Pith. sign in

REVIEW 4 major objections 4 minor 76 references

A Framework for Population-Based Stochastic Optimization on Abstract Riemannian Manifolds

T0 review · 4 major / 4 minor · reviewed 2026-08-14 · deepseek-v4-flash

Pith's one-line read A population-based stochastic search can reach the global optimum of any compact Riemannian manifold in finitely many steps, using only local geometry and no ambient embedding.

desk verdict Impressive geometric framework for manifold population-based optimization, but the headline finite-step global convergence claim relies on an exploration assumption the implementation doesn't guarantee. read the letter →

arxiv 1908.06783 v3 pith:OS5ZW7FS submitted 2019-08-19 math.OC

classification math.OC MSC 90C5690C1553B2062B10
keywords ManifoldOptimizationPopulation-basedStochasticDerivative-FreeinformationgeometrymixturedensitiesglobalconvergenceJacob'sladderRiemannian
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper claims that a population-based stochastic search can solve black-box optimization problems on abstract Riemannian manifolds — curved spaces treated on their own terms, without an ambient Euclidean space to embed them in. The central result is that on connected compact manifolds the algorithm's explored region fills the entire manifold in finitely many steps, under the stated condition that each iteration contributes an exploration point on the boundary of the region explored so far; the global optimum is then reached in finite time rather than merely approached as a stationary-point limit. The route runs through information geometry: probability densities are inherited onto the manifold locally through orientation-preserving diffeomorphisms (the exponential map being the special case), then patched into mixture densities whose parameter space forms a product statistical manifold. That product structure yields monotone improvement of expected fitness and a first-principles derivation of the search-coefficient update. If the claims hold, manifold optimization no longer needs completeness, global parametrizations, or logarithms between arbitrary points — and problems on spaces such as Jacob's ladder, a surface of infinite genus with no polynomial-constraint description, become tractable.

What carries the argument

The load-bearing object is the family of mixture densities $\mathcal{L}_V$ over totally bounded subsets $V$ of the manifold. Each component is a locally inherited density, pulled back through an orientation-preserving diffeomorphism from a coordinate patch (the Riemannian exponential map is the special case); these are glued into mixtures $\tilde{p}(x) = \sum_{\alpha} \phi_\alpha \tilde{p}_\alpha(x)$ on a finite cover of $V$. Theorems 5.1–5.3 show $\mathcal{L}_V = S_0 \times \tilde{S}_1 \times \cdots \times \tilde{S}_\Lambda$ as a product statistical manifold, with the mixture divergence decomposing into a coefficient part and a component part — this is what decouples coefficient updates from component updates. On the coefficient simplex the paper installs the modified Fisher metric $G = F^{-1} + \epsilon_0 I$ (Equation (48)), whose natural-gradient fixed point, Equation (54), is proportional to the relative expected fitnesses $E_\alpha$; the practical update (Equation (40)) is the limiting case. Finally, the convergence argument of Theorem 8.1 is carried by the exploration distribution: sampling on geodesic-sphere boundaries and rejecting interior points places each new exploration centroid at least a uniform distance $j_M$ from all previous ones, and compactness then forces the explored region to exhaust $M$ in finitely many steps.

What would settle it

Run Extended RSDFO on the 2-sphere with the Section 9 exploration rule — sample points uniformly on geodesic-sphere boundaries and reject those lying in the interior of the explored region — and count the iterations in which every sampled point is rejected, so that no boundary centroid is added. The paper explicitly allows this fallback ('if all the sampled points are rejected, then Extended RSDFO will not sample new boundary points'), and it violates the hypothesis of Theorem 8.1 that at least one exploration point is generated whenever the boundary is nonempty; the empirical frequency of such stalled rounds at increasing manifold dimension would show whether the finite-step guarantee holds for the implemented algorithm or only for the idealized boundary oracle.

Watch

Extended reading notes

Core claim

The central discovery is Theorem 8.1: on a compact connected Riemannian manifold $(M,g)$, if Extended RSDFO generates at least one exploration point from the boundary of the explored region whenever that boundary is nonempty, then the global convergence condition of Equation (57) holds within finitely many steps. Concretely, there is a finite integer $N$ such that $\sup_{\alpha \in \Lambda_N} E^\alpha_N = E^*$, the best attainable expected fitness, so the algorithm does not merely converge to a stationary point in the limit — the global optimum is attained in the explored region after finitely many iterations. The key structural claim supporting this is that the family of mixture densities $\mathcal{L}_V$ over a totally bounded subset $V$ of the manifold is a product statistical manifold of the mixture-coefficient simplex and the locally inherited component families, which lets the algorithm evolve mixture coefficients and component parameters independently while comparing solution quality across disjoint tangent spaces.

Load-bearing premise

The proof of Theorem 8.1 assumes that whenever the explored region is not yet the whole manifold, the algorithm can always create a new search point exactly on the boundary of that region, with every search ball at least a fixed positive radius — but the implemented exploration samples boundary points randomly and simply adds none if all samples fall inside the explored region, and under continuous sampling the probability of landing exactly on the boundary is zero.

