Local Aggregate Multiscale Process (LAMP)

• LAMP is a multiscale spatial modeling framework that aggregates local models instead of relying on global covariance functions.
• The method uses sequential holdout validation to first capture large-scale patterns before incorporating finer local structures.
• It offers scalable predictive performance and uncertainty quantification, enabling seamless integration with machine learning algorithms like random forests and neural networks.

Searching arXiv for the exact LAMP term and closely related spatial-process work to ground the article in current arXiv records. The Local Aggregate Multiscale Process (LAMP) is a spatial modeling framework proposed as a scalable, validation-based, machine-learning-compatible alternative to Gaussian processes (GPs)/kriging. Instead of defining spatial dependence through a global covariance function, LAMP represents a spatial process as a sum of scale-specific spatial components, and each component is itself obtained by aggregating many local models. The method is designed for large spatial datasets, for settings in which larger-scale patterns that are easier to learn are modeled first, followed by smaller-scale patterns, and for workflows in which holdout validation must be integrated directly with other machine learning algorithms such as random forests and neural networks (Murakami et al., 1 Oct 2025).

1. Definition and conceptual scope

LAMP is built around a distinction between covariance-based spatial modeling and locally weighted aggregation. In the paper’s formulation, whereas spatial dependence is typically modeled through covariance functions, LAMP relies on local weighting. The process is explicitly multiscale:

$$Z_{1:R}(s_i) = \sum_{r=1}^{R} Z_r(s_i), \qquad Z_r(s_i)\sim N(\hat z_r(s_i), \sigma_r^2(s_i)).$$

Here $r=1,\dots,R$ indexes spatial resolution, $Z_1$ is the largest-scale component, $Z_R$ is the smallest-scale component, and $R$ is not fixed in advance but learned by validation. The scale sequence is controlled by

$h_{r+1} = \delta h_r, \qquad 0<\delta<1,$

so progressively smaller bandwidths are used to represent progressively finer spatial structure (Murakami et al., 1 Oct 2025).

The phrase local aggregate has a precise operational meaning in this framework. At a given scale, LAMP first fits many local models centered at different kernel locations, and then aggregates those local models using a generalized product-of-experts type combination. The phrase multiscale means that this local-aggregation construction is repeated across a sequence of bandwidths, with the final process defined as the sum over all retained scales. This gives LAMP a hybrid character: it is both a spatial process model, because each scale carries a Gaussian predictive distribution, and an algorithmic predictor, because the number of scales and their contributions are selected by holdout validation rather than by explicit global covariance inversion (Murakami et al., 1 Oct 2025).

2. Mathematical formulation

The local model is defined for observations $y(s_i)$ at sites $s_i\in D\subset \mathbb R^2$. A local model centered at location $s_c$ is

$$y(s_i)\mid c \sim N\!\left(\mu_c,\ \frac{\sigma^2}{w_h(d_{i,c})}\right), \qquad \mu_c \sim N(0,\tau^2),$$

where $r=1,\dots,R$0 is the distance from site $r=1,\dots,R$1 to center $r=1,\dots,R$2, $r=1,\dots,R$3 is a spatial kernel weight, $r=1,\dots,R$4 is a local mean parameter, and $r=1,\dots,R$5 is a Gaussian prior variance for regularization. The paper considers the Gaussian kernel

$r=1,\dots,R$6

and the exponential kernel

$r=1,\dots,R$7

The penalized least-squares estimator of the local mean is

$r=1,\dots,R$8

The predictive distribution under a local model is written as

$r=1,\dots,R$9

with predictive variance

$Z_1$0

The same local model is also expressed in distributional form as

$Z_1$1

These local models are the elementary objects from which each scale is constructed (Murakami et al., 1 Oct 2025).

