acceptodds
Under review as a conference paper at ICLR 2027

Gradient Estimation View of Zeroth-Order Optimizers

Abstract

Zeroth-order (derivative-free) methods—including evolution strategies, finite differences, and sphere-sampling estimators—are often viewed as approximating gradients of smoothed objectives. But is this a universal property of ZO methods, or do only specific algorithms admit such a view? We answer this with an update-first reconstruction principle: given a sampler and update weights , we reconstruct an implicit smoothing measure such that the expected update equals the derivative of . The reconstruction equation cleanly separates where the algorithm samples () from which objective it is smoothing (). It covers classical likelihood-ratio (REINFORCE) estimators, finite-difference and sphere rules whose weak derivatives live on boundaries, and previously-unexplained schemes such as random-denominator differences—for which we give the first explicit smoothed objective. Worked examples (Gaussian ES, forward and central FD, sphere sampling, randomized Kiefer–Wolfowitz, and an Epanechnikov-kernel uniform ES) yield closed-form kernels that make each method's induced smoothing visible, comparable, and interpretable.

Primary area:
optimization
Keywords:
zeroth-order optimization, policy gradient methods, score function gradient estimator, gradient estimation, evolution strategies

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_oG68OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper develop a framework to reconstruct the objective smoothed function optimized by a given zeroth-order gradient estimate.

The method consists in identifying the sampling distribution of the perturbations qθq_\theta and the weighting function ψθ\psi_\theta, and integrate the reconstruction equation given by the authors to obtain the smoothing distribution pθp_\theta.

The introduced "reconstruction" methodology unifies several zeroth-order gradient estimators (Gaussian ES, randomized/fixed-denominator perturbations or deterministic finite differences) estimators and allows their comparison in Table 1.

The authors also illustrate the obtained "reconstructed" smoothing kernels in Figure 1 and their implication on the minimized fonction.

Critical strengths

The paper proposes an elegant unification of several zeroth-order gradient estimators under the lens of the objective they minimize.

The paper is clearly written and develops the reconstruction strategy very pedagogically.

Critical weaknesses

Despite the quality of the unified framework, the papers does not leads to any important knowledge breakthrough. No general property emerge from this unification, no innovative kernel is introduced, and the method is mainly applied to known smoothing kernels. The traditional "target-first" construction of gradient estimates is a way to discover new potential zeroth-order gradient estimators, but the impact of the presented "update-first" reconstruction is less clear (see Figure 1).

Moreover, the main point of these ZO gradient estimator is to be used to minimize a given objective function ff. The optimization aspect is almost absent of the paper (beside being evoked in the Discussion section).

Questions for authors
  • Q1) Could this paper be useful to cite and to rely on to refer to existing known bounds of the DFO different estimators? Beharas et al, 2022 "A Theoretical and Empirical Comparison of Gradient Approximations in Derivative-Free Optimization"
  • Q2) Could you be more precise in line 161?
  • Q3) Line 192. Isn't this comment (about the unbiasedness of the gradient estimate relative to the smoothed objective) also true for the estimates of equation (5), cf Nesterov & Spokiony 2017 equation (31).
  • Q4) Line 282. Maybe it's trivial, but could you explain why ∂θjpθ(x)=qθ(x)ψθj(x)\partial_{\theta_j} p_\theta (x) = q_\theta (x) \psi_{\theta_j} (x) requires the weights function to be centered under qθq_\theta?
  • Q5) Out of curiosity, how are the kernels for fixed denominator Gaussian and uniform zeroth order estimators compare with those in Figure 2 (a)?

Comments/Suggestions:

  • It is known from experts of the domain that one must be careful with the rescaling factor dd when it comes to ZO gradient estimate with perturbations sampled uniformly over the (unit) sphere (or one has to sample over the sphere of radius d\sqrt{d}. I would encourage the authors to precise this once in their article around Section 2.2
  • Maybe it could be good to add explicitly in Table 1 the gradient estimates corresponding to each pair of sampling and weight function.
  • Line 346,. "these full candidates" are the individual coordinate-wise candidates given in 344?
2 comments
AuthorsOpenReview ↗

Response to Reviewer oG68 (1/2)

Thank you for your careful review and constructive questions.

