Chapter 9: The Eigenvalues of a Kitchen
“The cooking of a society is a language in which it unconsciously translates its structure.” — Claude Lévi-Strauss, The Raw and the Cooked (1964)
RUNNING EXAMPLE — THE FUSION TASTING
The brief is one sentence long and the chef has read it forty times: “Build us a tasting that travels from Oaxaca to Osaka in nine courses.” She is standing at the pass with two dishes in front of her, one at each elbow. On the left, a mole negro — toasted chiles, charred onion, sesame, a whisper of chocolate, the long bitter-sweet bass note of burnt seeds, all of it carried on lard and lifted by a squeeze of lime. On the right, a dashi-poached snapper — kombu and katsuobushi, the clean glutamate hum of the broth, mirin’s restrained sweetness, a brightness of yuzu, the whole thing lean and translucent where the mole is opaque and dense.
Course five is supposed to be the hinge: the dish that is halfway between them. Not mole beside fish, not fish beside mole — halfway. She has tried the obvious thing, mole over snapper, and it was a collision: the mole’s lard and roasted bitterness flattened the dashi’s delicate umami into mud, and the dish tasted like neither parent and like no third thing either. It curdled. So now she is asking a different question. If these two dishes are two points, what is the straight line between them? And — the harder question, the one that took her years to even phrase — is the straight line in the kitchen the same as the straight line in the mouth?
She lines up the shared coordinates first. Both dishes have umami: dashi from the sea, mole from toasted seeds and a touch of fish-sauce-adjacent funk she has snuck in before. Both have acid: yuzu, lime. Both have a managed sweetness. Along those axes she can move smoothly — a poaching liquid that is two-thirds dashi and one-third a thin, strained, fat-free mole essence holds together, brightened with a citrus that is half yuzu and half lime. What will not interpolate is the fat and the char: lard and deep-toast bitterness have no counterpart on the Japanese side, and every step she takes along those directions tears the dish apart. The geodesic, she is discovering, runs through the dimensions the two cuisines share, and falls off a cliff along the dimensions they do not.
The Geometry of a Tradition
Up to this chapter we have treated a single dish as a point and a single transformation as a map. We now zoom out by an order of magnitude. The object of interest is no longer a plate but a tradition — Italian cooking, Japanese cooking, Oaxacan cooking — a body of thousands of recipes accumulated over centuries by millions of cooks who never met and never coordinated, yet who somehow produced a thing with a recognizable identity. You can taste a dish you have never encountered and say, with confidence, “that is Thai,” or “that is unmistakably French.” What is the mathematical content of that judgment?
The claim of this chapter is that a cuisine is a probability distribution over the flavor space $\mathcal{F}$ developed in Chapter 1, and that the tools for analyzing distributions — their means, their covariances, their principal axes, the geodesics between them — are exactly the tools for analyzing cuisines. A cuisine has a center of mass (its characteristic flavor profile), a spread (how adventurous or constrained it is), and a set of principal directions: the eigenvectors of its covariance matrix, the axes along which its dishes vary most. These eigenvectors are the title of the chapter. They are the answer to the question “what does this cuisine care about?” — and, as we will see, they are computable from a recipe database the way the principal components of a face dataset are computable from photographs.
[Modeling Axiom.] A cuisine $C$ is a probability distribution $p_C(\mathbf{f})$ over flavor space $\mathcal{F}$, estimated empirically from a corpus of recipes, each recipe contributing one (or several) flavor vectors $\mathbf{f} \in \mathcal{F}$. The cuisine is characterized to second order by its mean $\boldsymbol{\mu}_C = \mathbb{E}[\mathbf{f}]$ and covariance $\Sigma_C = \mathbb{E}[(\mathbf{f}-\boldsymbol{\mu}_C)(\mathbf{f}-\boldsymbol{\mu}_C)^\top]$. The eigenvectors of $\Sigma_C$ are the flavor eigenvectors of the cuisine; their eigenvalues report how much of the cuisine’s variation each direction explains.
This is the natural continuation of the book’s argument. Chapter 1 gave us the space; Chapter 8 gave us the recipe as a point constructed within it. A cuisine is what you get when you plot ten thousand such points and ask after the shape of the cloud. The shape is not round. It is a thin, oriented, structured ellipsoid — and its orientation is the cuisine’s identity.
