Chapter 3: Historical Precursors — Geometry Before Geometry

RUNNING EXAMPLE — Hiroshi’s Library

Hiroshi is a curator at a mid-size museum. He is building the wall text for a show that traces a single motif — the human figure in inward contemplation — across two and a half millennia. He has assembled a reading list: Vitruvius on proportion, Alberti on istoria, a translation of Zeami’s treatises that he has annotated heavily, Kant’s Critique of Judgment, Birkhoff’s Aesthetic Measure, Arnheim’s Art and Visual Perception, a selection from Bense and Moles, and a stack of recent papers on deep-learning-based image generation. He has been reading the list for months. He is struck by a peculiar fact: almost every author is trying to say the same thing, and almost none of them have the vocabulary for it. They have proportion, harmony, measure, balance, purposiveness, information, redundancy, order, complexity. They each name one coordinate of something larger. None of them, until the very end of the list, can write down the object that has the coordinates. Hiroshi’s show, he decides, will be about exactly that: the object that the tradition kept reaching toward and could not quite hold.

Introduction: Structure Before Formalism

The mathematical apparatus of differential geometry was developed in the nineteenth and early twentieth centuries. Aesthetic theory predates it by millennia. Yet the structural insights that geometric aesthetics formalizes — multi-dimensionality, directional content, transformation behavior under reframing, context-dependent metrics, the distinction between intrinsic and perspectival properties — have appeared throughout the history of aesthetics in various guises.

This chapter traces a genealogy of proto-geometric thinking. The claim is not that Vitruvius or Kant or Birkhoff secretly knew differential geometry. The claim is that they grappled with phenomena that resist scalar treatment and developed conceptual tools that, in retrospect, capture aspects of the geometric structure we now make explicit. Reading these thinkers through a geometric lens both illuminates their insights and reveals why the formulation had to wait for the arrival of learned representations in the 2010s.

The story has roughly four acts: the proportional tradition (Pythagoras through Vitruvius and Fibonacci), the philosophical turn (Kant), the first mathematical aesthetics (Birkhoff and the information-theoretic school), and the perceptual-structural tradition (Arnheim and his successors). Each contributed a fragment of the vocabulary. None could write down the manifold.

The Proportional Tradition: The First Geometry of Beauty

The oldest strand of Western aesthetic theory identifies beauty with proportion. The Pythagoreans, studying vibrating strings, observed that musical consonance corresponds to simple integer ratios: 2:1 (octave), 3:2 (fifth), 4:3 (fourth). This empirical observation, plausibly the first quantitative aesthetic finding in the Western tradition, generalized into a metaphysical commitment: beauty is ratio; the cosmos is beautiful because it is mathematically structured; the human soul is attuned to beauty because it participates in the same harmonic structure.

Plato, in the Timaeus, assigns the regular polyhedra to the classical elements and treats geometric proportion as the constitutive structure of the sensible world. In the Philebus, beauty is associated with symmetria — commensurability of parts — and with the proper mixing of measure (metron) into the infinite (apeiron). The recurring move is to locate aesthetic value not at a point on a line but in relations among magnitudes.

Vitruvius and the Canon of the Body

Vitruvius’s De Architectura (c. 30–15 BCE) extends the proportional program to building and to the human figure. The famous passage — a man inscribed in a circle and a square, arms and legs extended to touch both — is a claim about commensurability: the human body, correctly measured, exhibits simultaneous proportional relations to two distinct elementary forms. Leonardo’s Vitruvian Man of c. 1490 is a drawing of this claim.

What Vitruvius is doing, structurally, is specifying a small set of coordinates and a set of relations they must satisfy. The body is a foot long, a fathom wide, a finger thick, and these magnitudes stand in particular ratios to each other. A building whose columns, entablature, and intercolumniations satisfy parallel ratios participates in the same structural beauty. The aesthetic claim is not that any single dimension is “right” but that the joint configuration of dimensions falls into a constrained region of a space of possible configurations.

Fibonacci and the Golden Ratio

The golden ratio, $\varphi = (1+\sqrt{5})/2 \approx 1.618$, makes its formal entrance in Euclid as the “division in extreme and mean ratio.” Fibonacci’s Liber Abaci (1202) introduces the recursive sequence $F_{n+1} = F_n + F_{n-1}$ whose successive ratios converge to $\varphi$. Over the next centuries, a folklore accumulates around the golden ratio’s supposed aesthetic privilege: the Parthenon, Renaissance portrait composition, the nautilus, Le Corbusier’s Modulor.

