Pith. sign in

REVIEW 3 major objections 4 minor 51 references

K$^2$IE: Kernel Method-based Kernel Intensity Estimators for Inhomogeneous Poisson Processes

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Penalized least squares in an RKHS yields a kernel intensity estimator with all unit weights.

desk verdict Useful equivalence result, but the self-contained proof in Appendix B has an algebraic error in the gradient and the main proof uses a formal inverse operator; the theorem is true and fixable. read the letter →

arxiv 2505.24704 v1 pith:XBCSP37Z submitted 2025-05-30 stat.ML cs.LG

Hideaki Kim , Tomoharu Iwata , Akinori Fujino This is my paper
classification stat.MLcs.LG MSC 60G5562G0562M30
keywords K2IEinhomogeneousPoissonprocesskernelintensityestimatorreproducingHilbertspaceleastsquareslossequivalentRKHSrepresentertheoremrandomFourierfeatures
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

The paper tries to close the gap between two families of intensity estimators for inhomogeneous Poisson processes: classical kernel intensity estimators (KIEs), which add up smoothing kernels with unit weights, and RKHS-based estimators, which must fit dual coefficients by optimization. It introduces a penalized least squares objective with the intensity restricted to an RKHS and proves a specialized representer theorem: the optimal dual coefficient is identically one, so the estimator is exactly $\hat{\lambda}(x)=\sum_{n=1}^N h(x,x_n)$, where the equivalent RKHS kernel $h$ solves a Fredholm integral equation. Because the solution needs no dual-coefficient optimization, fitting reduces to solving that integral equation once, matching the speed of a classical KIE while keeping the RKHS edge-correction mechanism. If the result is right, the two 'kernel' traditions are the same method seen from different sides, and RKHS tools such as random features and closed-form integrals become available to classical KIE practice. The paper also reports synthetic experiments showing predictive accuracy on par with the standard kernel method-based estimator at a fraction of the fitting cost.

What carries the argument

The carrying object is the equivalent RKHS kernel $h$, which solves the Fredholm integral equation of the second kind $(1/\gamma)h(x,x')+\int_{\mathcal{X}} k(x,s)h(s,x')ds=k(x,x')$; it plays the role of an edge-corrected smoothing kernel that automatically accounts for the finite observation domain. The argument also rests on the path-integral representation of the RKHS norm as the integral of $\lambda(x)\lambda(s)$ against the inverse of the kernel integral operator, and on a degenerate random Fourier feature expansion that turns the integral equation into a closed-form $2M\times 2M$ linear system. The Mercer-based proof in the appendix avoids the unbounded operator inversion, which matters for the rigor of the theorem.

What would settle it

Solve the same penalized least squares problem by brute-force numerical minimization on a fine grid over a compact domain with a Gaussian kernel and fixed $\gamma$, and compare the grid solution to $\sum_n h(x,x_n)$ obtained by numerically solving equation (8); any discrepancy beyond discretization error would falsify Theorem 1. For the extrapolation claim, fit K2IE on data restricted to a subdomain and evaluate the estimator outside the training region against a brute-force minimizer on that extended region.

Watch

Extended reading notes

Core claim

On its own terms, the paper's discovery is Theorem 1: the solution of the penalized least squares problem $\min_{\lambda \in \mathcal{H}_k}\left\{-2\sum_{n=1}^N \lambda(x_n)+\int_{\mathcal{X}} \lambda(x)^2 dx+\frac{1}{\gamma}\|\lambda\|^2_{\mathcal{H}_k}\right\}$ is $\hat{\lambda}(x)=\sum_{n=1}^N h(x,x_n)$, where $h$ is the equivalent RKHS kernel defined by $(1/\gamma)h(x,x')+\int_{\mathcal{X}} k(x,s)h(s,x')ds=k(x,x')$. Thus the least squares RKHS estimator coincides with a classical kernel intensity estimator with unit weights, and the paper claims this is the first representer theorem for the penalized least squares Poisson intensity problem. The proof runs through a path-integral representation of the RKHS norm, with a Mercer-based alternative given in the appendix, and the paper extends the statement to extrapolation by broadening the support of the kernel.