Editorial extensions

If this is right

  • On any connected compact Riemannian manifold, Extended RSDFO's expected fitness is monotone non-decreasing across iterations (Proposition 6.1), and with boundary exploration it reaches the global optimum in finitely many steps rather than converging only to a stationary point (Theorem 8.1).
  • Because mixture coefficients and component parameters live on independent factors of a product statistical manifold, solution quality can be compared across disjoint tangent spaces — something the paper argues is impossible for single-centroid Riemannian SDFO methods such as Riemannian CMA-ES.
  • Extended RSDFO's computations are strictly local geodesic-ball computations, so the manifold need not be complete and no Riemannian logarithm map between arbitrary points is required; the paper shows the latter assumption is what cripples Riemannian PSO on large manifolds.
  • The experiments on the sphere, Grassmann manifolds, and Jacob's ladder indicate the method combines the global-exploration behaviour of Riemannian PSO with the local-optimum accuracy of Riemannian CMA-ES, at the price of extra function evaluations and boundary-point evaluations.
  • Jacob's ladder — a countably infinite connected sum of tori without a global ambient representation — constitutes a problem class that constraint-based Euclidean optimization cannot formulate, so any success there is evidence for intrinsic manifold optimization as such.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The finiteness argument is essentially topological: boundary centroids pinned at a uniform distance $j_M$ plus compactness forces exhaustion. It should transfer to any exploration rule that guarantees a point at least a fixed distance beyond the explored region — for example deterministic, low-discrepancy boundary sampling — which would also remove the measure-zero difficulty the paper's own rando
  • Reading the algorithm through the product structure $\mathcal{L}_V = S_0 \times \prod_\alpha \tilde{S}_\alpha$ suggests a design principle for population-based manifold optimizers: evolve coefficients by natural gradient on the simplex under $G = F^{-1} + \epsilon_0 I$ and evolve components separately. Deriving component-parameter updates from the same fixed-point condition would extend the paper'
  • A testable extension is to separate the two sources of difficulty the paper's Grassmann experiments conflate: the cost of estimating local component models in higher dimension, and the density of centroid coverage needed for exploration. Budgeting boundary samples rather than centroid count per iteration would isolate which one drives the sharp drop in success rate from $\mathrm{Gr}(2,4)$ to $\mat
  • The Jacob's ladder construction is reusable as a stress test for any intrinsic manifold optimizer: a countably infinite connected sum of tori with a glued exponential map is a manifold on which no ambient constraint formulation exists, so comparisons there measure genuine manifold behaviour rather than embedding artefacts.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The manuscript develops an information-geometric framework for population-based stochastic optimization on Riemannian manifolds and introduces Extended RSDFO, an algorithm that combines locally inherited probability densities into mixture densities over totally bounded subsets of the search space. The paper claims that the expected fitness of Extended RSDFO improves monotonically (Proposition 6.1) and that the algorithm converges globally in finitely many steps on connected compact Riemannian manifolds (Theorem 8.1). It also reports experiments on the sphere, Grassmannian manifolds, and a novel Jacob's-ladder test problem, comparing Extended RSDFO with Riemannian trust-region, Riemannian CMA-ES, and Riemannian PSO.

Significance. If the convergence result held for the implemented algorithm, it would be a notable contribution: finite-step global convergence on general compact Riemannian manifolds is a strong guarantee for a population-based derivative-free method, and the product statistical-manifold construction (Theorem 5.1, Remark 5.4) is a useful geometric framework. The paper also introduces a genuinely non-embeddable test manifold, Jacob's ladder, which helps motivate intrinsically manifold-native optimization. The experimental study is extensive and the paper gives credit where due by providing explicit geometric derivations and a relatively reproducible setup. However, the advertised global-convergence claim is not established for the algorithm as implemented, and the geometric derivation of the mixture-coefficient update contains a metric inconsistency; the significance is therefore real but conditional on resolving these issues.

