Appendix C: Experimental Methods

The geometric theory of flavor developed in this book rests on a chain of empirical claims: that foods map to chemical vectors, that those vectors project through receptors to a perceptual space of modest dimension, and that the resulting locations and distances are stable enough to be measured. A theory of this kind is only as good as the instruments that ground it. This appendix surveys the three measurement systems that supply the data behind the chapters above — gas chromatography–mass spectrometry, electronic tongue and nose arrays, and human sensory panels — and treats each one explicitly as a map in the geometric sense: a function from one space to another, with a domain, a codomain, and a characteristic pattern of what it preserves and what it destroys.

The organizing observation is that none of these three instruments measures the flavor vector $\mathbf{f} \in \mathcal{F}$ directly. GC-MS measures the chemical input that lives upstream of perception. Electronic sensors measure a coarse, non-human projection of that chemistry. Sensory panels measure perception itself, but slowly, expensively, and with human noise. Each is a different stage of the encoding chain $\text{chemistry} \to \text{receptors} \to \text{perception}$ introduced in Chapter 1 and formalized in Chapter 2. Reading the three together — chemistry from GC-MS, a cheap proxy from the e-tongue, and ground truth from the panel — is how the empirical components of $\mathcal{F}$ are recovered. The methods are complementary precisely because they sit at different points along the projection.

A note on register: this is a reference appendix. It defines terms, fixes notation, and lists procedures. The worked theory is in the body of the book; here we give the reader enough to evaluate that theory’s empirical footing and, where relevant, to reproduce it. Code accompanying these methods is described in Appendix D; the compound databases the methods feed are described in Appendix B.

C.1 GC-MS as the Encoder Mapping Foods to Chemical Vectors

Gas chromatography–mass spectrometry is the workhorse instrument of flavor chemistry, and in the vocabulary of this book it plays a precise role: it is the encoder that turns a physical food sample into a high-dimensional chemical vector. Where Chapter 2 treats the receptor array as the biological encoder that projects chemistry onto $\sim\!400$ receptor channels, GC-MS is the analytical encoder that operates one step earlier, resolving a complex volatile mixture into its constituent compounds and their abundances. It is the instrument that makes the chemical input space measurable in the first place.

The two-stage separation. The method composes two operations. Gas chromatography separates a volatile mixture in time: the sample is vaporized and carried by an inert gas through a long capillary column whose interior wall is coated with a stationary phase. Compounds partition between the moving gas and the stationary coating according to volatility and polarity, so they exit the column at different retention times. A compound’s retention time (or, more robustly, its retention index, normalized against a homologous series of $n$-alkanes following the convention of Kováts 1958) is a coordinate that locates it along the separation axis. Mass spectrometry then identifies what exits at each moment: the eluting compound is ionized, typically by electron impact at 70 eV, and fragmented into a reproducible pattern of charged fragments sorted by mass-to-charge ratio $m/z$. The resulting fragmentation spectrum is a fingerprint matched against libraries such as NIST or Wiley.

[Modeling Axiom.] Formally, GC-MS realizes a map $$ E_{\text{chem}} : \{\text{food samples}\} \longrightarrow \mathbb{R}^{N}_{\geq 0}, \qquad E_{\text{chem}}(s) = \mathbf{c} = (c_1, c_2, \dots, c_N), $$ where each coordinate $c_k$ is the abundance of compound $k$ and $N$ is the size of the detected compound vocabulary — on the order of $10^2$ to $10^3$ for a single complex food, drawn from a universe of $>10^4$ known food volatiles. The compound vector $\mathbf{c}$ is the chemical representation; it is the raw material from which the databases of Appendix B (FlavorDB, FooDB, VCF) are assembled, and it is the input to the receptor map of Chapter 2. The geometric content of the flavor-pairing hypothesis (Chapter 4; Ahn et al. 2011) is a statement about cosine similarity between such vectors $\mathbf{c}_i$ and $\mathbf{c}_j$ — the flavor network of Ahn and colleagues is, in this language, a thresholded similarity graph on the GC-MS compound vectors.