Load-bearing premise

The main-text proof treats the RKHS norm as an integral against the inverse of the kernel integral operator and freely inverts integral operators, so on a compact domain with a Gaussian or Matérn kernel — where those inverses are unbounded or distributional — the derivation is formal; only the Mercer-based appendix gives a rigorous path, and the extrapolation extension again assumes an operator inversion that is not justified.

Editorial extensions

If this is right

  • Under least squares loss, the RKHS intensity estimator is exactly a classical kernel intensity estimator with unit weights, so no gradient-based dual fitting is needed.
  • Edge correction becomes a property of the equivalent kernel $h$ rather than an explicit rescaling, which lets K2IE handle disjoint or irregular observation domains without extra correction terms.
  • With random Fourier features, the integral of the squared intensity $\int_{\mathcal{X}} \left(\sum_n h(x,x_n)\right)^2 dx$ is available in closed form, removing the need for Monte Carlo integration during cross-validation and prediction.
  • K2IE matches the accuracy of the standard kernel method-based estimator while being roughly an order of magnitude faster in the reported experiments.
  • Non-negativity of the estimated intensity is not guaranteed in general; unless the RKHS kernel is a non-negative inverse M-kernel, negative estimates in sparse regions require post-hoc clipping.

Reading between the lines

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

  • The unit-coefficient phenomenon likely extends to any loss whose gradient is affine in $\lambda$; if so, analogous 'equal-weight' representer theorems may connect RKHS methods to other classical smoothers, such as spline and density estimators.
  • On infinite domains, the equivalent kernel's Fourier form $\tilde{h} = \tilde{k}/(\gamma^{-1}+\tilde{k})$ implicitly defines a family of smoothing kernels tuned for squared-error loss; testing these against standard Gaussian smoothing kernels in ordinary kernel density estimation would probe the practical value of the identity outside point processes.
  • If the extrapolation result in Appendix C can be made rigorous, it would give a principled way to forecast intensities beyond the observation window, a regime where classical KIEs are undefined without a boundary model.
  • The closed-form cross-validation integral suggests that leave-one-out or $k$-fold criteria could be computed exactly for K2IE, making fully deterministic hyperparameter selection possible without stochastic thinning.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper proposes a penalized least-squares estimator for the intensity function of an inhomogeneous Poisson process, with the intensity modeled directly as an element of an RKHS. The central claim (Theorem 1) is that the minimizer has a representer-type expansion with all dual coefficients equal to one, i.e., the estimator coincides with a classical kernel intensity estimator using the equivalent RKHS kernel h(x,x') defined by the Fredholm integral equation (1/γ)h + ∫_X k h = k. The authors derive this result using a path-integral representation of the RKHS norm, provide an alternative Mercer-based proof in Appendix B, develop a random-Fourier-feature approximation for finite domains with closed-form edge-correction matrices, and report experiments on synthetic and real datasets showing competitive accuracy and large computational savings relative to the previous Flaxman et al. estimator.

Significance. If rigorously established, the equivalence between an RKHS least-squares estimator and a classical kernel intensity estimator with unit weights is a conceptually valuable and practically useful result: it connects two lines of intensity-estimation methodology, removes the need for dual-coefficient optimization, and enables analytic integration of the estimated intensity over subregions. The random-feature construction with closed-form A-matrices is a concrete algorithmic contribution, and the paper includes an ablation study, a real-world experiment, and a code link, all of which support reproducibility. However, the proof of the central theorem is not currently in a publishable state: the main-text proof is formal, and the supposedly rigorous Appendix B proof contains an incorrect gradient computation. The theorem itself is true and is easily repaired, but the manuscript must supply a correct, self-contained proof before the central claim can be accepted.