Modern empirical psychology has been, at best, ambivalent about the claim. Controlled rectangle-preference experiments (Fechner, 1876; and most replications since) show small and inconsistent effects, often swamped by framing, familiarity, and context. The literal claim — that $\varphi$ is the aesthetic constant of rectangles — is not well-supported by the evidence.

What is well-supported is a weaker, structural claim that the proportional tradition points toward: aesthetic responses to spatial composition are sensitive to proportional relationships, with preferences that vary across individuals, tasks, and cultures, and that concentrate in certain regions of the ratio axis. In the geometric vocabulary: the rectangle-preference landscape is a function on a one-dimensional ratio-space, it has non-trivial shape, and different subpopulations locate their preferences at different points. The golden-ratio folklore errs by asserting a single global maximum on this axis; the proportional tradition is correct to assert that the axis itself is aesthetically meaningful.

What the Proportional Tradition Saw and Missed

The proportional tradition’s structural insight is that aesthetic value is not a magnitude of a property but a relation among magnitudes. A beautiful building is not a building with a lot of some property $B$; it is a building whose proportions fall into the right region of a space of possible proportions. This is exactly a claim that aesthetic value lives over a multi-dimensional configuration space rather than on a single scalar axis.

What the tradition lacked was the vocabulary for that configuration space. The Pythagoreans had ratios, not manifolds. Vitruvius had specific numerical constraints, not a metric tensor. The golden-ratio folklore reached for a single privileged constant precisely because the tradition had no formal apparatus for describing families of aesthetic relations. The reduction of relational insight to a single magic number is what you get when you have the intuition that aesthetics is structural but no formal grammar for structure.

The proportional tradition is the first pre-geometric aesthetics. Chapter 9 (The Origin of the Aesthetic Metric) returns to it when we ask how the metric on the aesthetic manifold is chosen: the Pythagorean answer — that the metric is given by consonance, which is given by simple ratios, which is given by physical acoustics — is structurally the ancestor of the governed-discovery account we develop there.

Kant: Purposiveness Without Purpose

Immanuel Kant’s Critique of the Power of Judgment (1790) is the most sustained philosophical attempt before the twentieth century to identify the distinctive structure of aesthetic evaluation. Kant’s thesis: aesthetic judgments of beauty are disinterested, universal, and purposive without purpose (Zweckmäßigkeit ohne Zweck). A beautiful object exhibits the form of purposiveness — it looks as if it were made for something — without having a determinate concept of what that something is.

This is notoriously slippery prose. In the light of the present framework, it is also an astonishingly precise set of structural claims.

The Disinterested Judgment

Disinterested means that the aesthetic judgment does not arise from desire (the beautiful is not the useful), from cognition (the beautiful is not the true), or from the moral (the beautiful is not the good). Kant is carving out a judgment that is orthogonal in a specific structural sense to three other directions in evaluative space.

The geometric reading is clear: Kant is positing an axis of evaluation distinct from the utilitarian axis, the cognitive axis, and the moral axis. The aesthetic judgment is a projection onto that axis, not onto the others. The fact that the axes are distinct — that they can be separately named and separately interrogated — is itself a claim that evaluation space is multi-dimensional.

This is a proto-geometric move. A scalar framework of evaluation collapses all value to a single number; Kant is insisting on at least four dimensions (useful, true, good, beautiful), each with its own structure, and each deserving its own analysis. He does not have the tensor vocabulary to write this down, but the claim has the form of a direct-sum decomposition of evaluation space.

Purposiveness Without Purpose

The phrase purposiveness without purpose is Kant’s name for the structural feature he observes in beautiful objects: they exhibit a form of internal coherence — a mutual fitness of parts — that resembles what we see in purposive artifacts, but without presupposing a specific purpose that the parts are fit for. A flower is not a clock, but its petals, pistil, and stamen exhibit the same form of internal fitness that a clock’s gears do, without being fit for any specific external end in the way the clock’s gears are fit for telling time.

