Daryl Costello: Independent Researcher

Correspondence: Daryl.costello@outlook.com

Abstract

We present Ontogenetic Geometry, a unified mathematical framework integrating biological development, cognitive emergence, and evolutionary dynamics within a common geometric substrate. The framework rests on four interlocking formal structures. First, we demonstrate that both biological and cognitive ontogeny are naturally described as flows on fibre-bundle-structured state spaces: the base space encodes environmental and evolutionary contexts, while fibres parametrise developmental trajectories available within each context. This formulation subsumes and generalises Waddington’s epigenetic landscape and D’Arcy Thompson’s transformation theory within a rigorous differential-geometric setting. Second, we introduce the renormalization group (RG) flow as the canonical coarse-graining operator for developmental biology, providing a principled bridge between molecular-genetic and organ-level descriptions. Fixed points of the developmental RG flow correspond to conserved body plans and phylotypic stages; relevant, irrelevant, and marginal perturbations translate directly into macroevolutionary operator classifications. Third, we formalise an operator-stack architecture (a category-theoretic hierarchy of morphisms on nested state-space levels) that encodes hierarchical transformation rules governing both individual development and phylogenetic change. Standard evo-devo concepts, including heterochrony, heterotopy, modularity, and evolvability, receive precise algebraic expression within this architecture. Fourth, we construct the unified state space as a product manifold of developmental, cognitive, and evolutionary sub-manifolds, each operating at a characteristic timescale, and show that the classical recapitulation debate dissolves when Haeckel’s linear narrative is replaced by a multi-dimensional attractor geometry: transient convergence of developmental trajectories toward shared RG fixed-point attractors, followed by divergence under relevant perturbations. The framework yields testable predictions regarding power-law scaling of morphogenetic correlations at developmental phase transitions, conservation of gene-regulatory-network operator subalgebras across sister clades, and signatures of RG flow in infant cognitive development. We close by identifying implications for evo-devo synthesis, theoretical neuroscience, and AI alignment, arguing that artificial cognitive architectures implementing RG-structured hierarchies are better candidates for robust generalisation.

Keywords: ontogeny; evolutionary geometry; renormalization group flow; operator-stack architecture; cognitive development; morphogenetic fields; fibre bundles; attractor dynamics; evo-devo; phylogenetic trajectory

1. Ontogenetic Geometry: Introduction

The Classical Problem: Ontogeny and Phylogeny

Few problems in the history of biology have proved simultaneously so seductive and so treacherous as the relationship between individual development and the history of species. When Ernst Haeckel (1866) enunciated his biogenetic law, the proposition that ontogeny recapitulates phylogeny, he gave nineteenth-century comparative anatomy a powerful, if ultimately flawed, organising principle. On Haeckel’s reading, the embryo of any given species passes through stages that replicate, in order, the adult forms of its evolutionary ancestors: the human embryo, in this telling, is successively fish-like, amphibian-like, and reptilian before acquiring distinctively mammalian characters. The appeal of the idea was enormous; it promised to render the geological record of phylogenetic history legible in the bodies of living embryos, rendering museums of comparative embryology into direct windows onto deep evolutionary time.

The empirical collapse of the strong biogenetic law was both swift and decisive. Anatomical investigation showed that embryos of different vertebrate species, far from passing through adult stages of ancestral forms, actually pass through stages that resemble embryos of related species, a crucial distinction that Haeckel himself sometimes obscured or, in the notorious case of his comparative embryo drawings, actively misrepresented (Richardson et al., 1997). Karl Ernst von Baer had in fact articulated the more accurate generalisation several decades earlier (von Baer, 1828): embryos of different species within a major group (e.g., vertebrates) resemble one another more closely at early developmental stages than at later ones, and the developmental sequence proceeds from the general to the particular rather than from the ancestral to the derived. Von Baer’s law is empirically robust; Haeckel’s is not.

Yet the collapse of Haeckel’s programme left a conceptual vacuum at the heart of theoretical biology: in what precise sense, if any, does the history of a lineage constrain or shape the development of its members? The question is not merely of historical interest. Development is, on its face, a local, mechanistic process governed by molecular-genetic interactions unfolding in real time within a single organism. Evolution, by contrast, is a population-level, historical process extending across generations and shaped by differential reproductive success. The challenge is therefore not simply empirical but formalismological: what mathematical language, if any, can simultaneously encode the internal logic of developmental causation and the external, population-level dynamics of selection and drift? No framework within nineteenth- or early twentieth-century biology offered a satisfying answer. Von Baer’s law, accurate as an empirical observation, remained essentially a descriptive generalisation without deep mechanistic or mathematical underpinning. It is precisely here that a geometric reformulation opens a genuinely new theoretical space.

The Geometric Turn in Biology

The history of mathematical biology contains several partial anticipations of the geometric approach advocated here, each illuminating an important facet while lacking the algebraic completeness that would make it fully general. D’Arcy Wentworth Thompson’s landmark monograph On Growth and Form (Thompson, 1917) demonstrated that the morphological differences between related species could be captured by smooth coordinate transformations applied to a reference body plan. Thompson’s “Cartesian transformations” were visually striking and biologically suggestive, but they lacked a dynamical component: they described the endpoint of development in different species, not the process by which those endpoints are reached, and they provided no account of the selective or genetic mechanisms driving morphological divergence.

Conrad Hal Waddington’s epigenetic landscape (Waddington, 1957) supplied the dynamical element that Thompson’s transformations lacked. By imagining development as a ball rolling downhill across a rugged landscape of valleys and ridges (the “chreods”) Waddington captured the canalization of developmental trajectories (their resistance to perturbation) and the discreteness of cell-fate decisions in an intuitively compelling metaphor. The landscape remains a productive conceptual tool, and recent computational work has rendered it mathematically precise within the framework of quasi-potential theory for stochastic gene-regulatory networks (Li and Wang, 2013; Bhattacharya et al., 2011). Nevertheless, Waddington’s landscape, as standardly formulated, applies to a fixed organismal context; it does not readily accommodate the deformation of the landscape itself under evolutionary pressure, nor does it provide a language for the hierarchical organisation of developmental processes across biological scales.

René Thom’s catastrophe theory (Thom, 1972) attempted to supply precisely such a language by classifying the generic types of qualitative change (bifurcations) that a smooth potential function can undergo as its parameters are varied. Thom proposed that developmental transitions (gastrulation, neurulation, organogenesis) correspond to elementary catastrophes in the space of morphogenetic potentials. The programme was mathematically sophisticated and biologically provocative, but it stalled partly because its predictions were qualitative rather than quantitative and partly because it was formulated before the molecular details of developmental gene networks were available. Stuart Kauffman’s work on random Boolean networks and NK fitness landscapes (Kauffman, 1969, 1993) complemented Thom’s approach from the complexity-theory direction, demonstrating that gene networks at the boundary between order and chaos exhibit robust, evolvable behaviour, an early articulation of what would later be called the edge-of-chaos hypothesis. Nevertheless, Kauffman’s landscapes, like Waddington’s, lack a full algebraic structure connecting the microscopic (genetic) level to the macroscopic (morphological) level across multiple organisational scales.

What the existing geometric approaches collectively lack is a unified algebraic structure possessing two properties simultaneously: first, the capacity to encode the internal logic of development (the composition of developmental operations across hierarchical biological levels) in a manner that is mathematically tractable and structurally transparent; and second, the capacity to represent the external pressure of selection as a deformation of the geometric structure (the metric) on the relevant state spaces. The framework proposed in this paper is designed to supply precisely these two properties, by combining fibre-bundle differential geometry, category-theoretic operator composition, renormalization-group coarse-graining, and information geometry within a coherent theoretical whole.

Scope and Contributions

This paper makes four principal contributions to theoretical biology and the mathematical sciences of cognition and evolution. The first is a fibre-bundle formulation of developmental state space: we model the space of organismal states as a smooth manifold equipped with a Riemannian metric, and embed it as the total space of a fibre bundle over a base space of environmental and evolutionary contexts. This provides a principled geometric setting for Waddington’s landscape, Wolpert’s positional information, and Turing’s reaction-diffusion morphogenesis simultaneously. The second contribution is a category-theoretic operator-stack architecture that captures hierarchical biological and cognitive transformations: from genome to phenotype, and from sensory stimulation to abstract cognition, as compositions of morphisms in the category of developmental stages. Evo-devo mechanisms including heterochrony, heterotopy, and heterometry receive precise algebraic formulations within this architecture. The third contribution is a formal analogy between renormalization group (RG) flow in quantum field theory and coarse-graining in developmental biology, which we develop in sufficient detail to yield the concepts of developmental universality classes, developmental fixed points, and relevant versus irrelevant developmental operators, each with concrete biological interpretations. The fourth contribution is a geometric unification of ontogeny and phylogeny via shared attractor geometry on a product evolutionary manifold, which resolves the recapitulation debate and subsumes the major evo-devo theoretical frameworks within a common mathematical language.

The paper is organised as follows. Section 2 establishes the mathematical preliminaries: manifold theory, fibre bundles, information geometry, and category theory. Section 3 develops the geometric theory of biological ontogeny, formalising morphogenetic fields, attractors, gene-regulatory networks, and developmental phase transitions. Section 4 extends the framework to cognitive ontogeny, modelling Piagetian stage transitions and neural criticality geometrically. Section 5 presents the operator-stack architecture in full generality. Section 6 develops the RG-flow analogy for development in detail. Section 7 constructs the evolutionary manifold and analyses phylogenetic trajectories geometrically. Section 8 integrates all three strands: developmental, cognitive, and evolutionary geometry, into a unified framework and resolves the recapitulation debate. Section 9 discusses strengths, limitations, testable predictions, and relations to adjacent frameworks. Section 10 concludes.

The mathematical tools required for the framework are introduced in the following section, with an emphasis on physical intuition alongside formal precision.