W1. What the framework reveals

Despite the quality of the unified framework, the papers does not leads to any important knowledge breakthrough. No general property emerge from this unification, no innovative kernel is introduced, and the method is mainly applied to known smoothing kernels. The traditional "target-first" construction of gradient estimates is a way to discover new potential zeroth-order gradient estimators, but the impact of the presented "update-first" reconstruction is less clear (see Figure 1).

Our contribution is to identify the objective function that an implemented sampler and set of update weights implicitly differentiate, thereby providing an intuitive interpretation of existing algorithms. The kernel shapes in Fig. 2(a) help explain how these estimators behave.

The reconstructed kernels also provide new insights. Section 5.3 identifies the objectives implicitly differentiated by existing random-denominator estimators. Uniform perturbations yield a logarithmic kernel, and Gaussian perturbations yield a kernel involving the exponential integral. Fig. 2(a) also shows that, even when evaluation points are sampled from the same uniform distribution on an interval, different update weights yield logarithmic and Epanechnikov kernels.

Proposition F.1 further illustrates the value of kernel reconstruction. For a function ff with an LL-Lipschitz gradient, it uses the mean perturbation norm aj=E[∥U(j)∥]a_j=\mathbb E[\|U^{(j)}\|] under the reconstructed kernel. The bias b(μ)=E[g^μ]−∇f(μ)b(\mu)=\mathbb E[\widehat g_\mu]-\nabla f(\mu) satisfies

∥b(μ)∥≤ζrec:=L(∑j=1daj2)1/2.\|b(\mu)\|\leq \zeta_{\mathrm{rec}}:=L\left(\sum_{j=1}^d a_j^2\right)^{1/2}.

W2. Connection to optimization

Moreover, the main point of these ZO gradient estimator is to be used to minimize a given objective function ff. The optimization aspect is almost absent of the paper (beside being evoked in the Discussion section).

The appendix discusses the connection to optimization performance. If the paper is accepted, we will also include this discussion in the main text. For example, under the assumptions of Proposition F.1, if the original objective ff is λ\lambda-strongly convex, Appendix F shows that the distance from any point μ^\hat\mu at which the expected update is zero to the minimizer μ⋆\mu^\star of ff satisfies

∥μ^−μ⋆∥≤ζrecλ.\left\|\hat{\mu}-\mu^{\star}\right\| \leq \frac{\zeta_{\mathrm{rec}}}{\lambda}.

AuthorsOpenReview ↗

Response to Reviewer oG68 (2/2)

Q1. Existing bounds for derivative-free estimators

Could this paper be useful to cite and to rely on to refer to existing known bounds of the DFO different estimators? Beharas et al, 2022 "A Theoretical and Empirical Comparison of Gradient Approximations in Derivative-Free Optimization"

Thank you for pointing us to this paper. The results of Berahas et al. (2022) could help connect the smoothing bias discussed in Appendix F to the accuracy of gradient estimates obtained from finitely many function evaluations and to convergence conditions for optimization.

Q2. Line 161

Could you be more precise in line 161?

Assume that the Hessian of ff is LHL_H-Lipschitz in operator norm and that E∥Z∥4<∞\mathbb E\|Z\|^4<\infty. The estimator in Eq. (5) then satisfies