Sample preparation determines the codomain. A point that is easy to miss is that the column you choose and the extraction you perform define which chemical space you are measuring, and therefore the basis in which $\mathbf{c}$ is expressed. Headspace methods sample only what volatilizes above the food and so weight the vector toward the aroma-active fraction that actually reaches the olfactory epithelium — the right bias for a flavor study, since it aligns the chemical measurement with retronasal delivery (Chapter 1, §1.2). Solid-phase microextraction (SPME), in which a coated fiber is exposed to the headspace and then desorbed in the injector, is the dominant technique because it is solvent-free and concentrates trace volatiles. Solvent extraction and simultaneous distillation–extraction recover heavier, less volatile compounds but introduce artifacts and discard the lightest top notes. [Empirical.] The consequence is that two laboratories can publish different compound vectors for the same food, not because either is wrong but because they sampled different projections of the same underlying chemistry. Any cross-study reuse of GC-MS data — and the flavor networks of Chapter 4 reuse it heavily — must treat the extraction method as part of the coordinate system, not as a transparent window onto it.

What the encoder preserves and what it discards. GC-MS is exquisitely sensitive to identity and quantity and almost completely blind to salience. The instrument reports that a compound is present at, say, 2 parts per billion; it does not report whether 2 ppb is far above or far below that compound’s human detection threshold. This is the single most important caveat for geometric work, and it recurs as a limitation in Chapter 10 (§10.6). A faithful chemical vector for flavor purposes must be reweighted by perceptual relevance. Two standard corrections are used in this book:

• Odor activity value (OAV). For compound $k$ with measured concentration $c_k$ and orthonasal detection threshold $\theta_k$, define $\text{OAV}_k = c_k / \theta_k$. Compounds with $\text{OAV}_k < 1$ are, to a first approximation, below threshold and contribute little to perception; compounds with large OAV dominate. Reweighting $\mathbf{c}$ by OAV converts a chemical vector into a crude perceptual-relevance vector and is the cheapest available approximation to the receptor projection of Chapter 2.
• GC-Olfactometry (GC-O). A human assessor sniffs the column effluent at the detector outlet and reports the time intervals at which an odor is perceived and its quality. GC-O directly identifies which of the hundreds of separated peaks are odor-active, often revealing that the perceptually dominant compounds are minor peaks by mass. Aroma extract dilution analysis (AEDA), systematized by Grosch (2001) into the modern “molecular sensory science” program, extends this by serially diluting the extract until each odor disappears, yielding a flavor dilution factor that ranks compounds by potency. GC-O is, in effect, a single-channel human sensor spliced onto the analytical encoder — a bridge between §C.1 and §C.3.

[Empirical.] The OAV correction and GC-O routinely show that the high-dimensional chemical vector collapses, for perceptual purposes, onto a sparse subset: in many foods, fewer than two dozen “character-impact” compounds account for the bulk of the recognizable aroma (Grosch 2001; Dunkel et al. 2014). This sparsity is the empirical face of the dimensional collapse argued throughout the book. The chemistry lives in $10^3$ dimensions; the perceptually load-bearing chemistry lives in tens. GC-MS measures the former, and OAV/GC-O is how we recover an estimate of the latter without yet involving a full sensory panel.

C.2 Electronic Tongue and Nose as Approximate Flavor Sensors

If GC-MS is a precise but slow and laboratory-bound encoder, electronic tongue and nose systems are the opposite: fast, portable, inexpensive, and deliberately imprecise. They do not separate and identify individual compounds. Instead they produce a low-dimensional response vector that is a holistic, non-specific transduction of the sample — closer in spirit to the way a biological receptor array works than to the way an analytical chemist works. In the geometric framing they are best understood as learned approximate projections: maps from a sample to a short vector $\mathbf{r} \in \mathbb{R}^m$, with $m$ typically between 6 and 32, whose coordinates have no individual chemical meaning but whose geometry — distances and clusters — is calibrated to track some target.