Read geometrically, this is a claim about internal coherence as an aesthetic feature — a claim that can be, and in our empirical work has been, operationalized. Chapter 17 reports that among the channels that predict aesthetic rating in books, the internal-coherence features (pair_sim_mean, pair_sim_std — the mean and standard deviation of cosine similarities between embedded paragraph pairs within a book) carry 8.4σ signal and are statistically distinct from the spectral-divergence features. Internal coherence — the mutual fitness of the parts — is an empirically measurable aesthetic dimension.

And here the modality-specific sign flip is instructive. In books (Chapter 17, Phase 1), pair_sim_mean has ρ = +0.126 with rating: higher internal coherence, higher rating. Kant’s flower and Kant’s novel both reward internal fitness. In music (Chapter 21), the same feature has ρ = −0.076 with listen counts (p = 5×10⁻³³): music rewards contrast and deviation from internal self-similarity. Kant’s analysis of the beautiful object, built primarily on visual and natural-scene examples, identifies a feature that indeed generalizes — but with a sign that depends on the medium. The direction of purposiveness-without-purpose is modality-specific. Kant saw the feature. The geometric framework tells us how it transforms across strata.

The Universal Subjective

The third component of Kant’s analysis — that aesthetic judgments claim universal validity despite being subjective — is the deepest. Kant wants to say that when we judge something beautiful, we are not merely reporting our preference; we are making a claim that invokes the agreement of every other judge with the same faculties. And yet we cannot prove the judgment; we can only invite the other to look for herself.

This is a transformation-invariance claim. Kant is asserting that the aesthetic judgment should be invariant under the transformation from my perspective to any other rational being with aesthetic sensibility. That is, the judgment is not indexical; it attempts a rank that is stable across observers.

The companion volume’s Chapter 12 develops this for ethics under the name of the Bond Invariance Principle. The aesthetics version is the subject of our Chapter 12, whose empirical witness is the cross-lingual result. In our data (Chapter 17, Phase 3), 4,683 non-English books across 19 languages and 10 language families, LaBSE-encoded and projected into the English-corpus PCA-128 basis, reveal that core aesthetic features transform coherently across languages. Internal coherence (pair_sim_mean) has ρ = +0.712 averaged across language pairs; Mahalanobis-mean has ρ = +0.710; the divergence family (Hellinger, Bhattacharyya, JS) has ρ ≈ 0.67. The headline result — English–Finnish Hellinger ρ = +0.77 at n = 288, p = 8×10⁻⁵⁷ — is, as much as anything we have, the empirical witness to what Kant meant by the universal subjective: the aesthetic structure is stable under the transformation from one language’s readership to another’s.

Kant had the structural intuition. He could not test it, and the testing required both learned representation spaces and a corpus of cross-lingually paired works. The framework he pointed at was geometric all along.

Birkhoff: The First Mathematical Aesthetics

George David Birkhoff’s Aesthetic Measure (1933) is the first sustained attempt to reduce aesthetic evaluation to a computable formula. Birkhoff, a mathematician at Harvard who had produced major work in dynamical systems and ergodic theory, proposed:

$$M = \frac{O}{C},$$

where $M$ is the aesthetic measure of an object, $O$ is its order (the amount of regularity, symmetry, and rule-governed structure), and $C$ is its complexity (the amount of elementary material the perceiver must process). Beauty, Birkhoff argued, is high when order is high relative to complexity — when the observer’s perceptual effort is efficiently rewarded by the detection of structure.

Birkhoff applied the formula to polygonal shapes, ornamental tile patterns, musical melodies, and lines of poetry. His tables assigning specific $O$ and $C$ values to particular works read, today, as both ambitious and strange. But the formula’s structural content is important: aesthetic value is a ratio of two distinct quantities, not a magnitude of one.

What Birkhoff Got Right

Birkhoff identified two distinct dimensions of aesthetic evaluation. Order is what the perceptual tradition will later call pattern or redundancy or integration. Complexity is the raw quantity of the stimulus. A scalar aesthetics that knew nothing about this distinction would treat a simple square and an ornate rose window as comparable along a single axis. Birkhoff’s ratio says: no — they differ both in $O$ and in $C$, and their aesthetic measure depends on both.