major comments (4)
  1. [Section 8, Theorem 8.1; Sections 6.3.4 and 9] Theorem 8.1 is proved under the hypothesis that, whenever the explored region W_k is not all of M, the algorithm generates at least one exploration point on the boundary of W_k, and the proof additionally asserts a uniform positive lower bound j_M on all geodesic-ball radii. Neither condition is guaranteed by Algorithm 3 as described. The exploration procedure in Section 6.3.4 samples from boundaries of individual geodesic spheres and rejects points lying in interiors (Equation (46)); Section 9 explicitly states that if all sampled points are rejected, no new boundary point is added. Thus the theorem is a conditional statement about an idealized boundary-exploration oracle, not about the implemented algorithm. The abstract's unconditional phrasing, 'converges globally eventually in finitely many steps on connected compact Riemannian manifolds', is therefore unsupported for the implemented method. The proof's assertion j_M > 0 also needs an explicit assumption, since j_x ≤ inj(x) and the algorithm could in principle choose radii with no uniform positive lower bound.
  2. [Section 7, Equations (48), (52)-(55)] The metric used in the natural-gradient computation is not consistently defined. Equation (48) defines G := F^{-1} + ε0 I, but the Riemannian metric immediately below is gξ(Y,Z) := Σ yα zα (φ + ε0), which corresponds to the diagonal matrix diag(φ + ε0). For the simplex Fisher information matrix F, the matrix F^{-1} is not diagonal, so its inverse is not diag(1/(φ + ε0)). The inverse metric used in Equation (52) is therefore not the inverse of the matrix defined in Equation (48). The fixed-point derivation in Equations (53)-(55) and the claimed recovery of Equation (40) in Remark 7.2 rely on this ambiguity, so the 'first principles' derivation of the mixture-coefficient update needs to be corrected.
  3. [Section 7, Remark 7.2; Section 6.3.2, Equation (40)] The claim that Equation (40) is 'rigorously derived from first principles' in Remark 7.2 is circular in the present form. Equation (40) is introduced before Section 7 as the definition of the update, and the natural-gradient fixed point derived in Section 7 is, by construction, proportional to the expected fitnesses Eα. The metric in Equation (48) is explicitly selected so that its natural-gradient fixed point favors the interior point with coordinates proportional to relative fitness (see the bullet list in Section 7). The derivation therefore shows consistency between the chosen metric and the pre-existing update rule, but it does not independently derive Equation (40). This is a framing and justification issue that should be corrected.
  4. [Sections 6.3 and 7.2] The paper states 'without loss of generality' that f can be made strictly positive by translation, but the mixture-coefficient updates in Equations (40), (54), and (55) are not invariant under adding a constant to f. If f is replaced by f + C, then every Eα changes to Eα + C, and the normalized coefficients change unless C = 0. The same translation dependence appears in the minimization counterpart in Equation (56). This invalidates the 'without loss of generality' claim and further weakens the natural-gradient justification of the update rule.
minor comments (4)
  1. [Global] The manuscript contains many typographical artifacts and misspellings, including '/f_irst', 'de/f_ined', 'neighhbourhood', 'experiement', and 'Jaccob's ladder'; please proofread carefully.
  2. [Section 9, Tables 4-6] The relaxed exploration parameter ϵ_b is introduced in the experiments but is not analyzed in the theoretical sections; please clarify how sampling from boundaries of smaller geodesic spheres (ϵ_b < 1) interacts with the convergence assumptions in Theorem 8.1.
  3. [Section 6.3.5] The theoretical termination criterion (boundary of the explored region empty) differs from the practical termination criterion used in the experiments (all local RSDFO streams terminate); the paper should state explicitly that the convergence analysis applies only to the former.
  4. [Section 9.2] In the experimental setup, the local RSDFO core is described with budgets of 'parents' and 'offsprings', but these terms are not defined in the RSDFO framework of Section 3.1, which makes the setup harder to reproduce.

Circularity Check

1 steps flagged · score 4.0 of 10

The main circularity is in Section 7: the modified metric is chosen so that its natural-gradient fixed point equals the already-adopted proportional-fitness update (Eq. 40), which is then called a first-principles derivation. Theorem 8.1 is conditional and independent.

  1. self definitional [Section 7, Equation (48) through Remark 7.2; compare Equation (40) in Section 6.3.2]
    "Natural gradient ascent on the closure S0 under the modified metric, as we will discuss in the subsequent subsection, favours the interior point of the simplex with coordinates proportional to the relative weights of the vertices. ... if ... ϵ0 ... sufficiently small ... we retrieve Equation (40). This fixed point therefore directly reflects the relative expected fitness ..."

    Equation (40) was introduced directly as the coefficient update: 'it is natural to assign to the individual mixture coefficients φα a value proportional to Eα. In particular, in line 7 ... φ^{k+1}_α = E_α / Σ E_α.' The metric G = F^{-1} + ϵ0 I was then introduced with the stated purpose that natural gradient ascent on the simplex 'favours the interior point ... with coordinates proportional to the relative weights of the vertices.' Under this metric the stationarity condition (Eq. 53) is E_α/(φ_α + ϵ0) = c, which rearranges to φ_α ∝ E_α and, in the small-ϵ0 limit, is exactly Equation (40). Thus the 'derived' fixed point is not an independent consequence of information geometry; it is a restatement of the heuristic update that motivated the choice of metric.

full rationale

The paper's central information-geometric constructions (Sections 4 and 5) are largely self-contained: the pulled-back dualistic structures and the product statistical manifold result for mixture densities are derived from stated definitions, and no load-bearing self-citation chain appears. Theorem 8.1 is a conditional statement: if the algorithm generates a boundary exploration point whenever the explored region is a proper subset, compactness forces exhaustion of M in finitely many steps. That argument is independent, though the abstract omits the boundary-generation hypothesis. The implementation in Section 9 explicitly admits that the acceptance-rejection exploration can fail to produce boundary points ('if all the sampled points are rejected, then Extended RSDFO will not sample new boundary points'), so the theorem's hypothesis is not guaranteed by the implemented algorithm; this is a correctness gap rather than a circularity. The significant circular step is in Section 7: the modified metric is chosen precisely so that its natural-gradient fixed point equals the proportional-fitness update already adopted in Equation (40), and Remark 7.2 then labels this retrospective equivalence a first-principles derivation. Because that derivation is forced by the metric choice rather than by independent principles, the score is 4; the remaining convergence theorem and experimental comparisons still carry independent content.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The central claims rest on standard Riemannian and information-geometric background, on domain assumptions about the objective and manifold, and on an idealized boundary-exploration condition that is not guaranteed by the implemented algorithm. Several hyperparameters are hand-chosen in experiments. No new physical entities are introduced.

