Chapter 5: The Pairing Manifold
“There is no love sincerer than the love of food.” — George Bernard Shaw, Man and Superman (1903)
RUNNING EXAMPLE — THE RECOMMENDER’S MISFIRE
The app is called Concord, and on paper its logic is impeccable. Feed it an ingredient and it returns three pairings, ranked by the number of volatile aroma compounds shared with the input — the food-pairing hypothesis, productized. A user types “strawberry.” Concord scans its database, finds that strawberry and fresh basil share a long list of green, fruity esters, and confidently returns: strawberry and white chocolate (a known hit), strawberry and basil (a known hit), and — its top novel suggestion, flagged with a little sparkle icon — strawberry and fresh tomato. Chemically, the match is excellent. Strawberry and tomato share a dense cluster of the same volatiles; on the cosine-similarity score of Chapter 4, they are near-twins.
A tester makes the dish. It is wrong. Not disgusting — wrong in a more unsettling way. The tomato is sweet enough to read as fruit, the strawberry acidic and seedy enough to read as vegetable, and the brain, primed by every previous strawberry it has eaten, keeps reaching for a category and missing. The result is a near-fruit that is also a near-vegetable, familiar in every note and incoherent as a whole. The tester writes a one-word review: “uncanny.”
The same week, a rival app suggests blue cheese and honey — two ingredients that share almost no volatiles, that the food-pairing hypothesis would never connect — and the pairing lands. It is strange, and it works. Concord’s engineers are left with a problem that cosine similarity cannot touch: why does a high-similarity pairing fall into a valley of wrongness, while a low-similarity pairing climbs to a peak of delight? The answer is that “good pairing” does not live on the similarity axis at all. It lives on a manifold — and this chapter is a map of that manifold.
The Subspace of the Pleasant
Chapter 4 gave us the operations of pairing: a combination is, to first approximation, a sum of flavor vectors, $\mathbf{f}_{\text{pair}} = \mathbf{f}_1 + \mathbf{f}_2$, and similarity is the cosine between the vectors being summed. But it left a question hanging. Knowing how two flavors combine tells us where the combination lands in flavor space $\mathcal{F}$; it does not tell us whether we will like where it lands. The food-pairing hypothesis predicts that shared-compound pairings are common in Western cuisine, and they are — but commonness is not goodness, and Chapter 4’s own data showed East Asian cuisines pairing against the similarity grain. Similarity is a description of structure, not a verdict of value.
The verdict of value is a separate object with its own shape. Of all the points reachable in $\mathcal{F}$ by combining real ingredients, only a thin region is experienced as pleasant; most of $\mathcal{F}$ is empty, incoherent, or actively unpleasant, and the dishes humans actually enjoy cluster on a low-dimensional sheet curved through the ambient space. We call this sheet the hedonic manifold, $\mathcal{H} \subset \mathcal{F}$, and the central claim of this chapter is that it is a real, learnable, low-dimensional geometric object — and that the great puzzles of pairing (the uncanny valley, the triumph of the unexpected, the cultural variation of taste) are all statements about its shape.
[Modeling Axiom.] There exists a hedonic function $H : \mathcal{F} \to \mathbb{R}$ assigning to each flavor point a scalar pleasantness, and the hedonic manifold $\mathcal{H}$ is its high-value region — the set $\{\mathbf{f} : H(\mathbf{f}) > \theta\}$ for some threshold $\theta$, or more usefully its ridge, the locus where $H$ is locally maximal across the directions transverse to it. Pairing is the problem of choosing $\mathbf{f}_1, \mathbf{f}_2$ so that their sum lands on $\mathcal{H}$.
Note what this reframing does to Concord’s failure. Concord optimized similarity, a property of the summands; but pleasantness is a property of the sum’s location relative to $\mathcal{H}$. Strawberry-plus-tomato has high summand similarity yet sums into the valley of §5.3; blue-cheese-plus-honey has near-zero summand similarity yet sums squarely onto a peak of $\mathcal{H}$. The two metrics measure different things, and conflating them is the original sin of naïve computational pairing. The rest of this chapter develops $\mathcal{H}$ rigorously enough that the difference becomes calculable.
