗ ∥ 2 ≤ ϵ \lVert x_{k}-x_{*}\rVert_{2}\leq\epsilon ∥ x k − x ∗ ∥ 2 ≤ ϵ for all k ≥ K 0 ( r , ϵ ) k\geq K_{0}(r,\epsilon) k ≥ K 0 ( r , ϵ ) .
(iii) Let r > 0 r>0 r > 0 and ϵ ∈ ( 0 , 2 r ] \epsilon\in(0,2r] ϵ ∈ ( 0 , 2 r ] . For all α > 384 d λ − ϵ 2 + 20 λ − σ min ( Σ ) 2 \alpha>\frac{384d}{\lambda_{-}\epsilon^{2}}+\frac{20}{\lambda_{-}\sigma_{\min}(\Sigma)^{2}} α > λ − ϵ 2 384 d + λ − σ m i n ( Σ ) 2 20 , K ∈ N + K\in\mathbb{N}_{+} K ∈ N + , δ ∈ ( 0 , 1 ) \delta\in(0,1) δ ∈ ( 0 , 1 ) , N ∈ N + N\in\mathbb{N}_{+} N ∈ N + with
N ≥ 12 K δ ( 1 + α λ + σ max ( Σ ) 2 ) d / 2 exp ( r 2 2 σ min ( Σ ) 2 ) , N\geq\frac{12K}{\delta}\left(1+\alpha\lambda_{+}\sigma_{\max}(\Sigma)^{2}\right)^{d/2}\exp\left(\frac{r^{2}}{2\sigma_{\min}(\Sigma)^{2}}\right), N ≥ δ 12 K ( 1 + α λ + σ m a x ( Σ ) 2 ) d /2 exp ( 2 σ m i n ( Σ ) 2 r 2 ) , and x ^ 0 ∈ B r ( x ∗ ) \hat{x}_{0}\in B_{r}(x_{*}) x ^ 0 ∈ B r ( x ∗ ) , the MPPI iterates satisfy
P ( ∥ x ^ k − x ∗ ∥ 2 ≤ ϵ for all k with K 0 ( r , ϵ ) ≤ k ≤ K ) ≥ 1 − δ . \mathbb{P}\left(\lVert\hat{x}_{k}-x_{*}\rVert_{2}\leq\epsilon\text{ for all }k\text{ with }K_{0}(r,\epsilon)\leq k\leq K\right)\geq1-\delta. P ( ∥ x ^ k − x ∗ ∥ 2 ≤ ϵ for all k with K 0 ( r , ϵ ) ≤ k ≤ K ) ≥ 1 − δ .
Proof Ad (i). f f f is continuous and attains its global minimum f ∗ f_\ast f ∗ at x ∗ x_\ast x ∗ . Since H − λ − I d ⪰ 0 H-\lambda_- I_d \succeq0 H − λ − I d ⪰ 0 , the function y ↦ f ( y ) − λ − 2 ∥ y − x ∗ ∥ 2 = f ∗ + 1 2 ( y − x ∗ ) ⊤ ( H − λ − I d ) ( y − x ∗ ) y\mapsto f(y)-\frac{\lambda_-}{2}\lVert y-x_\ast\rVert^2=f_\ast+\frac{1}{2}(y-x_\ast)^\top(H-\lambda_-I_d)(y-x_\ast) y ↦ f ( y ) − 2 λ − ∥ y − x ∗ ∥ 2 = f ∗ + 2 1 ( y − x ∗ ) ⊤ ( H − λ − I d ) ( y − x ∗ ) is convex on R d \mathbb{R}^d R d , which gives (SF1) with m = λ − m=\lambda_- m = λ − . For ∥ y − x ∗ ∥ 2 > R \lVert y-x_\ast\rVert^2>R ∥ y − x ∗ ∥ 2 > R , we have f ( y ) − f ∗ ≥ λ − 2 ∥ y − x ∗ ∥ 2 ≥ λ − 2 R 2 f(y)-f_\ast\ge\frac{\lambda_-}{2}\lVert