9.1 A Cuisine as a Probability Distribution in Flavor Space
Begin with the cloud. Take every recipe in a national tradition, compute the flavor vector of each (Chapter 8: ingredients as basis vectors, the dish as their seasoned sum), and scatter the resulting points in $\mathcal{F}$. What you see is not a uniform fog filling the space, and it is not a single tight knot. It is a region — a connected, anisotropic, characteristically positioned cloud — and different cuisines occupy different regions, with partial overlap.
[Empirical.] The regions are real and they are separable. The flavor network of Ahn, Ahnert, Bagrow, and Barabási (2011) — built from 56,498 recipes drawn from five regional corpora and the volatile-compound profiles of their ingredients — showed that North American and Western European recipes occupy a different part of compound space than East Asian and Southern European recipes, separable along the axis of compound-sharing (we return to this in §9.6 and it connects to the Western/Eastern asymmetry of Chapter 4.3). The separation is not perfect and the clouds overlap, exactly as two Gaussians with distinct means but nonzero variance overlap. That is the point: a cuisine is a distribution, not a set, and distributions have tails. There exist French dishes that taste vaguely Japanese and Japanese dishes that taste vaguely French, and they live in the overlap region where $p_{\text{Fr}}(\mathbf{f})$ and $p_{\text{Jp}}(\mathbf{f})$ are both nonneglible.
The distributional picture immediately earns its keep by explaining things a list cannot. Authenticity becomes likelihood: a dish is “authentically Sichuanese” to the degree that $p_{\text{Sichuan}}(\mathbf{f})$ is high at its location — which is why a dish can violate every traditional ingredient rule and still read as authentic if it lands in the right region, and why canonically Sichuanese ingredients can read as wrong if the proportions push the result into the tail. Typicality becomes proximity to the mean $\boldsymbol{\mu}_C$: a simple tomato-basil-olive-oil-garlic plate feels more emblematically Italian than an elaborate regional specialty far out on a principal axis. And the endless arguments over whether this or that ingredient is “allowed” dissolve: a cuisine was never a rulebook. It is a density, and densities have no sharp edges.
9.2 PCA of Recipe Databases: The Flavor Eigenvectors of Each Cuisine
A distribution characterized to second order is characterized by its mean and its covariance, and the covariance is where the structure lives. Principal Component Analysis (PCA) — the eigendecomposition of $\Sigma_C$ — finds the orthogonal directions along which a cuisine’s dishes vary most, ordered by how much variance each explains. These are the cuisine’s flavor eigenvectors, and they are interpretable.
[Theorem.] For symmetric positive-semidefinite covariance $\Sigma_C$, the eigendecomposition $\Sigma_C \mathbf{v}_k = \lambda_k \mathbf{v}_k$ yields an orthonormal eigenbasis with real nonnegative eigenvalues. The eigenvectors $\mathbf{v}_1, \mathbf{v}_2, \dots$ (ordered by $\lambda_1 \ge \lambda_2 \ge \cdots$) are the principal flavor axes; $\lambda_k$ is the variance captured by axis $k$. The fraction $\lambda_k / \sum_j \lambda_j$ is the share of the cuisine’s total flavor variation that lies along $\mathbf{v}_k$. A cuisine whose first eigenvalue dominates ($\lambda_1 \gg \lambda_2$) varies mostly along a single axis; one with a flat eigenvalue spectrum spreads its dishes across many independent directions.
What do these axes look like in practice? Consider a stylized but representative reading. The first eigenvector of a classical French corpus tends to align with a richness axis — butter/cream/stock loading at one end, an acid-and-herb leanness at the other — because the largest single source of dish-to-dish variation in that tradition is how much fat-and-reduction a dish carries. The first eigenvector of a Japanese corpus tends to align with a dashi-to-not-dashi umami axis, because the presence or absence of the kombu-katsuobushi glutamate-inosinate base is the dominant organizing variable. A Mexican corpus loads heavily on a chile-and-toast axis: the depth of dried-chile, char, and roasted-seed character is the principal direction of variation, with the acid-and-fresh-herb pole (salsa verde, lime, cilantro) at the other end.