2. Ontogenetic Geometry: Mathematical Preliminaries

Manifolds, Fibre Bundles, and Developmental State Spaces

We begin by establishing the geometric setting that will underpin all subsequent analysis. Let M be a smooth (C^∞) manifold of dimension n, representing the organismal state space. Each point p ∈ M encodes a complete phenotypic and morphogenetic configuration of the organism at a given instant: the vector of gene-expression levels, the spatial distribution of morphogen concentrations, the topological organisation of tissues, and any other continuous descriptor of organismal state. The dimension n is, in principle, enormous (on the order of the number of biologically relevant molecular species) but the effective dimensionality of dynamics is typically much lower, owing to the strong coupling structure of gene-regulatory networks (GRNs) and the low-dimensional attractor geometry they support. We return to this point when discussing the renormalization group.

The organismal state space M does not exist in isolation: development unfolds within an environmental and evolutionary context that varies across taxa and ecological settings. To formalise this dependence, we introduce a fibre bundle structure. Let B denote the base manifold, whose points b ∈ B parametrise environmental and evolutionary contexts: for example, the temperature, nutritional environment, population-genetic parameters (effective population size, mutation rate), and clade-level developmental background. We define the total space E and the projection map π: E → B, such that for each context b ∈ B, the fibre F_b = π⁻¹(b) is a copy of the developmental state manifold appropriate to that context, the space of all possible developmental trajectories available to organisms in context b. The collection (E, B, π, F) constitutes the developmental fibre bundle.

π : E → B,    F_b = π⁻¹(b) ≅ M    ∀ b ∈ B (1)

A developmental trajectory is then a smooth curve γ: [0,T] → M within a fibre, generated as an integral curve of a smooth vector field V on M. The vector field V = Vⁱ(p, t) ∂/∂xⁱ encodes the instantaneous rate of change of organismal state, and its qualitative geometry (the arrangement of fixed points, limit cycles, and heteroclinic connections) determines the repertoire of developmental outcomes available to the organism. This vector field arises, in the biological setting, from the differential equations governing gene-regulatory dynamics, morphogen diffusion, and cell-mechanical interactions.

Riemannian Metric and Information Geometry

A smooth manifold alone, while providing a topological setting for developmental dynamics, does not carry a notion of distance. To compare developmental trajectories, measure the cost of phenotypic change, and define geodesics, we must equip M with a Riemannian metric. We therefore introduce a symmetric, positive-definite bilinear form g_ij(p) on the tangent space T_pM at each point, smoothly varying over M. The pair (M, g) is a Riemannian manifold, and the geodesic distance between two points measures the minimal developmental cost of transitioning between the corresponding phenotypic configurations.

A natural and principled choice for the metric g is provided by information geometry (Amari and Nagaoka, 2000). If we represent the developmental state at each point as a probability distribution over molecular configurations, as is natural given the stochastic nature of gene expression, then the Fisher–Rao metric furnishes a canonical Riemannian metric on the space of such distributions:

g_ij(θ) = 𝔼_p(x|θ) [∂_i log p(x|θ) · ∂_j log p(x|θ)] (2)

where θ are the parameters of the developmental probability distribution p(x|θ). The Fisher–Rao metric is the unique Riemannian metric on statistical manifolds that is invariant under sufficient statistics transformations, a property that makes it intrinsically coordinate-independent and therefore biologically natural. Geodesics on (M, g) then correspond to minimal-cost developmental paths: the developmental trajectories that minimise the total phenotypic change required to move between two states while remaining dynamically feasible. Waddington’s canalization; the tendency of development to follow stereotyped trajectories that are robust to perturbation, corresponds, in this geometric language, to geodesic stability: the perturbed trajectory returns to the geodesic because the latter is a local minimum of the path-length functional in the metric g.

Category Theory and Natural Transformations

The fibre-bundle and Riemannian frameworks capture the geometry of individual developmental trajectories, but they do not immediately provide a language for the compositional, hierarchical structure of developmental processes, the fact that development proceeds through a structured sequence of regulatory decisions, each depending on the outcomes of previous ones. Category theory provides the appropriate language for this compositional structure (Mac Lane, 1998).

We define the category 𝒟ev whose objects are developmental stages, equivalently, submanifolds M_t ⊂ M representing the set of organismal states reachable at developmental time t, and whose morphisms are developmental maps f: M_t → M_t’ for t < t’, encoding the causal transitions between stages. Composition of morphisms in 𝒟ev corresponds to the sequential composition of developmental processes; the identity morphisms correspond to the trivial, time-frozen developmental state. Regulatory programs (coordinated sets of gene-regulatory interactions that drive a specific developmental transition) are formalised as functors F: 𝒟ev → 𝒟ev, mapping the category of developmental stages to itself and preserving the compositional structure of developmental sequences.

Crucially, evolutionary modifications to regulatory programs (the primary currency of evo-devo) are captured by natural transformations. A natural transformation η: F ⟹ G between two functors (two regulatory programs) specifies, for each developmental stage M_t, a morphism η_t: F(M_t) → G(M_t) that commutes with all developmental maps, that is, a systematic, stage-by-stage modification of the developmental program that is coherent across all stages simultaneously. Heterochrony (timing changes), heterotopy (spatial changes), and heterometry (magnitude changes) are all species of natural transformation in this sense, as we shall formalise in Section 5.

With the geometric and algebraic foundations in place, we turn in the next section to the detailed analysis of biological ontogeny as geometric flow on the developmental manifold.

3. Ontogenetic Geometry: Biological Ontogeny as Geometric Flow

The Morphogenetic Field as a Vector Field

The concept of a morphogenetic field, a spatial domain within the developing embryo within which cells interact to produce a coordinated pattern, has been central to developmental biology since the pioneering experimental work of Spemann and Mangold in the 1920s. The concept received its canonical theoretical elaboration in two complementary frameworks: Alan Turing’s reaction-diffusion theory of morphogenesis (Turing, 1952), which showed that simple two-component systems of diffusing and reacting chemical species can spontaneously break spatial symmetry to generate periodic patterns, and Lewis Wolpert’s positional information theory (Wolpert, 1969), which proposed that cells interpret graded concentration profiles of morphogens as positional coordinates and respond by executing specific differentiation programs.

Within the geometric framework, both of these classical theories emerge as special cases of the general structure. We define the morphogenetic submanifold M_morph ⊂ M as the subspace of organismal state space coordinatised by morphogen concentration fields and their spatial gradients. The Turing reaction-diffusion system generates a vector field V_morph on M_morph:

∂_tu = f(u, v) + D_u∇²u,    ∂_tv = g(u, v) + D_v∇²v (3)

where u(x, t) and v(x, t) are morphogen concentration fields, f and g are reaction kinetics, and D_u, D_v are diffusion coefficients. The right-hand side of equation (3), viewed on the state space of (u, v) profiles in the appropriate function space, defines precisely the vector field V_morph. The celebrated Turing instability, the spontaneous emergence of spatial patterns from a homogeneous base state, corresponds to a bifurcation of V_morph: a fixed point (uniform state) loses stability and a spatially periodic solution (limit cycle in position space) emerges. Wolpert’s positional information corresponds to a coordinate chart on M_morph: the graded morphogen concentration profile u(x) defines a smooth map from the tissue domain to ℝ that assigns a unique positional value to each cellular location, functioning as a local coordinate in the morphogenetic state manifold.

Recent experimental validation of Turing-type patterning in mammalian limb development (Raspopovic et al., 2014) confirms that this geometric formalisation captures real biological dynamics. The spatial arrangement of digit primordia in the developing limb corresponds precisely to the attractor geometry of V_morph in the relevant parameter regime.

Attractors, Canalization, and Robustness

The most consequential structural feature of developmental dynamics, from both an empirical and a theoretical perspective, is the presence of robust attractors: regions of the developmental state space toward which trajectories converge, irrespective of small perturbations to initial conditions. Waddington’s epigenetic landscape gave vivid metaphorical expression to this property, but the underlying mathematics is that of Lyapunov stability theory for dynamical systems on manifolds.

Formally, we define a developmental attractor as a compact, positively invariant set Ω_α ⊂ M such that every trajectory γ(t) starting within an open neighbourhood U(Ω_α) satisfies lim_t→∞ dist(γ(t), Ω_α) = 0, where the distance is measured in the Riemannian metric g. The basin of attraction B(Ω_α) is the maximal such neighbourhood. Different attractors Ω_α, Ω_β, … correspond to different cell fates, tissue types, or final organismal morphologies, and the partition of M into attraction basins captures the discreteness of developmental outcomes. We define canalization formally as follows: a developmental trajectory γ: [0,T] → M is k-canalized if it remains within the ε-neighbourhood of an attractor basin under k independent perturbations each of magnitude δ, for some ε = ε(k, δ). The Lyapunov stability of the attractor controls the degree of canalization.

The gradient-flow interpretation of Waddington’s landscape corresponds to the special case in which V derives from a scalar potential V(φ), a quasi-potential in the language of stochastic dynamics, via the relation:

∂_tφ = −∇V(φ) (4)

where ∇ denotes the Riemannian gradient in the metric g. In this case the developmental attractors are precisely the local minima of V(φ), the basins of attraction are the domains of descent to each minimum, and the “ridges” of the landscape are the separatrices between basins. This gradient-flow picture is exact for Waddington’s metaphor and approximately correct for many GRN dynamics, though generic developmental vector fields are not precisely gradient flows, they may contain limit cycles and heteroclinic orbits that have no potential-function description (a fact with important implications for developmental plasticity, as discussed in Section 3.3).

Gene Regulatory Networks as Connection Forms

The role of gene-regulatory networks in development is to coordinate and transmit regulatory information across the developmental state space, to ensure that the developmental trajectory in one part of the embryo is appropriately coupled to trajectories in other parts, and that the developmental history of a cell lineage is preserved and communicated to descendant cells. In the geometric framework, this coordinating and transmitting function is most naturally formalised as a connection on the developmental fibre bundle.

We model the GRN as a connection 1-form A on the fibre bundle (E, B, π, F). In local coordinates, A = A_μ^a(p) dx^μ T_a, where T_a are generators of the structure group (the group of developmental symmetries, capturing the invariances of the GRN under gene-regulatory substitutions). The curvature 2-form of the connection encodes epistatic interactions:

F = dA + A ∧ A (5)

The two terms on the right-hand side of equation (5) represent distinct biological contributions. The term dA captures the linear (additive) component of gene-gene interactions, the contribution of individual regulatory edges to the overall network phenotype. The quadratic term A ∧ A captures epistatic interactions: the non-linear, synergistic or antagonistic coupling between pairs of regulatory genes that cannot be decomposed into additive contributions. When F = 0 (flat connection), the GRN exerts perfectly buffered, additive effects: each gene contributes independently, and the developmental trajectory is insensitive to the order in which regulatory decisions are made. Non-zero curvature, conversely, implies that regulatory information is path-dependent, that the phenotypic effect of a gene depends critically on the developmental context established by prior regulatory events. This is precisely the biological phenomenon of context-dependence or developmental epistasis.