y-x_\ast\rVert^2\ge\frac{\lambda_-}{2}R^2 f ( y ) − f ∗ ≥ 2 λ − ∥ y − x ∗ ∥ 2 ≥ 2 λ − R 2 , which is (SF2) with Δ = λ − 2 R 2 \Delta=\frac{\lambda_-}{2}R^2 Δ = 2 λ − R 2 . Moreover, f ( y ) − f ∗ ≤ λ + 2 ∥ y − x ∗ ∥ 2 f(y)-f_\ast\le\frac{\lambda_+}{2}\lVert y-x_\ast\rVert^2 f ( y ) − f ∗ ≤ 2 λ + ∥ y − x ∗ ∥ 2 for all y y y , which is (SF3) with M = λ + ≥ m M=\lambda_+\ge m M = λ + ≥ m . Finally, Δ / M = R λ − / ( 2 λ + ) ≤ R \sqrt{\Delta/M}=R\sqrt{\lambda_-/(2\lambda_+)}\le R Δ/ M = R λ − / ( 2 λ + ) ≤ R , so Lemma 19(iv) shows that s = R λ − / ( 2 λ + ) s=R\sqrt{\lambda_-/(2\lambda_+)} s = R λ − / ( 2 λ + ) is admissible.
Ad (ii). For R > 0 R>0 R > 0 , insert m = λ − m=\lambda_- m = λ − , Δ = λ − 2 R 2 \Delta=\frac{\lambda_-}{2}R^2 Δ = 2 λ − R 2 and s = R λ − / ( 2 λ + ) s=R\sqrt{\lambda_-/(2\lambda_+)} s = R λ − / ( 2 λ + ) from (i) into (104) , and denote the resulting right-hand side by B m f ( R ) B_{\mathrm{mf}}(R) B mf ( R ) . Then 4 ( r + R ) 2 Δ σ min ( Σ ) 2 = 8 ( r + R ) 2 λ − R 2 σ min ( Σ ) 2 \frac{4(r+R)^2}{\Delta\sigma_{\min}(\Sigma)^2}=\frac{8(r+R)^2}{\lambda_-R^2\sigma_{\min}(\Sigma)^2} Δ σ m i n ( Σ ) 2 4 ( r + R ) 2 = λ − R 2 σ m i n ( Σ ) 2 8 ( r + R ) 2 , and the term in square brackets is multiplied by 2 Δ = 4 λ − R 2 \frac{2}{\Delta}=\frac{4}{\lambda_-R^2} Δ 2 = λ − R 2 4 and is nonincreasing in R R R , since s s s is increasing in R R R . Hence B m f ( R ) B_{\mathrm{mf}}(R) B mf ( R ) is nonincreasing in R R R , and
lim R → ∞ B m f ( R ) = 32 d λ − ϵ 2 + 4 λ − σ min ( Σ ) 2 + 8 λ − σ min ( Σ ) 2 = 32 d λ − ϵ 2 + 12 λ − σ min ( Σ ) 2 . \lim_{R\to\infty}B_{\mathrm{mf}}(R)=\frac{32d}{\lambda_-\epsilon^2}+\frac{4}{\lambda_-\sigma_{\min}(\Sigma)^2}+\frac{8}{\lambda_-\sigma_{\min}(\Sigma)^2}=\frac{32d}{\lambda_-\epsilon^2}+\frac{12}{\lambda_-\sigma_{\min}(\Sigma)^2}. R → ∞ lim B mf ( R ) = λ − ϵ 2 32 d + λ − σ m i n ( Σ ) 2 4 + λ − σ m i n ( Σ ) 2 8 = λ − ϵ 2 32 d + λ − σ m i n ( Σ ) 2 12 .
