Beauty as structured cognitive work — the full mathematical core

This is the scientific layer of TITI, written for the reader who wants the actual machinery, not a slogan. It states the engine's aesthetic core the way the factory publishes everything else: in full, with its formulas and with an explicit frame for what it claims and what it does not. The central claim is a reframing. Beauty is not treated as a static scalar property of an object. It is modelled as the structured cognitive work performed by an object-induced potential field on the observer's predictive model, along a trajectory on an information manifold, under bounded dissipation and phase coherence. The model is presented as a coherent theoretical framework and a diagnostic instrument — not as a proven physical law of the brain. Its purpose is operational: to turn aesthetic experience into measurable design hypotheses — cognitive cost, structured mystery, phase alignment, interpretive capacity, and the value of the trajectory itself.

Ontology: object, subject, and a path through model space

An aesthetic object — an image, a piece of music, an interface, a building, a brand system, even a mathematical formula — is never scored as beautiful in itself. Instead the object induces a field on the observer's space of internal predictive models. The subject is represented not as a fixed preference vector but as a predictive cognitive system with internal model states. A model state is a point on a manifold,

m(t)\in\mathcal{M}

and the aesthetic experience is a trajectory of that state while the observer engages with the object:

\gamma(t):[0,T]\to\mathcal{M},\qquad \gamma(0)=m_{0},\ \ \gamma(T)=m_{T}

This is the first and load-bearing move: the aesthetic value is a property of the path the observer's model travels, not of the object alone.

The information manifold and the Fisher-Rao metric

The observer's internal models are taken to be a parametric family of distributions, so the model space carries the natural geometry of statistics. The metric is the Fisher information matrix:

g_{ij}(\theta)=\mathbb{E}_{x\sim p(x\mid\theta)}\!\left[\frac{\partial\log p(x\mid\theta)}{\partial\theta_{i}}\,\frac{\partial\log p(x\mid\theta)}{\partial\theta_{j}}\right]

This metric measures how hard it is for the observer to distinguish nearby predictive states. States that are close under the Fisher-Rao metric can be moved between at low cognitive cost; states that are far apart require more work. Aesthetic distance is therefore not flat — the same visual difference can be easy for one observer and hard for another because their internal manifolds have different curvature. The distance between two cognitive states is the geodesic length:

D_{g}(m_{a},m_{b})=\inf_{\gamma}\int_{0}^{T}\!\sqrt{g_{ij}(\gamma(t))\,\dot{\gamma}^{i}(t)\,\dot{\gamma}^{j}(t)}\;dt

The object-induced aesthetic potential

The object induces a scalar potential field over model space. At each model state its value represents the unresolved free-energy tension, predictive mismatch, or interpretive potential the object creates. It can be identified with the variational free energy of the observer model relative to the object input:

\Phi_{O}:\mathcal{M}\to\mathbb{R},\qquad \Phi_{O}(m)=\mathcal{F}(O,m)

This deliberately does not collapse into the Free Energy Principle. That principle asks how an organism minimises free energy to preserve homeostasis; here the question is how an object shapes a trajectory through the free-energy landscape such that the path is cognitively valuable, phase-coherent, and not trivially resolved. The engine does not reward the lowest possible free energy — it rewards structured movement through a rich landscape. The force the object exerts on interpretation is the metric gradient of the potential:

F_{\mathrm{aest}}=-\nabla_{g}\Phi_{O},\qquad (\nabla_{g}\Phi)^{i}=g^{ij}\frac{\partial\Phi}{\partial q^{j}}

A boring object has a shallow, trivial field; a chaotic object an incoherent one; a masterpiece a deep, structured, multi-attractor field.

Structured mystery, distinguished from noise

Surprise decays — a familiar masterpiece would lose all value if beauty were only surprise. So the model needs a static but structured component. Because interpretive layers are statistically dependent, a naive weighted sum would double-count; the corrected form uses the entropy chain rule:

H_{\mathrm{mystery}}(O,S)=H(L_{1})+\sum_{i=2}^{n}H\!\bigl(L_{i}\mid L_{i-1},\dots,L_{1}\bigr)

where the interpretation layers L_1, L_2, …, L_n carry compositional, symbolic, contextual, archetypal, cultural, and higher-order interpretive uncertainty (the symbols are written L_i, not H_i, to avoid an entropy-of-an-entropy reading). Each conditional term adds only the structured uncertainty that remains after the lower layers are accounted for, so nothing is double-counted. Mystery is sharply separated from noise. Noise is incompressible randomness that raises cost without adding coherent interpretation; mystery is compressible-but-not-yet-exhausted structure that invites further interpretation:

