SAND-Math: Critical Models & Synthetic Data

• SAND-Math is a comprehensive suite uniting discrete sandpile dynamics, variational PDEs, algebraic group theory, and ethnomathematical insights.
• It employs methods from finite-state transducers, mixed FEM discretizations, and quasi-variational inequalities to elucidate stability, wave formation, and mass flux in granular flows.
• Recent advances include synthetic data pipelines that generate challenging math problems for LLMs, enhancing problem difficulty calibration and interdisciplinary applications.

SAND-Math refers to a suite of mathematical concepts, models, and computational frameworks, spanning discrete dynamical systems (notably sandpile and chip-firing models), variational PDEs arising from granular flow, graph-theoretic and algebraic structures tied to self-organized criticality, ethnomathematical representations in cultural practices, and—recently—synthetic data generation pipelines for advanced mathematics question-answering with LLMs. The unifying thread is the emergence of complex, often critical, mathematical phenomena from local rules, structure, or data generation protocols reminiscent of granular avalanches.

1. Discrete Dynamical Sandpile Models: Kadanoff and Abelian Sandpile

The Kadanoff Sand Pile Model (KSPM) generalizes the Abelian Sandpile Model (ASM), introducing a parameter $D$ which controls toppling thresholds and the spread of grains. The state is a non-increasing, ultimately zero configuration $h=(h_0,h_1,\ldots)$, or equivalently in the height-difference (slope) representation $\sigma_i = h_i - h_{i+1}$.

Local Topple Rule: For fixed $D\ge2$, if $\sigma_i \ge D$ at column $i$, $D-1$ grains move from $i$ to columns $i+1$ to $i+D-1$:

• $h=(h_0,h_1,\ldots)$0 (if $h=(h_0,h_1,\ldots)$1)
• $h=(h_0,h_1,\ldots)$2
• $h=(h_0,h_1,\ldots)$3
• Other $h=(h_0,h_1,\ldots)$4 unchanged

Stability and Fixed Points: A configuration is stable if $h=(h_0,h_1,\ldots)$5 for all $h=(h_0,h_1,\ldots)$6. Due to the diamond property and finiteness, every initial state (e.g., $h=(h_0,h_1,\ldots)$7 grains at $h=(h_0,h_1,\ldots)$8) relaxes, uniquely, to a stable configuration $h=(h_0,h_1,\ldots)$9.

Iterating grain addition—each time adding one grain at $\sigma_i = h_i - h_{i+1}$0 and stabilizing—generates a sequence of avalanches, each captured by a firing sequence of column indices.

Wave Emergence and Analysis: For $\sigma_i = h_i - h_{i+1}$1, beyond a column at distance $\sigma_i = h_i - h_{i+1}$2 from the origin, the fixed point displays a periodic wave

$\sigma_i = h_i - h_{i+1}$3

modulo a possible single zero; generally, larger $\sigma_i = h_i - h_{i+1}$4 yields more intricate wave patterns. The spatial regularity can be formalized via a deterministic finite-state transducer, mapping temporal avalanche types to spatial patterns, establishing an analogy to renormalization and linking temporal dynamics to emergent spatial order (Perrot et al., 2013, Perrot et al., 2013).

Mathematical Significance: Compared to ASM ($\sigma_i = h_i - h_{i+1}$5), where fixed points are staircases, KSPM $\sigma_i = h_i - h_{i+1}$6 supports nontrivial wave structures beyond a logarithmic prefix, illustrating the deep connection between local rules and global self-organized criticality.

2. Variational and Quasi-Variational PDEs for Sandpile Growth

The evolution of sandpile surfaces with slope constraints leads to quasi-variational inequalities. For a domain $\sigma_i = h_i - h_{i+1}$7, with height $\sigma_i = h_i - h_{i+1}$8 and flux $\sigma_i = h_i - h_{i+1}$9, mass balance yields:

$D\ge2$0

with an angle-of-repose constraint $D\ge2$1. The admissible set depends nonlinearly on the evolving pile:

$D\ge2$2

Quasi-Variational Inequality (QVI): The standard formulation seeks $D\ge2$3, $D\ge2$4, such that

$D\ge2$5

Mixed Primal-Dual Regularization: To recover both $D\ge2$6 and $D\ge2$7, regularized formulations introduce power-law terms $D\ge2$8 and mollified slope constraints $D\ge2$9. Existence is established through finite-element discretization, weak convergence arguments, and passage to the limit in mesh, time, and regularization parameters (Barrett et al., 2012, Barrett et al., 2013).

Discretization and Algorithms:

• Raviart–Thomas (RT$\sigma_i \ge D$0) mixed FEM captures mass flux and surface with guaranteed convergence.
• Nonconforming $\sigma_i \ge D$1 finite elements (Crouzeix–Raviart) offer simpler implementational options while retaining accuracy and convergence guarantees.
• Splitting methods and Picard iteration solve nonlinear algebraic systems per time step. Numerical experiments validate these schemes, demonstrating their correctness and ability to recover both height and flux—including singular, measure-valued flows.

