Homological Foods: Topological Culinary Innovations

• Homological foods are a mathematical framework employing topological data analysis to identify persistent recipe patterns and combinatorial food structures.
• The methodology encodes recipes as binary vectors and constructs food quivers to uncover directed cycles that inform recursive culinary innovation.
• Optimization techniques and empirical case studies demonstrate the potential to generate novel food products with underexplored ingredient combinations.

Homological foods constitute a mathematical and data-driven framework for analyzing and generating novel food products based on topological principles, specifically persistent homology and the structure of directed cycles in graphs of commercially available foods. Two distinct strands define the term: one leverages topological data analysis (TDA) to map the “shape” of recipe combinatorics, while the other formalizes the idea of self-referential and cyclically-combined foods using recursion on product graphs, or “food quivers” (Escolar et al., 2024, Bosca, 1 Apr 2026).

1. Mathematical Representations of Recipes and Foods

Recipes are encoded as binary vectors $x \in \{0,1\}^M$ where $M$ is the total number of ingredients (in one study, $M=381$), with $x_j=1$ indicating ingredient $i_j$ is present. Pairwise similarity is quantified by cosine dissimilarity:

$$s_\mathrm{cos}(x, y) = \frac{x \cdot y}{\|x\| \|y\|}; \quad d(x, y) = 1 - s_\mathrm{cos}(x, y)$$

with $d(x, y) \in [0,1]$ quantifying flavor-combinatorial distance (Escolar et al., 2024).

In commercial product space, foods form the vertices of a directed graph, or “food quiver” $Q$. An arrow $A \rightarrow B$ exists if blending $A$ into $M$0 as a mix-in yields a valid supermarket product. Cycles in this quiver underpin the recursive structure of homological foods (Bosca, 1 Apr 2026).

2. Topological Data Analysis and Persistent Homology

To capture higher-order patterns beyond pairwise similarity, TDA constructs a Vietoris–Rips simplicial complex $M$1 for a set $M$2 of recipes at scale $M$3:

$M$4

As $M$5 increases, complexes grow via inclusion, forming a filtration. The $M$6-graded chain complexes $M$7 and boundary maps $M$8 support computation of homology groups $M$9, with $M=381$0th Betti number $M=381$1 enumerating $M=381$2-dimensional “holes” (Escolar et al., 2024).

Persistent homology tracks the birth and death of features through the filtration, yielding barcodes or persistence diagrams $M=381$3. For culinary data, $M=381$4 corresponds to clusters of similar recipes; $M=381$5 corresponds to cycles enclosing “holes”—regions in combinatorial space unpopulated by known recipes.

3. Homological Foods: Graph-Theoretic and Recursive Formulation

Homological foods arise from directed cycles in the food quiver $M=381$6 (Bosca, 1 Apr 2026). For a $M=381$7-cycle $M=381$8, one defines a $M=381$9-chain of affine recursions for each food’s key compositional fraction: $x_j=1$0

$x_j=1$1

A coupled system of recursions for the mass-fractions of ingredients results, with coefficients given by empirically measured mix-in fractions. The system’s contraction property ensures convergence to a unique fixed point; the limiting food is then the homological food associated to the cycle. The corresponding “$x_j=1$2-food” has well-defined limiting compositions regardless of initial composition (Bosca, 1 Apr 2026).

4. Culinary Innovation via Combinatorial Optimization on Topological Features

Persistent homology detects long-lived $x_j=1$3 features, each representing a cycle of recipes circling an underexplored “hole” in combinatorial space. Consider a representative $x_j=1$4-cycle $x_j=1$5 with recipes $x_j=1$6. The candidate ingredient pool is $x_j=1$7. The objective is to synthesize a new recipe $x_j=1$8 of fixed size $x_j=1$9 that is maximally “distant” from all known recipes, i.e., that lies within the topological hole: $i_j$0 The search is cast as a mixed-integer linear program with the epigraph trick, solvable by GLPK, and up to $i_j$1 distinct optima per cycle can be found (Escolar et al., 2024).

Empirical evaluation on $i_j$2 recipes ($i_j$3 ingredients) showed that out of $i_j$4 suggested 5-ingredient combinations, $i_j$5 matched existing recipes and $i_j$6 were strict sub-recipes, with a bias toward rare ingredient usage. Experimental case studies (cream-cheese biscuit variants) demonstrated palatability as validated by blinded sensory study ($i_j$7), with all biscuits scoring above acceptability threshold (Escolar et al., 2024).