Parallel transport along environmental parameter paths in B captures developmental buffering: as environmental parameters vary along a path in the base space, the connection A determines how the developmental trajectory in the fibre is transported; if the connection is flat, the trajectory is transported without distortion (complete buffering); if the curvature is non-zero, the trajectory acquires a holonomy angle proportional to the enclosed curvature. The holonomy of the GRN connection around closed loops in B, the accumulated rotational mismatch upon returning to the starting context, corresponds precisely to developmental plasticity and polyphenism: the ability of a single genotype to produce qualitatively different phenotypes in different environmental contexts (West-Eberhard, 2003). The holonomy group of the connection characterises the full range of environmentally induced phenotypic variation accessible to the organism, its reaction norm, expressed geometrically.

Phase Transitions in Development: Gastrulation and Neurulation as Geometric Events

The major transitions of embryonic development: fertilisation, cleavage, gastrulation, neurulation, organogenesis, are among the most dramatic events in all of biology, involving wholesale reorganisation of tissue architecture and gene-expression programs within remarkably short developmental windows. The geometric framework provides a unified language for these transitions as topological and dynamical events in the developmental manifold.

Gastrulation, in which the single-layered blastula is reorganised into a trilaminar structure (ectoderm, mesoderm, endoderm), corresponds to a bifurcation of the vector field V near a saddle-node point in M. Prior to gastrulation, the developmental trajectory is confined to a single attractor basin (the pluripotent epiblast state). The onset of gastrulation corresponds to the saddle-node annihilation, the collision and disappearance of an attractor-repeller pair, causing the developmental trajectory to be expelled from the original basin and captured by one of the newly accessible germ-layer attractor basins. The speed of gastrulation (its compressed temporal window) reflects the sharpness of the bifurcation: near the saddle-node, the vector field is locally approximately quadratic, and the transit time through the saddle is proportional to (λ₊)^−1/2, where λ₊ is the positive eigenvalue at the saddle point.

Neurulation, the folding of the neural plate to form the neural tube, involves a qualitative change in the topology of the developmental manifold M_morph itself, not merely a change in the vector field on a fixed manifold. Specifically, neurulation corresponds to a handle attachment: a topological surgery operation in which a 1-handle is attached to the existing manifold, increasing its genus by one and thereby opening up a new class of topologically inequivalent trajectories (the neural lineage trajectories). This topological enrichment of M_morph enables the existence of new attractor basins (the neural cell-fate attractors) that were simply not representable in the pre-neurulation manifold. Table 1 below summarises the major developmental transitions and their geometric interpretations.

Table 1. Major developmental transitions mapped to geometric events in the developmental state manifold M. Bifurcation types follow the classification of Strogatz (1994). Attractor changes are described in terms of basin-of-attraction topology.

Having established the geometry of biological ontogeny, we now extend the framework to the domain of cognitive development, where an analogous geometric structure governs the emergence of higher mental functions.

4. Ontogenetic Geometry: Cognitive Ontogeny

Neural State Space and Synaptic Weight Geometry

The geometry of cognitive development is, at the most basic level, the geometry of neural circuit dynamics and their modification through experience-dependent plasticity. We define the cognitive state space C as the manifold of neural activation patterns, in which each point represents a specific pattern of activity across all neurons in the nervous system. The tangent space at each point of C is spanned by the directions of possible instantaneous changes in neural activity; the global topology of C encodes the repertoire of cognitive states accessible to the organism.

The geometry of C: its metric, curvature, and attractor structure, is determined by the synaptic weight tensor W_ij, which specifies the strength and sign of synaptic connections between neurons i and j. As development proceeds and synaptic weights are modified by experience, Hebbian learning, and activity-dependent pruning, the geometry of C changes: attractor basins deepen or shift, new basins form, and previously accessible states become inaccessible. The natural gradient framework of Amari (1998) provides the appropriate Riemannian metric on the space of neural network parameters, namely the Fisher–Rao metric applied to the distribution of network outputs:

∂_tW = −g_F⁻¹ ∇_W L(W) (6)

where L(W) is the loss functional (a measure of behavioural or predictive error) and g_F⁻¹ is the inverse Fisher information metric. This natural gradient learning rule defines a geodesic flow on the Riemannian manifold of synaptic weights: the weight update follows the steepest descent path in the information-geometric sense, efficiently tracking the most informative direction of improvement. Biological learning, in this formulation, is literally a geodesic flow on the cognitive state manifold.

Cognitive Development as Trajectory on a Hierarchical Manifold

The most influential theory of cognitive development, Jean Piaget’s stage theory, proposes that children pass through four qualitatively distinct stages of cognitive organization: sensorimotor, preoperational, concrete operational, and formal operational, each characterised by a coherent and internally consistent cognitive structure that is qualitatively different from, and in some sense more powerful than, the preceding one. Within the geometric framework, we model Piagetian stage transitions as attractor transitions on the cognitive manifold C.

Each Piagetian stage corresponds to a stable attractor basin Ω_stage ⊂ C: a region of cognitive state space toward which the child’s cognitive dynamics converge and within which they remain stably for an extended developmental period. The cognitive dynamics within a given stage, the repertoire of cognitive operations and their compositions, are the limit-cycle and fixed-point structures within the corresponding attractor basin. A stage transition corresponds to a heteroclinic orbit in C: a trajectory that departs from one attractor basin, passes through a high-dimensional saddle region of C, and arrives at the next attractor basin. The fact that stage transitions are relatively abrupt (on developmental timescales) and followed by consolidation and exploration reflects the geometric fact that heteroclinic orbits are rapid passages through saddle regions, followed by convergence to the new attractor.

Vygotsky’s concept of the Zone of Proximal Development, the set of cognitive tasks achievable with external scaffolding but not yet independently, receives a precise geometric interpretation as a metric deformation. Scaffolding corresponds to a time-varying modification g(t) of the cognitive metric that reduces the geodesic distance between the child’s current attractor basin and the next: by temporarily flattening the inter-basin saddle and thereby reducing the cost of the heteroclinic passage, scaffolding makes the transition geometrically accessible, the ZPD is the set of cognitive states reachable under the deformed metric g(t) but not under the undeformed metric g(0). As the child internalises the scaffolded operation, the deformation becomes permanent: the new basin is incorporated into the cognitive manifold, and the scaffolding becomes unnecessary.

Criticality and the Edge-of-Chaos Hypothesis

A striking convergence between theoretical biology (Kauffman, 1993) and systems neuroscience (Beggs and Plenz, 2003) has established that information-processing systems operating near a critical phase transition (the boundary between ordered and chaotic dynamics) exhibit maximal computational capacity, dynamic range, and information transmission. In Kauffman’s NK Boolean networks, criticality occurs at a specific average connectivity K = 2, where the system is poised between a frozen (ordered) phase and a chaotic phase. In cortical networks, Beggs and Plenz demonstrated that spontaneous neural activity organises into cascades (neural avalanches) whose size distributions follow power laws, with exponents characteristic of a critical branching process, a signature of proximity to a phase transition.

Within the geometric framework, the critical point in the space of dynamical systems corresponds to a degenerate Riemannian metric. At a phase boundary in C, the metric tensor g develops a zero eigenvalue (det g → 0) reflecting the appearance of a flat (zero-curvature) direction in the cognitive state space. Physically, this degenerate direction corresponds to the anomalously long correlations that characterise critical systems: perturbations in the flat direction propagate without decay, enabling system-wide information integration. The power-law distributions of neural avalanches are the empirical signature of this metric degeneracy.

We propose the following proposition: cognitive development proceeds by tracking the critical manifold C_crit ⊂ C, the submanifold of degenerate metric, at each developmental stage. This critical tracking maximises the information-processing capacity of the neural system at each stage of development, ensuring that cognitive development is not merely a sequence of attractor transitions but a sequence of transitions that systematically approach and exploit the maximum-capacity regime. This proposal connects the geometric framework to the empirical literature on brain development, which shows progressive increases in the degree of scale-free, critical-regime activity through childhood and adolescence (Fransson et al., 2007).

Interplay of Biological and Cognitive Ontogeny

The biological developmental manifold M and the cognitive state manifold C are not independent structures: the geometry of C is determined by the biological organisation of the nervous system, which is itself a product of the biological developmental process described in Section 3. We formalise this dependence using the fibre-bundle structure: the cognitive manifold C is the fibre of the developmental bundle above the neural developmental context b_neural ∈ B, i.e., C = π⁻¹(b_neural). Changes in the biological developmental state: myelination of axonal tracts, synaptic pruning of overproduced connections, maturation of prefrontal circuits, correspond to movements in the base space B, which deform the fibre metric g|_C and thereby alter the attractor geometry of cognitive dynamics.

This coupling provides a precise geometric model of the well-established relationship between brain maturation and cognitive stage transitions. The protracted myelination of long-range cortico-cortical connections in human development, a biological process extending from birth to the third decade of life (Yakovlev and Lecours, 1967), corresponds, in the geometric framework, to a progressive reduction in the geodesic distance between the pre-frontal and posterior cortical attractor basins, enabling increasingly efficient integration of perceptual and executive cognitive operations. The formal-operational stage, in Piaget’s scheme, becomes geometrically accessible precisely when myelination has reduced the inter-basin geodesic cost below the threshold of cognitive attainability, a quantitative prediction amenable to empirical test.

The preceding analysis of biological and cognitive ontogeny as geometric flows motivates the need for a formal algebraic architecture capable of encoding the hierarchical structure of these flows across multiple organisational levels; this is the operator-stack architecture developed in the following section.