The geometric reading: Birkhoff posits a two-dimensional evaluation space with coordinates $(O, C)$, and defines aesthetic measure as a specific function on that space. The function $M = O/C$ is one choice of contraction; other choices (linear combinations, logarithmic weightings, context-dependent metrics) are possible and empirically different. But the move from scalar $M$ to a point $(O, C)$ in a plane is the right structural move.

Birkhoff also recognized that different artistic domains require different operationalizations of $O$ and $C$. His procedure for polygonal shapes is different from his procedure for melodies. This is a proto-stratification claim: the metric is locally defined on each domain and not straightforwardly portable across them. Chapter 8 makes this precise.

What Birkhoff Missed

Birkhoff’s formula has been, rightly, criticized. The values of $O$ and $C$ that he computes for specific objects depend heavily on hand-chosen counting conventions; the formula gives results that do not match informed aesthetic judgment in several well-studied cases; and the ratio $O/C$ collapses two independent dimensions back into a single number, reintroducing the scalar pathology at the moment of evaluation even after dimensional structure has been correctly identified.

The deeper problem is that Birkhoff’s $O$ and $C$ are hand features. They are numerical summaries of properties that a theorist has decided to measure. They are not learned from data about how actual perceivers respond. In our empirical work on books (Chapter 17), the difference between hand features and learned-representation geometric features is substantial: hand features recover a small fraction of the signal that learned-embedding-plus-geometric-channels recover.

Birkhoff’s aesthetics was the right structural move with the wrong substrate. He needed a representation space that the perceptual system itself uses. That substrate — learned representations — was nine decades away.

Information-Theoretic Aesthetics: Bense, Moles, and the Shannon Turn

After Shannon’s 1948 Mathematical Theory of Communication, a generation of theorists — Max Bense in Germany, Abraham Moles in France, Rudolf Arnheim in a related but distinct lineage — attempted to ground aesthetics in information theory. Bense’s Aesthetica (four volumes, 1954–60) proposed that aesthetic objects carry a computable information content and that aesthetic value relates to the interplay between this content and the observer’s expectations. Moles’s Théorie de l’information et perception esthétique (1958) formalized the idea that art operates in the zone of moderate information — too low, and the work is boring; too high, and the work is noise.

Shannon himself noted, in scattered remarks, that redundancy is characteristic of artistic language: natural text carries about one bit per character of information against a raw capacity of roughly five bits, and this redundancy — the structure that makes text both compressible and resistant to corruption — is structurally analogous to the internal coherence of aesthetic objects.

What the Information-Theoretic Tradition Saw

The information-theoretic aestheticians made two contributions that persist.

First, they identified redundancy as an aesthetic feature — an empirically measurable property of a work, related to but distinct from Birkhoff’s order. Redundancy is a statistical property of the work over its own corpus; it is not a hand-counted feature of the work in isolation. This is an important move. It begins to treat aesthetic features as properties of the work’s location in a representation space populated by other works, not as intrinsic properties of the work alone.

Second, they introduced the idea of aesthetic value as an optimum on a mid-range of information — not a monotone function of information content. The Moles curve, hand-drawn in his monograph, shows aesthetic value peaking at a moderate level of statistical surprise, falling off on both the low end (boredom, cliché) and the high end (noise, incomprehensibility). This is a stratification claim: the aesthetic manifold has interior regions and boundaries, and the value function is not monotone along the information-content axis.

In our own work, we can locate operationalizations of both claims. Redundancy in text corresponds, approximately, to pair_sim_mean (higher pair-wise similarity of embedded paragraphs = lower per-paragraph novelty = higher redundancy). In books, this correlates positively with rating (ρ = +0.126, 8.4σ, n=4,998). In music, it correlates negatively with listen counts (ρ = −0.076, p = 5×10⁻³³, n=24,801). Moles’s generic middle-of-the-information-curve claim is empirically wrong in its universal form — the sign of the redundancy gradient depends on the medium. But the structural move — that redundancy is an aesthetic dimension with a non-monotone relation to value — is correct.

What the Information-Theoretic Tradition Missed

Two limitations. First, Shannon’s information is defined relative to a probability distribution over symbols, and the choice of symbol alphabet profoundly affects the computation. Bense and Moles computed entropies over letter and phoneme distributions — shallow statistics that capture surface structure but miss semantics and deep form. An information-theoretic aesthetic computed over characters will be nearly the same for Moby-Dick and a same-length sample of 19th-century literary English. The work has to be embedded in a richer space for its aesthetic structure to become visible.

