Fourth-Order Hierarchical Array (FOHA)

• FOHA is a collection of fourth-order constructs defined via hierarchical frameworks across fields such as Bayesian modeling, tensor analysis, and array signal processing.
• The methodological frameworks include tree-indexed exchangeability, covariance sharing in ANOVA models, and level-polymorphic array designs, each enabling improved statistical and computational performance.
• FOHA underpins advanced techniques like forward symplectic integration for N-body dynamics and sparse array design, enhancing DOA estimation and reducing sensor coupling.

to=arxiv_search.query 娱乐总代理 代理娱乐_code=json {"query":"all:\"Fourth-Order Hierarchical Array\" OR all:FOHA", "max_results": 10, "sort_by": "submittedDate", "sort_order": "descending"} to=arxiv_search.query 大发快三开奖 菲龙码=json {"query":"id:(Wang et al., 27 Aug 2025) OR id:(Austin et al., 2013) OR id:(Rantala et al., 2020) OR id:(Volfovsky et al., 2012) OR id:({Š}inkarovs, 2020)", "max_results": 10, "sort_by": "relevance", "sort_order": "descending"} Fourth-Order Hierarchical Array (FOHA) is not a single universally standardized object. In current arXiv usage, the term and closely related fourth-order hierarchical constructions denote at least five distinct entities: a sparse linear array scheme based on the fourth-order difference co-array; a fourth-order interaction tensor in hierarchical array-normal ANOVA priors; a level-4 array in a hierarchy of arrays with levels; a hierarchical fourth-order forward symplectic integration framework for direct $N$-body dynamics; and, by specialization of hierarchical exchangeability theory, arrays indexed by a depth-4 tree or by four trees (Wang et al., 27 Aug 2025, Volfovsky et al., 2012, {Š}inkarovs, 2020, Rantala et al., 2020, Austin et al., 2013). The common motif is the combination of a fourth-order object with an explicitly structured hierarchy, but the meanings of “fourth-order” and “hierarchical” differ substantially across probability, Bayesian modeling, type theory, computational dynamics, and array signal processing.

1. Terminological scope and conceptual organization

The main usages can be organized by the mathematical object to which hierarchy is attached. In some settings the hierarchy is in the index set; in others it is in the prior covariance, shape type, time-step recursion, or sensor placement design. This immediately rules out any assumption that FOHA names a single transferable formalism across disciplines.

Context	FOHA meaning	Hierarchical structure
Exchangeability theory	Depth-4 tree-indexed array or four-tree array	Tree paths and tree-preserving rearrangements
Bayesian ANOVA	Fourth-order interaction tensor $(abcd)$	Shared per-factor covariance matrices
Arrays with levels	Level-4 array	Shape is a level-3 array of natural numbers
Direct $N$-body integration	Hierarchical fourth-order forward symplectic framework	Recursive time-step levels
Sparse array design	FODCA-based sparse linear array	Generator subarray plus two arithmetic-progression subarrays

A recurrent source of confusion is that “fourth-order” does not have a single meaning. In the exchangeability and ANOVA settings it refers to an array indexed by four coordinates or a fourth-order tensor; in the dependently typed array setting it refers to level 4 in a hierarchy of array shapes; in the FROST integrator it refers to fourth-order accuracy of the symplectic composition; and in the sparse-array setting it refers to fourth-order cumulants and the fourth-order difference co-array. This suggests that FOHA is best treated as a context-dependent designation rather than a canonical mathematical noun.

2. Tree-indexed and multi-tree FOHAs in hierarchical exchangeability

In the representation-theoretic setting, the natural FOHA specialization arises from arrays indexed by the leaves of an infinitary rooted tree of finite depth, or by products of several such trees. For a rooted tree of depth $r$, the paper models the leaf set as $\mathbb{N}^r$, with vertex set

$$\mathscr{A}(r)=\mathbb{N}^0\cup \mathbb{N}\cup \mathbb{N}^2\cup \cdots \cup \mathbb{N}^r,$$

where $\mathbb{N}^0=\{\emptyset\}$ is the root. For a vertex $\alpha$, the path from the root is $p(\alpha)$, and for two leaves $\alpha,\beta\in\mathbb{N}^r$ the quantity $(abcd)$0 counts common vertices on their paths (Austin et al., 2013).