major comments (3)
  1. [Section 3.1, proof of Theorem 1] The proof assumes the existence of an inverse integral operator K* with kernel k* for K·(x)=∫_X k(x,s)·(s)ds, and then uses the functional calculus with q*(x,s)=δ(x-s)+γ^{-1}k*(x,s). On a compact domain X with a continuous kernel such as the Gaussian or Matérn kernel, K is a compact operator whose eigenvalues tend to zero, so K* is unbounded and generally cannot be represented as an ordinary kernel function; the Dirac-δ manipulations in the proof are therefore formal. Since this is the only proof of Theorem 1 in the main text, the authors must either state and justify conditions under which the inverse-operator calculus is legitimate or replace the argument with the corrected Mercer proof (or a resolvent-based proof).
  2. [Appendix B, proof of Theorem 1] The gradient computation in Appendix B is incorrect. With Q_{nn'}=q(x_n,x_{n'}) and F(α)=-2∑_n λ(x_n)+||λ||²_{H_q}=-2 1^T Q α + α^T Q α, the correct derivative is ∂F/∂α_n = -2(Q1)_n + 2(Qα)_n, not -2∑_{n'}q(x_{n'},x_n)+2α_n∑_{n'}q(x_{n'},x_n) as printed. The printed stationarity equation is not the derivative of the objective; it was written so that α_n=1 appears to follow immediately. The correct stationarity condition is Q(α-1)=0, and since Q is positive definite for distinct data points, α=1 follows. Thus the theorem is true, but the only self-contained rigorous proof in the manuscript is invalid as written and must be corrected.
  3. [Appendix C, Proposition 2 (extrapolation)] The proof of the extrapolation extension again invokes inverse operators K* and H on R^d, while the objective in (11) and the kernel h* in (C2) contain the indicator 1_{s∈X}. The existence, mapping properties, and kernel representation of these inverse operators are not established, and the distributional manipulation with δ(x-s)1_{s∈X} is not justified. The claim may be true and provable by a resolvent or Fredholm-integral argument, but the manuscript does not supply such a proof. Please provide a rigorous proof of Proposition 2 or state precise conditions under which (C1) has a solution h for which the representer formula holds on all of R^d.
minor comments (4)
  1. [Section 4.2] The sentence "Figure 2 displays some estimation results on λ^{1.0}_{2D}(x) and λ^{0.8}_{2D}(x)" should refer to Figure 3, which is the figure showing 2D results; Figure 2 shows 1D estimates.
  2. [Section 2.1] There is a typo in "Siverman's rule-of-thumb"; it should be "Silverman's rule-of-thumb".
  3. [Equation (20)] The displayed formula for |L| appears to be missing an absolute-value bar around the integrand; please check that the notation is typeset correctly.
  4. [Section 3.2.1] The Fourier-domain expression in (14) is stated for shift-invariant kernels but the text does not specify the required integrability/decay conditions on k for the Fourier transform and the ratio k̃(ω)/(γ^{-1}+k̃(ω)) to be well defined; adding a short remark would be helpful.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the least-squares objective and the independent Mercer proof in Appendix B determine the unit coefficients; hyperparameters are cross-validated, not fitted.

full rationale

The central derivation is not circular. Theorem 1 solves the explicit penalized least-squares problem (11), and the equivalent kernel h is defined independently by the Fredholm equation (8); the data enter only through the sum over h(x,x_n). The unit dual coefficient follows from stationarity of the quadratic objective, not from a fitted parameter: in the Mercer proof of Appendix B, minimization over H_q yields alpha=1 because the normal equations reduce to Q alpha = Q 1 with Q positive definite, so the conclusion is a mathematical consequence of the objective rather than an input. The main-text proof invokes the path-integral representation from Kim (2021), a self-citation, but that citation is not load-bearing because Appendix B supplies a self-contained proof via Mercer's theorem and the classical representer theorem, without relying on the path-integral result. Hyperparameters beta and gamma are tuned by 5-fold cross-validation, which is standard model selection and does not amount to fitting the claimed unit coefficients. The acknowledged non-negativity limitation and the max(u,0) post-hoc correction do not smuggle in the theorem. The remaining concerns about formal inverse-operator calculus and the displayed gradient in Appendix B are rigor or correctness issues, not circularity; no prediction in the paper reduces by construction to its own inputs.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The method introduces no new entities; the equivalent kernel h is derived from the input kernel k and the domain. The free parameters are standard hyperparameters, while the main formal risk is the unbounded inverse operator used in the primary proof.