free parameters (5)
  • epsilon_0 in modified simplex metric G = F^{-1} + epsilon_0 I = not specified, arbitrarily small positive
    Introduced in Equation (48) to define a metric whose natural-gradient fixed point Equation (54) equals the fitness-proportional update Equation (40) in the small-epsilon_0 limit.
  • N_random and N_cull = 2 and 2 on S2/Grassmann; 6 and 3 on Jacob's ladder
    Chosen by hand; they control the number of new and retained centroids and therefore the exploration-exploitation balance.
  • exploration rate tau(k) = printed as 0.6 e^{0.015 k}, contradicting stated non-increasing requirement
    Controls the mixture between elitist sampling and exploration; the formula in Section 9 is inconsistent with its stated purpose.
  • exploration radius factor epsilon_b = 1.0, 0.4, or 0.5 in different experiments
    Scales the geodesic balls used for boundary exploration and changes performance, e.g., global-minimum rate on S2 rises from 128/200 to 154/200.
  • local RSDFO budgets and Monte Carlo sample sizes = parents/offsprings and MC samples listed per experiment
    Set per benchmark and affect the stochastic estimates of expected fitness used by the algorithm.
assumptions (5)
  • domain assumption M is a connected and orientable Riemannian manifold, and compact for the convergence theorem.
    Assumed in Sections 4 and 8; orientability is needed for volume-form constructions and compactness is used in Theorem 8.1.
  • domain assumption The objective f is bounded above, strictly positive by translation, and attains a global optimum.
    Used in Section 8 and Remark 7.3; positivity is required for the fitness-proportional mixture coefficient update.
  • ad hoc to paper Exploration generates a centroid exactly on the boundary of the explored region whenever the boundary is nonempty, with all geodesic ball radii uniformly bounded below by a positive constant.
    The proof of Theorem 8.1 relies on this. The acceptance-rejection procedure in Section 6.3.4 and the simplified implementation in Section 9 can reject all sampled points and add no boundary centroid.
  • domain assumption Local component families satisfy Conditions 1 and 2: distinct proper support and no functional dependency between components.
    Stated in Section 5.3.1; these conditions make the mixture parametrization injective and an immersion. Condition 1 can be met by choosing the open cover, and Condition 2 is automatic for unconstrained mixtures.
  • standard math Standard statistical-manifold background is taken as given: Fisher metric, dual connections, Bregman divergence, and the existence of a divergence generating a dualistic structure.
    Section 4.1 and Theorem 5.1 rely on results from Amari and Nagaoka, Eguchi, and Matumoto without reproving them.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Framework for Population-Based Stochastic Optimization on Abstract Riemannian Manifolds." pith.science (2026). https://pith.science/paper/OS5ZW7FS

@misc{pith2026190806783,
  author       = {Pith},
  title        = {Pith review of: A Framework for Population-Based Stochastic Optimization on Abstract Riemannian Manifolds},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OS5ZW7FS}},
  note         = {Machine review of arXiv:1908.06783}
}
read the original abstract

We present Extended Riemannian Stochastic Derivative-Free Optimization (Extended RSDFO), a novel population-based stochastic optimization algorithm on Riemannian manifolds that addresses the locality and implicit assumptions of manifold optimization in the literature. We begin by investigating the Information Geometrical structure of statistical model over Riemannian manifolds. This establishes a geometrical framework of Extended RSDFO using both the statistical geometry of the decision space and the Riemannian geometry of the search space. We construct locally inherited probability distribution via an orientation-preserving diffeomorphic bundle morphism, and then extend the information geometrical structure to mixture densities over totally bounded subsets of manifolds. The former relates the information geometry of the decision space and the local point estimations on the search space manifold. The latter overcomes the locality of parametric probability distributions on Riemannian manifolds. We then construct Extended RSDFO and study its structure and properties from a geometrical perspective. We show that Extended RSDFO's expected fitness improves monotonically and it's global eventual convergence in finitely many steps on connected compact Riemannian manifolds. Extended RSDFO is compared to state-of-the-art manifold optimization algorithms on multi-modal optimization problems over a variety of manifolds. In particular, we perform a novel synthetic experiment on Jacob's ladder to motivate and necessitate manifold optimization. Jacob's ladder is a non-compact manifold of countably infinite genus, which cannot be expressed as polynomial constraints and does not have a global representation in an ambient Euclidean space. Optimization problems on Jacob's ladder thus cannot be addressed by traditional (constraint) optimization methods on Euclidean spaces.

Figures

Figures reproduced from arXiv: 1908.06783 by the authors.