[Empirical.] These descriptions are interpretations of the kind of structure that recipe-corpus dimensionality reduction recovers; the specific axes depend on the corpus and the featurization, and honest practice reports the loadings rather than asserting the labels. The robust and repeatedly observed fact is that recipe corpora are low-rank: a handful of principal components — typically on the order of five to ten — captures the large majority of cross-recipe flavor variance within a cuisine. This is the same “eigenfaces” phenomenon that lets a few dozen components reconstruct a face dataset, and it is the cuisine-level echo of the dimensional collapse that has organized this whole book: the chemistry is enormous, the perception is tens of dimensions, and the within-cuisine variation is smaller still. A cuisine is a low-rank object embedded in a low-dimensional space. That is why it is learnable, and why a child raised in it acquires its grammar without being taught a single rule.
9.3 Signature Ingredients as Extreme Points
If the eigenvectors are the axes of a cuisine, the signature ingredients are the landmarks that define their direction. Soy sauce, fish sauce, cumin, dried chile, miso, garam masala, smen, nuoc cham — these are not merely common ingredients. They are the points whose presence most sharply distinguishes one cuisine’s distribution from the global mean, and geometrically they sit at the extreme points of the cuisine’s cloud, pulling its mean and tilting its principal axes toward themselves.
[Modeling Axiom.] An ingredient $i$ is a signature ingredient of cuisine $C$ to the degree that its flavor vector $\mathbf{f}_i$ is both (a) far from the global mean $\boldsymbol{\mu}_{\text{global}}$ and (b) far more prevalent in $C$ than in cuisines generally — formally, a high product of distinctiveness $\lVert \mathbf{f}_i - \boldsymbol{\mu}_{\text{global}}\rVert$ and lift $p_C(i)/p_{\text{global}}(i)$. Such an ingredient defines a direction in flavor space, $\hat{\mathbf{u}}_C \propto \mathbf{f}_i - \boldsymbol{\mu}_{\text{global}}$, and a cuisine’s identity is, to first approximation, its mean displacement along the directions of its signature ingredients.
This formalizes Elisabeth Rozin’s celebrated notion of the flavor principle: the idea that each cuisine is defined less by its raw ingredients than by a characteristic seasoning combination — soy sauce + sake + ginger for Japan, tomato + olive oil + garlic for southern Italy, chile + lime + cilantro for Mexico, onion + cumin + ginger + garlic for north India. In the geometric reading, a flavor principle is a fixed direction $\hat{\mathbf{u}}_C$ that nearly every dish in the cuisine is displaced along, regardless of the protein or vegetable at its center. The flavor principle is the cuisine’s first eigenvector made edible — the axis you cannot remove without the cuisine ceasing to be itself.
The extreme-point picture also explains the non-substitutability of signature ingredients. You can swap a vegetable in a curry and keep the curry; swap out the garam masala and you no longer have a curry at all, because you have collapsed the very direction that defined the dish’s membership in the distribution. Signature ingredients carry disproportionate identity-load precisely because they are the basis vectors of the cuisine’s distinguishing subspace — and removing a basis vector does not move you within a space, it shrinks the space.
9.4 Fusion as Interpolation: Geodesics Between Cuisine Distributions
Now we can state the fusion problem precisely. Two cuisines are two distributions, $p_A$ and $p_B$, with means $\boldsymbol{\mu}_A, \boldsymbol{\mu}_B$ and covariances $\Sigma_A, \Sigma_B$. A fusion dish is a point that lies, in some meaningful sense, between them. The naïve between is the linear midpoint of the means, $\tfrac{1}{2}(\boldsymbol{\mu}_A + \boldsymbol{\mu}_B)$. The chef’s hard-won knowledge is that the naïve between often curdles — and the geometry tells us why, and what to do instead.
[Modeling Axiom.] The correct interpolation between two distributions is not the linear average of their samples but the geodesic between the distributions under a metric that respects their structure. Under the Wasserstein (optimal-transport) metric on Gaussians, the geodesic from $(\boldsymbol{\mu}_A, \Sigma_A)$ to $(\boldsymbol{\mu}_B, \Sigma_B)$ interpolates the means linearly while interpolating the covariances along a curved path that keeps every intermediate $\Sigma(t)$ a valid (positive-semidefinite) covariance. The intermediate cuisine $C(t)$ is a genuine distribution at every step — not a mixture, not an average of two clouds, but a single coherent cloud that morphs from one shape into the other.
The distinction is the whole craft of fusion. A mixture — put mole on one half of the plate and dashi on the other — is the convex combination $t\,p_A + (1-t)\,p_B$, which is bimodal: it tastes like two things sitting next to each other, because it is two things sitting next to each other. A geodesic interpolation is unimodal: a single new thing whose flavor profile is genuinely intermediate. Good fusion travels the geodesic; bad fusion is a mixture wearing a single plate.