5.1 The Hedonic Manifold
Begin with the bare existence claim. If pleasant combinations filled $\mathcal{F}$ uniformly, cooking would again be the combinatorial nightmare of Chapter 1: every new pairing a coin-flip, no transfer of skill, no traditions converging. The opposite is observed. Across the roughly $10^{11}$ three-ingredient combinations available to a kitchen, the proportion any panel rates as pleasant is minute, and — crucially — the pleasant ones are clustered, not scattered. Knowing that A+B works and B+C works lets you predict A+C far better than chance, and predictability of this kind is the signature of low intrinsic dimension: the pleasant points lie near a sheet whose local tangent you can estimate from neighbors.
[Empirical.] Sensory science has mapped this sheet for a century under the name hedonic response, and the robust findings are geometric. First, hedonic response to a single dimension is rarely monotonic but typically an inverted-U: pleasantness rises with concentration to an optimum and falls thereafter (the classic sugar-in-water and salt-in-soup curves of Pangborn and later workers). An inverted-U in each dimension means $\mathcal{H}$ is bounded — there is such a thing as too sweet, too salty, too intense, in every direction at once. Second, the optima interact: preferred salt depends on sugar, preferred acid on fat, so the body is curved, not an axis-aligned box. Third, it is thin in the dimensions of intensity and balance and broad in the dimensions of character — you can roam widely across cuisines while staying pleasant, but cannot stray far toward “unbalanced” before falling off.
Putting these together: $\mathcal{H}$ is a bounded, curved, low-dimensional region of $\mathcal{F}$ whose boundary is where the inverted-U curves turn down. This already explains why balance (Chapter 6) is the master virtue of cooking — balance is the operation of returning a dish to the interior of $\mathcal{H}$ after some axis has overshot its optimum — and why expertise feels like a sense of direction rather than a list of rules. The expert cook has an internal model of $\mathcal{H}$’s local shape and can feel which way its surface tilts. We met this in Chapter 1 as the line cook’s geodesic; we can now name its destination — a path along the hedonic manifold toward its local ridge.
One subtlety must be flagged immediately, because the rest of the chapter turns on it. $\mathcal{H}$ is not a fixed feature of the world like a coastline. It is, in part, learned — it differs across people and cultures, and it deforms with experience. Whether its deepest structure is biological or cultural is the question of §5.2, and the answer is “both, in separable layers.”
5.2 Training-Set Bias: Biological or Cultural?
The machine-learning reader will already have heard, in the phrase “learned manifold,” an alarm bell labeled training-set bias — and it is exactly the right alarm. If $\mathcal{H}$ is estimated from the dishes a person has eaten, its shape encodes the distribution of that person’s culinary exposure, and a manifold fit to one training set will mispredict on data from another. The question of whether the hedonic manifold is biological or cultural is, geometrically, the question of which parts of $\mathcal{H}$ are innate priors and which are fit to the training set.
[Empirical.] The innate layer is real and small. Newborns, with no culinary training set whatsoever, show a fixed sign on a few axes: positive hedonic response to sweet and to umami-like (glutamate) stimuli, negative response to bitter and to high sour, and an unlearned appetite for salt that switches on around four months (Steiner’s facial-expression studies; Beauchamp and Mennella’s developmental work). These are evolutionarily legible — sweet and umami flag calories and protein, bitter and sour flag toxins and spoilage — and they fix the gross orientation of $\mathcal{H}$ in the taste subspace: roughly “more sweet/savory good, more bitter/sour bad,” before any learning.
[Empirical.] Almost everything else is fit to the training set. Coffee, chili heat, blue cheese, natto, durian, bitter greens, intense fermentation — every one of these starts on the unpleasant side of the innate prior and is moved to the pleasant side by exposure (Rozin’s work on the acquisition of chili-pepper liking is the canonical case: humans are the only species that learns to love a substance whose entire chemical purpose is to punish mammalian tongues). Mere repeated exposure shifts hedonic response (the mere-exposure effect, Zajonc), and culture supplies the exposure. The consequence is a manifold whose innate skeleton is shared across all humans but whose surface is sculpted by cuisine.