5. Ontogenetic Geometry: The Operator-Stack Architecture

Motivation: Hierarchical Biological Organisation

Biological systems are distinguished from most physical systems by the depth and regularity of their hierarchical organisation. From the genome to the phenotype, the causal structure of development passes through at least seven nested levels: the gene (nucleotide sequence), the transcript (mRNA), the protein (polypeptide), the cellular state (gene-expression profile and signalling), the tissue (collective cell-fate), the organ (functional architecture), and the organism (integrated physiology and morphology). Each level has its own characteristic dynamics, its own state space, and its own causal relationships, yet the levels are tightly coupled: the state at each level constrains and shapes the dynamics at the next. Population and evolutionary processes add further levels above the organism. A flat, single-level description of any of these processes necessarily loses the causal information encoded in the inter-level coupling; a fully reductionist, purely molecular description, while in principle complete, is computationally and conceptually intractable for understanding macroscopic developmental and evolutionary outcomes.

The operator-stack architecture is designed to resolve this tension by providing a formal algebraic structure that explicitly encodes the hierarchical causal relationships between levels, without either collapsing them to a single level or requiring exhaustive specification of every molecular detail. The architecture abstracts each inter-level transition as a morphism in the category 𝒟ev, and formalises the full developmental process as the composition of these morphisms, a mathematical structure that is simultaneously complete (it encodes all inter-level causal relationships) and modular (each level can be modified, constrained, or coarse-grained independently).

Formal Definition of the Operator Stack

We define the operator stack as an ordered sequence O = (O₁, O₂, …, O_n), where each O_k: M_k → M_k+1 is a morphism in the category 𝒟ev mapping the level-k state space to the level-(k+1) state space. The domains M₁, M₂, …, M_n form an ascending sequence of state spaces corresponding to successively higher organisational levels: M₁ is the genotype manifold (the space of possible GRN configurations), M₂ is the transcriptome manifold, M₃ the proteome manifold, and so on through cellular, tissue, organ, and organismal levels, up to M_n, the full phenotype manifold.

The composition rule is the standard morphism composition in 𝒟ev: the composite O_k+1 ∘ O_k: M_k → M_k+2 encodes the emergent properties arising from the transition from level k through level (k+1) to level (k+2), properties that are present at the higher level but not fully described at either of the lower levels individually. The full developmental operator is the total composition:

D = O_n ∘ O_n−1 ∘ … ∘ O₂ ∘ O₁ : M₁ → M_n (7)

The operator D in equation (7) is the formal expression of the genotype-phenotype (GP) map: it maps points in the genotype manifold M₁ (specific GRN configurations) to points in the phenotype manifold M_n (complete organismal phenotypes). The GP map is, in general, a highly nonlinear, non-injective, and dimensionally contractive mapping: many distinct genotypes map to the same phenotype (robustness), and the image D(M₁) ⊊ M_n is a proper subset of the full phenotype space (developmental constraint).

Diagram 1. Schematic representation of the operator-stack architecture. Each level M_k is a state space at a distinct organisational scale. Operators O_k map between adjacent levels; within-level dynamics are governed by the vector field V_k. The full developmental operator D is the total composition of inter-level operators.

Heterochrony and Heterotopy as Operator Modifications

A central achievement of evo-devo has been the systematic characterisation of the macroevolutionary mechanisms by which evolutionary changes in development produce morphological novelty and phylogenetic divergence. The three classical mechanisms (heterochrony, heterotopy, and heterometry) receive natural and precise algebraic formulations within the operator-stack framework.

Heterochrony, the evolutionary change in the relative timing or rate of developmental events, is formalised as a modification of the temporal parameter at which operator O_k is applied. Each operator O_k is applied at a developmental time t_k within the developmental timeline. Heterochrony modifies t_k → t_k + Δt_k, or equivalently changes the rate dt_k/dτ at which developmental time within the level-k fibre is traversed relative to global developmental time τ. This is a reparametrisation of the temporal coordinate within the fibre, a diffeomorphism of [0,T_k] to [0,T_k‘]. Classical heterochronic mechanisms; paedomorphosis (retention of juvenile features in adult descendants) and peramorphosis (addition of new adult stages beyond the ancestral endpoint), correspond, respectively, to truncation and extension of the temporal domain of O_k.

Heterotopy, the evolutionary change in the spatial location at which a developmental process occurs, corresponds to a modification of the spatial domain of the operator O_k. Formally, if O_k acts on the spatial subdomain U_k ⊂ M_k, heterotopy replaces U_k by a diffeomorphic image φ(U_k) under a diffeomorphism φ of the fibre F_b. The classical example is the ectopic expression of Pax6 in Drosophila (Halder et al., 1995), in which the eye-specifying operator is applied to a novel spatial domain (the leg imaginal disc) to produce ectopic eye structures: a dramatic heterotopic transformation captured in the operator framework as the replacement of the eye-disc domain U_eye by the leg-disc domain U_leg. Heterometry (the evolutionary change in the magnitude of a developmental operation) corresponds to rescaling of the operator O_k by a factor λ_k ∈ ℝ₊, equivalent to a conformal rescaling of the metric at level k: g_k → λ_k² g_k. Allometric scaling relationships between organ size and body size are the most familiar biological manifestations of heterometry, and the algebraic scaling parameter λ_k is directly related to the allometric exponent (Gould, 1966).

Operator Algebra and Evolutionary Constraint

The set of all valid operator compositions, subject to the biochemical, physical, and informational constraints of the developmental system, constitutes the operator algebra of development: 𝒜 = ⟨O₁, …, O_n | relations⟩. The “relations” in this presentation are the developmental constraints: the requirements of stoichiometric conservation, physical tissue mechanics, thermodynamic feasibility, and logical consistency of regulatory decisions (Maynard Smith et al., 1985). They correspond, geometrically, to constraints on which compositions of operators produce smooth, well-defined mappings between the corresponding state spaces.

Not all elements of 𝒜 are accessible to evolutionary variation: the evolutionary process is constrained by the developmental system to explore only a subalgebra 𝒜_evol ⊂ 𝒜, whose structure is determined by the phylogenetic history of the lineage. This subalgebra encodes phylogenetic inertia, the tendency of lineages to explore a restricted region of morphological space determined by their ancestral developmental program. The following proposition captures the key theoretical result: the topology of the set of reachable phenotypes Φ = D(M₁) ⊆ M_n is determined by the algebraic structure of 𝒜_evol. Specifically, connected components of Φ correspond to connected components of the representation space of 𝒜_evol; the dimension of Φ equals the dimension of the generating subalgebra; and the singularities of the GP map D at points of M₁ correspond to the zero-divisors of 𝒜_evol. The Jacobian of D encodes developmental bias, the tendency of the GP map to map genetic variation more effectively in some phenotypic directions than others, and its eigenvalue spectrum characterises the directions of preferred phenotypic variation (Maynard Smith et al., 1985; Wagner and Altenberg, 1996).

Cognitive Operator Stack

The operator-stack formalism extends naturally to the cognitive domain. We define the cognitive operator stack as O^cog = (O^cog₁, …, O^cog_m), where each O^cog_k: C_k → C_k+1 maps a lower-level representational space to a higher-level one — from sensory input spaces (C₁, encoding raw photoreceptor or mechanoreceptor activations) through intermediate representational spaces (C₂, …, C_m−1, encoding progressively more abstract feature representations) to high-level conceptual spaces (C_m, encoding abstract propositions, causal models, and meta-cognitive representations).

Each O^cog_k is a learned mapping (acquired through developmental experience and formal education) analogous to the learned weight matrix of a layer in a deep neural network. This analogy is more than superficial: deep neural networks literally implement cognitive operator stacks, with each layer implementing one operator in the sequence (LeCun, Bengio, and Hinton, 2015). The composition O^cog_m ∘ … ∘ O^cog₁ maps sensory inputs to abstract conceptual representations, just as the biological operator D maps genotypes to phenotypes. Training a deep neural network, by gradient descent on a loss functional, is, in the language of the framework, a gradient flow on the cognitive operator algebra: a systematic deformation of the operator stack in the direction of decreasing loss, analogous to natural selection acting on the developmental operator algebra in the direction of increasing fitness. We therefore propose a precise sense in which cognitive architecture recapitulates developmental architecture: both are implementations of the same abstract category-theoretic structure (the operator stack) instantiated at different levels of biological organisation.

The operator-stack architecture provides the algebraic scaffolding; what is still required is a principled account of how information at different levels of the stack relates across scales. The renormalization group, introduced in the following section, provides this account.

6. Ontogenetic Geometry: Renormalization Group Flow in Development

The RG Framework in Physics: A Brief Review

The renormalization group, introduced by Kenneth Wilson in a landmark series of papers (Wilson, 1971; Wilson and Kogut, 1974), represents one of the most profound conceptual advances in twentieth-century physics. Its central insight is that the apparent complexity of physical systems at short scales, the enormous number of microscopic degrees of freedom in a condensed-matter system or quantum field, need not be confronted directly; instead, one can systematically eliminate short-scale degrees of freedom while adjusting the parameters of the effective theory that governs the remaining long-scale degrees of freedom. The result is an effective field theory that accurately describes macroscopic behaviour without explicit reference to microscopic details.

Formally, the RG is a semigroup of transformations acting on the space of coupling constants 𝒢 of a field theory. The RG transformation R_s, parametrised by a scale factor s > 1, integrates out all degrees of freedom at length scales shorter than s times the ultraviolet cutoff, yielding a new effective action with renormalised coupling constants. The flow of couplings under iterated RG transformations defines a vector field on 𝒢 (the RG flow) whose fixed points are special theories that are exactly scale-invariant: they describe phenomena that look the same at all length scales, exhibiting power-law correlations. The classification of perturbations around a fixed point into relevant (growing under RG flow), irrelevant (shrinking), and marginal (invariant) operators is the mathematical foundation of the concept of universality: the macroscopic behaviour of a system is determined not by the specific microscopic details but by the fixed point to which the RG flow is attracted, together with the relevant perturbations that cause it to flow away from the fixed point. Systems with wildly different microscopic constitutions that flow to the same RG fixed point share the same macroscopic critical behaviour, the same universality class.