Second, and relatedly, the information-theoretic tradition still worked with a scalar output: a single number (entropy, redundancy) that would correlate with aesthetic value. The directional and tensorial structure of aesthetic evaluation — its rank-1-and-higher content — was not expressible in their framework. They had axes but not vectors.

Chapter 6 develops the tensor hierarchy. The information-theoretic aestheticians reached the bottom rung — rank-0 scalars over a two-dimensional order-complexity plane — and stopped there.

Arnheim and Visual Balance

Rudolf Arnheim’s Art and Visual Perception (1954; revised 1974) took a different route. Arnheim was a Gestalt psychologist and a scholar of visual art. His book reads as a catalog of perceptual-structural principles — balance, shape, form, growth, space, light, color, movement, dynamics, expression — each illustrated with schematic diagrams and concrete works.

Arnheim’s central insight is that visual perception is field-like: a composition is not a collection of isolated elements but a dynamic equilibrium of forces. A shape in the upper right of a canvas exerts a perceptual “weight”; a smaller shape in the lower left can balance it, and the balance is felt, not computed. Visual balance is a property of the whole, not of any part.

The Field-Theoretic Reading

Arnheim is describing what a physicist would call a field and what a differential geometer would call a section of a tensor bundle over the visual plane. Each point in the image carries a perceptual weight (a scalar), a direction of pull (a vector), and — for more complex figures — higher-rank structure. The composition is evaluated globally via an integral of these fields, and the aesthetic verdict depends on whether the integral has a particular form (balance) that the perceptual system registers.

This is the most sophisticated pre-computational aesthetic geometry in the tradition. Arnheim does not have the formal apparatus, but his diagrams show directed forces, curvatures, and tensorial interactions. When he writes about the “dynamics” of a composition, he means what a physicist would mean: the structure of forces that act on the observer’s attention as the eye moves through the frame.

What Arnheim Missed

Arnheim’s field theory was qualitative. He could not compute the weight vector at a given point of a given image, and he had no means to test whether his claims about visual forces matched measured viewer response. His theory is, structurally, a rank-1 vector-field account of the visual plane, and it is almost certainly directionally correct. But it lived as phenomenological description, not as measurable structure.

The arrival of convolutional and transformer-based vision models has changed this. A sufficiently trained vision encoder produces a representation space in which the position of an image induces a structured response that is substantially correlated with Arnheim’s described forces. Chapter 23 (Geometric Visual Art) develops the details. The framework that Arnheim reached for, without being able to hold it, has become tractable in the last decade.

What the Tradition Saw — And What It Could Not Have Seen

Let us summarize, honestly, what the pre-geometric tradition accomplished and where it necessarily stopped.

The proportional tradition identified the structural claim that aesthetic value is a function on a multi-dimensional space of configurations, not a magnitude of a single property. It lacked the formalism for that space.

Kant identified the distinctive structure of aesthetic judgment — disinterested, purposive-without-purpose, universal-subjective — and thereby argued that evaluation space has an aesthetic axis distinct from the utilitarian, cognitive, and moral axes. He also, in the universal-subjective component, articulated a transformation-invariance claim whose empirical test had to wait for learned representations and cross-lingual corpora.

Birkhoff made the first explicit move to a multi-dimensional evaluation plane, with his $(O, C)$ coordinates, and recognized that aesthetic measure is a non-trivial function over that plane. He used hand features and a fixed ratio; he lacked learned representations and a richer geometry.

The information-theoretic aestheticians made redundancy an explicit aesthetic dimension and identified a non-monotone relation between information content and aesthetic value. They worked with shallow entropies and still-scalar outputs.

Arnheim produced the most sophisticated pre-computational account: a field-theoretic description of visual composition as a dynamic equilibrium of directed forces. It was qualitative.

What could none of them have done? Three things.

