Nonstationary Support Product Forcing

• Nonstationary Support Product Forcing is a mechanism that governs a product's behavior by applying a support-sensitive factor to ignore or relax exceptions.
• In set theory, it is exemplified by full products and mod-finite quotients that alter cardinal collapse properties through support modifications.
• In Gaussian-process modeling, a compactly-supported sparse kernel multiplies a core covariance to force exact zeros and enable sparse, efficient computations.

Searching arXiv for the cited papers to ground the article in current records. “Nonstationary Support Product Forcing” is not a standard named family in the cited literature. The phrase instead points to a recurring structural idea: a product construction whose effective behavior is determined by a support notion that ignores, relaxes, or annihilates a class of exceptions. In the set-theoretic forcing literature, the closest results concern full-support products over $\omega$, reduced products modulo finite, Easton support iterations, and finite-support-style generalizations; these papers repeatedly stress that they do not explicitly treat nonstationary support products (Golshani et al., 2015). In probabilistic machine learning, the phrase closely matches a construction in which a nonstationary multiplicative factor forces exact zeros in a product covariance kernel, although that phrase is likewise not the paper’s own terminology (Risser et al., 2024).

1. Terminological scope and conceptual profile

In the set-theoretic material, “support” refers to which coordinates of a product or iteration are retained, relaxed, or treated as negligible. The strongest direct analogue in the cited papers is the reduced product modulo the Fréchet ideal on $\omega$, where all coordinates remain present but the order is relaxed to eventual coordinatewise extension. This is explicitly distinguished from finite-support products, Easton support iterations, and nonstationary support on uncountable index sets (Golshani et al., 2015).

In the Gaussian-process material, “support” is geometric rather than set-theoretic. A compactly-supported sparse kernel factor is multiplied by a core covariance, and the product kernel is exactly zero wherever the sparse factor is zero. The paper states that this construction “does match” the idea of “Nonstationary Support Product Forcing,” with the caveat that the phrase is not its own terminology (Risser et al., 2024).

This suggests two technically distinct but structurally analogous uses of the phrase. In one, support is an order-theoretic or ideal-theoretic constraint on forcing coordinates. In the other, support is a multiplicative geometric constraint on covariance. The analogy is exact at the level of mechanism: a product is governed by a support-sensitive factor that can dominate the global behavior.

2. Full products and mod-finite quotients at singular cardinals

A central set-theoretic instance begins with a singular strong limit cardinal $\kappa$ of countable cofinality and an increasing sequence of regular cardinals $\langle \kappa_n:n<\omega\rangle$ cofinal in $\kappa$. For each $n<\omega$, the paper uses

$\operatorname{Add}(\kappa_n,1),$

the Cohen forcing adding one subset of $\kappa_n$, and forms the full product

$P=\prod_{n<\omega}\operatorname{Add}(\kappa_n,1).$

Conditions are sequences $p=\langle p(n):n<\omega\rangle$ with $\omega$0 for every $\omega$1, ordered coordinatewise: $\omega$2 Because the index set is $\omega$3, “full support” means simply that all coordinates are specified (Golshani et al., 2015).

The paper also studies the quotient modulo finite,

$\omega$4

where $\omega$5 means

$\omega$6

This is not a finite-support product. It is the reduced product order modulo the Fréchet ideal on $\omega$7, so the forcing remembers all coordinates while comparison is relaxed to eventual domination in the product order (Golshani et al., 2015).

Two theorems organize the behavior of this construction. First, if $\omega$8, then forcing with

$\omega$9

collapses $\kappa$0 to $\kappa$1. More generally, the same collapse holds for arbitrary nontrivial separative forcing notions $\kappa$2 such that each $\kappa$3 is $\kappa$4-closed, $\kappa$5, and every decreasing sequence of length $\kappa$6 has a greatest lower bound (Golshani et al., 2015).

Second, if the sequence $\kappa$7 carries a scale of length $\kappa$8, then

$\kappa$9

In particular, forcing with the full product adds a generic filter for $\langle \kappa_n:n<\omega\rangle$0. The paper presents this as a new proof of a result due to Shelah, while also making the quotient analysis explicit (Golshani et al., 2015).