Developmental Coarse-Graining as RG Flow

The formal analogy between the RG framework and developmental biology is striking and, we argue, more than merely metaphorical. In developmental biology, there is an exact parallel between the microscopic/macroscopic distinction of physics and the molecular/morphological distinction of biology: the molecular degrees of freedom (individual mRNA concentrations, protein phosphorylation states, individual cell positions) are the “microscopic” variables, while the tissue-level and organ-level phenotypic variables are the “macroscopic” ones. The developmental process of differentiation and morphogenesis is precisely a process of integrating out fine-grained molecular variability to yield a robust, coarse-grained phenotypic outcome, exactly the function of the RG.

We define the developmental RG transformation R_s as follows: for a scale factor s > 1, R_s maps the fine-grained developmental state p ∈ M to a coarse-grained state p’ = R_s(p) ∈ M’, where M’ is a state manifold with fewer effective degrees of freedom. The coarse-graining operation consists in averaging (or marginalising) over cellular and molecular variability at spatial scales smaller than s times the cell diameter, yielding an effective description in terms of cell-population averages and tissue-level variables. The resulting effective GRN (the coarse-grained gene-regulatory interactions at the tissue level) has renormalised coupling constants that differ, in general, from the microscopic coupling constants:

g_ij^eff(s) = R_s[g_ij^micro] (8)

The flow of effective coupling constants g_ij^eff(s) as the scale parameter s is varied defines the developmental RG flow on the space of GRN parameter configurations. The semigroup property of the RG (R_s·t = R_s ∘ R_t) corresponds to the consistency of developmental descriptions at different scales: the tissue-level effective theory derived from the molecular theory by integrating out cellular variability must be consistent with the organ-level theory derived from the tissue-level theory by integrating out tissue variability. The semigroup property is therefore a formal statement of the coherence of hierarchical biological description.

Relevant, Irrelevant, and Marginal Developmental Operators

The most powerful consequence of the RG framework for developmental biology is the classification of developmental perturbations (genetic mutations, epigenetic modifications, or environmental influences) into relevant, irrelevant, and marginal operators, by analogy with the classification of physical operators around an RG fixed point. This classification determines which perturbations will have large macroscopic consequences and which will average out under developmental coarse-graining.

A developmental perturbation is relevant if it grows under the developmental RG flow: beginning as a small modification at the molecular level, it amplifies as coarse-graining proceeds, ultimately producing a large change in the coarse-grained (tissue- or organ-level) phenotype. Hox genes are the canonical example of relevant developmental operators: point mutations or homeotic transformations of Hox gene expression cause large, body-plan-level morphological changes (the transformation of entire body segments). Pax6, the master regulator of eye development, is another: ectopic Pax6 expression anywhere in the developing embryo nucleates the formation of ectopic eye-like structures, demonstrating that the eye-forming operator is relevant in virtually all developmental contexts. An irrelevant developmental perturbation is one that shrinks under the developmental RG: most synonymous coding mutations, regulatory single-nucleotide polymorphisms (SNPs) outside conserved binding sites, and molecular-level stochastic fluctuations in gene expression are irrelevant in this sense, they produce molecular-level variability that is averaged out under coarse-graining, leaving the tissue- and organ-level phenotype essentially unchanged. This is the developmental mechanism underlying genetic robustness and phenotypic buffering. Marginal operators occupy the boundary between relevance and irrelevance: small perturbations that neither grow nor shrink consistently under RG flow, but that can drift to either outcome depending on the precise developmental context. Marginal operators are associated with developmental criticality and may produce punctuated-equilibrium-like dynamics: prolonged periods of phenotypic stasis punctuated by rapid, apparently discontinuous transitions. Table 2 summarises these correspondences.

Table 2. Biological correlates of renormalization group operator types in developmental biology. RG classification follows Wilson (1971); biological examples are drawn from Carroll (2005) and Wolpert et al. (2015).

Universality and the Conservation of Body Plans

The concept of universality in the RG framework provides a natural explanation for one of the most striking phenomena in comparative developmental biology: the conservation of fundamental body plans (Bauplan) across vast phylogenetic distances, despite enormous divergence in adult morphology and molecular detail. Vertebrates: including fish, amphibians, reptiles, birds, and mammals, share a conserved set of developmental features at the phylotypic stage (also called the pharyngula), including the presence of somites, pharyngeal arches, a notochord, and a dorsal neural tube, despite the enormous morphological diversity of their adult forms (Raff, 1996). This conservation cannot be explained by recent common ancestry alone, since sufficient time has elapsed for virtually all molecular details to diverge extensively. The RG framework provides a principled explanation.

We state the following universality theorem (informal version): if two organisms with different molecular-level GRN parameters exhibit developmental RG flows that are attracted to the same fixed point in the space of GRN configurations, then their coarse-grained (tissue- and organ-level) developmental programs will be identical, up to irrelevant perturbations. The phylotypic stage corresponds to the trajectory of the developmental system in the vicinity of the RG fixed point: it is the stage at which the developmental dynamics are most nearly scale-invariant, exhibiting long-range correlations and collective behaviour that is independent of molecular details. Adult morphological divergence corresponds to the subsequent flow away from the fixed point under relevant perturbations, different relevant deformations of the same fixed-point theory produce different adult morphologies, just as different relevant perturbations of the same quantum-critical point produce different ordered phases.

This analysis gives a formal, quantitative content to the empirical observation of Raff (1996) that the phylotypic stage is the most conserved stage of vertebrate development: it is conserved because it corresponds to the RG fixed point, and the RG fixed point is an attractor of the developmental flow, it is approached by all systems in its universality class, regardless of their initial (molecular-level) conditions.

RG Flow and the Operator Stack

The operator-stack architecture of Section 5 and the RG-flow framework of the present section are not independent, they are complementary descriptions of the same hierarchical developmental process. The correspondence is as follows: the operator stack O = (O₁, …, O_n) is the discrete approximation of the continuous developmental RG flow. Each inter-level operator O_k+1 ∘ O_k corresponds to a single step of the RG transformation R_s at a characteristic scale s_k. The coupling constants of the developmental field theory, the parameters of the GRN, are encoded in the operator algebra 𝒜; the RG flow on 𝒜 determines how evolutionary innovations (modifications to the coupling constants at the molecular level) propagate through the operator stack to produce macroscopic phenotypic changes. An innovation that modifies a coupling constant classified as relevant at the molecular level will propagate upward through all levels of the operator stack, producing a large phenotypic change at the organismal level. An irrelevant modification will be attenuated at each level of the stack, vanishing before reaching the organismal level.

O_k+1 ∘ O_k ≈ R_sk : M_k → M_k+2 (9)

Equation (9) makes precise the relationship between the algebraic and the analytical descriptions of hierarchical developmental organisation: operator composition corresponds to a single step of the RG flow, and the full developmental operator D = O_n ∘ … ∘ O₁ corresponds to the full RG flow from the ultraviolet (molecular) to the infrared (organismal) fixed point. The universality of developmental fixed points in the RG framework translates into the robustness of body plans across lineages in the operator-stack framework: lineages that differ in the fine-grained structure of their operator stacks (their specific GRN parameters) but that share the same large-scale operator algebra structure (the same universality class) will produce the same coarse-grained developmental outcome (the same Bauplan).

RG Flow in Cognitive Development

The RG framework applies with equal naturalness to cognitive ontogeny. In the cognitive domain, the analogue of the ultraviolet (short-scale) degrees of freedom are the fine-grained sensory representations: the responses of individual retinal ganglion cells, cochlear hair cells, or somatosensory mechanoreceptors to specific stimulus features. The infrared (long-scale) degrees of freedom are the abstract conceptual representations encoded in prefrontal and association cortices, the neural correlates of logical propositions, causal models, and metacognitive judgments. The developmental progression from perceptual to abstract cognition is, in this reading, a cognitive RG flow from the ultraviolet to the infrared: a systematic coarse-graining of sensory representations that preserves relevant task-related information while discarding irrelevant sensory detail.

The hierarchy of representations in the visual cortex: from oriented edge detectors in V1 through intermediate-complexity feature detectors in V2 and V4 to object-identity representations in the inferotemporal cortex, is a particularly well-documented example of this cognitive RG flow. Each successive cortical area implements approximately one step of the RG transformation: it pools (coarse-grains) over the representations of the preceding area, retaining information about the invariant, large-scale structure of the stimulus (object identity) while discarding information about fine-grained, position- and scale-specific details. The progressive development of these representations through childhood, with higher-order areas maturing later than lower-order ones (Johnson, 2001), reflects the sequential construction of the cognitive operator stack from the bottom up: lower-level operators must be established before higher-level operators can be defined.

The criticality hypothesis of Section 4.3 connects directly to this RG picture: the critical manifold C_crit is precisely the set of cognitive states at the RG fixed point, states in which the representation is neither over-smoothed (too irrelevant) nor over-detailed (too ultraviolet), but poised at the scale-invariant critical point that maximises information transmission between representational levels. Cognitive development proceeds by tracking this critical point as the cognitive operator stack is progressively constructed.

Having developed the RG framework for individual development and cognition, we now turn to the evolutionary manifold, where the same geometric tools apply at the population and phylogenetic timescales.

7. Ontogenetic Geometry: Evolutionary Geometry

The Evolutionary Manifold

The geometry of biological evolution has been a subject of theoretical investigation since at least Sewall Wright’s introduction of the adaptive landscape (Wright, 1932), but a fully rigorous differential-geometric formulation has remained elusive. We introduce the evolutionary manifold ℰ as the space of all viable organismal lineages, each point e ∈ ℰ corresponding to a specific lineage characterised by its GRN parameter configuration, population-genetic parameters (effective population size N_e, mutation rate μ, recombination rate ρ), and clade-level developmental background. The manifold ℰ is high-dimensional and, in general, non-compact, it has no boundary corresponding to a maximum degree of evolutionary divergence.

The natural metric on ℰ is the information-theoretic distance between lineage distributions. We define the evolutionary metric tensor g_evo at a point e ∈ ℰ as the Fisher–Rao metric on the space of GRN probability distributions: the Kullback-Leibler divergence between the GRN parameter distributions of two infinitesimally separated lineages defines the squared infinitesimal distance in g_evo. This metric has several desirable properties: it is invariant under reparametrisation of the GRN parameters (a property shared by the Fisher metric generally), it respects the statistical structure of genetic variation, and it reduces to known phylogenetic distance measures (such as the Jukes-Cantor distance) in the limit of purely neutral molecular evolution.