Let α \alpha α be strictly larger than this limit. Then there is R > 0 R>0 R > 0 with α ≥ B m f ( R ) \alpha\ge B_{\mathrm{mf}}(R) α ≥ B mf ( R ) , and Lemma 27 , applied with this R R R , gives α ≥ α m f S ( r , ϵ ) \alpha\ge\alpha_{\mathrm{mf}}^S(r,\epsilon) α ≥ α mf S ( r , ϵ ) . Corollary 23 yields the claim.
Ad (iii). Analogously, insert (i) into (105) and denote the right-hand side by B f s ( R ) B_{\mathrm{fs}}(R) B fs ( R ) . Then 8 ( r + R ) 2 Δ σ min ( Σ ) 2 = 16 ( r + R ) 2 λ − R 2 σ min ( Σ ) 2 \frac{8(r+R)^2}{\Delta\sigma_{\min}(\Sigma)^2}=\frac{16(r+R)^2}{\lambda_-R^2\sigma_{\min}(\Sigma)^2} Δ σ m i n ( Σ ) 2 8 ( r + R ) 2 = λ − R 2 σ m i n ( Σ ) 2 16 ( r + R ) 2 , the term in square brackets is multiplied by 1 Δ = 2 λ − R 2 \frac{1}{\Delta}=\frac{2}{\lambda_-R^2} Δ 1 = λ − R 2 2 and is nonincreasing in R R R , so B f s ( R ) B_{\mathrm{fs}}(R) B fs ( R ) is nonincreasing in R R R and converges to 384 d λ − ϵ 2 + 4 + 16 λ − σ min ( Σ ) 2 \frac{384d}{\lambda_-\epsilon^2}+\frac{4+16}{\lambda_-\sigma_{\min}(\Sigma)^2} λ − ϵ 2 384 d + λ − σ m i n ( Σ ) 2 4 + 16 as R → ∞ R\to\infty R → ∞ . For α \alpha α strictly larger than this limit, there is R > 0 R>0 R > 0 with α ≥ B f s ( R ) ≥ α f s S ( r , ϵ ) \alpha\ge B_{\mathrm{fs}}(R)\ge\alpha_{\mathrm{fs}}^S(r,\epsilon) α ≥ B fs ( R ) ≥ α fs S ( r , ϵ ) by Lemma 27 , and Corollary 26 yields the claim, since the sample size condition (101) with M = λ + M=\lambda_+ M = λ + does not depend on R R R .
Details for Section 5
We recall how the cost (23) is rewritten as the quadratic function (25) . We identify a control sequence a = ( a 0 , … , a T − 1 ) ∈ ( R p ) T a=(a_0,\ldots,a_{T-1})\in(\mathbb{R}^p)^T a = ( a 0 , … , a T − 1 ) ∈ ( R p ) T with the stacked vector ( a 0 ⊤ , … , a T − 1 ⊤ ) ⊤ ∈ R p T (a_0^\top,\ldots,a_{T-1}^\top)^\top\in\mathbb{R}^{pT} ( a 0 ⊤ , … , a T − 1 ⊤ ) ⊤ ∈ R pT , and the state sequence ( s 0 , … , s T − 1 ) (s_0,\ldots,s_{T-1}) ( s 0 , … , s T − 1 ) with the stacked vector s : = ( s 0 ⊤ , … , s T − 1 ⊤ ) ⊤ ∈ R n T s:=(s_0^\top,\ldots,s_{T-1}^\top)^\top\in\mathbb{R}^{nT} s := ( s 0 ⊤ , … , s T − 1 ⊤ ) ⊤ ∈ R n T . Note that the final state s T s_T s T does not enter (23) .
