Chapter 2: The Receptor Basis
“The sense of smell is the sense of imagination.” — Jean-Jacques Rousseau, Émile, or On Education (1762)
RUNNING EXAMPLE — THE CILANTRO DIVIDE
Two friends sit at the same table in a Oaxaca market stall, sharing one plate of tacos al pastor. The same cilantro is chopped over both. Priya leans in: bright, green, citrus-and-pepper, the herb lifting the pork and pineapple into something whole. Across the table, Daniel takes one bite and sets his fork down. To him the same leaves taste of dish soap — a sharp, waxy, vaguely chemical wrongness that coats everything it touches. He is not being difficult, and he is not imagining it. He has tried, for years, to “learn” cilantro the way one learns olives or coffee. It does not take. The herb has never once tasted of anything but soap.
Two nervous systems, one stimulus, two incompatible reports. The naïve reaction is to declare taste “subjective” and stop — but that explains nothing, because the disagreement is not random. It is structured, heritable, and predictable from a single locus in the genome. Daniel carries a variant of an olfactory receptor, OR6A2, that is unusually well-tuned to the aldehydes that cilantro shares with soap. His receptor array reads the same molecules through a slightly different lens, and the difference propagates all the way up to a confident, stable, lifelong percept.
This chapter is about that lens — the bank of receptors that stands between the chemistry of food and the geometry of flavor. The cilantro divide is not a failure of objectivity. It is two people computing the same projection through two slightly different matrices, and arriving, correctly, at two different points in flavor space.
From Chemistry to Coordinates
Chapter 1 made a promise it did not keep. It asserted that flavor lives in a perceptual space $\mathcal{F}$ of a few tens of dimensions, recovered from receptor responses rather than read off the chemistry table, and asked the reader to take on faith that biology performs the reduction from ten thousand compounds to a navigable handful of axes. This chapter pays the debt: how a receptor turns a molecule into a number, how a few hundred such numbers form a distributed code, and how that code is compressed into the perceptual axes the rest of the book will manipulate.
The central object is the receptor basis — the set of measurements the nervous system actually takes of a chemical stimulus. Everything downstream is a transformation of this basis, and its mathematical character is, at first, almost too convenient: each receptor type behaves to good approximation as a linear functional on the space of stimuli, taking a high-dimensional chemical input and returning a single scalar, its activation. A few hundred such functionals, applied in parallel, produce a vector — the raw flavor vector $\mathbf{f}$ before perception interprets it, and the whole problem of gastronomy, geometrically, is the problem of what it means.
[Modeling Axiom.] We model the chemical stimulus as a vector $\mathbf{c} \in \mathbb{R}^N$, where $N \sim 10^4$ is the number of distinct tastant and volatile species and $c_j$ is the concentration of species $j$ at the receptor surface. The perceptual apparatus does not have access to $\mathbf{c}$. It has access only to the receptor responses $r_i = \rho_i(\mathbf{c})$, where each $\rho_i$ is the response function of receptor type $i$. The claim of this chapter is that the family $\{\rho_i\}$ — a few hundred functions — is the coordinate system of flavor, and that its structure (linear, combinatorial, compressible, and individually variable) explains the phenomena that a chemistry-first account leaves mysterious, including the cilantro divide. This is the same move the sibling Volume 4: Geometric Economics makes when it treats GDP as a contraction of a richer tensor — the observable is a projection of something higher-dimensional, and the discipline begins when you ask what the projection keeps and discards — except that here the projection is performed by a sheet of cells.
2.1 Taste Receptor Cells as Linear Functionals
Begin with the simplest case, the gustatory receptors of the tongue, where the linearity is cleanest. A sweet-sensitive taste receptor cell expresses the T1R2–T1R3 heterodimer, a G-protein-coupled receptor binding sugars, some amino acids, and many synthetic sweeteners. When a ligand binds, it triggers a cascade whose output is, over a working range, a graded depolarization — a number. The cell does not report which sugar bound; it reports how much binding occurred, summed across all that bound.
