Deep Reinforcement Learning for Fano Hypersurfaces

Marc Truter

math.AGMar 16, 2026

Deep Reinforcement Learning for Fano Hypersurfaces

Authors: Marc Truter

Abstract

We design a deep reinforcement learning algorithm to explore a high-dimensional integer lattice with sparse rewards, training a feedforward neural network as a dynamic search heuristic to steer exploration toward reward dense regions. We apply this to the discovery of Fano 4-fold hypersurfaces with terminal singularities, objects of central importance in algebraic geometry. Fano varieties with terminal singularities are fundamental building blocks of algebraic varieties, and explicit examples serve as a vital testing ground for the development and generalisation of theory. Despite decades of effort, the combinatorial intractability of the underlying search space has left this classification severely incomplete. Our reinforcement learning approach yields thousands of previously unknown examples, hundreds of which we show are inaccessible to known search methods.

Figures & tables

Figure 1. The $6$ -dimensional dynamic heuristic (deep reinforcement learning) search for terminal Fano $4$ -fold hypersurfaces projected onto $2$ -dimensions. While the quasismooth terminal points are fully classified, the search discovers previously unknown nonquasismooth terminal ones. See Figure 9 for full details.

Figure 2. Flowchart of the fixed heuristic search algorithm.

Figure 3. Flowchart of the dynamic heuristic search algorithm.

Figure 4. Exhaustive search of Fano 4-fold hypersurfaces with terminal singularities. In total $84,733$ terminal examples were found, $7,346$ quasismooth, and $77,387$ nonquasismooth. Each frame shows points in $\mathbb{Z}^{2}$ , obtained by projecting the original $\mathbb{Z}^{6}$ search space onto consecutive coordinate pairs via $(a_{1},\ldots,a_{6})\mapsto(a_{i},a_{i+1})$ .

Figure 5. Classification of $11,617$ quasismooth Fano $4$ -fold hypersurfaces with terminal singularities. Each frame shows points in $\mathbb{Z}^{2}$ , obtained by projecting the original $\mathbb{Z}^{6}$ search space onto consecutive coordinate pairs via $(a_{1},\ldots,a_{6})\mapsto(a_{i},a_{i+1})$ .

Figure 6. The cumulative number of terminal Fano $4$ -fold hypersurfaces found in the exhaustive search against degree.

Panel of Figure 7 — Figure 7. ( 7A ) The real locus of the affine cone of the member of $X_{6}\subset\mathbb{P}(1,2,3)$ whose coefficients are all equal to $1$ , which is quasismooth and hence admits only quotient singularities. ( 7B ) The real locus of the affine cone of the member of $X_{10}\subset\mathbb{P}(1,3,4)$ whose coefficients are all equal to $1$ , which is nonquasismooth; the hyperquotient singularity is visible as the line passing through the origin.

Hyperparameter	Values
MLP Neural Network Layers	(40,)
Activation function	LeakyReLU
LeakyReLU slope	0.01
Optimiser	Adam
Optimiser learning rate	0.001
TD discount factor, $\gamma$	0.2

Table 1. Hyperparameters used in the dynamic heuristic (deep reinforcement learning) search.

Figure 8. Fixed heuristic search. The search found $113,996$ nonquasismooth Fano $4$ -fold hypersurfaces with terminal singularities. Each frame shows points in $\mathbb{Z}^{2}$ , obtained by projecting the original $\mathbb{Z}^{6}$ search space onto consecutive coordinate pairs via $(a_{1},\ldots,a_{6})\mapsto(a_{i},a_{i+1})$ .

Figure 9. Dynamic heuristic (deep reinforcement learning) search. The search found $85,262$ nonquasismooth Fano $4$ -fold hypersurfaces with terminal singularities. Each frame shows points in $\mathbb{Z}^{2}$ , obtained by projecting the original $\mathbb{Z}^{6}$ search space onto consecutive coordinate pairs via $(a_{1},\ldots,a_{6})\mapsto(a_{i},a_{i+1})$ .