Figure 1
Figure 1. Illustration of parallel transport from x to y within normal neighbourhood Nx of x, described in Section 2.1. The orthonormal basis at x is parallel transported to an orthonormal basis at y. Finally, when restricting our attention to a metric ball within the pre-imagie of the normal neighbourhood, we retrieve a metric ball in M under normal coordinates. Denition 2.1. Given a point x ∈ M, the injectivity radius at x… view at source ↗
Figure 2
Figure 2. Illustration of Riemannian adaptation of optimization algorithms. For each iteration, Riemannian [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Illustration of an iteration of RSDFO. For each iteration, the algorithm initiates with [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figures from the paper (17 more)
Figure 4
Figure 4. Figure 4: A summary of the induced statistical dualistic geometry of locally inherited parametrized densities [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]
Figure 5
Figure 5. Figure 5: 19It is worth noting that for this illustration ( [PITH_FULL_IMAGE:figures/full_fig_p037_5.png]
Figure 5
Figure 5. Figure 5: Illustration (from left to right) of the “full" evolutionary step of Extended RSDFO. At each iteration, [PITH_FULL_IMAGE:figures/full_fig_p038_5.png]
Figure 6
Figure 6. Figure 6: Simplicial illustration of the simplices in evolutionary step of Extended RSDFO (Algorithm 3). [PITH_FULL_IMAGE:figures/full_fig_p043_6.png]
Figure 7
Figure 7. Figure 7: An simplied illustration of the exploration distribution sampling using acceptance-rejection [PITH_FULL_IMAGE:figures/full_fig_p046_7.png]
Figure 8
Figure 8. Figure 8: A simplied illustrated of the mixture distribution in the special case if the search region is su - [PITH_FULL_IMAGE:figures/full_fig_p048_8.png]
Figure 9
Figure 9. Figure 9: Simplicial illustration of the evolutionary step of Algorithm 3. The number on the arrows cor [PITH_FULL_IMAGE:figures/full_fig_p051_9.png]
Figure 10
Figure 10. Figure 10: A demonstration of natural gradient ascent under the metric [PITH_FULL_IMAGE:figures/full_fig_p054_10.png]
Figure 11
Figure 11. Figure 11: Heat map of f on S 2 Injectivity radius on S 2 is inj(x) = π for all x ∈ S 2 [67], and the Riemannian exponential map of S 2 is given by [5]: expx (v) = γv(t)| t=1 = cos (||v|| · t) x + sin (||v|| · t) v ||v|| [PITH_FULL_IMAGE:figures/full_fig_p063_11.png]
Figure 12
Figure 12. Figure 12: Experimental results of one of the experiments on [PITH_FULL_IMAGE:figures/full_fig_p065_12.png]
Figure 13
Figure 13. Figure 13: Illustration of objective function f on GrR(p, n), notice the x-axis is given by the Riemannian distance dGr of x relative to In×p. The injectivity of Grassmannian manifolds GrR(p, n) is π 2 for all x ∈ GrR(p, n) [58, 67], and the Riemannian exponential map of is give…
Figure 14
Figure 14. Figure 14: Experimental results of one of the experiments on [PITH_FULL_IMAGE:figures/full_fig_p068_14.png]
Figure 15
Figure 15. Figure 15: Illustration of Jacob’s ladder surface as an innite connected sum of tori. [PITH_FULL_IMAGE:figures/full_fig_p069_15.png]
Figure 16
Figure 16. Figure 16: Illustration of minor and major rotational angles in the Jacob’s ladder. The vertical smaller [PITH_FULL_IMAGE:figures/full_fig_p070_16.png]
Figure 17
Figure 17. Figure 17: fG: the “global" part of the objective function f on Jacob’s ladder. 0 2 20 40 2 60 Objective value of fL 80 \varphi 100 120 0 0 [PITH_FULL_IMAGE:figures/full_fig_p071_17.png]
Figure 19
Figure 19. Figure 19: Experimental results of one of the experiments on Jacob’s ladder. [PITH_FULL_IMAGE:figures/full_fig_p073_19.png]
Figure 20
Figure 20. Figure 20: Illustration of the “trap” of fL. The local optimal angles (θ ∗, ϕ∗) are marked by the points labelled with ×. The objective value increases along dotted lines on T2 (as well as T1 and the other tori), as the “objective land￾scape", described by [PITH_FULL_IMAGE:figu…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

76 extracted references · 73 canonical work pages

  1. [1]

    Injectivity radius estimates and sphere theorems

    Abresch, U., Meyer, W.T., 1997. Injectivity radius estimates and sphere theorems. Comparison geometry 30, 1–47

  2. [2]

    Trust-region methods on Rieman- nian manifolds

    Absil, P.A., Baker, C.G., Gallivan, K.A., 2007. Trust-region methods on Rieman- nian manifolds. Foundations of Computational Mathematics 7, 303–330

  3. [3]

    A collection of nonsmooth riemannian optimiza- tion problems, in: Nonsmooth Optimization and Its Applications

    Absil, P.A., Hosseini, S., 2019. A collection of nonsmooth riemannian optimiza- tion problems, in: Nonsmooth Optimization and Its Applications. Springer, pp. 1–15

  4. [4]

    Riemannian geometry of grassmann manifolds with a view on algorithmic computation

    Absil, P.A., Mahony, R., Sepulchre, R., 2004. Riemannian geometry of grassmann manifolds with a view on algorithmic computation. Acta Applicandae Mathem- atica 80, 199–220

  5. [5]

    Optimization algorithms on matrix manifolds

    Absil, P.A., Mahony, R., Sepulchre, R., 2009. Optimization algorithms on matrix manifolds. Princeton University Press

  6. [6]

    Methods of Information Geometry

    Amari, S., Nagaoka, H., 2000. Methods of Information Geometry. volume 191 of Translations of Mathematical monographs . Oxford University Press

  7. [7]

    Natural gradient works efficiently in learning

    Amari, S.I., 1998. Natural gradient works efficiently in learning. Neural compu- tation 10, 251–276

  8. [8]

    Information geometry of divergence functions

    Amari, S.i., Cichocki, A., 2010. Information geometry of divergence functions. Bulletin of the Polish Academy of Sciences: Technical Sciences 58, 183–195