Fitness as a Scalar Field and Natural Selection as Gradient Flow

Natural selection, in the geometric framework, is a force acting on the evolutionary manifold ℰ: it deforms trajectories on ℰ away from the geodesics of the neutral metric g_evo, biasing the evolutionary walk in the direction of increasing fitness. We formalise this by defining fitness W: ℰ → ℝ as a smooth scalar field on the evolutionary manifold. The value W(e) at a point e ∈ ℰ is the mean fitness of the lineage corresponding to e in its ecological context, the average number of offspring per individual per generation, appropriately averaged over the stochastic variation of the environment.

The fundamental equation of directional selection, in the language of the geometric framework, is that the mean phenotype evolves by covariant gradient ascent on W:

d⟨φ⟩/dt = g_evo^ij ∂_jW (10)

Equation (10) is the covariant (metric-respecting) gradient of the fitness function, analogous to the natural gradient in information geometry. The factor g_evo^ij (the inverse metric tensor) ensures that the direction of steepest fitness ascent is correctly identified in the curved geometry of the evolutionary manifold, it is not simply the direction of maximum partial derivative ∂_jW in a flat parameter space, but the direction of maximum rate of fitness increase per unit of evolutionary distance. This is the geometric content of Lande’s (1979) equation relating the evolutionary response to selection to the genetic variance-covariance matrix: the matrix G of quantitative genetics is precisely the discretised, empirically estimated form of g_evo⁻¹.

Genetic drift, the stochastic sampling of alleles in finite populations, adds a stochastic perturbation to the gradient flow of equation (10). In the geometric framework, drift corresponds to a Wiener process on (ℰ, g_evo) with diffusion coefficient proportional to 1/(2N_e), where N_e is the effective population size. Sewall Wright’s adaptive landscape is the potential function for the deterministic component of this stochastic gradient flow on ℰ: Wright’s “peaks” correspond to local maxima of W, and “valleys” to saddle points that stochastic drift must traverse to enable peak shifting.

Phylogenetic Trajectories as Geodesics Under Selective Pressure

Under neutral evolution, in the absence of directional selection, the evolutionary dynamics on (ℰ, g_evo) reduce to a stochastic geodesic flow: the mean phylogenetic trajectory is a geodesic of the evolutionary metric, and deviations from the geodesic are due solely to stochastic drift. This is consistent with the neutral theory of molecular evolution (Kimura, 1983): under neutrality, molecular evolution proceeds at a rate determined by the mutation rate, and the accumulated sequence divergence is proportional to evolutionary time, a property that corresponds, geometrically, to constant-speed travel along a geodesic of the evolutionary manifold.

Under directional selection, the phylogenetic trajectory deviates from the geodesic by a force term proportional to the covariant gradient of W. The selection curvature (the deviation of the actual trajectory from the neutral geodesic) measures the intensity and directionality of the selective pressure driving adaptive evolution. Macroevolutionary convergence (the independent evolution of similar morphological, physiological, or behavioural characters in distantly related lineages) corresponds to geodesic convergence in ℰ: different starting points in ℰ (different ancestral genotypes) that lie within the basin of attraction of the same fitness maximum will generate phylogenetic trajectories that converge on the same region of ℰ, producing similar evolved phenotypes through different molecular routes. The geometric framework thus provides a precise, quantitative account of convergent evolution, one of the most striking phenomena in macroevolution.

Embedding Ontogeny in Evolutionary Geometry

The developmental operator D: M₁ → M_n of equation (7) defines a smooth mapping from the genotype manifold to the phenotype manifold. This mapping induces a push-forward metric on the phenotype manifold: given the evolutionary metric g_geno on the genotype space M₁, the induced phenotypic metric is:

g_phen = D_*g_geno (11)

where D_* denotes the push-forward (direct image) of the metric under D. Equation (11) has a fundamental biological interpretation: it states that phenotypic distances (the metric structure of morphological space) are determined by genotypic distances translated through the developmental map. Two genotypes that are close in the genetic metric will produce phenotypes that are close in the phenotypic metric, but the scaling is non-uniform: in regions of the genotype space where the Jacobian of D has large eigenvalues, genetic distances are amplified into phenotypic distances; where the Jacobian is small, genetic distances are compressed. This amplification and compression constitutes developmental bias, the tendency for the GP map to facilitate variation in some phenotypic directions and constrain it in others. The key result follows: the evolutionary metric g_evo on the evolutionary manifold ℰ factors through the developmental operator D, in the sense that selection acts on phenotypes (and hence on g_phen) but is translated back to genotypes via the push-forward D_*. Developmental bias therefore shapes the evolutionary response to selection: lineages evolve most rapidly in the directions of greatest developmental amplification, regardless of the direction of fitness gradient in the phenotype space. The algebraic structure of the operator algebra 𝒜_evol determines the topology of the image D(M₁) and hence the set of phenotypes accessible to evolution, the evolutionary evolvability of the lineage is precisely the topological richness of this image.

With developmental, cognitive, and evolutionary geometry individually formalised, we are in a position to integrate them into a unified framework and address the classical questions of evo-devo from a novel geometric perspective.

8. Ontogenetic Geometry: Unification: Ontogeny, Cognition, and Phylogeny

The Unified State Space

The three geometric structures developed in Sections 3 through 7: the developmental manifold (M, g), the cognitive state manifold (C, g_F), and the evolutionary manifold (ℰ, g_evo), can now be integrated into a unified state space that simultaneously encodes developmental, cognitive, and evolutionary dynamics. We define the unified state space as the product manifold:

𝒰 = M × C × ℰ (12)

equipped with a joint Riemannian metric g_𝒰 that is, to leading order, a block-diagonal combination of the three component metrics, with off-diagonal coupling terms reflecting the inter-level interactions described in Sections 4.4 and 7.4. The dynamics on 𝒰 separate into three characteristic timescales: fast dynamics on M (the developmental timescale, ranging from hours to years within a single organism’s lifetime), intermediate dynamics on C (the learning timescale, ranging from milliseconds for individual synaptic events to years for consolidated conceptual development), and slow dynamics on ℰ (the evolutionary timescale, ranging from generations to millions of years for deep phylogenetic divergence).

The most important feature of the unified state space 𝒰 is that the dynamics at the three timescales are not independent, they are coupled through the off-diagonal terms of the metric g_𝒰. Specifically, changes in ℰ (evolutionary events: mutations fixed by selection, genetic drift, developmental system drift) modify the structure of the developmental operator stack, which in turn changes the metric on M. Changes in the metric on M: for example, through myelination, synaptic pruning, or hormonal developmental signals, deform the geometry of the cognitive sub-manifold C = π⁻¹(b_neural), altering the attractor geometry of cognitive dynamics and hence the characteristic cognitive capacities of organisms at each developmental stage. These couplings give the unified framework its explanatory power: it can account simultaneously for the developmental causes of cognitive capacity, the evolutionary origins of developmental programs, and the developmental constraints on evolutionary trajectories, within a single coherent geometric formalism.

Resolution of the Recapitulation Debate

The unified geometric framework provides a definitive and mathematically rigorous resolution of the longstanding recapitulation debate. Haeckel’s biogenetic law, in its strong form, asserts that ontogenetic stages linearly retrace phylogenetic history, that the embryo passes sequentially through adult stages of ancestral species. This claim is straightforwardly falsified empirically and is geometrically incoherent: it requires that the developmental trajectory in M pass through a sequence of distinct attractor basins in a fixed order determined by phylogenetic history, which would demand an implausible degree of coordination between developmental dynamics and the historical record of evolutionary events. No mechanism capable of implementing such a coordination has ever been proposed or empirically documented.

Von Baer’s law, that embryos of related species resemble each other more at early stages than at late stages, is, by contrast, a direct consequence of the developmental RG fixed-point structure. Early developmental stages, occurring in the vicinity of the RG fixed point (the phylotypic stage), are described by the scale-invariant theory associated with the fixed point: they are independent of the specific molecular details of the lineage and therefore similar across all taxa in the same universality class. Late developmental stages, by contrast, are driven by relevant perturbations that pull the system away from the fixed point; different lineages carry different relevant perturbations and therefore diverge progressively as development proceeds. Von Baer’s law is thus the phenomenological statement of the convergence of developmental trajectories toward a shared RG fixed point, followed by divergence under lineage-specific relevant operators.

The geometric resolution of the recapitulation debate is therefore as follows: “recapitulation,” properly understood, refers not to a linear replay of adult ancestral stages but to a transient convergence of developmental trajectories toward shared attractor basins, the RG fixed points inherited from common ancestors, followed by lineage-specific divergence as those trajectories are deflected by relevant perturbations corresponding to each lineage’s distinctive evolutionary modifications. The “ancestral stages” visible in embryos are not the adult forms of ancestors but the fixed-point attractor basins that all descendants of a common ancestor share, by virtue of sharing the same universality class of developmental dynamics. This is a precise, falsifiable, and geometrically natural account, one that unifies von Baer and Haeckel by showing that von Baer’s law is the correct geometric form of the intuition that Haeckel attempted but failed to capture.

Evo-Devo Synthesis

The geometric framework provides natural formalisations of the major conceptual contributions of the evo-devo research programme (Carroll, 2005; Kirschner and Gerhart, 2005; Wagner, 2011). We briefly enumerate these correspondences to demonstrate the comprehensive scope of the unification. Modularity, the semi-independence of developmental subsystems from one another, corresponds to a tensor product decomposition of the developmental manifold: M ≅ M₁ ⊗ M₂ ⊗ … ⊗ M_r, where each M_α is a module manifold governing a semi-independent developmental subsystem (forelimb, hindlimb, brain, etc.). The degree of modularity is measured by the off-diagonal coupling terms in g_𝒰 between module manifolds: strong modularity corresponds to weak coupling, and the evolution of modularity corresponds to the progressive minimisation of these coupling terms under selection for evolvability. Evolvability (the capacity of a lineage to generate heritable phenotypic variation) is the topological richness of the image D(M₁) of the developmental operator. Lineages with high evolvability have developmental operators whose images densely sample the phenotype manifold, enabling rapid phenotypic exploration under selection; lineages with low evolvability have operators whose images are confined to low-dimensional submanifolds, limiting the accessible phenotypic variation. Developmental constraint: the limitation of accessible phenotypic variation to a proper subspace of the full phenotype manifold, is the algebraic restriction 𝒜_evol ⊂ 𝒜, and its geometric expression is the confinement of the image D(M₁) to a submanifold of M_n. Robustness and plasticity: the complementary properties of phenotypic stability under genetic perturbation and sensitivity to environmental change, are encoded in the Lyapunov stability of developmental attractors and the curvature of the GRN connection, respectively, as formalised in Sections 3.2 and 3.3.