And here is the result the running-example chef discovered with her hands. The geodesic moves freely along the dimensions the two cuisines share — where both $\Sigma_A$ and $\Sigma_B$ have support, the interpolating covariance is well-conditioned and the intermediate dish is stable. Along dimensions where one cuisine has variance and the other has none — lard and deep-char in the mole, which have no counterpart in the dashi distribution — the optimal-transport map has nothing to transport to, and pushing along those directions takes you out of both distributions into a low-density no-man’s-land. That is the geometric definition of curdling: leaving the support of both parent distributions simultaneously. Fusion succeeds on the shared subspace and fails on the symmetric difference of the supports.
THE FUSION TASTING, REVISITED — WHY THE GEODESIC IS NOT THE STRAIGHT LINE
Return to course five. The chef’s first instinct — mole over snapper — was a linear average of ingredients: take half of dish A, half of dish B, superimpose. Geometrically she had computed $\tfrac{1}{2}(\mathbf{f}_{\text{mole}} + \mathbf{f}_{\text{dashi}})$ in ingredient space and assumed it would land between them in the mouth. It did not, because the straight line in ingredient space is not the geodesic in perceptual space (the metric of Chapter 3 is curved). The lard component had no partner in the dashi to transport onto; it simply piled up, dragging the dish off the support of both parents into the curdled region.
Her successful version — two-thirds dashi essence, one-third fat-stripped mole essence, citrus split between yuzu and lime — is a true geodesic. She interpolated only along the shared axes (umami, acid, restrained sweetness) and refused to interpolate along the unshared ones (she removed the lard rather than averaging it). The dish that resulted was not mole and not dashi and not a collision of the two. It was a single point on the curve $C(t)$ at $t \approx 0.5$, coherent because she had transported along the metric and not along the ingredient list. The geometry told her what to keep and what to leave on the cutting board.
9.5 Cultural Evolution as Gradient Descent
A cuisine’s distribution is not static. Over centuries it drifts, and the drift is not random: cuisines move, slowly, toward regions of flavor space that are better adapted to their constraints. This is an optimization process running on a timescale of generations, with the population of cooks as the search agents and dishes as the candidate solutions — and the natural mathematical description is gradient descent on a fitness landscape.
[Modeling Axiom.] A cuisine evolves by descending (ascending, in fitness terms) a landscape $\Phi_C(\mathbf{f})$ whose value at each point combines several pressures: availability (how cheaply and reliably the local environment supplies the ingredients implied by $\mathbf{f}$), preservation and safety (whether the implied techniques keep food edible in the local climate), nutrition (whether the diet meets caloric and micronutrient needs — Chapter 11), and hedonics (how pleasant the result is, the hedonic gradient of Chapter 3.6). The cuisine’s mean $\boldsymbol{\mu}_C(t)$ moves down the combined gradient $-\nabla \Phi_C$, subject to the high curvature and frozen-accident path-dependence of cultural transmission.
The pressures are legible in the world’s cuisines if you read them as terms in $\Phi_C$. Availability explains why coastal cuisines load on the umami axis through seafood and fish-sauce while inland cuisines reach umami through aged dairy, cured meat, or fermented beans — the same perceptual destination from different local stocks. Preservation explains the global ubiquity of fermentation, salting, smoking, and chile-as-antimicrobial in hot climates, where the safety term in $\Phi_C$ steepens and pushes cuisines toward acidified, salted, and capsaicin-laden regions (the much-cited correlation between climate and spice use; the antimicrobial hypothesis is debated, but the gradient it would imply is exactly the observed drift). Hedonics explains why the descent does not stop at the merely edible: cuisines keep moving toward the pleasant even after safety and nutrition are satisfied.
[Empirical.] Cultural evolution is gradient descent, not gradient computation: it is local, greedy, and path-dependent. Cuisines get trapped in local optima — basins they cannot leave because every small change tastes worse to a palate trained on the current optimum, even though a distant, better region exists. This is why cuisines are simultaneously well-adapted and arbitrary: well-adapted because they have descended a real gradient for centuries, arbitrary because the basin they descended into was set by frozen accidents of geography and history. The metric that defines “small change” is the learned hedonic metric of the population, and that metric is itself shaped by the current distribution — a coupled dynamics (the distribution shapes the metric that shapes the motion of the distribution) that we connect to the learned metrics of Volume 6: Geometric Cognition, where the same circularity governs how expertise reshapes the space it navigates.