Show all 76 references
  1. [9]

    Benchmarkfcns toolbox

    Ansari Ardeh, M., 2019. Benchmarkfcns toolbox. Retrieved July 3, 2019, from http://benchmarkfcns.xyz/

  2. [10]

    On the use of evolution strategies for optimization on spher- ical manifolds, in: International Conference on Parallel Problem Solving from Nature, Springer

    Arnold, D.V., 2014. On the use of evolution strategies for optimization on spher- ical manifolds, in: International Conference on Parallel Problem Solving from Nature, Springer. pp. 882–891

  3. [11]

    Uniqueness of the /f_isher–rao metric on the space of smooth densities

    Bauer, M., Bruveris, M., Michor, P.W., 2016. Uniqueness of the /f_isher–rao metric on the space of smooth densities. Bulletin of the London Mathematical Society 48, 499–506

  4. [12]

    Convergence analysis of evolutionary algorithms that are based on the paradigm of information geometry

    Beyer, H.G., 2014. Convergence analysis of evolutionary algorithms that are based on the paradigm of information geometry. Evolutionary Computation 22, 679–709

  5. [13]

    Statistics on riemannian manifolds: asymptotic distribution and curvature

    Bhattacharya, A., Bhattacharya, R., 2008. Statistics on riemannian manifolds: asymptotic distribution and curvature. Proceedings of the American Mathem- atical Society 136, 2959–2967

  6. [14]

    Algebraic models of smooth manifolds

    Bochnak, J., Kucharz, W., 1989. Algebraic models of smooth manifolds. Inven- tiones mathematicae 97, 585–611. 78

  7. [15]

    Oriented bounding box computation using particle swarm optimization

    Borckmans, P.B., Absil, P.A., 2010. Oriented bounding box computation using particle swarm optimization

  8. [16]

    Borckmans, P.B., Ishteva, M., Absil, P.A., 2010. A modi/f_ied particle swarm op- timization algorithm for the best low multilinear rank approximation of higher- order tensors, in: International Conference on Swarm Intelligence, Springer. pp. 13–23

  9. [17]

    Manopt, a matlab toolbox for optimization on manifolds

    Boumal, N., Mishra, B., Absil, P.A., Sepulchre, R., 2014. Manopt, a matlab toolbox for optimization on manifolds. The Journal of Machine Learning Research 15, 1455–1459

  10. [18]

    Geometry of the Fisher-Rao metric on the space of smooth densities on a compact manifold

    Bruveris, M., Michor, P.W., 2016. Geometry of the Fisher-Rao metric on the space of smooth densities on a compact manifold. arXiv preprint arXiv:1607.04550

  11. [19]

    Projected gradient methods for linearly con- strained problems

    Calamai, P.H., Moré, J.J., 1987. Projected gradient methods for linearly con- strained problems. Mathematical programming 39, 93–116

  12. [20]

    Geometric modeling in probability and statistics

    Calin, O., Udrişte, C., 2014. Geometric modeling in probability and statistics. Springer

  13. [21]

    Riemannian Geometry

    do Carmo, M., 1992. Riemannian Geometry. Mathematics (Boston, Mass.), Birkhäuser. URL: https://books.google.co.uk/books?id=uXJQQgAACAAJ

  14. [22]

    Statistical Decision Rules and Optimal Inference

    Cencov, N.N., 1982. Statistical Decision Rules and Optimal Inference. Trans- lations of mathematical monographs, American Mathematical Society. URL: https://books.google.co.uk/books?id=kTJJtQAACAAJ

  15. [23]

    Comparison theorems in Riemannian geometry

    Cheeger, J., Ebin, D.G., Ebin, D.G., 1975. Comparison theorems in Riemannian geometry. volume 9. North-Holland Publishing Company Amsterdam

  16. [24]

    The cma-es on rieman- nian manifolds to reconstruct shapes in 3-d voxel images

    Colutto, S., Fruhauf, F., Fuchs, M., Scherzer, O., 2010. The cma-es on rieman- nian manifolds to reconstruct shapes in 3-d voxel images. IEEE Transactions on Evolutionary Computation 14, 227–245

  17. [25]

    Particle swarm optimization, in: Proceedings of the IEEE international conference on neural networks, Citeseer

    Eberhart, R., Kennedy, J., 1995. Particle swarm optimization, in: Proceedings of the IEEE international conference on neural networks, Citeseer. pp. 1942–1948

  18. [26]

    Geometry of minimum contrast

    Eguchi, S., et al., 1992. Geometry of minimum contrast. Hiroshima Mathematical Journal 22, 631–647

  19. [27]

    In/f_inite- genus riemann surfaces

    Feldman, J., Knörrer, H., Trubowitz, E., 1995. In/f_inite- genus riemann surfaces. Canadian Mathematical Society 3, 91–111

  20. [28]

    Minimizing a differentiable function over a differential manifold

    Gabay, D., 1982. Minimizing a differentiable function over a differential manifold. Journal of Optimization Theory and Applications 37, 177–219

  21. [29]

    Topologie des feuilles génériques

    Ghys, É., 1995. Topologie des feuilles génériques. Annals of Mathematics , 387– 422. 79

  22. [30]