3. Algebraic and Group-Theoretical Aspects: Sandpile Groups and Limits

Sandpile groups $\sigma_i \ge D$2 encode the recurrent stable states of a sandpile process on finite graphs $\sigma_i \ge D$3 (with sink). These groups are canonically isomorphic to the cokernel of the reduced Laplacian:

$\sigma_i \ge D$4

and their order gives the number of spanning trees on $\sigma_i \ge D$5 by the Matrix-Tree Theorem.

Limits and Canonical Maps: Considering rectangular domains $\sigma_i \ge D$6 nested via monomorphisms $\sigma_i \ge D$7 and their dual epimorphisms $\sigma_i \ge D$8, the inverse limit of the extended sandpile groups is shown to be

$\sigma_i \ge D$9

with $i$0 the space of $i$1-valued harmonic functions. This is a real two-torus in the planar case (Shkolnikov, 11 Jun 2025).

Combinatorial Consequences: For rectangle sizes $i$2 and scaling $i$3, the cardinalities of the sandpile groups (spanning tree counts) satisfy

$i$4

4. Graph-Theoretic and Ethnomathematical Connections

Ethnomathematical practices, such as sand drawings in Oceania and Africa, embody fundamental mathematical concepts—Eulerian cycles, symmetry groups, and algorithmic construction—often predating formal Western codifications.

Continuity, Cyclicity, and Symmetry: Drawings are realized as continuous (piecewise-differentiable) maps $i$5, corresponding to Euler circuits/trails in planar graphs. Cultural artifacts frequently realize cycles $i$6 and exhibit actions of dihedral symmetry groups $i$7.

Comparison to Classical European Graph Theory: The tracing of sand-figures matches Euler's celebrated solution to the Königsberg bridges problem. Indigenous patterns manifest Eulerian or Hamiltonian circuits and are analyzed with the same theoretical apparatus, further supporting curriculum enrichment via ethnomathematics (Wang et al., 2024).

5. Synthetic SAND-Math Datasets for Mathematical Reasoning with LLMs

"SAND-Math" (Synthetic Augmented Novel and Difficult Mathematics) extends the scope to automated generation of challenging mathematics problems—targeted at training LLMs—and provides an iterative pipeline for data creation, filtering, difficulty elevation ("Difficulty Hiking"), and curation.

Pipeline Structure:

• Initial problem generation by LLM (DeepSeek-R1), yielding raw problems and solutions.
• Self-consistency filtering: problems are retained only if solutions are self-consistent across independent samples.
• De-duplication and benchmark decontamination using semantic similarity and Llama-judged checks.
• Difficulty filtering: problems unsolved by a strong model (Qwen2.5-32B) are preferentially retained.
• Difficulty hiking: selected problems are algorithmically "hiked" in difficulty by requiring novel theorem usage or cross-domain integration within the problem construct.

Dataset and Evaluation:

• The final SAND-Math dataset comprises 8,842 novel, decontaminated, graded (difficulty-rated) math problems.
• Experimental results: augmentation with SAND-Math yields increases in pass@1 accuracy (up to $i$82.85 points on AIME25 over previous best synthetic sets).
• Difficulty hiking step raises mean problem rating from 5.02 to 5.98 and increases the percentage of problems at $i$9 from 47.2% to 76.8%.
• Released at https://huggingface.co/datasets/amd/SAND-MATH (Manem et al., 28 Jul 2025).

6. Statistical and Number-Theoretic SAND-Math: Digital-Sum Phenomena

S(um)anD(ifference) numbers and primes are pairs $D-1$0 such that the digital sum $D-1$1 in a fixed numerical base $D-1$2. When both $D-1$3 and $D-1$4 are prime, these are "SanD primes." Asymptotic analysis demonstrates:

• The number of SanD numbers $D-1$5 grows as $D-1$6 in decimal
• The number of SanD primes $D-1$7 has the form $D-1$8 (base 10), or $D-1$9 in binary

This parallels the twin-prime constant in growth, but slow "quasi-fractal" fluctuations in the digital sum accelerate convergence irregularity, traced to Delange's periodic, nowhere-differentiable corrections to $i$0 statistics (Dyson et al., 2019).

7. Significance, Applications, and Future Perspectives

SAND-Math unifies critical-state phenomena in distributed dynamical systems, offers a rigorous PDE and algebraic framework for granular flows and self-organized criticality, elucidates deep connections to group theory, combinatorics, and number theory, and now extends to LLM-oriented data generation and computational ethnomathematics. The field bridges theoretical mathematics, computational modeling, data-driven AI, and mathematical anthropology. Open directions include multi-round difficulty hiking for synthetic data, further algebraic study of sandpile group limits, and continued cross-fertilization with topological, stochastic, and ethnomathematical perspectives.