States as an affine function of the controls. Iterating (22) gives the variation-of-constants formula
s t = A t s 0 + ∑ k = 0 t − 1 A t − 1 − k B a k , t = 0 , … , T − 1. s_t=A^t s_0+\sum_{k=0}^{t-1}A^{t-1-k}Ba_k,\qquad t=0,\ldots,T-1. s t = A t s 0 + k = 0 ∑ t − 1 A t − 1 − k B a k , t = 0 , … , T − 1.
In stacked form, this reads s = Φ T s 0 + Γ T a s=\Phi_Ts_0+\Gamma_Ta s = Φ T s 0 + Γ T a with
Φ T : = ( I n A A 2 ⋮ A T − 1 ) ∈ R n T × n , Γ T : = ( 0 0 ⋯ 0 0 B 0 ⋯ 0 0 A B B ⋱ ⋮ ⋮ ⋮ ⋱ ⋱ 0 0 A T − 2 B ⋯ A B B 0 ) ∈ R n T × p T , (106) \Phi_T:=\begin{pmatrix} I_n \\ A \\ A^2 \\ \vdots\\ A^{T-1} \end{pmatrix}\in\mathbb{R}^{nT\times n},\qquad
\Gamma_T:=\begin{pmatrix}
0 & 0 & \cdots& 0 & 0\\
B & 0 & \cdots& 0 & 0\\
AB & B & \ddots& \vdots& \vdots\\
\vdots& \ddots& \ddots& 0 & 0\\
A^{T-2}B & \cdots& AB & B & 0
\end{pmatrix}\in\mathbb{R}^{nT\times pT}, \tag*{(106)} Φ T := I n A A 2 ⋮ A T − 1 ∈ R n T × n , Γ T := 0 B A B ⋮ A T − 2 B 0 0 B ⋱ ⋯ ⋯ ⋯ ⋱ ⋱ A B 0 0 ⋮ 0 B 0 0 ⋮ 0 0 ∈ R n T × pT , (106)
i.e. the ( t , k ) (t,k) ( t , k ) -th block of Γ T \Gamma_T Γ T , t , k = 0 , … , T − 1 t,k=0,\ldots,T-1 t , k = 0 , … , T − 1 , is A t − 1 − k B A^{t-1-k}B A t − 1 − k B if k < t k<t k < t and 0 0 0 otherwise.
The cost as a quadratic form. Let Q ˉ : = I T ⊗ Q ∈ R n T × n T \bar{Q} := I_T \otimes Q \in\mathbb{R}^{nT \times nT} Q ˉ := I T ⊗ Q ∈ R n T × n T , R ˉ : = I T ⊗ R ∈ R p T × p T \bar{R} := I_T \otimes R \in\mathbb{R}^{pT \times pT} R ˉ := I T ⊗ R ∈ R pT × pT and S ˉ : = I T ⊗ S ∈ R n T × p T \bar{S} := I_T \otimes S \in\mathbb{R}^{nT \times pT} S ˉ := I T ⊗ S ∈ R n T × pT be the block diagonal matrices with diagonal blocks Q Q Q , R R R and S S S , respectively. Then (23) becomes
J T ( s 0 , a ) = s ⊤ Q ˉ s + a ⊤ R ˉ a + 2 s ⊤ S ˉ a . J_T(s_0,a) = s^\top\bar{Q}s + a^\top\bar{R}a + 2s^\top\bar{S}a. J T ( s 0 , a ) = s ⊤ Q ˉ s + a ⊤ R ˉ a + 2 s ⊤ S ˉ a .