Architecture of the electronic nose. An e-nose is an array of $m$ partially selective gas sensors, each responding broadly and with overlapping specificity to many volatiles. Common transduction principles include metal-oxide semiconductor (MOS) sensors, whose conductivity changes on adsorption of reducing gases; conducting-polymer sensors; and quartz crystal microbalance sensors, which register mass loading as a frequency shift. [Modeling Axiom.] Each sensor $j$ implements an approximately linear functional on the headspace composition, $$ r_j \approx \sum_{k} w_{jk}\, c_k, $$ where $w_{jk}$ is the (broad, non-zero across many $k$) affinity of sensor $j$ for compound $k$. This is exactly the combinatorial-coding structure of olfaction from Chapter 2 (Malnic et al. 1999): a single sensor responds to many compounds, and a single compound excites many sensors, so identity is carried by the pattern across the array, not by any one channel. The e-nose is therefore an engineered caricature of the olfactory receptor sheet — with $m \sim 16$ channels standing in for the $\sim\!400$ biological receptor types. It captures the form of the biological encoder while drastically reducing its resolution.

Architecture of the electronic tongue. The e-tongue is the analogous instrument for the non-volatile, dissolved tastant fraction. It is an array of potentiometric or voltammetric electrodes with cross-selective membranes, each responding to a broad swath of ions and small molecules in solution. The same bilinear response model applies, with the array tracking the taste-relevant chemistry (the five-dimensional simplex of Chapter 1, §1.1, plus astringency and metallic notes) rather than the volatile aroma. Combined “electronic tongue + nose” platforms attempt to span both the taste and aroma blocks of the full flavor vector of §1.4, though texture, temperature, and the trigeminal axes (§1.3) remain almost entirely outside their reach.

Calibration: the geometry is borrowed, not intrinsic. The raw response vector $\mathbf{r}$ is meaningless on its own; its value comes entirely from a supervised calibration against a reference. A model — partial least squares regression, a support vector machine, or a small neural network — is trained to map $\mathbf{r}$ onto a target, which may be a chemical concentration measured by GC-MS, a sensory panel score from §C.3, or a discrete class label (authentic vs. adulterated, fresh vs. spoiled). The pattern-recognition formulation traces to Gardner and Bartlett (1994), who first framed the e-nose explicitly as an array plus a classifier. [Empirical.] Within the region of flavor space spanned by its training set, a well-calibrated e-nose can reproduce panel rankings of, for example, coffee roast level or olive-oil rancidity with correlations that rival a trained panel at a fraction of the cost and time. The instrument has learned a local chart of the perceptual manifold.

The crucial limitation is geometric and is worth stating sharply. [Empirical.] An electronic tongue or nose is reliable only locally — within the neighborhood of flavor space covered by its calibration samples — and degrades unpredictably outside it. Because its coordinates are non-specific learned combinations, it has no principled way to extrapolate: a novel off-flavor that excites the array in a direction unseen during training produces a confident but meaningless reading. This is the same failure mode that afflicts any embedding evaluated off its training manifold (Chapter 5, §5.4; Chapter 10), and it is why these instruments excel at quality control and adulteration screening — repeated measurement within a known region — and fail at de novo flavor description. They are local linearizations of the perceptual map, not global encoders of it. Used within their charts they are excellent cheap proxies; pushed past the chart boundary they report fiction with full confidence.

C.3 Sensory Panel Evaluation as Ground Truth

Neither GC-MS nor the electronic sensors measures perception. The chemical vector $\mathbf{c}$ is upstream of the receptors; the sensor vector $\mathbf{r}$ is a non-human caricature of them. The flavor vector $\mathbf{f}$ that this book is about — the perceptual location of a food in the space of human experience — is accessible by only one instrument: the trained human sensory panel. The panel is the ground truth against which the other two methods are calibrated, and against which the geometric predictions of the book (pairing quality, perceived distance, balance) must ultimately be tested. It is also the noisiest, slowest, and most expensive instrument, and the discipline of sensory science — codified in the standard reference of Lawless and Heymann (2010) and, for the descriptive method specifically, in Stone and Sidel’s Quantitative Descriptive Analysis (Stone et al. 1974) — exists precisely to render its output reproducible enough to function as a measurement at all.