At scale $Z_1$2, LAMP uses $Z_1$3 local models centered at $Z_1$4. The scale-$Z_1$5 local model is

$Z_1$6

The scale-specific process is then obtained by aggregating the $Z_1$7 local models using generalized product of experts (gPoE):

$Z_1$8

The resulting aggregated process is again Gaussian,

$Z_1$9

with predictive mean

$Z_R$0

and predictive variance

$Z_R$1

The number of kernel centers is selected by a coverage heuristic. For a one-dimensional interval of length $Z_R$2,

$Z_R$3

and for a two-dimensional setting, if $Z_R$4 is the diagonal length of the bounding square containing all sample sites,

$Z_R$5

Kernel centers are chosen as sample sites closest to $Z_R$6-means cluster centers, and each local model is estimated using local samples satisfying

$Z_R$7

The full response model is

$Z_R$8

This places a linear term, an optional nonlinear machine-learning term, and the multiscale spatial process in a single additive predictor (Murakami et al., 1 Oct 2025).

3. Sequential estimation and validation-based training

The paper reports that simultaneous optimization of

$Z_R$9

was unstable and computationally inefficient. LAMP therefore uses a sequential learning procedure that estimates

$R$0

in that order. Larger-scale structure is treated as easier to learn, smaller-scale structure as progressively more local and complex, and the nonlinear term $R$1 as the most complex component. This suggests a residual-fitting or curriculum-learning view of the algorithm, although the paper’s formal description remains in terms of sequential holdout validation (Murakami et al., 1 Oct 2025).

For the linear model

$R$2

training uses two consecutive holdout-validation procedures. The training fraction is

$R$3

the validation fraction is $R$4, and predictive performance is scored by validation SSE:

$R$5

The first holdout-validation stage determines the multiscale process and the terminal resolution $R$6. It initializes with

$R$7

chooses $R$8 as a sufficiently large initial bandwidth, and repeatedly alternates between a regression update

$R$9

a residual update

$h_{r+1} = \delta h_r, \qquad 0<\delta<1,$0

and fitting a new scale-$h_{r+1} = \delta h_r, \qquad 0<\delta<1,$1 LAMP to those residuals. A newly added scale is accepted if it decreases validation SSE. Bandwidths are refined according to

$h_{r+1} = \delta h_r, \qquad 0<\delta<1,$2

with

$h_{r+1} = \delta h_r, \qquad 0<\delta<1,$3

and training stops when validation SSE fails to improve for

$h_{r+1} = \delta h_r, \qquad 0<\delta<1,$4

successive finer resolutions (Murakami et al., 1 Oct 2025).

Because the first stage is sequential rather than jointly optimized, the paper adds a second holdout-validation stage that rescales the learned scale contributions:

$h_{r+1} = \delta h_r, \qquad 0<\delta<1,$5

The parameters $h_{r+1} = \delta h_r, \qquad 0<\delta<1,$6 and $h_{r+1} = \delta h_r, \qquad 0<\delta<1,$7 are chosen by minimizing validation SSE for the adjusted sum of scales. If $h_{r+1} = \delta h_r, \qquad 0<\delta<1,$8 and $h_{r+1} = \delta h_r, \qquad 0<\delta<1,$9, then $y(s_i)$0 for all $y(s_i)$1, so no adjustment occurs (Murakami et al., 1 Oct 2025).

For nonlinear LAMP, the paper fits linear LAMP first and then adds a machine-learning term. In the implemented example, $y(s_i)$2 is a random forest. The resulting fitted predictor before the nonlinear augmentation is

$y(s_i)$3

A second validation step then tests whether adding $y(s_i)$4 improves predictive SSE. This is the basis of the claim that LAMP training is easily integrated with other machine learning algorithms (Murakami et al., 1 Oct 2025).

4. Prediction, uncertainty, and computational profile

Prediction at a new location $y(s_i)$5 proceeds by evaluating each scale-specific aggregate and summing them. For each scale,

$y(s_i)$6

and

$y(s_i)$7

The adjusted multiscale predictor is

$y(s_i)$8

Hence the final linear predictor is

$y(s_i)$9

and the final nonlinear predictor is

$s_i\in D\subset \mathbb R^2$0

This is the operational prediction rule used throughout the paper (Murakami et al., 1 Oct 2025).