3. Collapse mechanisms, coding, and the role of support sensitivity

The collapse theorem factors through the mod-finite quotient. The natural map

$\langle \kappa_n:n<\omega\rangle$1

is a projection in the forcing sense, and $\langle \kappa_n:n<\omega\rangle$2 is $\langle \kappa_n:n<\omega\rangle$3-strategically closed. Since $\langle \kappa_n:n<\omega\rangle$4, the argument uses a closure-plus-size mechanism rather than a chain-condition argument: under $\langle \kappa_n:n<\omega\rangle$5, a strategically closed forcing of size $\langle \kappa_n:n<\omega\rangle$6 with suitable greatest-lower-bound play adds a new sequence of ordinals of length $\langle \kappa_n:n<\omega\rangle$7, and then collapses $\langle \kappa_n:n<\omega\rangle$8 onto the least length of a new ordinal sequence added by the forcing. Because $\langle \kappa_n:n<\omega\rangle$9, the collapse lifts from $\kappa$0 to the full product $\kappa$1 (Golshani et al., 2015).

Under the scale hypothesis, the mechanism is more explicit. If

$\kappa$2

is a scale in $\kappa$3, the paper forms

$\kappa$4

Then $\kappa$5, and $\kappa$6 is cofinal in $\kappa$7 under eventual domination. Writing $\kappa$8 for the Cohen generic subset added at coordinate $\kappa$9, the paper defines $n<\omega$0 by blockwise coding into successive coordinate generics. The crucial density argument is that for every ground-model $n<\omega$1, there is $n<\omega$2 such that $n<\omega$3. Since there are only $n<\omega$4 many such $n<\omega$5’s modulo eventual equality, the full product collapses $n<\omega$6 to $n<\omega$7 (Golshani et al., 2015).

For the topic at hand, the relevant point is not that the paper studies nonstationary support directly, but that replacing exact coordinatewise order by order modulo a small ideal of exceptions can radically change the forcing. The paper states explicitly that it does not discuss nonstationary support products, because the index set is just $n<\omega$8; nevertheless, the quotient modulo finite is presented as a countable-index analogue of forcing constructions where the order is insensitive to a prescribed ideal of exceptions (Golshani et al., 2015).

A broader support-sensitive contrast appears in the study of Easton support iterations of Prikry-type forcing notions. That paper is not about nonstationary support products per se, but it emphasizes that the behavior of normal measures and ultrapower embeddings under Easton support is dramatically different from what was previously known for nonstationary support and full support iterations (Gitik et al., 2023). In particular, it states that a phenomenon such as

$n<\omega$9

is impossible with nonstationary support, whereas Easton support can allow it (Gitik et al., 2023). A separate expository paper on forcing techniques related to finite support iteration likewise does not explicitly mention nonstationary support, stationary support, revised countable support, or support rules defined in terms of stationary or nonstationary subsets; its relevance is instead through generalized support via template families, correctness diagrams, direct limits, ultrapowers, Boolean ultrapowers, and restrictions to elementary submodels (Brendle, 2021).

4. Nonstationary multiplicative support in Gaussian-process kernels

A second, non-set-theoretic realization of the phrase appears in Gaussian-process modeling. The proposed covariance has the product form

$\operatorname{Add}(\kappa_n,1),$0

where $\operatorname{Add}(\kappa_n,1),$1 is any positive definite kernel, stationary or nonstationary, and $\operatorname{Add}(\kappa_n,1),$2 is a specially constructed compactly-supported, nonstationary, sparsity-inducing kernel (Risser et al., 2024).

The sparse factor is defined by

$\operatorname{Add}(\kappa_n,1),$3

with

$\operatorname{Add}(\kappa_n,1),$4

Each $\operatorname{Add}(\kappa_n,1),$5 is a compactly supported smooth bump function: $\operatorname{Add}(\kappa_n,1),$6 The stationary compactly supported term $\operatorname{Add}(\kappa_n,1),$7 is chosen from the Wendland family (Risser et al., 2024).

Nonstationarity enters in two layers. First, $\operatorname{Add}(\kappa_n,1),$8 is itself nonstationary because the bump functions depend on absolute location via $\operatorname{Add}(\kappa_n,1),$9 and $\kappa_n$0, not only on separation $\kappa_n$1. The rank-one terms $\kappa_n$2 can create covariance between regions where the same $\kappa_n$3 is simultaneously nonzero, regardless of Euclidean distance; nearby points can also have zero covariance if they do not share support. Second, the core can itself be nonstationary, for example through a Paciorek–Schervish-type kernel with spatially varying signal standard deviation $\kappa_n$4, spatially varying local length-scale $\kappa_n$5, or a location-dependent anisotropy matrix decomposed as

$\kappa_n$6

in locally anisotropic applications (Risser et al., 2024).