\le (L_H/6)\sigma^2\mathbb E\|Z\|^4.$$ In the second-order Taylor expansions of $f(\mu+\sigma Z)$ and $f(\mu-\sigma Z)$ about $\mu$, the constant and quadratic terms cancel upon subtraction. Each remainder is at most $(L_H/6)\sigma^3\|Z\|^3$ in absolute value. After dividing by $2\sigma$ and multiplying by $Z$, the remainder has norm at most $(L_H/6)\sigma^2\|Z\|^4$. The leading term is $ZZ^\top\nabla f(\mu)$, whose expectation is $\nabla f(\mu)$ because $\mathbb E[ZZ^\top]=I_d$ as assumed in Eq. (5). If the paper is accepted, we will replace “sufficiently smooth” with “the Hessian of $f$ is Lipschitz continuous in operator norm.” ## Q3. Does line 192 also apply to Eq. (5)? > Line 192. Isn't this comment (about the unbiasedness of the gradient estimate relative to the smoothed objective) also true for the estimates of equation (5), cf Nesterov & Spokiony 2017 equation (31). Thank you for pointing this out. Yes, this also holds for Eq. (5) with Gaussian perturbations. Nesterov and Spokoiny (2017) give the corresponding finite-scale identity and estimator. We will make the connection to Eq. (5) explicit at line 192 and cite them. ## Q4. Why must the weights be centered under the sampler? > Line 282. Maybe it's trivial, but could you explain why $\partial_{\theta_j} p_\theta (x) = q_\theta (x) \psi_{\theta_j} (x)$ requires the weights function to be centered under $q_\theta$? Since $p_\theta$ is a probability density, its total mass is $1$ for every $\theta$. Integrating the reconstruction equation over $x$ therefore gives $\mathbb E_{X\sim q_\theta}[\psi_\theta(X)]=\partial_\theta\int p_\theta(x)\,dx=0$, provided differentiation can be interchanged with integration. ## Q5. Fixed-denominator kernels in Fig. 2(a) > Line 282. Maybe it's trivial, but could you explain why $\partial_{\theta_j} p_\theta (x) = q_\theta (x) \psi_{\theta_j} (x)$ requires the weights function to be centered under $q_\theta$? Both fixed-denominator estimators are shown in Fig. 2(a). The Gaussian version corresponds to “Gaussian ES” and has a Gaussian smoothing kernel. The version with perturbations sampled uniformly from an interval corresponds to “Uniform numerator” and has an Epanechnikov smoothing kernel. ## Comments/Suggestions > It is known from experts of the domain that one must be careful with the rescaling factor $d$ when it comes to ZO gradient estimate with perturbations sampled uniformly over the (unit) sphere (or one has to sample over the sphere of radius $\sqrt{d}$. I would encourage the authors to precise this once in their article around Section 2.2 > Maybe it could be good to add explicitly in Table 1 the gradient estimates corresponding to each pair of sampling and weight function. >Line 346,. "these full candidates" are the individual coordinate-wise candidates given in 344? Thank you for your comments. The phrase "these full candidates" in line 346 refers to the individual coordinate-wise candidates introduced in line 344. If the paper is accepted, we will make this reference explicit.
Reviewer_TUT1OpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

Classical stochastic smoothing, zeroth order methods focus on an outline that is target first (as stated by the authors). This starts by choosing a specific mathematical smoothing distribution (target) to perturb the inputs of a non-differentiable function (like a black box algorithm), thus rendering it differentiable. The authors of the submitted draft resort to an update first approach. The authors analyze extant zeroth order optimization algorithms (update) and infer the smoothing measure those algorithms are optimizing. This is an attempt to formalize / unify extant approaches and enable apples to apples comparison between different approaches. They try to generalize the idea of stochastic and other smoothing approaches, and try to explain their findings with exotic signed smoothing distributions and mass drift.

Critical strengths

The biggest potential strength is the attempt to unify the zeroth order methods under one umbrella, where they can be compared in an apples to apples manner. This would enable a direct comparison of different approaches, understand their underlying behavior, why some approaches work for some problems and more importantly, why some dont.

Critical weaknesses
  1. The submission focuses on establishing a framework for extant approaches and does not define any new approaches or even provide prescriptive aids to design new and better zeroth order methods. This is reflected in the paucity of application based results in the mainn section of the paper.

  2. Many of the concepts of "signed distributions", "mass drift", etc are lacking grounding and interpretability. They come across as devices used to glue together the edge cases of a theory.

  3. The authors ignore adaptive and momentum-based methods, where their approach does not apply at all.

Questions for authors
  1. Please extend the update first approach to adaptive algorithms. Whilst target first is fine, many approaches are distribution first such as CMA and its extensions. One can hardly claim to unify a field if that requires ignoring half of the field.

  2. Please provide interpretations / explanations for concepts such as "signed distributions", "mass drift", etc.

1 comment
AuthorsOpenReview ↗

Response to Reviewer TUT1

Thank you for your careful review and constructive questions.