Implications for AI and Artificial Cognitive Architectures

The cognitive operator-stack architecture of Section 5.5 establishes a precise formal connection between deep neural networks and biological developmental programs. Both are implementations of the same abstract categorical structure (a graded sequence of morphisms between hierarchically organised representation spaces) and both are shaped by a version of the same optimisation process: gradient descent on a loss functional in the case of machine learning, and natural selection on a fitness landscape in the case of biological evolution. This formal correspondence has non-trivial implications for artificial intelligence research and, in particular, for the AI alignment problem.

If cognitive architecture is fundamentally an RG-structured hierarchy, as the framework proposes, then artificial cognitive systems exhibiting robust generalisation should implement representations that are organised as successive coarse-grainings of the input space, precisely the structure that effective deep learning architectures already approximate, and that the theory of convolutional networks makes explicit (LeCun, Bengio, and Hinton, 2015). More speculatively, we conjecture that alignment failures, failures of AI systems to generalise in ways that humans find appropriate or safe, may reflect mismatches between the cognitive RG fixed points of human and artificial cognitive systems. If human cognition has been shaped by biological development to approach a specific RG fixed point (a specific scale-invariant cognitive attractor encoding human values, social knowledge, and causal world models) then artificial systems trained on human-generated data but with a different architectural prior (a different developmental operator stack) will approach a different RG fixed point, one that may diverge from the human fixed point under relevant perturbations corresponding to novel or adversarial inputs. The AI alignment problem, on this reading, is the problem of engineering artificial cognitive systems whose developmental RG flow is attracted to the same fixed point as human cognition, a geometric criterion for alignment that is, at least in principle, formally precise and empirically testable.

The preceding synthesis invites a critical assessment of the framework’s strengths and limitations, and of the empirical predictions it generates, the task taken up in the following discussion section.

9. Ontogenetic Geometry: Discussion

Strengths and Limitations of the Framework

The principal strength of the Ontogenetic Geometry framework is its conceptual scope: it provides, for the first time, a single mathematical language: differential geometry, category theory, and renormalization group theory, that simultaneously and coherently describes biological development, cognitive emergence, and evolutionary dynamics. This is not merely an aesthetic achievement; it enables cross-domain reasoning that would be impossible within any of the constituent frameworks taken separately. The identification of Waddington’s canalization with geodesic stability, of developmental phase transitions with bifurcation theory, of evo-devo mechanisms with operator modifications, and of phylotypic conservation with RG fixed points are examples of cross-domain insights that become available only within the unified geometric setting.

The framework also generates structurally natural testable predictions (detailed in Section 9.2) and connects to established mathematical physics in ways that make the existing analytical toolkit of theoretical physics available to developmental and evolutionary biology. The renormalization group, in particular, is one of the most powerful analytical frameworks in the physical sciences; bringing it to bear on developmental biology opens the possibility of new quantitative methods for analysing the scale-dependence of GRN dynamics, classifying developmental perturbations by their macroscopic consequences, and predicting the conserved structure of body plans from the fixed-point structure of developmental field equations.

The primary limitation of the framework, as currently formulated, is that it operates largely at the level of formal analogy: the fibre-bundle, RG, and operator-stack structures are proposed as the correct mathematical setting for developmental biology, but the explicit quantitative models (the specific manifolds M, metrics g, vector fields V, and connection forms A) have not yet been fully instantiated from empirical data. Doing so will require parameterisation from large-scale developmental datasets: single-cell RNA-sequencing time courses that resolve the trajectory of individual cells through the developmental state space, spatial transcriptomics data that map the morphogenetic field as a vector field over the embryo, and comparative genomics data that characterise the evolutionary metric on the space of GRN configurations. Significant recent progress in all of these experimental domains (Cao et al., 2019; Lange et al., 2022) makes this programme of empirical parameterisation increasingly feasible. A further theoretical challenge is the extension of the framework to properly account for non-equilibrium dynamics: biological development is an intrinsically non-equilibrium process, and the quasi-potential description of Waddington’s landscape is an approximation that breaks down for developmental trajectories far from steady state. A fully non-equilibrium geometric framework, incorporating the tools of stochastic thermodynamics and non-equilibrium statistical mechanics, is a natural direction for future development.

Testable Predictions

A theoretical framework of the scope proposed here earns its keep not merely by conceptual unification but by generating falsifiable predictions that distinguish it from its competitors. We identify three principal predictions, each with a specific experimental methodology by which it could be tested.

Prediction 1: Power-law scaling of morphogenetic field correlations near developmental phase transitions. The identification of gastrulation and neurulation with bifurcations in the developmental vector field, and the identification of these bifurcations with approach to the developmental RG fixed point, implies that spatial correlations in the morphogenetic field (the correlated fluctuations of morphogen concentrations and gene-expression levels across neighbouring cells) should exhibit power-law scaling at developmental phase transitions, with an exponent determined by the universality class of the transition. This prediction is testable using spatial transcriptomics data at single-cell resolution, which can measure spatial correlations in gene-expression profiles across the embryo at successive developmental stages. We predict that the correlation length diverges and the spatial correlation function follows a power law at the onset of gastrulation and neurulation.

Prediction 2: Conservation of GRN operator algebra subalgebra structure across sister clades. The identification of phylotypic conservation with RG fixed-point structure implies that sister clades (pairs of related lineages that diverged from a common ancestor) should share the same subalgebra structure 𝒜_evol at levels of the operator stack corresponding to the conserved phylotypic stage, while differing in the operator algebra structure at levels corresponding to adult-stage divergence. This prediction is testable using comparative genomics data: the regulatory network structure (the topology of transcription-factor binding interactions) at the phylotypic stage should be more conserved between sister clades than the regulatory network structure at adult stages, even after controlling for overall sequence divergence. Comparative single-cell RNA-seq data across vertebrate species at matched developmental stages (Farrell et al., 2018) provides a natural experimental system for testing this prediction.

Prediction 3: Cognitive development trajectories in human infants should exhibit signatures of RG flow. The identification of cognitive development with a cognitive RG flow from UV (perceptual) to IR (conceptual) representations implies that the neural activity patterns of human infants, as measured by EEG or MEG, should exhibit progressive development of scale-free (1/f) power spectra characteristic of approach to the cognitive RG fixed point, with the timescale of this progression correlating with the maturation of the corresponding cortical areas. Specifically, the scale-free exponent of neural activity should approach the critical value (1/f noise, exponent ≈ 1) progressively from lower to higher cortical areas across the first years of life. This prediction is testable using longitudinal EEG studies of infant brain development and can be compared with structural MRI data characterising the spatial and temporal pattern of cortical myelination.

Relationship to Other Frameworks

The Ontogenetic Geometry framework does not emerge from a vacuum: it is informed by and builds upon several existing theoretical approaches to developmental and evolutionary biology. Salazar-Ciudad and Jernvall’s (2010) dynamical systems models of tooth morphogenesis are among the most sophisticated existing attempts to connect gene-regulatory dynamics to morphological outcomes through explicit mathematical modelling; the geometric framework subsumes their approach by providing the manifold-theoretic setting within which their specific ODE models represent particular vector fields and attractor structures. Christoph Adami’s information-theoretic approach to evolution (Adami, 2002) identifies biological fitness with the information content of the genome about the environment; within the geometric framework, this corresponds to the mutual information between the genotype manifold and the evolutionary manifold, measured in the Fisher-Rao metric. The topological data analysis approach to gene expression of Carlsson (2009) and collaborators extracts topological features (connected components, loops, higher-dimensional holes) from high-dimensional gene-expression data; within the geometric framework, these topological features are naturally interpreted as features of the developmental manifold M: its Betti numbers, homology groups, and homotopy type. Control-theoretic models of development (Bhattacharya et al., 2011) analyse GRN dynamics as optimal control problems; the geometric framework provides the natural Riemannian setting in which to formulate these control problems as geodesic pursuit problems on (M, g). The present framework therefore does not invalidate these existing approaches but provides a common geometric language within which each can be situated, compared, and generalised.

10. Ontogenetic Geometry: Conclusions

Conclusions

We have presented Ontogenetic Geometry as a unified mathematical framework integrating biological development, cognitive emergence, and evolutionary dynamics within a common geometric substrate. The four principal contributions of the paper are as follows. First, we have shown that biological and cognitive ontogeny are naturally described as flows on fibre-bundle-structured state spaces, providing a rigorous differential-geometric setting for classical conceptual frameworks: Waddington’s epigenetic landscape, Wolpert’s positional information, and Turing’s reaction-diffusion morphogenesis, while simultaneously connecting them to the information geometry of learning and neural coding. Second, we have demonstrated that the category-theoretic operator-stack architecture provides a precise algebraic encoding of hierarchical biological and cognitive transformations, within which the major mechanisms of evo-devo (heterochrony, heterotopy, heterometry, modularity, developmental bias) receive formal, manipulable definitions. Third, we have developed the formal analogy between renormalization group flow in quantum field theory and coarse-graining in developmental biology to the point where it yields the concepts of developmental universality classes, developmental RG fixed points, and relevant, irrelevant, and marginal developmental operators with concrete biological content, most notably, the identification of the phylotypic stage with the trajectory of the developmental system near the RG fixed point and the explanation of body-plan conservation as developmental universality. Fourth, we have constructed the unified state space 𝒰 = M × C × ℰ and shown that ontogeny and phylogeny are not fundamentally distinct processes but different-timescale flows on the same unified geometric space, driven by the same operator-stack machinery but at different rates and under different optimisation pressures.