free parameters (3)
  • regularization hyperparameter gamma = grid search over [0.1, 100] (synthetic) or [0.001, 1] (real) with 5-fold CV
    Controls the trade-off between data fit and RKHS norm; directly defines the equivalent kernel h and is selected on data.
  • inverse scale hyperparameter beta = grid search over [0.1, 100] * beta_bar
    Bandwidth of the Gaussian kernel; chosen by cross-validation and affects the estimated intensity.
  • number of random features 2M = 500 (ablation in Appendix D)
    Determines approximation accuracy of the equivalent kernel; fixed by hand, with an ablation showing performance improves with M.
assumptions (4)
  • standard math Mercer's theorem: continuous positive semidefinite kernel on compact X has an eigenfunction expansion with orthonormal eigenfunctions in L2(X).
    Used in the alternative proof of Theorem 1 (Appendix B).
  • domain assumption The intensity function lambda is assumed to lie directly in the RKHS H_k.
    This is the modeling assumption in Eq. (11); it is what makes the solution non-negative only by accident.
  • domain assumption The first variation of the objective functional is well-defined and the minimizer satisfies the Euler-Lagrange equation.
    The path integral proof in Section 3.1 uses functional derivatives without stating smoothness or compactness conditions.
  • ad hoc to paper The inverse operator K^* of the integral operator K exists (as used in the path integral representation).
    The main proof states 'Let K* be its inverse operator'; for compact operators on L2(X) the inverse is unbounded, making this a formal assumption.

how reviews work

0 comments
Cite this review

Pith. "Pith review of K$^2$IE: Kernel Method-based Kernel Intensity Estimators for Inhomogeneous Poisson Processes." pith.science (2026). https://pith.science/paper/XBCSP37Z

@misc{pith2026250524704,
  author       = {Pith},
  title        = {Pith review of: K$^2$IE: Kernel Method-based Kernel Intensity Estimators for Inhomogeneous Poisson Processes},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XBCSP37Z}},
  note         = {Machine review of arXiv:2505.24704}
}
abstract

Kernel method-based intensity estimators, formulated within reproducing kernel Hilbert spaces (RKHSs), and classical kernel intensity estimators (KIEs) have been among the most easy-to-implement and feasible methods for estimating the intensity functions of inhomogeneous Poisson processes. While both approaches share the term "kernel", they are founded on distinct theoretical principles, each with its own strengths and limitations. In this paper, we propose a novel regularized kernel method for Poisson processes based on the least squares loss and show that the resulting intensity estimator involves a specialized variant of the representer theorem: it has the dual coefficient of unity and coincides with classical KIEs. This result provides new theoretical insights into the connection between classical KIEs and kernel method-based intensity estimators, while enabling us to develop an efficient KIE by leveraging advanced techniques from RKHS theory. We refer to the proposed model as the kernel method-based kernel intensity estimator (K$^2$IE). Through experiments on synthetic datasets, we show that K$^2$IE achieves comparable predictive performance while significantly surpassing the state-of-the-art kernel method-based estimator in computational efficiency.

Figures

Figures reproduced from arXiv: 2505.24704 by the authors.

Figure 1
Figure 1. Examples of smoothing kernels g(x, xn), smoothing kernels with edge-correction g(x, xn)/ν(x), and equivalent kernels h(x, xn) for data points xn ∈ {−1.8, 0, 1.8}. Gaussian, Laplace, and Cauchy RKHS/smoothing kernels are e −|x−xn| 2 , e −|x−xn| , and 1 1+|x−xn| 2 , respectively. The regularization hyper-parameter, the number of random features, and the observation domain were set as (γ, 2M, X) = (2, 500, [−2, 2]). re… view at source ↗
Figure 2
Figure 2. Examples of the estimated intensity functions on 1 [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Examples of the estimated intensity functions on 2 [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

51 extracted references · 50 canonical work pages

  1. [1]

