# Crystal-Structured Discrete Latents

- ID: 2783
- Canonical URL: https://synthcore.org/idea/2783/crystal-structured-discrete-latents
- API JSON: https://synthcore.org/api/idea/2783.json
- API Markdown: https://synthcore.org/api/idea/2783.md
- Verification status: unverified
- Source: [arXiv:2608.27949](https://arxiv.org/abs/2608.27949)
- Category: architecture
- Solves: generalization, sample-efficiency, stability
- ML areas: embedding, quantization, moe-routing, regularization
- Math tags: representation-theory, combinatorics, algebra, graph-theory
- Ratings: usefulness 5/10; difficulty 5/10; novelty 8/10

## Idea description

Use reverse plane partitions of a minuscule heap as the discrete codebook for a VQ-VAE or discrete sequence model. Codes are not arbitrary indices: each code is an order-preserving array, and crystal raising/lowering operators define a sparse, semantically structured neighborhood graph for augmentation, routing, and metric regularization.

## Mathematical statement

In the minuscule case, Proposition 4.4 identifies the crystal of the highest-weight representation with ordered tensor products, \(\{b_1\otimes\cdots\otimes b_s\in B(\overline{\Lambda}_r)^{\otimes s}\mid b_1\leq\cdots\leq b_s\}\cong B(s\overline{\Lambda}_r)\). The paper realizes these states as reverse plane partitions \(\rho\) on a minuscule heap \(\mathcal P_r\): an RPP is an integer-valued order-preserving map \(\rho:\mathcal P_r\to\{0,\ldots,s\}\), so \(x\leq y\Rightarrow\rho(x)\leq\rho(y)\). Each heap cell has a color \(\pi(x)\in I_0\). The crystal raising operator \(f_i\) adds one to the largest entry among cells in the color fiber \(\pi^{-1}(i)\) whenever the result remains a valid RPP; \(e_i\) subtracts one from the corresponding removable entry, and returns \(\mathbf 0\) when no move exists. To determine which cell is active, read the color fiber bottom-to-top, write a right parenthesis for each addable cell and a left parenthesis for each removable cell, and repeatedly cancel adjacent \(()\) pairs; the surviving symbol selects the operator-active cell. For type \(A_{n-1}\), the heap is an \(r\times(n-r)\) rectangle, and the paper states that RSK gives a crystal isomorphism between polyhedral and Lusztig data for \(B(s\overline{\Lambda}_r)\). We exploit exact closure of the RPP state space and the operator-defined adjacency graph, not the representation-theoretic labels.

## Key formulas

- $$\{b_{1}\otimes\cdots\otimes b_{s}\in B(\overline{\Lambda}_{r})^{\otimes s}\mid b_{1}\leq\cdots\leq b_{s}\}\cong B(s\overline{\Lambda}_{r}).$$
- $$\rho:\mathcal P_r\to\{0,1,\ldots,s\},\qquad x\leq y\ \Longrightarrow\ \rho(x)\leq\rho(y).$$
- $$f_i(\rho)=\rho+\mathbf 1_{x_i^*},\qquad e_i(\rho)=\rho-\mathbf 1_{y_i^*},$$
- $$\mathcal L(\rho)=\text{unmatched parentheses after canceling all adjacent }()\text{ pairs in the bottom-to-top signature of }\pi^{-1}(i).$$

## Implementation notes

Use the module as the codebook and neighborhood graph of a VQ-VAE, discrete autoencoder, or small MoE router. First, restrict to type \(A_{n-1}\), choose a rectangular heap \(\mathcal P_r=\{1,\ldots,r\}\times\{1,\ldots,n-r\}\), and choose a small height \(s\), such as 3--8. Each code is an integer matrix \(\rho\) satisfying the selected row/column monotonicity convention and entries in \([0,s]\). Enumerate all valid RPPs for small shapes, or sample them by dynamic programming, and assign each valid RPP a learned vector embedding. (1) The encoder produces logits over cells or a continuous matrix \(z\). (2) Project \(z\) to the nearest valid RPP under squared error, using exhaustive search for the MVP or an isotonic projection for larger shapes. (3) Use the selected code embedding in place of the ordinary VQ code, with a straight-through estimator. (4) For each color \(i\), implement the paper's bottom-to-top signature: mark cells as addable or removable under the RPP inequalities, emit parentheses, cancel adjacent \(()\) pairs with a stack, and choose the surviving active cell for \(f_i\) or \(e_i\). (5) Add a local smoothness loss: for sampled valid moves, \(\|E(\rho)-E(f_i\rho)\|_2^2\) should be smaller than the distance to randomly chosen unrelated codes; optionally train a predictor to map the embedding of \(\rho\) to the embedding of \(f_i\rho\). At training time, augment each selected code with one random valid \(f_i\) or \(e_i\) move and require decoder consistency. The RPP constraints and operator transitions are computed exactly; embedding distances, active-color sampling, and the projection approximation are empirical design choices. The first cheap experiment should use CIFAR-10 with a 4x4 VQ-VAE and compare against an equal-size unconstrained codebook at equal latent bitrate. Measure reconstruction FID, codebook perplexity, dead-code rate, robustness to code perturbations, and downstream linear-probe accuracy. Success means lower dead-code rate and better reconstruction or perturbation robustness without increasing encoder FLOPs. A second small test can use a toy MoE with crystal-neighbor augmentation and measure routing entropy and expert-load variance under input noise.

## Disclaimer

AI-generated research hypothesis, automatically tested. Not peer-reviewed.