Inserting s = Φ T s 0 + Γ T a s = \Phi_Ts_0 + \Gamma_Ta s = Φ T s 0 + Γ T a and using the symmetry of Q ˉ \bar{Q} Q ˉ , we get
s ⊤ Q ˉ s = a ⊤ Γ T ⊤ Q ˉ Γ T a + 2 a ⊤ Γ T ⊤ Q ˉ Φ T s 0 + s 0 ⊤ Φ T ⊤ Q ˉ Φ T s 0 , s^\top\bar{Q}s = a^\top\Gamma_T^\top\bar{Q}\Gamma_Ta + 2a^\top\Gamma_T^\top\bar{Q}\Phi_Ts_0 + s_0^\top\Phi_T^\top\bar{Q}\Phi_Ts_0, s ⊤ Q ˉ s = a ⊤ Γ T ⊤ Q ˉ Γ T a + 2 a ⊤ Γ T ⊤ Q ˉ Φ T s 0 + s 0 ⊤ Φ T ⊤ Q ˉ Φ T s 0 ,
2 s ⊤ S ˉ a = 2 a ⊤ S ˉ ⊤ Γ T a + 2 a ⊤ S ˉ ⊤ Φ T s 0 = a ⊤ ( Γ T ⊤ S ˉ + S ˉ ⊤ Γ T ) a + 2 a ⊤ S ˉ ⊤ Φ T s 0 . 2s^\top\bar{S}a = 2a^\top\bar{S}^\top\Gamma_Ta + 2a^\top\bar{S}^\top\Phi_Ts_0 = a^\top(\Gamma_T^\top\bar{S} + \bar{S}^\top\Gamma_T)a + 2a^\top\bar{S}^\top\Phi_Ts_0. 2 s ⊤ S ˉ a = 2 a ⊤ S ˉ ⊤ Γ T a + 2 a ⊤ S ˉ ⊤ Φ T s 0 = a ⊤ ( Γ T ⊤ S ˉ + S ˉ ⊤ Γ T ) a + 2 a ⊤ S ˉ ⊤ Φ T s 0 .
Collecting the terms that are quadratic in a a a , bilinear in ( a , s 0 ) (a,s_0) ( a , s 0 ) , and quadratic in s 0 s_0 s 0 gives (25) with
H T : = R ˉ + Γ T ⊤ Q ˉ Γ T + Γ T ⊤ S ˉ + S ˉ ⊤ Γ T ∈ R p T × p T , H_T := \bar{R} + \Gamma_T^\top\bar{Q}\Gamma_T + \Gamma_T^\top\bar{S} + \bar{S}^\top\Gamma_T \in\mathbb{R}^{pT \times pT}, H T := R ˉ + Γ T ⊤ Q ˉ Γ T + Γ T ⊤ S ˉ + S ˉ ⊤ Γ T ∈ R pT × pT ,
F T : = ( Γ T ⊤ Q ˉ + S ˉ ⊤ ) Φ T ∈ R p T × n , Y T : = Φ T ⊤ Q ˉ Φ T ∈ R n × n . (107) F_T := \left(\Gamma_T^\top\bar{Q} + \bar{S}^\top\right)\Phi_T \in\mathbb{R}^{pT \times n}, \qquad Y_T := \Phi_T^\top\bar{Q}\Phi_T \in\mathbb{R}^{n \times n}. \tag*{(107)} F T := ( Γ T ⊤ Q ˉ + S ˉ ⊤ ) Φ T ∈ R pT × n , Y T := Φ T ⊤ Q ˉ Φ T ∈ R n × n . (107)
Since Q Q Q and R R R are symmetric, so are H T H_T H T and Y T Y_T Y T .