That summation is the key. Suppose a stimulus contains glucose at concentration $c_{\text{glu}}$ and fructose at $c_{\text{fru}}$, each with an affinity $a_{\text{glu}}, a_{\text{fru}}$ for the sweet receptor. To first order, below saturation, the receptor’s response is
$$ r_{\text{sweet}} = a_{\text{glu}}\, c_{\text{glu}} + a_{\text{fru}}\, c_{\text{fru}} + \cdots = \sum_j a_{\text{sweet},\,j}\, c_j = \langle \mathbf{a}_{\text{sweet}},\, \mathbf{c}\rangle. $$
This is a linear functional: an inner product between a fixed weight vector $\mathbf{a}_{\text{sweet}}$ — the receptor’s affinity profile across all chemical species — and the stimulus vector $\mathbf{c}$. The receptor is a row of a matrix, and to say a food “tastes sweet” is to say its projection onto $\mathbf{a}_{\text{sweet}}$ is large.
[Empirical.] Linearity is an approximation, and a knowable one. Real dose–response curves are sigmoidal (Hill functions): flat at threshold, roughly linear through the middle decade of concentration, saturating at the top. Over the saturating regime the functional is no longer linear, and Chapter 3 shows that the compressive shape of these curves is precisely the source of the logarithmic Weber–Fechner metric. For building the basis, the relevant fact is that within the operating range where most eating happens, each receptor type acts as a fixed linear measurement. The five gustatory qualities of Chapter 1 are five such functionals — five rows — and their affinity vectors $\mathbf{a}_{\text{sweet}}, \mathbf{a}_{\text{salt}}, \dots$ are nearly orthogonal in chemical space, responding to disjoint classes of molecule (sugars versus sodium versus protons versus bitter alkaloids versus glutamate).
The picture to hold is deliberately humble. A receptor is not a detector of a Platonic quality; it is a dot product. Sweetness is not in the sugar; it is the angle between the sugar’s chemical signature and a weight vector that evolution installed in a cell membrane. Change that vector — by mutation, as in §2.5 — and the same molecule yields a different number. The objectivity of flavor and its individual variation are the same fact seen from two sides: an objective computation performed by a subjective matrix.
2.2 Combinatorial Coding in Olfaction
Taste is the easy half precisely because its functionals are nearly orthogonal — five clean, separable measurements. Olfaction, which supplies most of flavor, is built on the opposite principle.
A single aroma molecule does not activate a single olfactory receptor; it activates a pattern across many, each at a graded level set by its affinity. And a single receptor responds to many molecules. Linalool, the floral-citrus volatile of coriander leaf and lavender, lights up a constellation. The map from molecules to receptors is many-to-many and dense.
[Empirical.] This is the combinatorial code, established directly by Malnic, Hirono, Sato, and Buck (1999): each odorant is recognized by a unique combination of receptors, and each receptor participates in recognizing many odorants. Humans express roughly 400 functional olfactory receptor types (Buck and Axel’s gene family, 1991; the human subset refined by later genomic work), and a typical odorant engages on the order of tens of them. The identity of a smell is not which receptor fires but the pattern across the whole array.
The geometric reading is, for a reader of this series, familiar. A distributed code in which each item is a pattern across a shared basis, and each basis element helps represent many items, is exactly a word embedding: the meaning of a word is spread across hundreds of dimensions, each contributing to thousands of words. The olfactory system arrived at the same design a few hundred million years earlier. An odorant is an embedding vector — its coordinates are receptor activations — and that space encodes similarity: smells that activate overlapping receptor patterns smell alike, the cosine between their receptor vectors predicting perceived similarity (a fact we exploit for pairing in Chapter 4).
This explains a number that would otherwise be a puzzle. With only ~400 receptor types, how can the system distinguish the vast catalogue of discriminable smells? The same way 300 embedding dimensions distinguish hundreds of thousands of words: combinatorially. If each receptor merely signalled present/absent, $400$ binary channels would index $2^{400}$ patterns — astronomically more than there are molecules; graded activation enlarges the capacity further. The combinatorial code is not a limitation worked around but a representation of enormous capacity built from a modest basis — the reason the chemical explosion of Chapter 1 can be read by a finite nose.