The relevant invariance group is $(abcd)$1, the set of bijections $(abcd)$2 such that

$(abcd)$3

for all $(abcd)$4. Equivalently, each $(abcd)$5 recursively permutes the children of every vertex while preserving the parent-child relation. An array $(abcd)$6 in a standard Borel space $(abcd)$7 is hierarchically exchangeable if

$(abcd)$8

for all $(abcd)$9.

For a single depth-4 tree, FOHA is the specialization $N$0: an array $N$1 indexed by leaves $N$2. The hierarchical de Finetti theorem states that there exists a measurable function $N$3 and an i.i.d. family $N$4 of Uniform$N$5 random variables such that

$N$6

For depth $N$7, if the path to a leaf $N$8 is $N$9, then

$r$0

Each leaf value is therefore a measurable function of i.i.d. latent variables attached to the vertices along its path.

For four separate trees, FOHA becomes a fourth-order array $r$1 with each index a leaf in its own tree $r$2 of finite depth $r$3. The invariance group is $r$4, acting coordinatewise. The hierarchical Aldous-Hoover representation gives a measurable function $r$5 and i.i.d. Uniform$r$6 latent variables $r$7 on all vertex tuples such that

$r$8

The inputs include the global root term $r$9, single-path terms, pairwise path-combination terms, triple path-combination terms, and quadruple path-combination terms. This differs from the classical non-hierarchical Aldous-Hoover form because the latent variables are indexed by path nodes at multiple depths rather than only by raw indices and their subsets.

The representation is not unique. As in classical de Finetti/Aldous-Hoover theory, measure-preserving transformations may be applied to the latent Uniform coordinates without changing the law of the array. The paper does not pursue a finer canonical form beyond this standard equivalence.

3. FOHA as a fourth-order interaction tensor in hierarchical ANOVA priors

In Bayesian analysis of cross-classified data, FOHA denotes the fourth-order interaction array in a four-factor ANOVA decomposition. For categorical factors $\mathbb{N}^r$0 with $\mathbb{N}^r$1 levels, the model is

$\mathbb{N}^r$2

with $\mathbb{N}^r$3 i.i.d. normal$\mathbb{N}^r$4 (Volfovsky et al., 2012).

The FOHA object of interest is the fourth-order interaction tensor

$\mathbb{N}^r$5

whose entries represent departures beyond all main effects, all two-way interactions, and all three-way interactions. In the paper’s Bayesian formulation, priors are placed directly on the effects without enforcing sum-to-zero constraints, and inference is conducted on identifiable cell means. Sum-to-zero constraints may be imposed if desired, but the paper notes that this is not necessary for Bayesian inference on cell means.

The key structural device is the array-normal prior with separable Kronecker covariance. For the fourth-order interaction,

$\mathbb{N}^r$6

with

$\mathbb{N}^r$7

Interaction magnitudes are scaled by a gamma precision parameter:

$\mathbb{N}^r$8

For FOHA this yields

$\mathbb{N}^r$9

The hierarchical aspect lies in the fact that the per-factor covariance matrices $$\mathscr{A}(r)=\mathbb{N}^0\cup \mathbb{N}\cup \mathbb{N}^2\cup \cdots \cup \mathbb{N}^r,$$0 are shared across all effects involving the corresponding factor. They are learned from main effects and lower-order interactions and then reused to regularize higher-order interactions. The paper characterizes this as “borrowing of strength”: if levels of a factor show similar coefficients in main effects and lower-order interactions, the same covariance structure shrinks corresponding higher-order interaction entries toward one another.

Posterior inference is performed by Gibbs sampling. In the balanced case, the full conditional for the FOHA tensor is

$$\mathscr{A}(r)=\mathbb{N}^0\cup \mathbb{N}\cup \mathbb{N}^2\cup \cdots \cup \mathbb{N}^r,$$1

$$\mathscr{A}(r)=\mathbb{N}^0\cup \mathbb{N}\cup \mathbb{N}^2\cup \cdots \cup \mathbb{N}^r,$$2

and

$$\mathscr{A}(r)=\mathbb{N}^0\cup \mathbb{N}\cup \mathbb{N}^2\cup \cdots \cup \mathbb{N}^r,$$3

For unbalanced designs, the paper proposes a data-augmentation procedure that imputes missing cell means so that the balanced-data full conditionals can be reused.