    P., Murray, I., and MacKay, D

    Adams, R. P., Murray, I., and MacKay, D. J. Tractable nonparametric B ayesian inference in P oisson processes with G aussian process intensities. In International Conference on Machine Learning , pp.\ 9--16, 2009

    work page 2009

  2. [2]

    V., Damoulas, T., and Cripps, S

    Aglietti, V., Bonilla, E. V., Damoulas, T., and Cripps, S. Structured variational inference in continuous C ox process models. In Advances in Neural Information Processing Systems 32 , 2019

    work page 2019

  3. [3]

    A personal perspective on the history of the numerical analysis of F redholm integral equations of the second kind

    Atkinson, K. A personal perspective on the history of the numerical analysis of F redholm integral equations of the second kind. In The Birth of Numerical Analysis , pp.\ 53--72. World Scientific, 2010

    work page 2010

  4. [4]

    Sparse and low-rank multivariate H awkes processes

    Bacry, E., Bompaire, M., Ga \" ffas, S., and Muzy, J.-F. Sparse and low-rank multivariate H awkes processes. Journal of Machine Learning Research , 21 0 (50): 0 1--32, 2020

    work page 2020

  5. [5]

    Latent network structure learning from high-dimensional multivariate point processes

    Cai, B., Zhang, J., and Guan, Y. Latent network structure learning from high-dimensional multivariate point processes. Journal of the American Statistical Association , 119 0 (545): 0 95--108, 2024

    work page 2024

  6. [6]

    G., Bradburn, M

    Clark, T. G., Bradburn, M. J., Love, S. B., and Altman, D. G. Survival analysis part i: basic concepts and first analyses. British Journal of Cancer , 89 0 (2): 0 232--238, 2003

    work page 2003

  7. [7]

    Cox, D. R. Regression models and life-tables. Journal of the Royal Statistical Society: Series B (Methodological) , 34 0 (2): 0 187--202, 1972

    work page 1972

  8. [8]

    Cronie, O., Moradi, M., and Biscio, C. A. A cross-validation-based statistical theory for point processes. Biometrika , 111 0 (2): 0 625--641, 2024

    work page 2024

Show all 51 references
  1. [9]

    P., Byron, M

    Cunningham, J. P., Byron, M. Y., Shenoy, K. V., and Sahani, M. Inferring neural firing rates from spike trains using G aussian processes. In Advances in Neural Information Processing Systems 20 , 2007

    2007

  2. [10]

    Daley, D. J. and Vere-Jones, D. An Introduction to the Theory of Point Processes . Springer-Verlag, New York, 1988

    1988

  3. [11]

    A., Lii, K.-S., and Politis, D

    Davis, R. A., Lii, K.-S., and Politis, D. N. Remarks on some nonparametric estimates of a density function. Selected Works of Murray Rosenblatt , pp.\ 95--100, 2011

    2011

  4. [12]

    A kernel method for smoothing point process data

    Diggle, P. A kernel method for smoothing point process data. Journal of the Royal Statistical Society: Series C (Applied Statistics) , 34 0 (2): 0 138--147, 1985

    1985

  5. [13]

    J., Moraga, P., Rowlingson, B., and Taylor, B

    Diggle, P. J., Moraga, P., Rowlingson, B., and Taylor, B. M. Spatial and spatio-temporal log- G aussian C ox processes: extending the geostatistical paradigm. Statistical Science , 28 0 (4): 0 542--563, 2013

    2013

  6. [14]

    and Opper, M

    Donner, C. and Opper, M. Efficient B ayesian inference of sigmoidal G aussian C ox processes. Journal of Machine Learning Research , 19: 0 1--34, 2018

    2018

  7. [15]

    W., and Sejdinovic, D

    Flaxman, S., Teh, Y. W., and Sejdinovic, D. Poisson intensity estimation with reproducing kernels. In Artificial Intelligence and Statistics , pp.\ 270--279. PMLR, 2017

    2017

  8. [16]