9.6 The Convergence Theorem: Why Distant Cuisines Independently Find Acid + Fat + Umami
We arrive at the chapter’s most striking claim, and the one that most needs care. Cuisines that never contacted one another — that descended independent gradients on opposite sides of the planet — nonetheless arrive at strikingly similar high-level flavor principles. Acid to brighten, fat to carry and round, umami to deepen, salt to amplify, a bitter or aromatic edge to provide contrast: this combination, in some local guise, appears in nearly every mature cuisine on Earth. Samin Nosrat’s organizing quartet — salt, fat, acid, heat — is presented in Salt, Fat, Acid, Heat (2017) not as a Western theory but as the shared substrate she found beneath every tradition she cooked in. Why should independent descents converge?
[Conjecture.] (Culinary Convergence.) If many cuisines descend independent gradients on the same fitness landscape $\Phi$ — same human receptor complement (Chapter 1), same broad metabolic needs, same physics of cooking — then their attractors cluster, because they are local optima of one shared function. The convergence is therefore not coincidence and not diffusion; it is the signature of a common landscape sampled by independent searches. The shared peak at “acid + fat + umami + salt + aromatic-contrast” is a global feature of $\Phi$ rooted in invariant human biology, and any sufficiently long independent descent finds it.
The argument has the same shape as evolutionary convergence in biology — the camera eye evolved independently many times because optics is a fixed problem with a small set of good solutions — and as the convergence results that recur across the Geometric Series, where independent optimizers on a shared landscape are predicted to find the same attractors. Here the shared landscape is the human sensory system itself. Acid works everywhere because the sour receptors and their interaction with sweetness and fat are the same in every human mouth. Fat works everywhere because it carries the volatiles (Chapter 6.3) and because the reward system is tuned to its caloric density. Umami works everywhere because the T1R1/T1R3 receptor, and its synergy with the salt that amplifies it (Chapter 1.1) and the nucleotides that potentiate it, is a human universal. The convergence is the biology showing through the culture.
[Empirical.] The claim must be stated carefully to remain honest, and the same large-scale data that motivates it also bounds it. Convergence is real at the level of high-level principles (every cuisine balances brightness, richness, depth, and contrast) while divergence is equally real at the level of implementation and pairing topology — and the Ahn et al. (2011) network is the cleanest evidence of the latter. Western cuisines tend to pair ingredients that share volatile compounds (the food-pairing hypothesis, Chapter 4), clustering near the diagonal of compound-similarity; several East Asian cuisines tend to pair ingredients that do not share compounds, sitting off-diagonal. Both strategies satisfy the same need for a balanced, multi-axis flavor vector; they differ in the geometric route by which they fill the dimensions. Convergence in the destination, divergence in the path — and the geometry predicts both, assigning destination to the shared landscape and route to the local basin.
This is the chapter’s resolution of an old debate. Are cuisines fundamentally the same (universalist) or fundamentally different (relativist)? The geometric answer is that they share a landscape and differ in their basins — the universalism is in $\Phi$, the relativism is in the path of descent. Both camps were describing the same object from opposite ends. The cuisines of the world are independent solutions to one well-posed optimization problem, and their family resemblance is the fingerprint of the human sensory system on every kitchen that ever existed.
From Distributions to Embeddings
We have now treated a cuisine as a probability distribution, found its eigenvectors, located its signature ingredients as extreme points, interpolated between cuisines along geodesics, and explained their convergence as the shadow of a shared landscape. Each of these moves required a flavor vector for every ingredient and dish — and so far we have helped ourselves to those vectors, taking the embedding of Chapter 8 as given. Chapter 10 confronts the embedding itself: how the flavor-compound databases (FlavorDB, FooDB) are turned into vectors, how word2vec-style models learn ingredient embeddings directly from recipe co-occurrence, and how the vector arithmetic that powers analogies — soy sauce is to Japanese as fish sauce is to Thai — is the same arithmetic that powers the eigenvectors and geodesics of this chapter, now learned rather than posited. The eigenvalues of a kitchen were always there in the recipes. The next chapter is about teaching a machine to read them. The geometry was always real; the scalars — the bare ingredient lists, the rulebooks of authenticity — were always insufficient.