Minimiser. If H T H_T H T is positive definite, completing the square in (25) , i.e. using
a ⊤ H T a + 2 a ⊤ F T s 0 = ( a + H T − 1 F T s 0 ) ⊤ H T ( a + H T − 1 F T s 0 ) − s 0 ⊤ F T ⊤ H T − 1 F T s 0 , a^\top H_Ta + 2a^\top F_Ts_0 = (a + H_T^{-1}F_Ts_0)^\top H_T(a + H_T^{-1}F_Ts_0) - s_0^\top F_T^\top H_T^{-1}F_Ts_0, a ⊤ H T a + 2 a ⊤ F T s 0 = ( a + H T − 1 F T s 0 ) ⊤ H T ( a + H T − 1 F T s 0 ) − s 0 ⊤ F T ⊤ H T − 1 F T s 0 ,
gives
J T ( s 0 , a ) = J T ∗ ( s 0 ) + ( a − a ∗ ( s 0 ) ) ⊤ H T ( a − a ∗ ( s 0 ) ) for all a ∈ R p T . (108) J_T(s_0,a) = J_T^*(s_0) + (a-a_*(s_0))^\top H_T(a-a_*(s_0)) \qquad\text{for all } a \in\mathbb{R}^{pT}. \tag*{(108)} J T ( s 0 , a ) = J T ∗ ( s 0 ) + ( a − a ∗ ( s 0 ) ) ⊤ H T ( a − a ∗ ( s 0 )) for all a ∈ R pT . (108)
with a ∗ ( s 0 ) = − H T − 1 F T s 0 a_*(s_0) = -H_T^{-1}F_Ts_0 a ∗ ( s 0 ) = − H T − 1 F T s 0 and J T ∗ ( s 0 ) = s 0 ⊤ ( Y T − F T ⊤ H T − 1 F T ) s 0 J_T^*(s_0) = s_0^\top(Y_T-F_T^\top H_T^{-1}F_T)s_0 J T ∗ ( s 0 ) = s 0 ⊤ ( Y T − F T ⊤ H T − 1 F T ) s 0 . In particular, a ∗ ( s 0 ) a_*(s_0) a ∗ ( s 0 ) is the unique solution of (24) and J T ∗ ( s 0 ) J_T^*(s_0) J T ∗ ( s 0 ) is its optimal value. If H T H_T H T is not positive definite, then there is v ≠ 0 v \ne0 v = 0 with v ⊤ H T v ≤ 0 v^\top H_Tv \le0 v ⊤ H T v ≤ 0 , and λ ↦ J T ( s 0 , a + λ v ) \lambda\mapsto J_T(s_0,a+\lambda v) λ ↦ J T ( s 0 , a + λ v ) is a polynomial of degree at most two with nonpositive leading coefficient, so a minimiser of (24) cannot be unique.
Recall from the main text that our MPPI results depend on spectral properties of H T H_T H T , which are in turn determined by those of Q Q Q , R R R , S S S , and A A A , B B B . For convenience, we provide some elementary bounds that are well-known in the control theory literature.
Lemma 29 Let W : = ( Q S S ⊤ R ) ∈ R ( n + p ) × ( n + p ) W := \begin{pmatrix} Q & S \\ S^\top& R \end{pmatrix} \in\mathbb{R}^{(n+p)\times(n+p)} W := ( Q S ⊤ S R ) ∈ R ( n + p ) × ( n + p ) be positive definite. Then H T H_T H T is positive definite, and
λ min ( H T ) ≥ λ min ( W ) , \lambda_{\min}(H_T) \ge\lambda_{\min}(W), λ m i n ( H T ) ≥ λ m i n ( W ) , λ max ( H T ) ≤ λ max ( W ) ( 1 + ∥ Γ T ∥ 2 ) ≤ λ max ( W ) ( 1 + ( ∑ j = 0 T − 2 ∥ A j B ∥ ) 2 ) . \lambda_{\max}(H_T) \le\lambda_{\max}(W)(1+\lVert\Gamma_T\rVert^2) \le\lambda_{\max}(W)\left(1+\left(\sum_{j=0}^{T-2}\lVert A^jB\rVert\right)^2\right). λ m a x ( H T ) ≤ λ m a x ( W ) ( 1 + ∥ Γ T ∥ 2 ) ≤ λ m a x ( W ) 1 + ( j = 0 ∑ T − 2 ∥ A j B ∥ ) 2 . In particular, if A A A is Schur stable, then λ max ( H T ) / λ min ( H T ) \lambda_{\max}(H_T)/\lambda_{\min}(H_T) λ m a x ( H T ) / λ m i n ( H T ) is bounded uniformly in T T T .