5. Classification of Homological Foods by Cycle Structure

Homological foods are classified by the minimal directed cycles in the associated food quiver: | Cycle Length $i_j$8 | Homological Food Type | Example | |------------------|----------------------|----------------------------------------| | 1 | mono-$i_j$9 food | $$s_\mathrm{cos}(x, y) = \frac{x \cdot y}{\|x\| \|y\|}; \quad d(x, y) = 1 - s_\mathrm{cos}(x, y)$$0-Oreo (self-loop Oreo→Oreo) | | 2 | bi-$$s_\mathrm{cos}(x, y) = \frac{x \cdot y}{\|x\| \|y\|}; \quad d(x, y) = 1 - s_\mathrm{cos}(x, y)$$1 food | $$s_\mathrm{cos}(x, y) = \frac{x \cdot y}{\|x\| \|y\|}; \quad d(x, y) = 1 - s_\mathrm{cos}(x, y)$$2-M& M Cookie, $$s_\mathrm{cos}(x, y) = \frac{x \cdot y}{\|x\| \|y\|}; \quad d(x, y) = 1 - s_\mathrm{cos}(x, y)$$3-Crunchy Cookie M&M | | $$s_\mathrm{cos}(x, y) = \frac{x \cdot y}{\|x\| \|y\|}; \quad d(x, y) = 1 - s_\mathrm{cos}(x, y)$$4 | $$s_\mathrm{cos}(x, y) = \frac{x \cdot y}{\|x\| \|y\|}; \quad d(x, y) = 1 - s_\mathrm{cos}(x, y)$$5-$$s_\mathrm{cos}(x, y) = \frac{x \cdot y}{\|x\| \|y\|}; \quad d(x, y) = 1 - s_\mathrm{cos}(x, y)$$6 food | 3-cycle: Oreo→Ice cream→Cake→Oreo |

In each case, the limiting product composition is determined by solving the recursively-coupled mass-fraction equations induced by the corresponding cycle and empirically measured mix-in factors (Bosca, 1 Apr 2026).

6. Limitations and Open Problems

Several key limitations are identified:

• The one-hot encoding omits information on ingredient proportions, preparation methods, and molecular flavor profiles (Escolar et al., 2024).
• The cosine similarity used is not a metric, complicating geometric interpretations.
• Higher-dimensional holes ($$s_\mathrm{cos}(x, y) = \frac{x \cdot y}{\|x\| \|y\|}; \quad d(x, y) = 1 - s_\mathrm{cos}(x, y)$$7 and above) remain challenging to compute and interpret in culinary space.
• Empirical validation remains limited to specific case studies; broader sensory and market evaluation is an open direction.
• The effect of intersecting cycles in the food quiver $$s_\mathrm{cos}(x, y) = \frac{x \cdot y}{\|x\| \|y\|}; \quad d(x, y) = 1 - s_\mathrm{cos}(x, y)$$8 on convergence rates and limiting compositions is conjectural, motivating further study of the spectral sheaf/Laplacian associated to $$s_\mathrm{cos}(x, y) = \frac{x \cdot y}{\|x\| \|y\|}; \quad d(x, y) = 1 - s_\mathrm{cos}(x, y)$$9 (Bosca, 1 Apr 2026).

7. Prospects and Extensions

Homological and persistent-homology–based methods facilitate systematic exploration of combinatorially novel, yet coherent, food combinations—an approach that augments culinary innovation by charting and filling “holes” in recipe space. Potential extensions include:

• Integration of flavor-chemical networks or multimodal embeddings (e.g., image–text) to augment recipe representations.
• Incorporation of ingredient amounts, preparation steps, and sensory data to enrich both the dissimilarity measures and topological features.
• Graph-spectral and sheaf-theoretic perspectives, to predict limits and convergence rates for homological foods arising from complex, intersecting cycles in $d(x, y) \in [0,1]$0.

This suggests a paradigm in which topological and combinatorial principles guide both the analysis of extant culinary practices and the algorithmic synthesis of unprecedented ingredient ensembles, thus enabling data-driven gastronomic creativity (Escolar et al., 2024, Bosca, 1 Apr 2026).