Stitching Non-Negative Fields in Square-Root Space

Many large spatial maps must stay non-negative. Examples are the rate at which wildfires ignite across a continent and the density of buildings in a satellite mosaic. These maps are often too large to fit in memory as one model, so practitioners split the domain into overlapping tiles, fit or predict each tile, and blend the results. Two failures follow. Each tile sees only truncated data near its own edge, which leaves seams. And the tile models are usually unconstrained, so they produce negative rates or densities. Practitioners then clip those values to zero, which breaks the total count. We propose to treat the tiles as pieces of a single global model rather than as separate models to be glued together afterward.

Each tile outputs an unconstrained "amplitude" vector. The tiles' amplitudes are blended with smooth window functions that sum to one, and the blended result is squared. The square makes the stitched field non-negative everywhere by construction. Because neighbouring tiles' amplitudes can cancel before squaring, the overlap zones can carve sharp structure instead of averaging it away. With cosine random features, the integral of the field (the expected count) has a closed form, and the matrices involved are sparse with a block-banded pattern, since only adjacent tiles interact. The result is a seam-free, exactly integrable intensity model whose memory grows with the number of tiles rather than quadratically in the total number of features.

Core Claim

The starting point is the squared neural family for Poisson point processes of Exact, Fast and Expressive Poisson Point Processes via Squared Neural Families. A Poisson point process is the standard model for random event locations governed by a rate field. Ong et al. write that rate as the squared norm of a two-layer network, $\lambda(x)=\alpha\|V\psi(x)\|_2^2$, where $\psi$ is a fixed hidden layer and $V$ is learned. This makes the integral of $\lambda$ closed-form and the fit convex when $\psi$ is fixed. Their model is global: one window, one feature map, one Gram matrix.

We localize it. Let tiles $i=1,\dots,T$ have smooth windows $w_i(x)\ge 0$ that are supported on the tile plus its overlap margin and satisfy $\sum_i w_i(x)=1$. Each tile has its own output matrix $V_i$ acting on cosine features $\psi(x)$. The stitched intensity is

$$\lambda(x)=\alpha\Big\|\sum_{i} w_i(x)\,V_i\,\psi(x)\Big\|_2^2 ,$$

and its integral over the observation window $\Omega$ is

$$\int_\Omega \lambda(x)\,dx=\alpha\sum_{i\sim j}\operatorname{tr}\!\big(V_i^\top V_j K_{ij}\big),\qquad K_{ij}=\int_\Omega w_i(x)w_j(x)\,\psi(x)\psi(x)^\top dx .$$

Here $i\sim j$ ranges over pairs of tiles whose windows overlap, including $i=j$. With raised-cosine windows and cosine features, each $K_{ij}$ is an integral of products of trigonometric functions over a box, so it has a closed form.

The key observation is that this is still a squared neural family. Its feature map is the stack of locally windowed features $w_i(x)\psi(x)$. It therefore inherits Ong et al.'s exact likelihood and the optimization structure they report for fixed hidden layers. Because $K_{ij}=0$ for non-neighbouring tiles, the Gram matrix is block-banded. A 2D grid has at most nine nonzero blocks per tile row, and each event's log-likelihood term touches at most four tiles.

What is new is the order of operations: blend first, then square. Blending square roots of densities rather than densities themselves is the geometry that Nonparametric Fisher Geometry with Application to Density Estimation identifies as natural. In that geometry, normalized densities become points on a unit sphere under the square-root map. Blending in this space also produces cross terms between neighbouring tiles. Those cross terms are exactly the subtractive interactions that Subtractive Mixture Models via Squaring: Representation and Learning shows can make squared mixtures far more compact than additive ones.

Relation to Prior Work

Several papers motivate the gap. K2IE: Kernel Method-based Kernel Intensity Estimators for Inhomogeneous Poisson Processes gives a closed-form kernel intensity estimator with built-in boundary correction. It goes negative in sparse regions, relies on clipping, and is evaluated on a single window. Uncertainty-aware Continuous Implicit Neural Representations for Remote Sensing Object Counting decodes continuous density fields from image tiles. Its authors explicitly list strict non-negativity as future work, and they do not stitch tiles. Both show that fast or flexible spatial estimators currently give up non-negativity or seam-free tiling.

Other papers constrain the design. Continuous Patch Stitching for Block-wise Image Compression shows that seams come from artificial edge handling, and that exact equivalence with full-domain processing is achievable when the overlap is derived from the model's support. Our windows play the role of that derived overlap. Unlike that work, we need no architectural surgery, because the global model is defined first and the tiles are its supports. The same paper's restriction to unconstrained RGB regression is what we lift. Estimating Orientation Distribution Functions with Probability Density Constraints and Spatial Regularity couples neighbouring voxels with a smoothness penalty under hard non-negativity constraints and solves by block coordinate descent. We borrow that block-wise solver pattern, but replace both the inequality constraints and the penalty: squaring provides non-negativity, and shared amplitudes provide coupling. Efficient and Private Marginal Reconstruction with Local Non-Negativity supplies evidence that principled non-negativity beats clipping. It reports a 44x average workload-error reduction over baselines, in a discrete setting. Our methods differ from it entirely.