Proof. Let a ∈ R p T a \in\mathbb{R}^{pT} a ∈ R pT , let s t = s ( t ; 0 , a ) s_t = s(t;0,a) s t = s ( t ; 0 , a ) , so that s = Γ T a s = \Gamma_Ta s = Γ T a , and let z t : = ( s t ⊤ , a t ⊤ ) ⊤ z_t := (s_t^\top,a_t^\top)^\top z t := ( s t ⊤ , a t ⊤ ) ⊤ . By (25) with s 0 = 0 s_0=0 s 0 = 0 and (23) ,
a ⊤ H T a = J T ( 0 , a ) = ∑ t = 0 T − 1 z t ⊤ W z t . a^\top H_Ta = J_T(0,a) = \sum_{t=0}^{T-1} z_t^\top Wz_t. a ⊤ H T a = J T ( 0 , a ) = t = 0 ∑ T − 1 z t ⊤ W z t .
Hence a ⊤ H T a ≥ λ min ( W ) ∑ t = 0 T − 1 ( ∥ s t ∥ 2 + ∥ a t ∥ 2 ) ≥ λ min ( W ) ∥ a ∥ 2 a^\top H_Ta \ge\lambda_{\min}(W)\sum_{t=0}^{T-1}(\|s_t\|^2+\|a_t\|^2) \ge\lambda_{\min}(W)\|a\|^2 a ⊤ H T a ≥ λ m i n ( W ) ∑ t = 0 T − 1 ( ∥ s t ∥ 2 + ∥ a t ∥ 2 ) ≥ λ m i n ( W ) ∥ a ∥ 2 , which gives positive definiteness and the lower bound, and a ⊤ H T a ≤ λ max ( W ) ( ∥ a ∥ 2 + ∥ Γ T a ∥ 2 ) a^\top H_Ta \le\lambda_{\max}(W)(\|a\|^2+\|\Gamma_Ta\|^2) a ⊤ H T a ≤ λ m a x ( W ) ( ∥ a ∥ 2 + ∥ Γ T a ∥ 2 ) , which gives the first upper bound. Let L T ∈ R T × T L_T \in\mathbb{R}^{T\times T} L T ∈ R T × T be the lower shift matrix, i.e. ( L T ) t , k = 1 (L_T)_{t,k}=1 ( L T ) t , k = 1 if t = k + 1 t=k+1 t = k + 1 and 0 0 0 otherwise. Then Γ T = ∑ j = 0 T − 2 L T j + 1 ⊗ A j B \Gamma_T=\sum_{j=0}^{T-2}L_T^{j+1}\otimes A^jB Γ T = ∑ j = 0 T − 2 L T j + 1 ⊗ A j B by (106) , and since ∥ L T j + 1 ∥ ≤ 1 \|L_T^{j+1}\|\le1 ∥ L T j + 1 ∥ ≤ 1 and ∥ X ⊗ Y ∥ = ∥ X ∥ ∥ Y ∥ \|X\otimes Y\|=\|X\|\|Y\| ∥ X ⊗ Y ∥ = ∥ X ∥∥ Y ∥ , we get ∥ Γ T ∥ ≤ ∑ j = 0 T − 2 ∥ A j B ∥ \|\Gamma_T\|\le\sum_{j=0}^{T-2}\|A^jB\| ∥ Γ T ∥ ≤ ∑ j = 0 T − 2 ∥ A j B ∥ . Finally, if A A A is Schur stable, then there are C A ≥ 1 C_A\ge1 C A ≥ 1 and ρ A ∈ ( 0 , 1 ) \rho_A\in(0,1) ρ A ∈ ( 0 , 1 ) with ∥ A j ∥ ≤ C A ρ A j \|A^j\|\le C_A\rho_A^j ∥ A j ∥ ≤ C A ρ A j for all j j j , so ∑ j = 0 T − 2 ∥ A j B ∥ ≤ C A ∥ B ∥ / ( 1 − ρ A ) \sum_{j=0}^{T-2}\|A^jB\|\le C_A\|B\|/(1-\rho_A) ∑ j = 0 T − 2 ∥ A j B ∥ ≤ C A ∥ B ∥/ ( 1 − ρ A ) for all T T T .