    Cases for the nugget in modeling computer ex- periments

    Gramacy, R.B., Lee, H.K., 2012. Cases for the nugget in modeling computer ex- periments. Statistics and Computing 22, 713–722

  23. [31]

    The cma evolution strategy: a comparing review, in: Towards a new evolutionary computation

    Hansen, N., 2006. The cma evolution strategy: a comparing review, in: Towards a new evolutionary computation. Springer, pp. 75–102

  24. [32]

    The cma evolution strategy: A tutorial

    Hansen, N., 2016. The cma evolution strategy: A tutorial. arXiv preprint arXiv:1604.00772

  25. [33]

    Evaluating the cma evolution strategy on multimodal test functions, in: International Conference on Parallel Problem Solving from Nature, Springer

    Hansen, N., Kern, S., 2004. Evaluating the cma evolution strategy on multimodal test functions, in: International Conference on Parallel Problem Solving from Nature, Springer. pp. 282–291

  26. [34]

    Hansen, N., Ostermeier, A., 1996. Adapting arbitrary normal mutation distribu- tions in evolution strategies: The covariance matrix adaptation, in: Evolutionary Computation, 1996., Proceedings of IEEE International Conference on, IEEE. pp. 312–317

  27. [35]

    Invariant bayesian estimation on manifolds

    Jermyn, I., 2005. Invariant bayesian estimation on manifolds. Ann. Stat. 33, 583–605

  28. [36]

    A uni/f_ied view of the theory of directional statistics, 1975-1988

    Jupp, P., Mardia, K., 1989. A uni/f_ied view of the theory of directional statistics, 1975-1988. International Statistical Review/Revue Internationale de Statistique , 261–294

  29. [37]

    Kern, S., Müller, S.D., Hansen, N., Büche, D., Ocenasek, J., Koumoutsakos, P.,

  30. [38]

    Geometry of diffeomorph- ism groups, complete integrability and geometric statistics

    Khesin, B., Lenells, J., Misiołek, G., Preston, S., 2013. Geometry of diffeomorph- ism groups, complete integrability and geometric statistics. Geometric and Func- tional Analysis 23, 334–366

  31. [39]

    Über riemannsche mannigfaltigkeiten mit positiver krümmung

    Klingenberg, W., 1961. Über riemannsche mannigfaltigkeiten mit positiver krümmung. Commentarii Mathematici Helvetici 35, 47–54

  32. [40]

    Estimation of distribution algorithms: A new tool for evolutionary computation

    Larrañaga, P., Lozano, J.A., 2001. Estimation of distribution algorithms: A new tool for evolutionary computation. volume 2. Springer Science & Business Me- dia

  33. [41]

    The riemannian structure of euclidean shape spaces: a novel environment for statistics

    Le, H., Kendall, D.G., et al., 1993. The riemannian structure of euclidean shape spaces: a novel environment for statistics. The Annals of Statistics 21, 1225– 1271

  34. [42]

    Learning riemannian metrics, in: Proceedings of the Nine- teenth conference on Uncertainty in Arti/f_icial Intelligence, Morgan Kaufmann Publishers Inc

    Lebanon, G., 2002. Learning riemannian metrics, in: Proceedings of the Nine- teenth conference on Uncertainty in Arti/f_icial Intelligence, Morgan Kaufmann Publishers Inc.. pp. 362–369

  35. [43]

    Introduction to smooth manifolds

    Lee, J.M., 2001. Introduction to smooth manifolds. Springer. 80

  36. [44]

    Riemannian manifolds: an introduction to curvature

    Lee, J.M., 2006. Riemannian manifolds: an introduction to curvature. volume

  37. [45]

    Gradient /f_low of the stochastic relaxation on a generic exponential family, in: AIP Conference Proceedings, AIP

    Malagò, L., Pistone, G., 2015. Gradient /f_low of the stochastic relaxation on a generic exponential family, in: AIP Conference Proceedings, AIP. pp. 353–360

  38. [46]

    Statistics of directional data

    Mardia, K.V., 1975. Statistics of directional data. Journal of the Royal Statistical Society: Series B (Methodological) 37, 349–371. URL: https:// rss.onlinelibrary.wiley.com/doi/abs/10.1111/j.2517-6161.1975.tb01550.x, doi:doi: 10.1111/j.2517-6161.1975.tb01550.x

  39. [47]

    Any statistical manifold has a contrast function—on the c3-functions taking the minimum at the diagonal of the product manifold

    Matumoto, T., 1993. Any statistical manifold has a contrast function—on the c3-functions taking the minimum at the diagonal of the product manifold. Hiroshima Math. J. 23, 327–332. URL: https://projecteuclid.org:443/euclid.hmj/ 1206128255

  40. [48]

    Manifolds of differentiable mappings

    Michor, P.W., 1980. Manifolds of differentiable mappings. volume 3. Birkhauser

  41. [49]

    On the betti numbers of real varieties

    Milnor, J., 1964. On the betti numbers of real varieties. Proceedings of the Amer- ican Mathematical Society 15, 275–280

  42. [50]

    Numerical optimization

    Nocedal, J., Wright, S., 2006. Numerical optimization. Springer Science & Busi- ness Media

  43. [51]

    Affine differential geometry: geometry of affine immersions

    Nomizu, K., Sasaki, T., 1994. Affine differential geometry: geometry of affine immersions. Cambridge University Press