Research Program

The first component is the model and its integrals. We will derive closed-form $K_{ij}$ for raised-cosine windows with cosine random features over axis-aligned boxes. We will verify them against dense quadrature, and characterize the block-banded structure and its memory cost against a global squared family with the same total feature count.

The second component is fitting. We will fit by projected gradient descent following Ong et al., using a block coordinate scheme. Each tile's $V_i$ is updated using only its own events and its neighbours' current amplitudes, which bounds memory by one tile neighbourhood. We will also test full joint fitting on problems small enough to permit it, to check that block updates reach the same optimum.

The third component is an analysis of what the cross terms buy. We will construct synthetic intensities with sharp ridges and voids that cross tile seams. We will compare the method against the same model with squaring moved inside the blend, $\sum_i w_i\|V_i\psi\|^2$, which has a block-diagonal Gram matrix and no cross terms.

The fourth component is a transfer test to image counting. We will replace UNIC's density decoder output with a squared amplitude, blended across overlapping RSOC tiles. Here integrals are grid sums, so closed-form integration is not the selling point. The test asks only whether non-negativity by construction plus amplitude blending improves counts.

Baselines

The naive reference is independent tiles with hard stitching. A separate squared-family model is fit on each non-overlapping tile, and the tile models are concatenated. This is non-negative but seamed and edge-biased. Beating it on held-out negative log-likelihood (NLL) is necessary but not interesting.

The first practical baseline is independent overlapping tiles with post-hoc blending. Each tile's squared model is fit on its own extended footprint, and the intensities are averaged with the same windows. This is what tiling practitioners would do, and it remains non-negative and integrable. To count as beating it, we need lower NLL and lower integrated squared error (ISE) concentrated in seam bands, at equal features per tile.

The strongest existing method is the global squared neural family of Exact, Fast and Expressive Poisson Point Processes via Squared Neural Families, run with the largest feature count that fits the same GPU memory budget as our tiled model. The only reported figures the file gives are on NASA FIRMS (100 million events): fitting in under 8 hours versus an estimated 88 days for a naive kernel method. No reported NLL value was found on Lacuna for any of its benchmarks. We will therefore rerun it ourselves with identical features, event splits, and tuning budget. Beating it means lower held-out NLL at equal memory, or equal NLL at substantially lower memory and wall-clock.

K2IE from K2IE: Kernel Method-based Kernel Intensity Estimators for Inhomogeneous Poisson Processes, with clipping, is the fast closed-form competitor. It reports 0.14 seconds on 2D synthetic data but no numeric ISE value that we found on Lacuna. We run it with its bandwidth tuned on the same validation split.

The ablation isolating our key ingredient is the square-then-blend variant, jointly fit. It shares windows, features, and data, and differs only in whether cross-tile amplitude interference exists. For RSOC, the baseline is UNIC itself. The file reports it reduces MAE by 4.87% relative to ASPDNet, but no absolute MAE is reported on Lacuna. We will retrain UNIC under its released protocol and compare against our decoder change under identical backbone and training budget.

Why not just…?

Why not just clip an unconstrained tiled model, or put a softplus on each tile and average? Clipping biases counts and kills gradients. Mullins et al. report principled non-negativity beating clipping heuristics by a wide margin, though in a discrete setting. A softplus or exponential output with gridded quadrature is a real competitor for binned image counting, where integration is a sum. For continuous point-process likelihood, the exponential link loses the exact integral and the convex fit that Ong et al. rely on. We include a softplus-per-tile, grid-quadrature variant on the synthetic Cox processes. If it matches our NLL at equal compute, the closed-form argument is moot in 2D and we will say so.

Why not just fit one global squared model with more features? That is our strongest baseline. Its dense Gram matrix grows quadratically in feature count, while ours is block-banded. The fixed-memory comparison settles whether localization helps beyond saving memory. Localized features may also simply express fine spatial detail more cheaply than global random features.

Why not just square inside each tile and blend the intensities? This is the most serious threat. It is simpler, block-diagonal, and equally non-negative and integrable. Our hypothesis is that it blurs sharp structures crossing seams, because a convex blend of two local fits cannot cancel mass the way interfering amplitudes can. The seam-crossing synthetic benchmark is designed to decide this. If the variant matches us there and on the real datasets, the honest conclusion is that joint fitting of additive local mixtures is sufficient. We would report it as such.

Proposed Experiments

On synthetic 2D Gaussian Cox processes in the style used by K2IE, plus our seam-crossing ridge and void intensities, the true intensity is known. We report ISE overall and within bands around tile boundaries, held-out NLL, and the fraction of the domain where any baseline goes negative. The decisive check is that seam-band ISE for blend-then-square sits near its interior ISE, while the independent-tile and square-then-blend baselines show a seam-band excess.

On bei, clmfires, and copper, we report held-out NLL over repeated random event splits as tile count varies at fixed total memory. On NASA FIRMS, we report NLL, peak memory, and wall-clock against the global squared model. The decisive check is whether tiling lets us use many more effective features within the memory the global model needs, and turns that into lower NLL.

On RSOC buildings, we report MAE and RMSE for UNIC versus UNIC with the squared amplitude head and amplitude blending, at matched training budget. We also report how often UNIC predicts negative density. The claim fails if (a) the global model matches us at equal memory, or (b) square-then-blend matches us on the seam benchmark. Either outcome would redirect the work toward the simpler design.