This resolves Chapter 4’s Western/Eastern asymmetry at the level of value rather than structure. The two traditions did not discover different chemistry; they fit $\mathcal{H}$ to different training distributions and so curved its learned surface in different directions — one toward shared-compound (similarity) pairings, the other toward complementary ones. Neither is right; both are well-fit to their own data. It also explains Concord’s deeper error: its database was overwhelmingly Western, so the manifold it implicitly learned was oriented toward shared-compound pairings — precisely the orientation that, pushed too far, walks into the uncanny valley of §5.3.
[Conjecture.] The biological layer is low-dimensional and rigid; the cultural layer is higher-dimensional and plastic. If so, the principal components of pairing preference (§5.5) should split into a small number of culture-invariant axes (carrying the innate sweet/bitter/intensity structure) and a larger number of culture-specific axes (carrying learned character preferences). This is testable by cross-cultural factor analysis and, to our knowledge, not yet decisively done.
5.3 The Uncanny Valley of Flavor
We can now name Concord’s misfire precisely. The visual uncanny valley, identified by Masahiro Mori (1970) for robots and animated faces, is a dip in the affective response curve as a function of human-likeness: a clearly artificial face is fine, a perfectly human face is fine, but a face that is almost human — human enough to invoke the category, wrong enough to violate it — drops into a trough of revulsion. The curve is non-monotonic in similarity-to-the-familiar, with a valley just short of the familiar.
[Conjecture.] There is a flavor analogue, and it has the same geometric cause. Let $\mathbf{f}^\star$ be a familiar, strongly-categorized flavor point — strawberry, say, a location the eater has visited thousands of times and bound tightly to a label and an expectation. Define hedonic response as a function of distance $d(\mathbf{f}, \mathbf{f}^\star)$ from that anchor. Far away, the eater forms no expectation of strawberry and judges the flavor on its own terms: fine. At $\mathbf{f}^\star$: the familiar pleasure. But in the near shell — close enough that the strawberry category fires and sets up a prediction, far enough that the prediction is violated — hedonic response dips into a valley. The flavor is a near-strawberry that fails the strawberry test, and the prediction error reads as wrongness.
Strawberry-and-tomato lands exactly there. The high summand similarity Concord prized is precisely what places the sum in the near shell of the strawberry anchor: tomato is similar enough to drag the combination toward the strawberry category without arriving there. The brain commits to “strawberry,” receives seedy, off-sweet, faintly savory evidence, and logs a prediction error. [Empirical.] This is the flavor face of the broader principle Prescott (2015), Spence (2015), and others document under expectation and sensory incongruity: a stimulus that violates a strongly held flavor expectation is rated worse than the same stimulus encountered with no expectation at all (the strawberry yogurt dyed the wrong color, the white wine dyed red of Chapter 12). Expectation is a prior centered on $\mathbf{f}^\star$; the valley is where likelihood and prior collide.
The geometry yields an immediate design rule, and it is the inverse of Concord’s. To avoid the valley, a novel pairing should sum either to a point comfortably inside a familiar category (so the prediction is confirmed) or comfortably outside all strong anchors (so no prediction is invoked). The dangerous zone is the near shell — the moat of partial familiarity around every well-known flavor. Cosine similarity, by rewarding partial overlap, aims a recommender straight into the moat.