2.3 The Binding Affinity Matrix
We now have two halves — gustatory functionals and the olfactory combinatorial code — and a single object unifies them. Stack every receptor’s affinity profile as a row, one row per receptor type and one column per chemical species, and the result is the binding affinity matrix $A$:
$$ A \in \mathbb{R}^{M \times N}, \qquad A_{ij} = \text{affinity of receptor } i \text{ for molecule } j, $$
where $M \approx 400 + 5 + \dots \sim 400$ receptor channels and $N \sim 10^4$ chemical species. The entire transduction step — a dish’s chemistry converted into the raw neural representation everything downstream operates on — is one matrix–vector product:
$$ \mathbf{r} = A\,\mathbf{c}, \qquad \mathbf{r} \in \mathbb{R}^{M},\ \mathbf{c} \in \mathbb{R}^{N}. $$
[Modeling Axiom.] We model peripheral chemosensory transduction as the bilinear map $(\mathbf{a}, \mathbf{c}) \mapsto \langle \mathbf{a}, \mathbf{c}\rangle$ realized by the rows of $A$, valid within the sub-saturating regime of each receptor. $A$ is the receptor basis in the literal sense: its rows are the directions in chemical space along which the organism can measure, and any component of $\mathbf{c}$ lying in the null space of $A$ — any chemical variation orthogonal to every receptor’s affinity profile — is, by definition, odorless and tasteless. The kernel of $A$ is every way food can differ chemically without differing in flavor at all.
Two structural facts about $A$ organize the rest of the book. First, $A$ is wide and rank-deficient: $M \ll N$, so the map from chemistry to receptor response is massively non-injective. Distinct stimuli that $A$ maps to the same $\mathbf{r}$ are metamers — perceptually identical despite chemical difference, the flavor analogue of the color metamers that let three phosphors reproduce any hue. Metamerism is what makes artificial flavoring possible (Chapter 13): a flavorist need not reproduce a strawberry’s full chemistry, only a stimulus landing on the same row-space image. Second, $A$ is sparse and structured, not random: its columns cluster, because the molecules of food are products of a limited biochemical repertoire (the recurring fatty-acid, amino-acid, and terpene chemistries of Chapter 1), and that clustering makes the next section’s compression possible.
The bilinear form also tells us exactly where linearity fails, and the failures are gastronomically the interesting part. When two molecules compete for one receptor’s binding site, one suppresses the other. When a tastant modifies the receptor’s conformation — as the miracle-fruit protein miraculin does, switching the sweet receptor to acid activation — the matrix entry itself changes with context. These off-diagonal and state-dependent terms, accumulated upward, become the curvature of the flavor metric in Chapter 3. The clean map $\mathbf{r} = A\mathbf{c}$ is the tangent-plane approximation; the cooking happens in the corrections.
2.4 Perceptual Compression: The Flavor PCA
We have arrived at the chapter’s quantitative spine — a chain of three dimensionalities, each smaller than the last:
$$ \underbrace{N \sim 10^{4}}_{\text{volatile + tastant species}} \ \xrightarrow{\ A\ }\ \underbrace{M \sim 400}_{\text{receptor channels}} \ \xrightarrow{\ \text{neural reduction}\ }\ \underbrace{D \sim 10\text{–}30}_{\text{perceptual dimensions}}. $$
The first arrow is the binding affinity matrix of §2.3 — a 25-fold compression at the receptor surface. The second, performed in the brain between the olfactory bulb and the orbitofrontal cortex, is a further 10-to-40-fold compression down to the modest number of axes a trained taster can discriminate. This is the reduction Chapter 1 named without explaining, and the right name for it is principal-component analysis on the receptor responses — a flavor PCA.
[Empirical.] The empirical warrant is now substantial. When experimenters collect human ratings of many odorants across many descriptors and reduce the resulting matrix, a small number of components capture most of the variance, the first and most robust axis being pleasantness (Khan et al., 2007), followed by a handful more (intensity, floral/fruity versus burnt/pungent). Independent attempts to estimate the intrinsic dimensionality of olfactory perceptual space converge on roughly ten meaningful dimensions, far below 400. The receptor space is high-dimensional; the occupied part of it — the perceptual manifold of Chapter 1 — is low-dimensional, because receptor responses to real foods are heavily correlated and PCA discards the correlated bulk.
Why must this be so? Because $A$ is structured, not random. The molecules of food co-occur in structured ways — esters travel with esters, pyrazines with pyrazines — so their receptor signatures are correlated, and the covariance $\Sigma_r = \operatorname{Cov}(\mathbf{r})$ over natural foods has a sharply decaying eigenvalue spectrum. The top $D \approx 10$–$30$ eigenvectors of $\Sigma_r$ are the flavor eigenvectors, the perceptual coordinates; everything below them is noise the system is right to discard.