W1. Guidance for designing new methods

The submission focuses on establishing a framework for extant approaches and does not define any new approaches or even provide prescriptive aids to design new and better zeroth order methods. This is reflected in the paucity of application based results in the mainn section of the paper.

Our contribution is to identify the objective function that an implemented sampler and set of update weights implicitly differentiate, thereby providing an intuitive interpretation of existing algorithms. The kernel shapes in Fig. 2(a) help explain how these estimators behave.

Proposition F.1 illustrates the value of reconstructing the kernels. For a function ff with an LL-Lipschitz gradient, the proposition uses the mean perturbation norm aj=E[∥U(j)∥]a_j=\mathbb E[\|U^{(j)}\|] under the reconstructed kernel. The bias b(μ)=E[g^μ]−∇f(μ)b(\mu)=\mathbb E[\widehat g_\mu]-\nabla f(\mu) satisfies

∥b(μ)∥≤L(∑j=1daj2)1/2.\|b(\mu)\|\leq L\left(\sum_{j=1}^d a_j^2\right)^{1/2}.

W2 and Q2. Interpretation of the concepts

Many of the concepts of "signed distributions", "mass drift", etc are lacking grounding and interpretability. They come across as devices used to glue together the edge cases of a theory.

Please provide interpretations / explanations for concepts such as "signed distributions", "mass drift", etc.

We agree that these concepts need clearer explanations in the manuscript and will revise this discussion if the paper is accepted. The term signed measure is used in prior works, such as (Mohamed et al., 2020). Here, a signed measure describes where probability mass increases and where it decreases as a parameter changes. Mass drift refers to a nonzero net change in probability mass implied by the update rule.

(Mohamed et al., 2020) Mohamed et al., 2020, Monte Carlo Gradient Estimation in Machine Learning, Journal of Machine Learning Research

W3. Adaptive and momentum-based methods

The authors ignore adaptive and momentum-based methods, where their approach does not apply at all.

The proposed reconstruction may extend to adaptive and momentum-based methods. In particular, an adaptation mechanism may change the reconstructed kernel over the course of the algorithm. The reconstruction can be applied to the underlying gradient estimator. The interpretation and convergence of the full history-dependent update are outside the scope of this paper.

Q1. Extension to adaptive algorithms such as CMA-ES

Please extend the update first approach to adaptive algorithms. Whilst target first is fine, many approaches are distribution first such as CMA and its extensions. One can hardly claim to unify a field if that requires ignoring half of the field.

We agree that our claim was too broad and will revise the wording. Prior theoretical work on CMA-ES has studied simplified variants by isolating or removing individual mechanisms. Our framework may apply to a variant with mean and rank-μ\mu covariance updates, known as pure rank-μ\mu update CMA-ES, for which the expected update can be interpreted as a natural-gradient direction at the current iteration. Establishing the connection to the rank weights actually computed in a finite population and analyzing the full update with evolution paths and step-size control require further work.

Reviewer_QYc4OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper studies zeroth-order estimators that can be expressed through a sampling distribution qθq_\theta and update weights ψθ\psi_\theta. It proposes reconstructing an implicit smoothing measure pθp_\theta such that the expected update equals a derivative of J(θ)=Ex∼pθ[f(x)]J(\theta)=E_{x\sim p_\theta}[f(x)]. This provides an alternative interpretation of these estimators as exact derivatives of smoothed objectives, rather than approximate derivatives of the original function.

The framework recovers familiar examples and provides explicit smoothing kernels for schemes such as random-denominator differences, for which the authors claim that such objectives were previously unidentified. The general reconstruction principle is not restricted to location–scale families, although the paper uses this structure to derive its explicit integration templates.

An important qualification is that reconstruction proceeds one parameter component at a time. The resulting candidates need not be compatible with a single common smoothing measure or objective.

Critical strengths
  1. The distinction between sampling and smoothing distributions provides a useful unifying perspective on zeroth-order estimators. The explicit kernels for random-denominator differences are particularly interesting and may help explain differences in the estimators’ smoothing behavior.
Critical weaknesses
  1. The scope and assumptions of the explicit reconstruction results could be clearer. The exposition assumes location–scale structure for the unknown smoothing measure, whereas the reader starts with a given sampler and update rule. Stating sufficient conditions directly in terms of qθq_\theta and ψθ\psi_\theta, including conditions for a valid common reconstruction, would make the applicability easier to assess.