Uncertainty quantification is handled by resampling-based predictive simulation rather than by exact posterior inference in the GP sense. For the linear model, the paper proposes drawing

$s_i\in D\subset \mathbb R^2$1

then drawing each scale as

$s_i\in D\subset \mathbb R^2$2

combining them as

$s_i\in D\subset \mathbb R^2$3

resampling residual noise $s_i\in D\subset \mathbb R^2$4 from fitted residuals, and finally forming

$s_i\in D\subset \mathbb R^2$5

For nonlinear LAMP, the corresponding predictive sample is

$s_i\in D\subset \mathbb R^2$6

The paper states that LAMP provides a pragmatic predictive uncertainty mechanism, but does not claim exact posterior inference for a globally coherent covariance model (Murakami et al., 1 Oct 2025).

The core computational claim is that LAMP avoids the main GP bottleneck, namely explicit inversion of the $s_i\in D\subset \mathbb R^2$7 covariance matrix. Local models at a given scale are independent during fitting, so that stage is readily parallelizable. Empirically, computation time grows approximately linearly with $s_i\in D\subset \mathbb R^2$8. In the simulation timing comparison, for $s_i\in D\subset \mathbb R^2$9, GP: 225.7 seconds and LAMP: 7.4 seconds. The paper also notes that LAMP was still slower than SPDE, NNGP, and GAM in their implementation, and suggests parallelization for very large samples (Murakami et al., 1 Oct 2025).

5. Empirical performance and substantive application

The linear Monte Carlo experiments use synthetic data generated from

$s_c$0

with $s_c$1. Sample sizes are

$s_c$2

and bandwidths are

$s_c$3

The paper interprets $s_c$4 as a small-scale process and $s_c$5 as a large-scale process. Baselines are GP, low-rank GP / GAM-like model, conjugate NNGP, SPDE, and LAMP, with performance measured by RMSE and MAE. The reported conclusion is that only LAMP predicted well in both large-scale and small-scale scenarios. For large-scale truth, GP performed best when feasible, but for $s_c$6, LAMP achieved the smallest RMSE and MAE. For small-scale truth, LAMP performed best overall, a result attributed to its sequential scale-learning procedure. In coefficient estimation, LAMP’s estimates stayed close to the truth, similarly to GP (Murakami et al., 1 Oct 2025).

The nonlinear experiments use

$s_c$7

with

$s_c$8

The models compared are SPDE, GAM, RF, LAMP, and LAMP-RF. The paper reports that SPDE and LAMP, being linear, missed the nonlinear effects, RF improved over them, and LAMP-RF achieved the best RMSE and MAE across cases. On data generated from the linear model, LAMP-RF matched LAMP, so adding RF did not hurt when nonlinearity was absent (Murakami et al., 1 Oct 2025).

The main empirical application is to log residential land price in 2007, with

$s_c$9

observations in the Tokyo metropolitan area. Covariates include StaDist, TokyoDist, and land-use proportions Agri, Forest, Waste, and River; for RF-based models, longitude and latitude are also included. The compared models are LM, RF, GP with Gaussian and exponential kernels, LAMP with Gaussian and exponential kernels, and LAMP-RF with Gaussian and exponential kernels. A central substantive result is that LAMP selected 40 resolutions, with bandwidths from 77.03 km down to 1.27 km, indicating pronounced multiscale structure. The scale-wise processes were grouped into large scale ($$y(s_i)\mid c \sim N\!\left(\mu_c,\ \frac{\sigma^2}{w_h(d_{i,c})}\right), \qquad \mu_c \sim N(0,\tau^2),$$0), medium scale ($$y(s_i)\mid c \sim N\!\left(\mu_c,\ \frac{\sigma^2}{w_h(d_{i,c})}\right), \qquad \mu_c \sim N(0,\tau^2),$$1), and small scale ($$y(s_i)\mid c \sim N\!\left(\mu_c,\ \frac{\sigma^2}{w_h(d_{i,c})}\right), \qquad \mu_c \sim N(0,\tau^2),$$2). In 10-fold cross-validation using RMSE, MAE, and CRPS, the best overall model was LAMP-RF with exponential kernel. The paper also states that LAMP and LAMP-RF had lower CRPS than GP, indicating better predictive distributions as well as better means (Murakami et al., 1 Oct 2025).