The empirical results reported in the paper are for three-factor settings, but they are used to motivate the four-factor extension. When order-consistent interactions exist, the hierarchical array prior substantially reduces average squared error relative to OLS and a standard nonhierarchical Bayes prior, with most of the gain coming from better estimation of higher-order interactions. In the order-consistent case with $$\mathscr{A}(r)=\mathbb{N}^0\cup \mathbb{N}\cup \mathbb{N}^2\cup \cdots \cup \mathbb{N}^r,$$4, interval widths for cell means were approximately $$\mathscr{A}(r)=\mathbb{N}^0\cup \mathbb{N}\cup \mathbb{N}^2\cup \cdots \cup \mathbb{N}^r,$$5 under HA versus $$\mathscr{A}(r)=\mathbb{N}^0\cup \mathbb{N}\cup \mathbb{N}^2\cup \cdots \cup \mathbb{N}^r,$$6 under SB, with approximately $$\mathscr{A}(r)=\mathbb{N}^0\cup \mathbb{N}\cup \mathbb{N}^2\cup \cdots \cup \mathbb{N}^r,$$7 coverage. The paper’s stated implication is that the same mechanism should be even more valuable for FOHA, because the fourth-order tensor is the least directly informed by data.

4. FOHA as a level-4 array in a hierarchy of arrays with levels

In dependently typed array theory, FOHA is a level-4 array rather than a fourth-order tensor in the ANOVA sense. The underlying formalism is based on containers. A container $$\mathscr{A}(r)=\mathbb{N}^0\cup \mathbb{N}\cup \mathbb{N}^2\cup \cdots \cup \mathbb{N}^r,$$8 denotes a functor with extension

$$\mathscr{A}(r)=\mathbb{N}^0\cup \mathbb{N}\cup \mathbb{N}^2\cup \cdots \cup \mathbb{N}^r,$$9

The hierarchy is generated by the diamond operation and by iterating a shape container so that an array of level $\mathbb{N}^0=\{\emptyset\}$0 has a shape that is a level-$\mathbb{N}^0=\{\emptyset\}$1 array of natural numbers ({Š}inkarovs, 2020).

The paper’s alternative encoding makes this shape relation definitional:

$\mathbb{N}^0=\{\emptyset\}$2

Accordingly, a FOHA is characterized by

$\mathbb{N}^0=\{\emptyset\}$3

so a level-4 shape is a level-3 array of natural numbers. Arrays with a fixed shape are represented by $(abcd)$35 and indexing is function application: $(abcd)$36

The index type is dependent on the shape and carries static bounds information. In the alternative encoding, BFin stores a natural number together with an irrelevant proof that it is below the corresponding bound, while FlatIx assembles these bounded components into a valid multidimensional index. Mathematically, for a level-$\mathbb{N}^0=\{\emptyset\}$4 array with shape $\mathbb{N}^0=\{\emptyset\}$5, the index type satisfies

$\mathbb{N}^0=\{\emptyset\}$6

so every index is in bounds by construction.

A concrete FOHA example in the paper uses a level-3 shape corresponding intuitively to a $\mathbb{N}^0=\{\emptyset\}$7 grid of natural numbers with content vector $\mathbb{N}^0=\{\emptyset\}$8. The resulting FOHA has four axes with bounds $\mathbb{N}^0=\{\emptyset\}$9. An array

$\alpha$0

can then be defined by a function on such bounded indices; the example takes

$\alpha$1

The principal distinction from ordinary rank-polymorphic array languages is the introduction of level-polymorphic operators. Standard operations such as map, reduce, and reshape are supported, but the formalism also defines ranked cuts and nesting across the hierarchy of shapes:

$\alpha$2

$\alpha$3

For a FOHA, this makes it possible to perform partial selections over hyperplanes determined by the level-3 shape, not merely over the last few axes. The paper presents this as a major gain in expressiveness, while also noting the extra proof burden imposed by dependent typing and by the alternative encoding required to avoid extensionality problems.