This is the precise sense in which the flavor PCA mirrors modern representation learning. A Matryoshka embedding — the nested-truncation scheme that makes text embeddings adjustable in width — works because the most important information is packed into the leading coordinates, so truncating degrades gracefully rather than catastrophically. The same holds for flavor: a wine described in three dimensions (a novice’s “red, dry, fruity”) is a truncation of the same vector $\mathbf{f}$ an expert reads in twenty, lossy but not random, keeping the high-variance axes. Expertise, in this view, is the recovery of trailing principal components — the sommelier attends to directions the novice’s reduction discards. We make this the formal definition of a “trained palate” in Chapter 12, connecting it to the dimensionality-reduction pipeline of Volume 6: Geometric Cognition.
THE CILANTRO DIVIDE, REVISITED — A ROTATED ROW
We can now locate the disagreement precisely. Priya and Daniel receive the same chemical vector $\mathbf{c}$ — the same dish, the same chopped coriander, the same aldehydes (notably the unsaturated decadienals and decenals that coriander shares with the volatiles of soap). The percept differs because their affinity matrices differ in one row. Daniel’s variant of OR6A2 has an affinity profile $\mathbf{a}^{(D)}_{\text{OR6A2}}$ that weights those soap-shared aldehydes more heavily than Priya’s $\mathbf{a}^{(P)}_{\text{OR6A2}}$. The single altered row tilts his receptor vector $\mathbf{r}^{(D)} = A^{(D)}\mathbf{c}$ toward the region of olfactory space already occupied, in his lifelong experience, by detergents and degreasers.
Crucially, the rest of the pipeline is working correctly. Daniel’s flavor PCA, his hedonic axis, his memory associations — all faithfully report that the dish has landed near “soap,” because on his basis it has. The divide is not in the brain’s interpretation; it is one perturbed measurement at the very first matrix, propagated upward by an otherwise identical computation. Two correct readings of one stimulus through two matrices that differ by a single row.
2.5 Individual Variation as Rotations of Flavor Space
The cilantro example is one instance of a general fact: the receptor basis is not a species constant. It varies from person to person, because the genes encoding receptors are polymorphic, and each common polymorphism is a small re-parameterization of flavor space.
The textbook case is TAS2R38, a bitter receptor for which two common haplotypes — PAV (taster) and AVI (non-taster) — differ at three amino-acid positions. PAV carriers perceive the synthetic compounds PTC and PROP, and the glucosinolates of cruciferous vegetables (the bitterness of raw broccoli, kale, Brussels sprouts), as intensely bitter; AVI homozygotes perceive them as nearly tasteless. The polymorphism predicts vegetable preference, coffee and alcohol consumption, and the compensatory use of salt and fat in cooking. The affinity profile $\mathbf{a}_{\text{TAS2R38}}$ for these ligands is, quite literally, edited by the genotype.
[Modeling Axiom.] We model heritable receptor variation as a linear transformation of the receptor basis. If $A$ is the reference affinity matrix, an individual’s basis is $A' = R\,A$ for some near-identity transformation $R$, and their receptor vector for any stimulus is $\mathbf{r}' = R\,\mathbf{r}$. For a single-receptor polymorphism such as TAS2R38 or OR6A2, $R$ is close to the identity — it rescales or reorients one coordinate while leaving the others fixed — and its action on flavor space is, to good approximation, a rotation in the plane spanned by the affected axis and its perceptual neighbors. Personal flavor spaces are not different spaces; they are the same space viewed in rotated frames.
This framing earns its keep by being the right kind of small. Two people do not inhabit incommensurable flavor worlds — that would predict chaos, not the structured, heritable, replicable disagreements we observe. They inhabit one shared space seen through frames that differ by modest rotations. Most of the basis agrees (we concur overwhelmingly about sweetness, saltiness, the smell of vanilla), so cuisines and shared meals are possible at all. But along the few axes where polymorphism concentrates — bitter (TAS2R38 and the broader TAS2R family), the aldehyde axis OR6A2 touches, the androstenone receptor OR7D4 that renders pork either savory or urinous by genotype (Keller et al., 2007) — individual frames diverge, and the divergence is exactly a rotation mapping one person’s “herb” onto another’s “soap.”
A consequence worth stating plainly: there is no privileged, “objective” flavor of cilantro, yet the matter is not arbitrary. There is a reference basis (the species-typical $A$) and a well-defined group of small transformations $\{R\}$ the population samples. The “subjectivity” of flavor is not the absence of structure; it is a group action on a shared manifold — the same gauge-theoretic move the series makes elsewhere, where relativism resolves into a precise symmetry group acting on an objective space. We return to it when personalized flavor design (Chapter 13) asks us to invert an individual’s rotation to season a dish for a specific palate.