The deepest claim of the framework is perhaps this: the classical distinction between ontogeny and phylogeny (between the development of the individual and the evolution of the lineage) is not an absolute distinction at the level of the underlying mathematics, but rather a distinction of timescale and coarse-graining. Both processes are flows on the unified state space 𝒰; both are governed by operator stacks whose composition encodes hierarchical causal structure; and both are subject to the same renormalization group logic that determines which features of the microscopic structure are amplified into macroscopic consequences and which are washed out by developmental and evolutionary coarse-graining. The geometry is the same; only the time is different.

We close with a forward-looking observation. The convergence of developmental biology, theoretical neuroscience, and artificial intelligence research, three fields that might, a decade ago, have seemed separated by insurmountable disciplinary distances, is one of the most significant intellectual developments of our time. Developmental biology provides the mechanistic foundations for the biological architecture of cognition; theoretical neuroscience characterises the computational properties of that architecture; and AI research both builds artificial instantiations of operator-stack architectures and provides theoretical tools (gradient learning, representation theory, information geometry) that illuminate the biological originals. The geometric language proposed in this paper is, we believe, well-suited to serve as the common theoretical medium within which these three fields can engage with maximum precision and mutual benefit. A unified geometric programme for the life sciences, one that treats development, cognition, and evolution as facets of a single geometric reality, is within reach, and the present contribution is intended as a step toward its realisation.

1  The term “ontogenetic geometry” is intended to recall D’Arcy Thompson’s programme while distinguishing the present framework from it by its emphasis on dynamics (flows) rather than statics (transformations) and by its provision of an algebraic structure.

2  We use the term “operator” in the sense of category theory and functional analysis throughout, not in the narrower sense of quantum mechanics. A developmental operator is any morphism in the category 𝒟ev; it need not be linear.

3  The analogy between RG flow and developmental coarse-graining has been noted informally by several authors; see Goldenfeld and Woese (2011) for a discussion of universal laws in biology from the RG perspective. The present paper is, to our knowledge, the first to develop this analogy systematically in the context of developmental biology and to connect it to the operator-stack architecture.

4  The universality theorem stated in Section 6.4 is informal: a rigorous proof would require specification of the precise regularity conditions on the GRN dynamics and the RG transformation, which lie beyond the scope of this paper. We regard the theorem as a structurally well-motivated conjecture awaiting rigorous mathematical treatment.

5  The AI alignment implication outlined in Section 8.4 is speculative and is offered as a research direction rather than a concluded result. The empirical content of the claim (that human cognitive RG fixed points can be characterised and compared with those of artificial systems) depends on experimental programs that do not yet exist.

References

[1]  Adami, C. (2002). What is complexity? BioEssays, 24(12), 1085–1094. https://doi.org/10.1002/bies.10192

[2]  Alberch, P., Gould, S. J., Oster, G. F., and Wake, D. B. (1979). Size and shape in ontogeny and phylogeny. Paleobiology, 5(3), 296–317.

[3]  Amari, S. (1998). Natural gradient works efficiently in learning. Neural Computation, 10(2), 251–276.

[4]  Amari, S., and Nagaoka, H. (2000). Methods of Information Geometry. American Mathematical Society / Oxford University Press.

[5]  Beggs, J. M., and Plenz, D. (2003). Neuronal avalanches in neocortical circuits. Journal of Neuroscience, 23(35), 11167–11177.

[6]  Bhattacharya, S., Zhang, Q., and Bhattacharya, M. (2011). A deterministic map of Waddington’s epigenetic landscape for cell fate specification. BMC Systems Biology, 5(1), 85.

[7]  Cao, J., Spielmann, M., Qiu, X., Huang, X., Ibrahim, D. M., Hill, A. J., et al. (2019). The single-cell transcriptional landscape of mammalian organogenesis. Nature, 566(7745), 496–502.

[8]  Carlsson, G. (2009). Topology and data. Bulletin of the American Mathematical Society, 46(2), 255–308.

[9]  Carroll, S. B. (2005). Endless Forms Most Beautiful: The New Science of Evo Devo. W. W. Norton.

[10]  D’Arcy Thompson, W. (1917). On Growth and Form. Cambridge University Press.

[11]  Farrell, J. A., Wang, Y., Riesenfeld, S. J., Shekhar, K., Regev, A., and Schier, A. F. (2018). Single-cell reconstruction of developmental trajectories during zebrafish embryogenesis. Science, 360(6392), eaar3131.

[12]  Fransson, P., Åden, U., Blennow, M., and Lagercrantz, H. (2007). The functional architecture of the infant brain as revealed by resting-state fMRI. Cerebral Cortex, 21(1), 145–154.

[13]  Goldenfeld, N., and Woese, C. (2011). Life is physics: evolution as a collective phenomenon far from equilibrium. Annual Review of Condensed Matter Physics, 2, 375–399.

[14]  Gould, S. J. (1966). Allometry and size in ontogeny and phylogeny. Biological Reviews, 41(4), 587–638.

[15]  Gurdon, J. B. (1962). The developmental capacity of nuclei taken from intestinal epithelium cells of feeding tadpoles. Journal of Embryology and Experimental Morphology, 10, 622–640.

[16]  Haeckel, E. (1866). Generelle Morphologie der Organismen. Georg Reimer Verlag.

[17]  Halder, G., Callaerts, P., and Gehring, W. J. (1995). Induction of ectopic eyes by targeted expression of the eyeless gene in Drosophila. Science, 267(5205), 1788–1792.

[18]  Johnson, M. H. (2001). Functional brain development in humans. Nature Reviews Neuroscience, 2(7), 475–483.

[19]  Kauffman, S. A. (1969). Metabolic stability and epigenesis in randomly constructed genetic nets. Journal of Theoretical Biology, 22(3), 437–467.

[20]  Kauffman, S. A. (1993). The Origins of Order: Self-Organization and Selection in Evolution. Oxford University Press.

[21]  Kimura, M. (1983). The Neutral Theory of Molecular Evolution. Cambridge University Press.

[22]  Kirschner, M., and Gerhart, J. (1998). Evolvability. Proceedings of the National Academy of Sciences USA, 95(15), 8420–8427.

[23]  Kirschner, M. W., and Gerhart, J. C. (2005). The Plausibility of Life: Resolving Darwin’s Dilemma. Yale University Press.

[24]  Lande, R. (1979). Quantitative genetic analysis of multivariate evolution, applied to brain:body size allometry. Evolution, 33(1), 402–416.

[25]  Lange, M., Bergen, V., Klein, M., Setty, M., Reuter, B., Bakhti, M., et al. (2022). CellRank for directed single-cell fate mapping. Nature Methods, 19, 159–170.

[26]  LeCun, Y., Bengio, Y., and Hinton, G. (2015). Deep learning. Nature, 521(7553), 436–444.

[27]  Li, C., and Wang, J. (2013). Quantifying Waddington landscapes and paths of non-adiabatic cell fate decisions for differentiation, reprogramming and transdifferentiation. Journal of The Royal Society Interface, 10(89), 20130787.

[28]  Mac Lane, S. (1998). Categories for the Working Mathematician (2nd ed.). Springer.

[29]  Maynard Smith, J., Burian, R., Kauffman, S., Alberch, P., Campbell, J., Goodwin, B., et al. (1985). Developmental constraints and evolution. Quarterly Review of Biology, 60(3), 265–287.

[30]  Raff, R. A. (1996). The Shape of Life: Genes, Development, and the Evolution of Animal Form. University of Chicago Press.

[31]  Raspopovic, J., Marcon, L., Russo, L., and Sharpe, J. (2014). Digit patterning is controlled by a Bmp-Sox9-Wnt Turing network modulated by morphogen gradients. Science, 345(6196), 566–570.

[32]  Richardson, M. K., Hanken, J., Gooneratne, M. L., Pieau, C., Raynaud, A., Selwood, L., and Wright, G. M. (1997). There is no highly conserved embryonic stage in the vertebrates: implications for current theories of evolution and development. Anatomy and Embryology, 196(2), 91–106.

[33]  Rutherford, S. L., and Lindquist, S. (1998). Hsp90 as a capacitor for morphological evolution. Nature, 396(6709), 336–342.

[34]  Salazar-Ciudad, I., and Jernvall, J. (2010). A computational model of teeth and the developmental origins of morphological variation. Nature, 464(7288), 583–586.

[35]  Thom, R. (1972). Stabilité Structurelle et Morphogenèse. W. A. Benjamin. [English translation: Structural Stability and Morphogenesis, 1975, Addison-Wesley.]

[36]  Turing, A. M. (1952). The chemical basis of morphogenesis. Philosophical Transactions of the Royal Society B, 237(641), 37–72.

[37]  von Baer, K. E. (1828). Über Entwicklungsgeschichte der Thiere: Beobachtung und Reflexion. Bornträger.

[38]  Wagner, G. P. (2011). Homology, Genes, and Evolutionary Innovation. Princeton University Press.

[39]  Wagner, G. P., and Altenberg, L. (1996). Perspective: complex adaptations and the evolution of evolvability. Evolution, 50(3), 967–976.

[40]  Waddington, C. H. (1957). The Strategy of the Genes. Allen and Unwin.

[41]  West-Eberhard, M. J. (2003). Developmental Plasticity and Evolution. Oxford University Press.

[42]  Wilson, K. G. (1971). Renormalization group and critical phenomena. I. Renormalization group and the Kadanoff scaling picture. Physical Review B, 4(9), 3174–3183.

[43]  Wilson, K. G., and Kogut, J. (1974). The renormalization group and the epsilon expansion. Physics Reports, 12(2), 75–199.

[44]  Wolpert, L. (1969). Positional information and the spatial pattern of cellular differentiation. Journal of Theoretical Biology, 25(1), 1–47.

[45]  Wright, S. (1932). The roles of mutation, inbreeding, crossbreeding, and selection in evolution. Proceedings of the Sixth International Congress of Genetics, 1, 356–366.

[46]  Yakovlev, P. I., and Lecours, A. R. (1967). The myelogenetic cycles of regional maturation of the brain. In A. Minkowski (Ed.), Regional Development of the Brain in Early Life (pp. 3–70). Blackwell.