5. FOHA as a hierarchical fourth-order forward symplectic framework

In direct-summation gravitational dynamics, FOHA denotes a hierarchical, strictly forward fourth-order integration framework implemented by FROST. It organizes particles into an array of time-step levels $\alpha$4 with

$\alpha$5

and composes level-wise, synchronous kick/drift operations so that all inter-level force calculations are pair-wise. The stated consequence is manifest momentum conservation, while fourth-order accuracy is obtained by adding a force-gradient term that keeps all substeps positive (Rantala et al., 2020).

The single-level method starts from a separable Hamiltonian $\alpha$6 with $\alpha$7 and $\alpha$8. The update rules are explicit:

• Drift: $\alpha$9.
• Kick: $p(\alpha)$0.
• Gradient-kick at the middle kick only:

$p(\alpha)$1

FROST uses strictly positive coefficients:

$p(\alpha)$2

The middle kick is modified by the potential

$p(\alpha)$3

where $p(\alpha)$4 implements $p(\alpha)$5. The paper emphasizes that all substeps are positive, so no negative time-slicing is required.

The hierarchical scheme, denoted HHS-FSI, partitions particles at a pivot $p(\alpha)$6 into slow and fast subsets,

$p(\alpha)$7

and recursively repeats the split on $p(\alpha)$8 at $p(\alpha)$9. On each level the Hamiltonian is decomposed as

$\alpha,\beta\in\mathbb{N}^r$0

with $\alpha,\beta\in\mathbb{N}^r$1 placed at the slow level (“HOLD”). One pivot step uses the forward pattern $\alpha,\beta\in\mathbb{N}^r$2, recursing on the fast subset during the two half-step segments and computing inter-level Newtonian or gradient accelerations synchronously. Because every inter-level kick is pair-wise and equal-and-opposite, the scheme is momentum-conserving up to floating-point round-off for any number of levels.

The reported numerical properties are specific. In binary and few-body tests, linear momentum errors satisfy $\alpha,\beta\in\mathbb{N}^r$3 and angular momentum errors satisfy $\alpha,\beta\in\mathbb{N}^r$4. In million-body star cluster simulations, the paper reports energy conservation $\alpha,\beta\in\mathbb{N}^r$5 over approximately $\alpha,\beta\in\mathbb{N}^r$6 Myr, angular momentum error $\alpha,\beta\in\mathbb{N}^r$7, and center-of-mass velocity error $\alpha,\beta\in\mathbb{N}^r$8. The abstract summarizes the method as conserving energy to a level of $\alpha,\beta\in\mathbb{N}^r$9 while momentum errors are practically negligible.

The implementation is MPI-parallelized CUDA C. All $(abcd)$00 operations—Newtonian forces, gradient forces, and time-step assignment—run on GPUs via synchronous kernels using shared-memory tiles. FROST exhibits strong scaling up to

$(abcd)$01

and the paper states that direct-summation simulations beyond $(abcd)$02 particles are possible on systems with several hundred and more GPUs. Because of the hierarchical integration architecture, inclusion of Kepler solvers, regularized integrators, and post-Newtonian corrections at the fastest levels is described as straightforward.

6. FOHA as a sparse array design based on the fourth-order difference co-array

The usage in array signal processing is the only one for which “Fourth-Order Hierarchical Array” appears as the explicit title of a paper. Here FOHA is a sparse linear array constructed from different forms of the fourth-order difference co-array (FODCA), with the goals of increasing degrees of freedom, reducing redundancy, and suppressing mutual coupling in direction-of-arrival estimation (Wang et al., 27 Aug 2025).

The data model assumes $(abcd)$03 non-Gaussian, mutually uncorrelated, far-field narrowband sources impinging on a sparse linear array with sensor set $(abcd)$04. Using circular fourth-order cumulants, the paper identifies two virtual-lag constructions:

$(abcd)$05

$(abcd)$06

The full fourth-order difference co-array is the union of these two forms. If the nonnegative support merges to a maximal continuous interval $(abcd)$07, then

$(abcd)$08

The general FOHA construction partitions the sensor set into three disjoint subarrays,

$(abcd)$09

The generator $(abcd)$10 is arbitrary; $(abcd)$11 and $(abcd)$12 are arithmetic progressions:

$(abcd)$13

The parameters are governed by spans $(abcd)$14 induced by the generator and by the corresponding second- and third-order co-arrays. Under the hole-free conditions