2.6 Supertasters and Non-Tasters: Effective Dimensionality
Polymorphisms that rotate the basis are not the only form of individual variation. A second form changes not the orientation of the receptor frame but its resolution — the number of axes along which a person can resolve meaningful differences. This is the axis along which supertasters, medium tasters, and non-tasters differ.
[Empirical.] Bartoshuk and colleagues established that humans vary several-fold in the density of fungiform papillae — the mushroom-shaped structures on the tongue that house taste buds. Supertasters (disproportionately PAV/PAV at TAS2R38, and disproportionately women) carry dense papillae and report bitterness, sweetness, oral burn, and fat as more intense; non-tasters carry sparse papillae and a flatter, lower-contrast taste world. Because the trigeminal and gustatory systems are anatomically intertwined at the papilla, the heightened response spills across modalities, so supertasters also feel chili heat and astringency more sharply.
The geometric reading is not that supertasters inhabit a higher-dimensional space than non-tasters — the receptor types are the same. It is that they make fuller use of the available dimensions. Return to the flavor PCA of §2.4: the perceptual dimensions a person can exploit are the receptor-response directions whose signal rises above their own noise floor — the eigenvalues of their personal $\Sigma_r$ that exceed threshold. A supertaster, with greater receptor density and gain, has more eigenvalues clearing the bar: more effective dimensions. A non-taster’s bitter axis, low in gain, may sink below threshold and collapse, leaving a flavor space of lower effective rank in that region.
[Modeling Axiom.] We define a person’s effective flavor dimensionality as the count of suprathreshold eigenvalues of the individual covariance $\Sigma_r$ — the perceptual axes whose discriminable variance exceeds their perceptual noise. Supertasting raises gain, lifting marginal eigenvalues above threshold and raising effective dimensionality; specific anosmias and non-taster haplotypes lower particular eigenvalues below threshold, collapsing dimensions. Two palates can share the same receptor basis yet differ in effective rank — and a dish optimized for one rank can be illegible or overwhelming at another.
This reframes a familiar kitchen tension as a statement about rank. The chef who loves bitter amari, raw kale, and aggressive single-origin coffee may be a non-taster for whom these are pleasantly mild — operating in a lower-rank space where extra intensity is needed to register at all. The supertaster guest at the same table, resolving every bitter and burn at high gain, finds the food punishing. Neither is wrong; they evaluate the dish at different effective dimensionalities. Good cooking for a crowd is, in part, the problem of finding a point that reads as balanced across a distribution of effective ranks — which we pose formally as robustness of the seasoning fixed point in Chapter 6.
What the Basis Buys Us
This chapter has built the coordinate system the rest of the volume assumes. A receptor is a linear functional, a dot product between an affinity profile and a stimulus. The full bank of receptors is the binding affinity matrix $A$, a wide, structured, rank-deficient map compressing ten thousand chemical species into a few hundred channels; its null space is the odorless, its row space the perceptible, its metamers the foundation of artificial flavor. A second compression, the flavor PCA performed in cortex, reduces those channels to the ten-to-thirty perceptual axes of $\mathcal{F}$ in which $\mathbf{f}$ finally lives. And the apparatus is individually parameterized: polymorphisms rotate the basis (the cilantro divide), while receptor density and gain set the effective rank (supertasters and non-tasters), so flavor is at once objective and personal.
Two threads now run forward. The first is the metric. We have located flavor space and labeled its axes but not said how far apart two flavors are — and the sigmoidal dose–response curves and competitive binding of §2.1 and §2.3 already warn the answer will be neither linear nor flat. Chapter 3 equips $\mathcal{F}$ with a Riemannian metric: logarithmic in intensity (Weber–Fechner), curved by cross-modal interaction, locally variable, so the distance between two dishes becomes a precise and tasteable quantity. The second thread is pairing: once flavors are vectors with a notion of angle, which ingredients belong together becomes a question about cosine similarity in receptor space — the food-pairing hypothesis of Chapter 4, at bottom a claim about the geometry of the very embedding this chapter has constructed.
The combinatorial wall of Chapter 1 has now been dismantled twice: once in principle, by the promise of a low-dimensional space, and once in mechanism, by the receptors that build it. What remains is to measure the distances.