  2. The practical implications remain limited. Although the paper provides smoothing interpretations and bias bounds, it leaves estimator-specific variance analysis outside its scope. Consequently, it does not establish when one estimator should be preferred over another or demonstrate how the framework leads to improved estimator design.

  3. The regularity conditions for the random-denominator gradient interpretation need clarification. Bounded measurable ff guarantees that the reconstructed objective JJ exists, but does not guarantee a finite derivative or expected update. For example, with f(x)=1{x≥0}f(x)=\mathbf{1}_{\{x\geq0\}} and ε∼Unif[−σ,σ]\varepsilon\sim\mathrm{Unif}[-\sigma,\sigma], the paired estimator at μ=0\mu=0 equals 1/(2∣ε∣)1/(2|\varepsilon|) and has infinite expectation, while J(0)J(0) is finite. Although the weak-derivative interpretation may remain valid, it does not establish pointwise unbiasedness for a finite gradient. Could the authors state sufficient conditions for this stronger interpretation and distinguish it explicitly from the weak-derivative identity?

Questions for authors
  1. Can the authors state sufficient conditions on the given sampler and update weights that guarantee a valid location–scale reconstruction? How restrictive are these conditions, including compatibility across parameter components?

  2. Can the reconstructed kernels support a concrete estimator-selection or design result; for example, identifying a setting where one estimator achieves a better bias–variance tradeoff under a fixed evaluation budget?

2 comments
AuthorsOpenReview ↗

Response to Reviewer QYc4 (1/2)

Thank you for your careful review and constructive questions.

W1. Scope and assumptions of reconstruction

The scope and assumptions of the explicit reconstruction results could be clearer. The exposition assumes location–scale structure for the unknown smoothing measure, whereas the reader starts with a given sampler and update rule. Stating sufficient conditions directly in terms of qθq_\theta and ψθ\psi_\theta, including conditions for a valid common reconstruction, would make the applicability easier to assess.

Appendix E, Eqs. (52)–(54), gives the coordinate-wise reconstruction and compatibility conditions for location–scale families. Equation (56) gives a compatibility condition in general parameterizations. A common candidate must also be nonnegative and have unit mass to define a probability measure. We will clarify this distinction and point readers to both sets of conditions from Section 4. The following answer to Q1 states the location–scale conditions directly in terms of the given sampler and weights.

W2. Practical implications

The practical implications remain limited. Although the paper provides smoothing interpretations and bias bounds, it leaves estimator-specific variance analysis outside its scope. Consequently, it does not establish when one estimator should be preferred over another or demonstrate how the framework leads to improved estimator design.

The framework identifies the objective whose gradient the expected update represents. In Figure 2(a), the uniform random-denominator kernel has a logarithmic singularity at the center. For the step function in W3, this explains why the smoothed objective has an infinite slope at the discontinuity. Estimator-specific variance requires a separate analysis of the sampling and weighting rule, as discussed in Section 6. Since our aim is to interpret existing update rules, we leave a detailed variance analysis for future work.

W3. Regularity of the random-denominator interpretation

The regularity conditions for the random-denominator gradient interpretation need clarification. Bounded measurable guarantees that the reconstructed objective exists, but does not guarantee a finite derivative or expected update. For example, with f(x)=1x≥0f(x)=\mathbf{1}_{{x\geq0}} and ε∼Unif[−σ,σ]\varepsilon\sim\mathrm{Unif}[-\sigma,\sigma], the paired estimator at μ=0\mu=0 equals 1/(2∣ε∣)1/(2|\varepsilon|) and has infinite expectation, while J(0)J(0) is finite. Although the weak-derivative interpretation may remain valid, it does not establish pointwise unbiasedness for a finite gradient. Could the authors state sufficient conditions for this stronger interpretation and distinguish it explicitly from the weak-derivative identity?

We agree that bounded measurability ensures that the smoothed objective is finite. In the step-function example, the paired estimator has infinite expectation at the discontinuity. A sufficient condition for pointwise unbiasedness is that ff be bounded and Lipschitz continuous.