THE RECOMMENDER’S MISFIRE, REVISITED — A VALLEY, NOT A SCORE
Picture Concord’s recommendations plotted not as a ranked list but as points on the hedonic manifold $\mathcal{H}$, with the familiar “strawberry” anchor $\mathbf{f}^\star_{\text{straw}}$ marked. White chocolate sums to a point well outside any anchor — new territory, judged on its merits, pleasant: on $\mathcal{H}$. Basil likewise lands in open country. But tomato sums into the near shell around $\mathbf{f}^\star_{\text{straw}}$ — into the moat — exactly where $H$ dips below the surrounding surface. Concord, ranking by summand similarity, had no coordinate for “distance from a strong anchor,” so it could not see the moat at all; the very feature that scored tomato highest (shared volatiles with strawberry) is the feature that dropped it into the valley. The fix is not a better similarity score but a second axis — distance to the nearest strong category — and a penalty on the near shell of each anchor. With that axis added, the sparkle icon moves off tomato and onto something safely far: rhubarb, black pepper, or the blue cheese of §5.4.
5.4 Novelty as Distance from the Training Manifold
If the near shell is a trap, the far country is the opportunity. The pairings we call inspired — chocolate and chili, blue cheese and honey, strawberry and black pepper, white chocolate and caviar (Blumenthal’s much-cited dish, built explicitly on shared amines) — are not near-misses of the familiar. They are deliberate excursions to parts of $\mathcal{H}$ that the eater’s training set never covered, and that is the source of their power.
[Modeling Axiom.] Define the novelty of a pairing as the distance from its sum to the eater’s estimated training manifold $\hat{\mathcal{H}}_{\text{train}}$ — the region densely populated by dishes that eater has experienced. Novelty is high when the sum lands far from prior experience. The hedonic payoff of novelty is then a product of two factors: the point must be novel (far from $\hat{\mathcal{H}}_{\text{train}}$) and yet on the true manifold $\mathcal{H}$ (genuinely pleasant). High novelty off $\mathcal{H}$ is mere weirdness; high novelty on $\mathcal{H}$ is delight.
This factorization is the whole secret of the surprising pairing, and it explains why blue cheese and honey succeed where strawberry and tomato fail despite sharing no volatiles. The two are complementary in the orthogonality sense of Chapter 4 — blue cheese is pungent, salty, fatty, microbially funky; honey is sweet, floral, viscous — so their sum fills several dimensions of $\mathcal{F}$ at once rather than piling redundant weight on a few. The sum is far from any common training point (most eaters have no “blue-cheese-and-honey” category, hence high novelty) yet lands on a genuine ridge of $\mathcal{H}$: the salt-sweet and pungent-floral contrasts are exactly the high-complementarity balance the manifold rewards. And because no strong anchor sits nearby, there is no prediction to violate — open country, not a moat. Novel, on-manifold, anchor-free: all three conditions met.
[Empirical.] Chocolate and chili (a pairing with deep Mesoamerican roots) works the same way and adds the trigeminal dimension of Chapter 1: chili contributes capsaicin heat, an axis orthogonal to everything chocolate occupies, so the sum is genuinely novel along $g_{\text{heat}}$ while remaining on-manifold in taste and aroma. The general pattern, observed repeatedly in the inventive-pairing literature, is that the celebrated surprises are complementary (orthogonal, filling new dimensions) rather than similar (redundant) — the opposite of the food-pairing hypothesis’s headline claim, and consistent with Chapter 4’s finding that complementarity, not similarity, is the richer principle.
[Conjecture.] There is an optimal novelty distance — a culinary zone of proximal development. Too little novelty is boring (the dish adds nothing to the training set); too much lands off-manifold or in incoherence; the sweet spot is the largest step that still lands on $\mathcal{H}$. This predicts an inverted-U of appreciation versus novelty, with the peak’s location shifting outward with expertise: experienced eaters, whose $\hat{\mathcal{H}}_{\text{train}}$ covers more ground, need larger steps to feel the same novelty and tolerate excursions that overwhelm novices. The familiar arc by which a palate “matures” is, on this account, the outward migration of the novelty optimum as the training manifold grows.
5.5 Principal Components of Pairing
How many dimensions does $\mathcal{H}$ actually have? The flavor space $\mathcal{F}$ of Chapter 1 ran to a few tens of dimensions, but the manifold of good pairings curved within it should be lower still — and the practical question is how few axes carry most of the variance in human pairing preference. If three to five suffice, the entire apparatus of the chapter reduces to a problem a person, or an app, can actually reason about.