$(abcd)$15

the FODCA is hole-free, and the largest nonnegative consecutive segment has length

$(abcd)$16

Two explicit configurations are developed. In FOHA(NA), the generator $(abcd)$17 is a two-level nested array; in FOHA(CNA), it is a concatenated nested array. For both cases the paper provides closed-form sensor placements, proves the hole-free conditions, and gives algorithms for choosing subarray sizes to maximize DOFs for fixed $(abcd)$18. Qualitatively, FOHA(CNA) achieves higher DOFs than several named baselines for $(abcd)$19, while FOHA(NA) surpasses all but FOHA(CNA) for $(abcd)$20. For $(abcd)$21, the paper reports DOFs of $(abcd)$22 for FOHA(NA) and $(abcd)$23 for FOHA(CNA), compared with $(abcd)$24 for FO-Fractal(NA) and $(abcd)$25 for SD-FOSA(CNA-NA).

Mutual coupling is modeled by a banded symmetric Toeplitz matrix $(abcd)$26 with entries determined by sensor separation, and leakage is quantified by

$(abcd)$27

A central analytical result links the leakage of the generator subarray $(abcd)$28 to the leakage of the full FOHA and yields

$(abcd)$29

provided the coupling limit $(abcd)$30. The interpretation given in the paper is that the two wide-spacing subarrays contribute negligibly to off-diagonal coupling while increasing the Frobenius norm of the full coupling matrix, thereby lowering overall leakage.

The paper also studies redundancy through

$(abcd)$31

with lower bound

$(abcd)$32

For FOHA(CNA),

$(abcd)$33

and the paper reports that FOHA(CNA) achieves the lowest redundancy among the tested designs for $(abcd)$34.

Signal reconstruction conditions are given in necessary-and-sufficient form through an LCM-based criterion on sensor positions, and FOHA-specific conditions are then derived for the NA and CNA generators. The DOA pipeline consists of estimating fourth-order cumulants, mapping quartets of sensor indices to virtual lags, applying spatial smoothing on the virtual array, and then using subspace-based estimators such as MUSIC on the smoothed virtual manifold. In the numerical simulations summarized in the paper, FOHA(CNA) consistently achieves the lowest RMSE under both uncoupled and coupled conditions, with FOHA(NA) second-best.

The sparse-array FOHA is therefore a fourth-order hierarchical design in a literal engineering sense: hierarchy is built into the physical sensor layout through a generator plus two wide-spacing subarrays, while fourth-order statistics create the virtual aperture that drives high-DOF DOA estimation.

7. Comparative interpretation

Across these literatures, the role of hierarchy changes, but it always serves to constrain or organize otherwise high-dimensional fourth-order structure. In hierarchical exchangeability, the hierarchy is an invariance group acting on leaves of trees. In hierarchical ANOVA priors, it is a covariance-sharing mechanism that regularizes a fourth-order interaction tensor. In arrays with levels, it is the recursive dependence of shape on lower-level arrays of natural numbers. In FROST, it is a recursive scheduling of slow and fast particle subsets across time-step levels. In FODCA-based sparse sensing, it is a generator-based physical construction that synthesizes a large fourth-order virtual aperture.

A second cross-cutting theme is that FOHA-like constructions typically mediate between local structure and global interaction. Tree-path latents combine vertexwise randomness into leaf values; Kronecker covariances propagate factor-level similarity into four-way interaction estimates; level-polymorphic operators let shape-of-shape structure govern hyperplane selection; synchronous slow-fast kicks preserve pair-wise conservation while composing a global fourth-order integrator; and FODCA unions combine multiple cumulant-derived lag forms into a single hole-free virtual array. This suggests a common architectural principle: fourth-order complexity is rendered tractable by imposing a recursively organized latent, geometric, algebraic, or computational scaffold.

The principal misconception to avoid is that results from one FOHA literature transfer automatically to another. The tree-indexed FOHA of hierarchical exchangeability is a statement about distributional invariance and measurable representation; the ANOVA FOHA is a Bayesian prior on an interaction tensor; the level-4 FOHA is a typed data structure; the FROST FOHA is an integration algorithm; and the sparse-array FOHA is a sensor design for fourth-order cumulant processing. Their shared terminology reflects structural analogy, not formal equivalence.