  44. [52]

    On an intrinsic analysis of statistical estimation, in: Multivariate Analysis: Future Directions 2

    Oller, J.M., 1993. On an intrinsic analysis of statistical estimation, in: Multivariate Analysis: Future Directions 2. Elsevier, pp. 421–437

  45. [53]

    Information-geometric op- timization algorithms: A unifying picture via invariance principles

    Ollivier, Y., Arnold, L., Auger, A., Hansen, N., 2017. Information-geometric op- timization algorithms: A unifying picture via invariance principles. The Journal of Machine Learning Research 18, 564–628

  46. [54]

    Probabilities and statistics on Riemannian manifolds: A geo- metric approach

    Pennec, X., 2004. Probabilities and statistics on Riemannian manifolds: A geo- metric approach. Ph.D. thesis. INRIA

  47. [55]

    Intrinsic statistics on riemannian manifolds: Basic tools for geometric measurements

    Pennec, X., 2006. Intrinsic statistics on riemannian manifolds: Basic tools for geometric measurements. Journal of Mathematical Imaging and Vision 25, 127

  48. [56]

    Riemannian geometry

    Petersen, P., 2006. Riemannian geometry. volume 171. Springer

  49. [57]

    Geometry of leaves

    Phillips, A., Sullivan, D., 1981. Geometry of leaves. Topology 20, 209–218

  50. [58]

    Geometry of PDEs and mechanics

    Prástaro, A., 1996. Geometry of PDEs and mechanics. World Scienti/f_ic

  51. [59]

    Principles of mathematical analysis

    Rudin, W., et al., 1964. Principles of mathematical analysis. volume 3. McGraw- hill New York

  52. [60]

    Riemannian Geometry

    Sakai, T., 1996. Riemannian Geometry. volume 149. American Mathematical Soc

  53. [61]

    The geometry of Hessian structures

    Shima, H., 2007. The geometry of Hessian structures. World Scienti/f_ic. 81

  54. [62]

    Geometry of Hessian manifolds

    Shima, H., Yagi, K., 1997. Geometry of Hessian manifolds. Differential Geometry and its Applications 7, 277–290

  55. [63]

    The statistical theory of shape

    Small, C.G., 2012. The statistical theory of shape. Springer Science & Business Media

  56. [64]

    A comprehensive introduction to differential geometry, publish or perish

    Spivak, M., 1979. A comprehensive introduction to differential geometry, publish or perish

  57. [65]

    Virtual library of simulation experiments: Test functions and datasets

    Surjanovic, S., Bingham, D., 2017. Virtual library of simulation experiments: Test functions and datasets. Retrieved July 3, 2019, from http://www.sfu.ca/~ssurjano

  58. [66]

    Su una congettura di nash

    Tognoli, A., 1973. Su una congettura di nash. Annali della Scuola Normale Superiore di Pisa-Classe di Scienze 27, 167–185

  59. [67]

    Riemannian consensus for manifolds with bounded curvature

    Tron, R., Afsari, B., Vidal, R., 2013. Riemannian consensus for manifolds with bounded curvature. IEEE Transactions on Automatic Control 58, 921–934

  60. [68]

    The self-intersections of a smooth n-manifold in 2n-space

    Whitney, H., 1944a. The self-intersections of a smooth n-manifold in 2n-space. Annals of Math 45, 180

  61. [69]

    The singularities of a smooth n-manifold in (2n- 1)-space

    Whitney, H., 1944b. The singularities of a smooth n-manifold in (2n- 1)-space. Ann. of Math 45, 247–293

  62. [70]

    Fitness expectation maximization, in: International Conference on Parallel Problem Solving from Nature, Springer

    Wierstra, D., Schaul, T., Peters, J., Schmidhuber, J., 2008a. Fitness expectation maximization, in: International Conference on Parallel Problem Solving from Nature, Springer. pp. 337–346

  63. [71]

    Natural evolution strategies, in: 2008 IEEE Congress on Evolutionary Computation (IEEE World Congress on Computational Intelligence), IEEE

    Wierstra, D., Schaul, T., Peters, J., Schmidhuber, J., 2008b. Natural evolution strategies, in: 2008 IEEE Congress on Evolutionary Computation (IEEE World Congress on Computational Intelligence), IEEE. pp. 3381–3387

  64. [72]

    On the convergence of a class of estimation of distribution algorithms

    Zhang, Q., Muhlenbein, H., 2004. On the convergence of a class of estimation of distribution algorithms. IEEE Transactions on evolutionary computation 8, 127–136. 82 Appendix A. Riemannian Adaptation of Optimization Algorithms in the lit- erature In this section we illustrate ...

  65. [75]

    Rx(⃗0) =x, where⃗0∈TxM and

  66. [76]

    personal best

    DRx(⃗0) = idTxM . This establishes the sufficient conditions to locally preserve the gradient atx∈M. This generalization is similar to the use of orientation-preserving diffeomorphism in Section 4.2 85 Appendix A.3. Riemannian Particle Swarm Optimizaiton The adaptation of Particl...

  67. [176]

    Springer Science & Business Media

  68. [2004]

    Natural Computing 3, 77–112

    Learning probability distributions in continuous evolutionary algorithms– a comparative review. Natural Computing 3, 77–112

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.