Panel of Figure 10 — Figure 10. ( 10A ) A histogram showing the distribution of distances from points found in the fixed but not dynamic search to their closest neighbours. ( 10B ) A histogram showing the distribution of distances from points found in the dynamic but not fixed search to their closest neighbours. The dynamic search finds points at greater distances than the fixed search, corresponding to points that are increasingly inaccessible to the latter. ( 10C ) A graph showing the number of nonquasismooth terminal examples found against the number of steps taken by the search. As expected, the dynamic search finds fewer terminal points in the same number of steps as the fixed one, since more steps are spent in unprofitable territory in order to reach terminal points at greater distances.

RL for Combinatorial Optimization

Reinforcement learning for combinatorial optimization learns sequential decisions and search strategies for difficult discrete problems.

See all 106 papers

Jun 22, 2026

Hierarchical Reinforcement Learning for Sparse-Reward Search in Commutative Algebra

Applying machine learning techniques to solving long-standing mathematical conjectures can be particularly challenging due to their extreme reward sparsity. As an illustrative example, we consider Kalai's algebraic Hirsch conjecture and recast the construction of its counterexamples as a sparse-reward reinforcement learning problem on graphs. We propose a constrained options-based HRL framework with an equivariant graph neural network policy, which allows us to learn useful temporal abstractions for this task. We evaluate our approach over a wide range of degrees and demonstrate that it consistently outperforms classical RL algorithms as well as greedy search. By exploiting the hierarchical structure of the problem, we effectively provide a first-of-its-kind application of HRL to a problem in commutative algebra.

Giorgi Butbaia, Paul Orland, Coco Huang +7

Department of Mathematics, California Institute of Technology, Pasadena, CA · Department of Mathematics, Temple University, Philadelphia, PA · Department of Mathematics, University of Kansas, Lawrence, KS +1

Hierarchical RL Sparse-Reward RL

Ask PDF

Sep 17, 2026

Special Lagrangian cones in Deep Learning

We introduce a matrix generalization of the cone of Harvey and Lawson and prove that it is an exact special Lagrangian manifold. We further show that it belongs to a family of exact special Lagrangian manifolds that foliate the balanced manifold arising in deep learning.

Tejas Kotwal, Govind Menon

Institute for Advanced Study, Princeton

Topic pending

Ask PDF

Jun 13, 2026

Discovering Lattice Reduction Strategies via Self-Play

The Lenstra-Lenstra-Lovász (LLL) algorithm is a seminal contribution to computer science used for lattice basis reduction, yet its polynomial-time outputs produce bases that are far from optimal as the dimension grows. We show that deep reinforcement learning can discover strictly superior, generalizable reduction strategies by interacting with the primitive action space of LLL. We formulate lattice reduction as a single-player Markov Decision Process (MDP) and train a deep residual network using an AlphaZero-style self-play pipeline augmented with adaptive-horizon MCTS (Monte Carlo Tree Search), which couples multi-step network predictions with an entropy-gated expansion mechanism. The resulting policy, DeltaStar, is trained exclusively on small

8

-dimensional

q

-ary lattices and requires fewer primitive row operations than LLL. Crucially, it generalizes zero-shot to unseen moduli and higher dimensions up to

n=32

without retraining.

Mohamed Malhou, Kristin Lauter, Ludovic Perret

FAIR, Meta Superintelligence Labs & Sorbonne Université CNRS, LIP6 Paris, France · EPITA, EPITA Research Lab (LRE) Le Kremlin-Bicêtre, France

Reinforcement Learning RL for Combinatorial Optimization Policy Learning

Ask PDF

Abstract

RL for Combinatorial Optimization

Explore similar work

Hierarchical Reinforcement Learning for Sparse-Reward Search in Commutative Algebra

Special Lagrangian cones in Deep Learning

Discovering Lattice Reduction Strategies via Self-Play