Finally, we provide a proof of Corollary 8 .
Proof [Proof of Corollary 8 ] By (108) , f ( a ) = J T ∗ ( s 0 ) + 1 2 ( a − a ∗ ) ⊤ ( 2 H T ) ( a − a ∗ ) f(a)=J_T^*(s_0)+\frac{1}{2}(a-a_*)^\top(2H_T)(a-a_*) f ( a ) = J T ∗ ( s 0 ) + 2 1 ( a − a ∗ ) ⊤ ( 2 H T ) ( a − a ∗ ) , so f f f is the quadratic objective of Lemma 28 with d = p T d=pT d = pT , x ∗ = a ∗ x_*=a_* x ∗ = a ∗ , H = 2 H T H=2H_T H = 2 H T , λ − = 2 λ min ( H T ) \lambda_-=2\lambda_{\min}(H_T) λ − = 2 λ m i n ( H T ) and λ + = 2 λ max ( H T ) \lambda_+=2\lambda_{\max}(H_T) λ + = 2 λ m a x ( H T ) . Moreover, Σ = I T ⊗ Σ c \Sigma=I_T\otimes\Sigma_c Σ = I T ⊗ Σ c has the same eigenvalues as Σ c \Sigma_c Σ c , so σ min ( Σ ) = σ min ( Σ c ) \sigma_{\min}(\Sigma)=\sigma_{\min}(\Sigma_c) σ m i n ( Σ ) = σ m i n ( Σ c ) and σ max ( Σ ) = σ max ( Σ c ) \sigma_{\max}(\Sigma)=\sigma_{\max}(\Sigma_c) σ m a x ( Σ ) = σ m a x ( Σ c ) . Since
32 d λ − ϵ 2 + 12 λ − σ min ( Σ ) 2 = 16 p T λ min ( H T ) ϵ 2 + 6 λ min ( H T ) σ min ( Σ c ) 2 , \frac{32d}{\lambda_-\epsilon^2}+\frac{12}{\lambda_-\sigma_{\min}(\Sigma)^2}=\frac{16pT}{\lambda_{\min}(H_T)\epsilon^2}+\frac{6}{\lambda_{\min}(H_T)\sigma_{\min}(\Sigma_c)^2}, λ − ϵ 2 32 d + λ − σ m i n ( Σ ) 2 12 = λ m i n ( H T ) ϵ 2 16 pT + λ m i n ( H T ) σ m i n ( Σ c ) 2 6 ,
384 d λ − ϵ 2 + 20 λ − σ min ( Σ ) 2 = 192 p T λ min ( H T ) ϵ 2 + 10 λ min ( H T ) σ min ( Σ c ) 2 , \frac{384d}{\lambda_-\epsilon^2}+\frac{20}{\lambda_-\sigma_{\min}(\Sigma)^2}=\frac{192pT}{\lambda_{\min}(H_T)\epsilon^2}+\frac{10}{\lambda_{\min}(H_T)\sigma_{\min}(\Sigma_c)^2}, λ − ϵ 2 384 d + λ − σ m i n ( Σ ) 2 20 = λ m i n ( H T ) ϵ 2 192 pT + λ m i n ( H T ) σ m i n ( Σ c ) 2 10 ,
and α λ + = 2 α λ max ( H T ) \alpha\lambda_+=2\alpha\lambda_{\max}(H_T) α λ + = 2 α λ m a x ( H T ) , Lemma 28 (ii) gives (i) and Lemma 28 (iii) gives (ii).