    C., Bailey, T

    Gatrell, A. C., Bailey, T. C., Diggle, P. J., and Rowlingson, B. S. Spatial point pattern analysis and its application in geographical epidemiology. Transactions of the Institute of British Geographers , pp.\ 256--274, 1996

    1996

  9. [17]

    A., and Roberts, S

    Gunter, T., Lloyd, C., Osborne, M. A., and Roberts, S. J. Efficient B ayesian nonparametric modelling of structured point processes. In Uncertainty in Artificial Intelligence , 2014

    2014

  10. [18]

    R., Reynaud-Bouret, P., and Rivoirard, V

    Hansen, N. R., Reynaud-Bouret, P., and Rivoirard, V. Lasso and probabilistic inequalities for multivariate point processes. Bernoulli , 21 0 (1): 0 83--143, 2015

    2015

  11. [19]

    and Arjas, E

    Heikkinen, J. and Arjas, E. Modeling a P oisson forest in variable elevations: A nonparametric B ayesian approach. Biometrics , 55 0 (3): 0 738--745, 1999

    1999

  12. [20]

    A simple sequentially rejective multiple test procedure

    Holm, S. A simple sequentially rejective multiple test procedure. Scandinavian Journal of Statistics , pp.\ 65--70, 1979

    1979

  13. [21]

    Hubbel, S. P. and Foster, R. B. Diversity of canopy trees in a neotropical forest and implications for conservation. Tropical Rain Forest: Ecology and Management , pp.\ 25--41, 1983

    1983

  14. [22]

    John, S. T. and Hensman, J. Large-scale C ox process inference using variational F ourier features. In International Conference on Machine Learning , volume 80, pp.\ 2362--2370. PMLR, 2018

    2018

  15. [23]

    Jones, M. C. Simple boundary correction for kernel density estimation. Statistics and Computing , 3: 0 135--146, 1993

    1993

  16. [24]

    Fast B ayesian inference for G aussian C ox processes via path integral formulation

    Kim, H. Fast B ayesian inference for G aussian C ox processes via path integral formulation. In Advances in Neural Information Processing Systems 34 , 2021

    2021

  17. [25]

    Inverse M -kernels for linear universal approximators of non-negative functions

    Kim, H. Inverse M -kernels for linear universal approximators of non-negative functions. In Advances in Neural Information Processing Systems 37 , 2024

    2024

  18. [26]

    Fast B ayesian estimation of point process intensity as function of covariates

    Kim, H., Asami, T., and Toda, H. Fast B ayesian estimation of point process intensity as function of covariates. In Advances in Neural Information Processing Systems 35 , 2022

    2022

  19. [27]

    Kingma, D. P. and Ba, J. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 , 2014

    2014 arXiv

  20. [28]

    and Kozlowski, J

    Kowalczuk, Z. and Kozlowski, J. Integrated squared error and integrated absolute error in recursive identification of continuous-time plants. In UKACC International Conference on Control'98 (Conf. Publ. No. 455) , volume 1, pp.\ 693--698. IET, 1998

    1998

  21. [29]

    Lai, C. D. and Xie, M. Stochastic Ageing and Dependence for Reliability . Springer Science & Business Media, 2006

    2006

  22. [30]

    and Gy o rffy, B

    L \'a nczky, A. and Gy o rffy, B. Web-based survival analysis tool tailored for medical research (kmplot): development and implementation. Journal of Medical Internet Research , 23 0 (7): 0 e27633, 2021

    2021

  23. [31]

    Variational inference for G aussian process modulated P oisson processes

    Lloyd, C., Gunter, T., Osborne, M., and Roberts, S. Variational inference for G aussian process modulated P oisson processes. In International Conference on Machine Learning , volume 37, pp.\ 1814--1822. PMLR, 2015

    2015

  24. [32]

    Mercer, J. Xvi. functions of positive and negative type, and their connection the theory of integral equations. Philosophical Transactions of the Royal Society of London. Series A, containing papers of a mathematical or physical character , 209 0 (441-458): 0 415--446, 1909

    1909

  25. [33]