[Empirical.] The evidence says three to five, and converges from several directions. Sensory hedonic studies, factor-analyzed, repeatedly recover a small number of dominant preference axes; the recurring candidates are:
- Overall intensity (how strong vs. mild) — the first principal component of almost every hedonic dataset, the “volume knob” of a dish, with its inverted-U optimum.
- Balance (how evenly weight is distributed across active taste/aroma axes vs. dominated by one) — closely tied to the seasoning fixed point of Chapter 6.
- Complexity (how many distinguishable components are present) — the axis along which novelty (§5.4) buys appreciation, with its own inverted-U.
- Contrast / complementarity (orthogonal fill, the salt-sweet, fat-acid, hot-cool tensions) — the axis on which blue-cheese-and-honey scores high.
- A familiarity / comfort axis (proximity to the training manifold), which trades against complexity and carries most of the cultural variation of §5.2.
[Conjecture.] These five — intensity, balance, complexity, contrast, familiarity — are the approximate principal components of the hedonic manifold, in the sense that a $H$-model built on them recovers most of the explainable variance in pairing-preference data, and that adding a sixth axis yields diminishing returns. This is the pairing analogue of the “flavor PCA” of Chapter 2 and the cuisine eigenvectors of Chapter 9, and it is, like them, a falsifiable claim about effective dimension rather than a metaphysical one. We state it as a conjecture because the decisive cross-cultural factor analysis (which would also test §5.2’s split between rigid biological and plastic cultural axes) remains to be run on a large enough panel.
Two of these axes are recognizably the innate layer (intensity and balance track the inverted-U optima fixed in §5.2), and the rest are largely learned (complexity, contrast, and familiarity are sculpted by exposure) — so the five-component picture lines up with the biological/cultural decomposition rather than cutting across it, a small piece of evidence that the decomposition is real and not an artifact of how we drew the axes. It also tells the app builder where to spend effort: intensity and balance can be modeled once for everyone, but contrast and familiarity must be fit per-culture, or better per-user — which is exactly where Concord, fitting a single Western-skewed manifold, went wrong.
5.6 A Kernel for Pleasantness
We close by assembling the pieces into a formal model of the hedonic function $H$, the object the running example needed and Concord lacked. We want $H(\mathbf{f}_1, \mathbf{f}_2)$ — the pleasantness of pairing two flavors — to be computable from the geometry, to reproduce the inverted-U, the uncanny valley, and the novelty-on-manifold reward, and to reduce to nothing more exotic than inner products and distances. The natural home for such an object is a kernel: a similarity-like function on flavor space, the device that powers kernel methods in machine learning.
[Modeling Axiom.] Write the paired point as $\mathbf{f} = \mathbf{f}_1 + \mathbf{f}_2$ (Chapter 4’s additive model, with the curvature corrections of §4.6 folded into $\mathbf{f}$). Model the hedonic function as a sum of three geometric terms:
$$ H(\mathbf{f}_1,\mathbf{f}_2) \;=\; \underbrace{-\,(\mathbf{f}-\boldsymbol{\mu})^{\!\top} A\,(\mathbf{f}-\boldsymbol{\mu})}_{\text{(i) on-manifold / inverted-U}} \;+\; \underbrace{\gamma\,\big[\,1 - \cos(\mathbf{f}_1,\mathbf{f}_2)\,\big]}_{\text{(ii) complementarity reward}} \;-\; \underbrace{\sum_{k} \beta_k\,K_{\text{shell}}\!\big(d(\mathbf{f},\mathbf{a}_k)\big)}_{\text{(iii) uncanny-valley penalty}}. $$
Read the three terms against the chapter. Term (i) places a broad inverted-paraboloid (more honestly, a kernel ridge) centered on the manifold’s interior $\boldsymbol{\mu}$, with the positive-definite matrix $A$ encoding the per-axis optima and interactions of §5.1 — the inverted-U in every direction at once, its principal axes the components of §5.5. Term (ii) rewards complementarity: large when $\mathbf{f}_1$ and $\mathbf{f}_2$ are orthogonal (cosine near 0), vanishing when parallel, capturing §5.4’s finding that orthogonal fill, not redundant overlap, earns the inspired-pairing bonus. Term (iii) is the uncanny-valley moat: a sum of localized penalty kernels $K_{\text{shell}}$, one per strong anchor $\mathbf{a}_k$, each shaped to bite hardest in the near shell — not at the anchor (where the prediction is confirmed), not far away (where no prediction fires), but in the ring just short of it.