Panel design. A descriptive analysis panel — the gold standard for characterizing rather than merely preferring a flavor — consists of 8 to 15 assessors selected and screened for acuity, then extensively trained (often 40 to 120 hours) on a common vocabulary. The training has an explicitly geometric purpose: it establishes a shared basis for $\mathcal{F}$. Through consensus, the panel develops a list of perceptual attributes (for a tomato: “ripe-fruity,” “green,” “acidic,” “sweet,” “umami,” and so on), each anchored by physical reference standards that fix the meaning and the zero point of that attribute’s axis. [Modeling Axiom.] The panel thereby agrees on a coordinate frame $\{a_1, \dots, a_p\}$ — typically $p$ on the order of 10 to 40 attributes — and each sample is scored as a vector $$ \hat{\mathbf{f}} = (a_1, a_2, \dots, a_p) \in \mathbb{R}^p, $$ the empirical estimate of the perceptual flavor vector whose theoretical structure occupies Chapter 1. Replicate scoring across assessors and sessions makes $\hat{\mathbf{f}}$ a measurement with quantifiable error rather than an impression.

Scaling. How assessors convert a perception into a number is a measurement problem in its own right, and the choice of scale is the choice of how the metric of Chapter 3 is sampled empirically. The dominant instruments are:

• Line scales / structured category scales. The assessor marks intensity on a continuous line anchored at the ends (e.g., “absent” to “extreme”); the position is digitized to a number. Simple and stable, but treated as interval data it implicitly assumes a linear perceptual metric, which the Weber–Fechner and Stevens results of Chapter 3 tell us is false. In practice the distortion is tolerable over the narrow intensity ranges typical of a single product category.
• Magnitude estimation. The assessor assigns numbers whose ratios reflect perceived intensity ratios, with no fixed upper anchor. This scaling is the one that directly exposes Stevens’ power law (§3.3; Stevens 1957): plotting log perceived magnitude against log physical concentration recovers the power-law exponent for each modality. It is the scale of choice when the curvature of the flavor metric is itself the object of study.
• Labeled magnitude scales (e.g., the gLMS). A quasi-logarithmic scale with semantic anchors (“barely detectable” … “strongest imaginable”) spaced according to empirically measured perceived intensities (Bartoshuk et al. 2004). The gLMS is designed to permit valid across-subject comparison, which matters for the supertaster/non-taster variation of Chapter 2 (§2.5–§2.6): a non-taster and a supertaster can be placed on the same axis without the supertaster’s compressed dynamic range distorting the group geometry.

[Empirical.] The choice of scale is not cosmetic. Because perceived intensity is a nonlinear (approximately power-law or logarithmic) function of concentration, a panel using a linear line scale and a panel using magnitude estimation will report different distances between the same two samples — they are sampling the same curved metric with differently shaped rulers. Geometric reuse of panel data, especially the iso-intensity surfaces of §3.3 and the hedonic gradients of §3.6, must record which scale produced the numbers.