    R., and Waagepetersen, R

    M ller, J., Syversveen, A. R., and Waagepetersen, R. P. Log G aussian C ox processes. Scandinavian Journal of Statistics , 25 0 (3): 0 451--482, 1998

    1998

  26. [34]

    Statistical models for earthquake occurrences and residual analysis for point processes

    Ogata, Y. Statistical models for earthquake occurrences and residual analysis for point processes. Journal of the American Statistical Association , 83 0 (401): 0 9--27, 1988

    1988

  27. [35]

    On estimation of a probability density function and mode

    Parzen, E. On estimation of a probability density function and mode. The Annals of Mathematical Statistics , 33 0 (3): 0 1065--1076, 1962

    1962

  28. [36]

    Polyanin, A. D. and Manzhirov, A. V. Handbook of Integral Equations . CRC press, 1998

    1998

  29. [37]

    and Recht, B

    Rahimi, A. and Recht, B. Random features for large-scale kernel machines. In Advances in Neural Information Processing Systems 20 , 2007

    2007

  30. [38]

    Smoothing counting process intensities by means of kernel functions

    Ramlau-Hansen, H. Smoothing counting process intensities by means of kernel functions. The Annals of Statistics , pp.\ 453--466, 1983

    1983

  31. [39]

    Rathbun, S. L. and Cressie, N. Asymptotic properties of estimators for the parameters of spatial inhomogeneous P oisson point processes. Advances in Applied Probability , 26 0 (1): 0 122--154, 1994

    1994

  32. [40]

    and Smola, A

    Scholkopf, B. and Smola, A. J. Learning with Kernels: Support Vector Machines, Regularization, Optimization, and Beyond . MIT press, 2018

    2018

  33. [41]

    Sch \"o lkopf, B., Herbrich, R., and Smola, A. J. A generalized representer theorem. In International Conference on Computational Learning Theory , pp.\ 416--426. Springer, 2001

    2001

  34. [42]

    and Dellaportas, P

    Sellier, J. and Dellaportas, P. Sparse spectral B ayesian permanental process with generalized kernel. In International Conference on Artificial Intelligence and Statistics , pp.\ 2769--2791. PMLR, 2023

    2023

  35. [43]

    and Cristianini, N

    Shawe-Taylor, J. and Cristianini, N. Kernel Methods for Pattern Analysis . Cambridge University Press, 2004

    2004

  36. [44]

    Silverman, B. W. Density Estimation for Statistics and Data Analysis . Routledge, 2018

    2018

  37. [45]

    Teng, M., Nathoo, F., and Johnson, T. D. Bayesian computation for log- G aussian C ox processes: A comparative analysis of methods. Journal of Statistical Computation and Simulation , 87: 0 2227--2252, 2017

    2017

  38. [46]

    S., and Sejdinovic, D

    Tsuchida, R., Ong, C. S., and Sejdinovic, D. Exact, fast and expressive P oisson point processes via squared neural families. In Proceedings of the AAAI Conference on Artificial Intelligence , volume 38, pp.\ 20559--20566, 2024

    2024

  39. [47]

    Empirical Processes in M-estimation , volume 6

    van de Geer, S. Empirical Processes in M-estimation , volume 6. Cambridge University Press, 2000

    2000

  40. [48]

    Spline Models for Observational Data , volume 59

    Wahba, G. Spline Models for Observational Data , volume 59. SIAM, 1990

    1990

  41. [49]

    Walder, C. J. and Bishop, A. N. Fast B ayesian intensity estimation for the permanental process. In International Conference on Machine Learning , volume 70, pp.\ 3579--3588. PMLR, 2017

    2017

  42. [50]

    and Seeger, M

    Williams, C. and Seeger, M. Using the N ystr \"o m method to speed up kernel machines. In Advances in Neural Information Processing Systems 13 , 2000

    2000

  43. [51]

    Quasi- M onte C arlo feature maps for shift-invariant kernels

    Yang, J., Sindhwani, V., Avron, H., and Mahoney, M. Quasi- M onte C arlo feature maps for shift-invariant kernels. In International Conference on Machine Learning , pp.\ 485--493. PMLR, 2014

    2014

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.