H_{\mathrm{mystery}}=H_{\mathrm{total}}-H_{\mathrm{noise}}-H_{\mathrm{resolved}}

Generative interpretive capacity

Let the set of stable interpretations an object generates in a subject be the interpretation space, each stable interpretation an attractor of the observer's cognitive dynamics. A naive ratio of interpretation volume to generator complexity diverges when the generator is trivial — falsely making random association triggers look infinitely valuable. The regularised form fixes this and rewards coherent richness:

\mathrm{GIC}^{*}(O,S)=\frac{\mathrm{Vol}\bigl(\Omega(O,S)\bigr)\cdot \mathrm{Coh}(\Omega)}{K_{\mathrm{gen}}(O)+\varepsilon}

The coherence factor penalises a random association explosion: high value demands many interpretations and structured coherence between them, not many unrelated associations. It is normalised by a characteristic interpretation-distance scale τ so the exponent is dimensionless, with the average taken over all pairs of stable interpretations:

\mathrm{Coh}(\Omega)=\exp\!\left(-\frac{1}{\tau}\,\mathbb{E}_{i,j}\bigl[D_{g}(x_{i}^{*},x_{j}^{*})\bigr]\right),\qquad \mathbb{E}_{i,j}[\,\cdot\,]=\frac{1}{\mathrm{Vol}(\Omega)^{2}}\int_{\Omega}\!\int_{\Omega}(\cdot)\,d\mu(x_{i}^{*})\,d\mu(x_{j}^{*})

Phase coherence — timing, where timing exists

Each dynamic aesthetic feature is written as a complex amplitude-and-phase signal. Phase is not metaphor: it is admitted only where genuine timing, rhythm, or oscillation exists — music, animation timing, scroll rhythm, saccadic timing, interaction latency, reveal timing, attention cycles.

z_{k}(t)=A_{k}(t)\,e^{i\theta_{k}(t)}

Representing object and subject dynamic states in a Hilbert space, normalised phase coherence is their bounded inner-product alignment:

\mathrm{PC}(O,S,t)=\frac{\bigl|\langle z_{O}(t),z_{S}(t)\rangle\bigr|}{\lVert z_{O}(t)\rVert\,\lVert z_{S}(t)\rVert},\qquad 0\le \mathrm{PC}\le 1

For dynamic systems this measures real timing alignment — in an interface, high phase coherence means motion, reveal, and response arrive exactly when the user's attention is ready to receive them. For a static artifact phase is not applied directly: it enters only through the trajectory of attention the artifact induces — the order in which the eye reads the composition and its layers unfold over time. Where no such attentional sequencing is measurable, phase coherence is replaced by structural alignment rather than forced onto a still image.

The Hamiltonian formulation

With the cognitive model position as generalised coordinate, the generalised cognitive momentum is the velocity weighted by the manifold's local curvature — so a small conceptual movement in a highly curved region can carry large momentum:

p_{i}=g_{ij}(q)\,\dot{q}^{j}

Kinetic cognitive energy is the energy of model movement — high values correspond to rapid restructuring of the observer's predictions. It is the same energy written two equivalent ways: with the metric acting on velocities (covariant g_{ij}), and — after lowering the velocity to the conjugate momentum above — with the inverse metric acting on momenta (contravariant g^{ij}). The raised index in the momentum form is exactly what undoes the lowered index in p_i:

T(q,\dot{q})=\tfrac{1}{2}\,g_{ij}(q)\,\dot{q}^{i}\dot{q}^{j}\qquad\text{(velocity form, covariant metric)}

T(q,p)=\tfrac{1}{2}\,g^{ij}(q)\,p_{i}\,p_{j}\qquad\text{(momentum form, inverse metric)}

The aesthetic potential combines the object field with mystery and phase coherence; the negative signs mean high mystery and high phase coherence create attractive wells that pull cognition into deeper engagement:

V(q,O,S)=\Phi_{O}(q)-\kappa\,H_{\mathrm{mystery}}(O,S,q)-\rho\,\mathrm{PC}(O,S,q)

The aesthetic Hamiltonian is the sum of kinetic and potential terms — the diagnostic energy of the aesthetic state:

\mathcal{H}_{\mathrm{aest}}(q,p)=\tfrac{1}{2}\,g^{ij}(q)\,p_{i}p_{j}+\Phi_{O}(q)-\kappa\,H_{\mathrm{mystery}}(O,S,q)-\rho\,\mathrm{PC}(O,S,q)