Reliability. A measurement that cannot be reproduced is not a measurement, and a panel’s central methodological burden is demonstrating that it is. The standard checks decompose the variance of $\hat{\mathbf{f}}$ into signal and several distinct noise terms:

• Discrimination. Can the panel reliably tell apart samples that are different? Tested with difference protocols — the triangle test (identify the odd sample of three), duo-trio, and tetrad tests — analyzed against the binomial null of guessing.
• Agreement (reproducibility across assessors). Do assessors place a sample at the same location? Quantified by intraclass correlation coefficients and by the assessor $\times$ sample interaction term in a three-way ANOVA (assessor, sample, replicate). A large interaction means assessors disagree about direction, not just scale, and signals an unstable basis.
• Repeatability (within assessor). Does an assessor score blind replicates of the same sample consistently? This is the per-axis measurement noise.
• Panel-level consensus. Generalized Procrustes analysis (GPA; Gower 1975) rotates, scales, and translates each assessor’s individual configuration of samples onto a common consensus space, then reports the residual as the disagreement that survives removal of idiosyncratic scale and orientation. GPA is, quite literally, the alignment of many individual perceptual coordinate frames into one — the empirical procedure that justifies treating $\mathcal{F}$ as shared across people despite the personal rotations of §2.5.

[Empirical.] Properly run, descriptive panels achieve good repeatability and consensus, and their multivariate output — the sample $\times$ attribute matrix of $\hat{\mathbf{f}}$ vectors — is exactly the form on which the principal-component analyses of Chapter 9 (the eigenvalues of a cuisine) operate. The panel is thus not merely a validator but a primary instrument for discovering the perceptual dimensions: the principal components of a well-designed panel’s scores are an empirical estimate of the axes of $\mathcal{F}$ itself.

Affective testing is a different instrument. A distinction that the geometry makes sharp: descriptive panels measure where a sample sits in $\mathcal{F}$ (intensity along attribute axes), while consumer affective panels measure whether people like it — a scalar (or low-dimensional) projection of the hedonic gradient of §3.6 and the hedonic manifold of Chapter 5. Affective tests use large untrained consumer samples (often $n > 100$) and hedonic scales (the 9-point hedonic scale; preference ranking; check-all-that-apply). The two must not be conflated: a trained panel can tell you a cheese is high on “barnyard” and “ammoniacal” with great precision and tell you nothing about whether anyone wants to eat it. Mapping descriptive coordinates onto consumer liking — preference mapping — is the empirical fitting of the hedonic function $H(\mathbf{f})$ of §5.6, and it is one of the most direct industrial uses of the geometric picture.

C.4 Reading the Three Together

The three instruments are aligned along the same projection chain that organizes the whole book, and their proper use is to triangulate the stages of that chain. GC-MS measures the chemical input $\mathbf{c}$ in high resolution but is blind to salience. The electronic nose and tongue measure a cheap, local, non-human projection $\mathbf{r}$ that approximates the receptor stage and is reliable only on its calibration manifold. The sensory panel measures the perceptual output $\hat{\mathbf{f}}$ — the genuine ground truth — at high cost and with human noise that sensory science is built to control. The standard empirical workflow of modern flavor science, and the one assumed by the analyses in this volume, runs the chain in both directions: GC-MS and GC-O identify the character-impact compounds; an e-nose calibrated against panel scores provides rapid routine measurement within a product category; and the panel periodically re-anchors the whole system to human perception. [Empirical.] Where all three agree, a claim about flavor geometry rests on chemistry, on a fast proxy, and on perception at once; where they disagree, the disagreement localizes which projection in the chain $\mathbf{c} \to \mathbf{r} \to \hat{\mathbf{f}}$ is responsible — which is itself a measurement of the structure this book exists to describe.

For the databases that aggregate GC-MS output into reusable compound vectors, see Appendix B. For the mathematical background on PCA, Procrustes alignment, and the metric structures these methods sample, see Appendix A. For code that ingests these data and computes the embeddings and distances of Chapters 4, 9, and 10, see Appendix D. The connection to the neural projection that ultimately produces $\hat{\mathbf{f}}$ — the gustatory and olfactory pathways as a biological dimensionality-reduction pipeline — is developed in Chapter 12 and, at the level of cross-modal cognition, in Volume 6: Geometric Cognition. The instruments of this appendix are how that pipeline is read from the outside. They do not measure the geometry directly; they measure its shadows at three different distances from the light, and the theory is what reconciles the shadows into a single shape.