First, they could not embed works into a rich, learned representation space of the sort that a modern encoder (LaBSE for text, MERT for music, large vision transformers for image) produces. Without such a space, there is no substrate on which to compute distances, divergences, and geodesics. The proportional tradition had ratios among chosen magnitudes. Kant had concepts. Birkhoff had hand features. The information-theoretic aestheticians had shallow entropies. Arnheim had diagrams. None of these is a space in the sense that differential geometry requires.

Second, they could not access cross-cultural or cross-linguistic invariance empirically. Kant asserted universal-subjective validity as a philosophical claim; he could not test it against ρ = 0.71 across six language families. The tradition debated whether beauty is universal; the debate was resolvable only with both the representation space and the corpora, neither of which existed before the 2010s.

Third, they could not build computational systems that embedded aesthetic structure into decisions at scale. The failure modes that Chapter 1 identified — engagement collapse, critical-aggregation collapse, canon drift — are modern failures because they require modern systems. The tradition could not anticipate the specific stakes that make a geometric framework not merely interesting but urgent.

Our Contribution in Context

What we offer in this book is not a new aesthetic theory. It is the formalization of an old one. The proportional tradition, Kant, Birkhoff, Bense, Moles, Arnheim, Zeami — these writers were reaching for a structural account of aesthetic judgment that scalar frameworks cannot provide and that requires differential-geometric vocabulary to state precisely.

Our specific contributions are three.

First: we embed works into learned representation spaces (LaBSE for text, MERT for music) and compute geometric structure in those spaces — divergences, internal coherence, trajectory geometry, spectral channels. The representation space is the substrate that the tradition lacked. Chapters 4–7 develop this.

Second: we test the structural claims against empirical data. The Kantian universal-subjective becomes the cross-lingual invariance result (Chapter 17, Phase 3). Birkhoff’s order-complexity plane becomes the spectral-plus-coherence channel decomposition (Chapter 17, Phase 1). The information-theoretic middle-of-the-curve becomes the sign-flipped internal-coherence result (Chapters 17, 21). The proportional tradition’s ratio-structure becomes the within-genre feature importances. Each historical intuition finds either confirmation, refinement, or honest refutation in the data.

Third: we honestly report where the framework’s predictive power is modest. Eighty-five percent of raw book signal is genre confound; ninety-one percent of raw music signal is genre confound. The within-genre residual is small (R in the 0.09–0.19 range) but significant. The geometric framework does not claim to predict aesthetic verdicts with high accuracy. It claims to give a richer, more honest, and more auditable decomposition of what is going on than the scalar alternatives. And that decomposition, tested against historical intuition, matches the tradition’s insights more precisely than the tradition could have tested them itself.

A framework that is mathematically rich, empirically modest, and honest about its residuals is a framework that the pre-geometric tradition would, we think, have welcomed. Kant’s epistemic restraint, Birkhoff’s interest in measurable aesthetics, and Arnheim’s patient observation are the governing spirits of this book. The mathematics is new. The questions are old.

Bridge to Part II

Part I has motivated the geometric turn. Chapter 1 introduced the three fragments — Hōjōki, Zeami’s flower, Duchamp’s urinal — and the three failures of scalar aesthetics. Chapter 2 developed the failures in empirical detail, using Goodreads, Metacritic, Billboard, and the 85%-genre-confound result. This chapter has traced the long pre-geometric history of the structural insight and located our contribution honestly within it.

Part II builds the apparatus. Chapter 4 is the mathematical preliminaries — manifolds, tensors, metrics, connections, stratified spaces — stated at a level accessible to readers willing to work and referenced as a lookup for readers who already know the vocabulary. Chapter 5 develops the aesthetic manifold proper. Chapter 6 develops the tensor hierarchy. Chapter 7 returns to a single work and walks it through five levels of geometric elaboration, parallel to the kidney-allocation chapter of Geometric Ethics.

Hiroshi’s show, when it opens, is organized not as a chronological tour but as a tour of a conceptual space. On the first wall: proportion. On the second: inner coherence. On the third: cross-perspective transformation. On the fourth: information and redundancy. On the fifth: visual force. On the sixth — the wall Hiroshi has been waiting to write — is the single diagram that tries, however imperfectly, to show that all of these are coordinates of the same object. The diagram is a manifold. The wall text reads: Beauty has had many names. They have all been naming parts of the same shape.

❖

The tradition saw the axes.

What it could not see, until now, was the space they coordinate.