To see this, let ε\varepsilon be the scalar perturbation drawn from either a uniform distribution on an interval centered at zero or a zero-mean Gaussian distribution. Write R=∣ε∣R=|\varepsilon| and Ar(μ)=(2r)−1∫μ−rμ+rf(x) dxA_r(\mu)=(2r)^{-1}\int_{\mu-r}^{\mu+r}f(x)\,dx. For both uniform and Gaussian perturbations, the reconstructed objective is J(μ)=E[AR(μ)]J(\mu)=\mathbb E[A_R(\mu)]. For each r>0r>0, Ar′(μ)=[f(μ+r)−f(μ−r)]/(2r)A_r'(\mu)=[f(\mu+r)-f(\mu-r)]/(2r) is continuous, and its absolute value is bounded by the Lipschitz constant of ff, uniformly in rr. By the mean value theorem the same bound holds for the difference quotients of ArA_r. Dominated convergence therefore gives J′(μ)=E[AR′(μ)]J'(\mu)=\mathbb E[A_R'(\mu)] at every μ\mu. The right-hand side is the finite expectation of the paired estimator. We will state this pointwise result separately from the distributional identity.

AuthorsOpenReview ↗

Response to Reviewer QYc4 (2/2)

Q1. Sufficient conditions for location–scale reconstruction

Can the authors state sufficient conditions on the given sampler and update weights that guarantee a valid location–scale reconstruction? How restrictive are these conditions, including compatibility across parameter components?

The following conditions are sufficient for location–scale reconstruction.

  1. For each location component j=1,…,dj=1,\ldots,d, assume that q((t,x−j)∣μ,σ)ψμj((t,x−j))q((t,x_{-j})\mid\mu,\sigma)\psi_{\mu_j}((t,x_{-j})) is locally integrable in tt and the following tail integral exists: p(j)(x∣μ,σ)=∫xj∞q((t,x−j)∣μ,σ) ψμj((t,x−j)) dtp^{(j)}(x\mid\mu,\sigma)=\int_{x_j}^{\infty}q((t,x_{-j})\mid\mu,\sigma)\,\psi_{\mu_j}((t,x_{-j}))\,dt

  2. All candidates p(j)p^{(j)} must agree almost everywhere with a common nonnegative density pp that integrates to one and has the location–scale form p(x∣μ,σ)=σ−dp0 ⁣(x−μσ).p(x\mid\mu,\sigma)=\sigma^{-d}p_0\!\left(\frac{x-\mu}{\sigma}\right).

  3. If the scale parameter is also updated, the same density must satisfy q(x∣μ,σ)ψσ(x)=∂σp(x∣μ,σ)q(x\mid\mu,\sigma)\psi_\sigma(x)=\partial_\sigma p(x\mid\mu,\sigma).

Given qq and ψ\psi, we can check these conditions by writing z=(x−μ)/σz=(x-\mu)/\sigma. Suppose that, for each jj, the standardized product

\psi_{\mu_j}(\mu+\sigma z)$$ depends only on $z$, and denote it by $a_j(z)$. If the tail integrals $k_j(z)=\int_{z_j}^{\infty}a_j(t,z_{-j})\,dt$ exist and agree almost everywhere with a common nonnegative density $k$ of unit mass, then $p(x\mid\mu,\sigma)=\sigma^{-d}k(z)$ satisfies the reconstruction equations for all location components. If the scale is also updated, we check under appropriate regularity conditions that $q\psi_\sigma=\partial_\sigma p$. ## Q2. Estimator selection and design > Can the reconstructed kernels support a concrete estimator-selection or design result; for example, identifying a setting where one estimator achieves a better bias–variance tradeoff under a fixed evaluation budget? Proposition F.1 provides a criterion for comparing bias bounds at fixed smoothing scale. When $\nabla f$ is $L$-Lipschitz, the bound is $\|b(\mu)\|\le L(\sum_j a_j^2)^{1/2}$, where $a_j=\mathbb E\|U^{(j)}\|$ under the reconstructed kernel for component $j$. Smaller values of this bound favor kernels that concentrate more closely around the search point under this criterion. Comparing mean squared error at a fixed evaluation budget also requires the variance of each estimator.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.