The support-forcing mechanism is exact. Each bump satisfies

$\kappa_n$7

and the Wendland term satisfies

$\kappa_n$8

Therefore,

$\kappa_n$9

if the pair is outside the Wendland support and the two points do not simultaneously belong to the support pattern of any common bump-sum component $P=\prod_{n<\omega}\operatorname{Add}(\kappa_n,1).$0. Since the full kernel is the product $P=\prod_{n<\omega}\operatorname{Add}(\kappa_n,1).$1,

$P=\prod_{n<\omega}\operatorname{Add}(\kappa_n,1).$2

That is the paper’s key support-forcing condition (Risser et al., 2024).

5. Bayesian inference, sparsity discovery, and exact computation

The Gaussian-process construction is embedded in a fully Bayesian model. Observations satisfy

$P=\prod_{n<\omega}\operatorname{Add}(\kappa_n,1).$3

with independent $P=\prod_{n<\omega}\operatorname{Add}(\kappa_n,1).$4. For observed locations $P=\prod_{n<\omega}\operatorname{Add}(\kappa_n,1).$5,

$P=\prod_{n<\omega}\operatorname{Add}(\kappa_n,1).$6

where $P=\prod_{n<\omega}\operatorname{Add}(\kappa_n,1).$7 is diagonal and $P=\prod_{n<\omega}\operatorname{Add}(\kappa_n,1).$8 has entries $P=\prod_{n<\omega}\operatorname{Add}(\kappa_n,1).$9. Posterior inference uses

$p=\langle p(n):n<\omega\rangle$0

The sparsity-discovery mechanism is carried in particular by the bump-amplitude prior

$p=\langle p(n):n<\omega\rangle$1

so each bump can be turned on or off by the posterior (Risser et al., 2024).

The computational significance of the support-forcing product is that compact support in the covariance itself yields a sparse covariance matrix, enabling exact GP inference with sparse linear algebra rather than approximate GP likelihoods. For a sparse matrix with $p=\langle p(n):n<\omega\rangle$2 nonzero entries, the paper gives $p=\langle p(n):n<\omega\rangle$3 for Lanczos log-determinant estimation and $p=\langle p(n):n<\omega\rangle$4 for conjugate gradients. It also notes that covariance construction itself remains expensive because the sparsity pattern changes with the hyperparameters, so each likelihood evaluation still requires batchwise dense kernel computation before sparsification (Risser et al., 2024).

In the largest space-time temperature application, the paper reports 256 A100 GPUs, about 66 seconds per likelihood, roughly 37 seconds covariance computation, 9 seconds MINRES solve, 2 seconds log-determinant estimate, and observed sparsity about $p=\langle p(n):n<\omega\rangle$5. More generally, it reports sparsity as low as $p=\langle p(n):n<\omega\rangle$6 for million-point problems. These are consequences of exact zeros forced by the support factor, not of sparse precision approximations such as Vecchia or NNGP (Risser et al., 2024).

Mathematically, validity is obtained by constructive closure. The paper states that as long as $p=\langle p(n):n<\omega\rangle$7 and $p=\langle p(n):n<\omega\rangle$8 are positive definite, their product is positive definite, using the standard Schur/Hadamard product closure argument. It also proves that the resulting kernel has

$p=\langle p(n):n<\omega\rangle$9

continuous derivatives at zero, where $\omega$00 has $\omega$01 continuous derivatives at the origin and $\omega$02 has $\omega$03 continuous derivatives at zero (Risser et al., 2024).

6. Limits of the term and interpretive significance

Several misconceptions are explicitly excluded by the cited papers. The singular-cardinal forcing paper should not be cited as a direct theorem about nonstationary support, because its index set is $\omega$04, and its two order notions are full support and product modulo finite (Golshani et al., 2015). The Easton-support paper is about iterations, not products, and its value for nonstationary support is mainly by contrast and analogy (Gitik et al., 2023). The expository paper on finite-support techniques does not discuss nonstationary support products or iterations in any direct sense (Brendle, 2021). Conversely, the Gaussian-process paper does match the idea of “Nonstationary Support Product Forcing,” but more precisely it is a data-driven nonstationary compact-support product kernel rather than a standard named family (Risser et al., 2024).

Taken together, the papers isolate a common structural lesson. In the set-theoretic setting, quotienting a full product by a small ideal of exceptions can determine the real forcing content, and pcf structure can identify the quotient with a single Cohen forcing at the successor cardinal. In the Gaussian-process setting, multiplying by a learned compactly-supported nonstationary factor can determine the real covariance content, forcing exact zeros and sparse computation. This suggests that the phrase “Nonstationary Support Product Forcing” is best understood not as a settled technical label, but as a description of a mechanism: a product whose global behavior is governed by a support-sensitive factor, whether that factor is an ideal of negligible coordinates or a nonstationary multiplicative support kernel.