A convenient choice for the shell penalty makes the valley explicit:
$$ K_{\text{shell}}(d) \;=\; \frac{d^{2}}{\sigma^{2}}\,\exp\!\Big(\!-\frac{d^{2}}{\sigma^{2}}\Big), $$
which is zero at the anchor ($d=0$), rises to a maximum penalty at $d=\sigma$ (the moat radius), and decays back to zero for $d \gg \sigma$ (open country). This single factor reproduces Mori’s non-monotone curve in §5.3: confirmation at the center, revulsion in the ring, indifference far out.
The model now explains the running example as arithmetic rather than a list of scores. Strawberry + tomato: term (ii) is small (cosine near 1, little complementarity reward — exactly the redundancy Concord mistook for a virtue), and term (iii) is large and negative because the sum sits in the moat of the strawberry anchor; the valley penalty dominates and $H$ goes negative. Blue cheese + honey: term (ii) is large (near-orthogonal, high complementarity), term (i) is satisfied (the sum lands in the manifold interior, salt-sweet balance respected), and term (iii) is negligible (no anchor nearby); all three terms cooperate and $H$ is high. The two pairings cosine similarity could not tell apart are now separated by the sign of $H$.
[Conjecture.] $H$ as written is personal and cultural: $\boldsymbol{\mu}$, $A$, the anchor set $\{\mathbf{a}_k\}$, and the moat radius $\sigma$ are all fit to an individual’s training manifold (§5.2), while a rigid biological core (the signs on the sweet, bitter, and intensity axes of $A$) is shared across all eaters. A recommender that fits this kernel per-user — rather than scoring shared compounds globally, as Concord did — should avoid the uncanny valley and find on-manifold novelty, and the gap between the two approaches is a clean, falsifiable prediction this chapter hands to Part IV.
Where the Manifold Leads
We have turned “good pairing” from a verdict into a geometry. The pleasant combinations of food occupy a bounded, curved, low-dimensional hedonic manifold $\mathcal{H} \subset \mathcal{F}$; its gross orientation is fixed by a small biological prior and its surface sculpted by a larger cultural training set; the moat of partial familiarity around each strong anchor is the uncanny valley; novelty pays when it steps off the training manifold but stays on the true one; three to five principal components carry most of the preference variance; and the hedonic function $H$ that ties it together is a kernel built from an inverted-U, a complementarity reward, and a valley penalty. The recommender’s misfire was, in the end, a missing coordinate — distance from the anchor — and the missing coordinate was a feature of the manifold’s shape.
This sets up Chapter 6 directly. If $\mathcal{H}$ has an interior and a boundary, then balance is the operation of moving a dish toward the interior, and seasoning — salt, acid, fat, heat — is the set of operators that perform that move. The seasoning fixed point of §6.6 is precisely a point where small perturbations leave you on the manifold: a local maximum of the $H$ we have just built. Further out, Chapter 9 will show that a cuisine is a probability distribution concentrated on its own region of $\mathcal{H}$, and that this chapter’s principal components are the cuisine eigenvectors of that one; and Volume 6: Geometric Cognition will recognize the anchor-and-moat structure of §5.3 as the same predictive-coding geometry that governs perception in every modality, flavor merely being the most delicious instance. The manifold was always there, holding the difference between the pairing that delights and the one that unsettles. Cosine similarity could not see it. The geometry can. The scalars were always insufficient.