6. Interpretation, strengths, limitations, and use cases

LAMP’s principal strengths, as presented in the paper, are scalability, multiscale adaptivity, machine-learning compatibility, predictive performance, interpretability, and uncertainty modeling. Scalability follows from avoiding explicit covariance inversion and fitting local models independently. Multiscale adaptivity follows from learning large scales first and then adding finer scales. Machine-learning compatibility follows from validation-based fitting rather than likelihood-based optimization. Interpretability is illustrated by the Tokyo decomposition, where large-, medium-, and small-scale components were read as distinct spatial structures. This suggests that LAMP’s scale decomposition can expose interpretable residual spatial organization that would be harder to isolate in a single-scale covariance model (Murakami et al., 1 Oct 2025).

The paper also makes the relationship to existing spatial methodology explicit. It states that LAMP is inspired by geographically weighted regression (GWR) and can thus be regarded as an extension of GWR. The difference is that GWR estimates local coefficients directly, whereas LAMP uses geographically weighted local models as building blocks of a multiscale latent spatial process. Relative to GP/kriging, the key distinction is again that GP/kriging models spatial dependence through a covariance function, while LAMP models it through local weighting and aggregation across scales (Murakami et al., 1 Oct 2025).

Several limitations are also explicit. The paper states that LAMP is not a classical covariance-based dependence model, that its uncertainty quantification is simulation/resampling-based and algorithmic, and that some hyperparameter choices are heuristic, including the number of centers, prior variance updating, local-neighborhood threshold, and initial bandwidth. It also notes sequential fitting bias, which motivates the second holdout-validation recalibration, and states that the framework is currently based on squared-error loss. Extensions to Poisson, logistic, and other losses are listed as future work. A practical implication is that LAMP is best matched to problems where large samples, multiscale structure, and integration with machine-learning predictors are central, rather than to settings in which exact covariance-based inference is the primary objective (Murakami et al., 1 Oct 2025).

7. Related literature and acronym ambiguity

The acronym LAMP is overloaded on arXiv, and several works using the same acronym are unrelated to the Local Aggregate Multiscale Process. In “Learning controllable Adaptive simulation for Multi-resolution Physics”, LAMP means Learning controllable Adaptive simulation for Multi-resolution Physics, and the method is a learned adaptive mesh-based simulator rather than a spatial process model (Wu et al., 2023). In “LAMP: Extracting Text from Gradients with LLM Priors”, LAMP is a text reconstruction attack in federated learning (Balunović et al., 2022). In “LAMP: Extracting Locally Linear Decision Surfaces from LLM World Models”, LAMP means Linear Attribution Mapping Probe, a black-box LLM auditing method (Chen et al., 17 May 2025). These acronymic overlaps are purely terminological, not methodological.

Several adjacent literatures are nonetheless relevant. “Regionalization of Multiscale Spatial Processes using a Criterion for Spatial Aggregation Error” develops a multiscale spatial process defined jointly over point support and areal support through a multiscale Karhunen–Loève expansion, with

$$y(s_i)\mid c \sim N\!\left(\mu_c,\ \frac{\sigma^2}{w_h(d_{i,c})}\right), \qquad \mu_c \sim N(0,\tau^2),$$3

That framework is a mathematically precise local-to-aggregate process model, but it addresses change of spatial support and regionalization rather than local weighted aggregation of experts (Bradley et al., 2015). “Local-Aggregate Modeling for Big-Data via Distributed Optimization” introduces a Local-Aggregate Model for supervised learning on tensor-valued neuroimaging covariates, fitted by ADMM, but it is not a multiscale stochastic process (Hu et al., 2014). “Precision Aggregated Local Models” studies covariance-aware aggregation of local Gaussian-process experts, which is local and aggregate but not explicitly multiscale in the spatial-process sense (Edwards et al., 2020). “Multi-Scale Process Modelling and Distributed Computation for Spatial Data” is closer in spirit to LAMP, because it uses an additive multiscale spatial process with localized nonstationarity and distributed computation, but it remains an SPDE/GMRF construction rather than a local-weighted aggregate process in the specific LAMP form (Zammit-Mangion et al., 2019).

Within this landscape, the 2025 LAMP paper is distinctive for combining five elements in one framework: local aggregation instead of covariance specification, multiscale additive decomposition, sequential residual learning from large to small scales, holdout-validation optimization, and straightforward integration with modern machine-learning learners (Murakami et al., 1 Oct 2025).