In the ideal, non-dissipative case the dynamics follow Hamilton's equations; this is the idealisation of pure contemplation, with no fatigue and no disturbance:

\dot{q}^{i}=\frac{\partial\mathcal{H}_{\mathrm{aest}}}{\partial p_{i}},\qquad \dot{p}_{i}=-\frac{\partial\mathcal{H}_{\mathrm{aest}}}{\partial q^{i}}

The open, dissipative balance — the central balance equation

Real cognition is not ideal. Introducing a dissipative cost rate and a structured input flux turns the system into an open one, governed by the open-system balance equation at the heart of the model (an analogy to a thermodynamic balance, not a claim of physical thermodynamics):

\frac{d\mathcal{H}_{\mathrm{aest}}}{dt}=-\lambda\,C(t)+\mathcal{I}_{\mathrm{input}}(t)

Beauty persists when the object supplies enough structured input to compensate for cognitive dissipation without overwhelming coherence. From this single balance, the cognitive regimes follow as cases.

Cognitive regimes

Boredom is not familiarity; it is dissipative imbalance — the object stops paying for the cost of attention and the trajectory collapses into a trivial minimum:

\mathcal{I}_{\mathrm{input}}(t)<\lambda\,C(t)\ \Rightarrow\ \dot{q}\to 0,\ p\to 0,\ H_{\mathrm{mystery}}\to 0

Overload is not high complexity alone; it is loss of coherent interpretation — too much information too fast, the observer cannot phase-lock, interpretations destabilise:

\mathcal{I}_{\mathrm{input}}(t)\gg\lambda\,C(t)\ \Rightarrow\ \mathrm{PC}\to 0,\ \mathrm{Coh}(\Omega)\to 0

Aesthetic flow is balanced open-system dynamics: input and cost in balance, mystery still structured, phase coherence high, the observer held inside the aesthetic field. Durable beauty is the same balance surviving repeated exposure — it lasts not because it produces endless surprise but because it holds stable structured mystery and coherent interpretive capacity:

\frac{d\mathcal{H}_{\mathrm{aest}}}{dt}\approx 0,\qquad H_{\mathrm{mystery}}>0,\ \ \mathrm{PC}\to 1,\ \ \mathrm{GIC}^{*}\gg 0

The full aesthetic action functional

Collecting the terms, the realised aesthetic value over a perceptual interval is the trajectory-integrated cognitive work, multiplied by coherent interpretive capacity:

\mathcal{B}_{v2.6}[\gamma]=\mathrm{GIC}^{*}(O,S)\int_{0}^{T}\!\left[\alpha\!\left(-\frac{d\Phi_{O}(\gamma(t))}{dt}\right)+\beta\,\frac{H_{\mathrm{mystery}}(O,S,t)}{T_{\mathrm{char}}}+\eta\,\nu_{\mathrm{res}}\,\mathrm{PC}(O,S,t)-\lambda\,C(t)\right]dt

Each term carries a dimension that resolves to an information-flux rate (nats per second): the potential-flow term, the structured-mystery release rate, the phase-coherence sampling rate, against the cognitive cost rate. The integral is total aesthetic work in nats; the interpretive-capacity multiplier is dimensionless.

term	meaning	engineering observable
$\Phi_{O}$	object-induced prediction-mismatch field	rate of model update over artifact states
$H_{\mathrm{mystery}}$	structured, unresolved interpretive depth	unresolved semantic depth that rewards further reading
$\mathrm{PC}$	timing alignment of object and subject dynamics	alignment of UI motion/reveal with user behaviour
$\lambda C(t)$	cognitive cost / dissipation	latency, hesitation, backtracking, overload
$\mathrm{GIC}^{*}$	coherent interpretive capacity per complexity	stable journey capacity per implementation complexity

The strongest compact statement of the model: beauty is the trajectory-integrated cognitive work generated by an object-induced information potential field, multiplied by coherent interpretive capacity and constrained by phase alignment and cognitive dissipation.

References

Amari, S. & Nagaoka, H. — Methods of Information Geometry (Fisher-Rao metric, statistical manifolds) — https://doi.org/10.1090/mmono/191
Friston, K. — The free-energy principle: a unified brain theory? (variational free energy) — https://doi.org/10.1038/nrn2787
Itti, L. & Baldi, P. — Bayesian surprise attracts human attention (surprise and attention) — https://doi.org/10.1016/j.visres.2008.09.007