G(z) = J_g(z)ᵀ J_g(z).

Advanced Technical Expansion | Charts, Graph Limits, Topology and Latent-Space Geometry

The first part of this article establishes the receiver rule: do not ask whether an embedding “looks right”; ask which geometry it is intended to preserve. This expansion develops the machinery behind that rule. It moves from local charts and tangent spaces to graph-consistency assumptions, spectral operators, topological summaries, decoder-induced metrics and falsifiable evaluation. The purpose is not to make every reader a differential geometer. It is to make manifold claims precise enough that they can fail in informative ways.

A. A manifold is locally simple and globally allowed to be difficult

A d-dimensional smooth manifold is locally modelled by open subsets of R^d. Around each point there is a chart: a coordinate map that turns a neighbourhood of the manifold into ordinary numerical coordinates. The power of this definition is that it does not require one global coordinate system.

The sphere illustrates the reason. Latitude and longitude are useful over much of the globe but have singular behaviour at the poles and a seam in longitude. No single flat chart can cover the sphere without some defect. An atlas uses several overlapping charts instead. On overlaps, transition maps tell us how one coordinate description converts into another.

Representation learning often behaves as if every latent space should be one global Euclidean chart. That assumption is convenient and sometimes sufficient. It becomes restrictive when the underlying state has periodicity, topology, singularities or multiple regimes. An atlas-style latent representation can be more faithful than forcing the world into one coordinate sheet.

B. The tangent space is the local linear approximation that makes calculus possible

At a smooth point on a manifold, the tangent space is a linear vector space that approximates nearby directions of motion. On a curved surface, the surface itself is not flat globally, but a sufficiently small neighbourhood looks approximately like its tangent plane.

Many algorithms quietly depend on this approximation. LLE assumes a point can be reconstructed from neighbours in a locally linear patch. Local PCA estimates tangent directions. Intrinsic-dimension methods infer how many independent directions of variation exist nearby. If neighbourhoods are too large, curvature invalidates the linear approximation. If they are too small, noise dominates.

This gives neighbourhood size a geometric interpretation. It is not merely a hyperparameter. It chooses the scale at which we claim the surface is approximately linear.

C. A Riemannian metric says how infinitesimal movement should be measured

A Riemannian metric assigns an inner product to each tangent space. In local coordinates it is represented by a positive-definite matrix G(z). For an infinitesimal displacement dz, squared length is approximately dzᵀG(z)dz.

If G(z)=I everywhere, ordinary Euclidean distance is locally appropriate. If the metric varies, the same coordinate step can have different physical or observational meaning in different regions. This is exactly what happens under a nonlinear decoder: the Jacobian stretches some latent directions more than others.

The metric is therefore part of representation meaning. Coordinates alone do not tell us distance unless the metric is specified or implied.

D. Geodesics are shortest paths under the chosen metric, not necessarily straight coordinate lines

A geodesic locally follows the shortest or straightest route permitted by the manifold and metric. On a plane, straight Euclidean lines are geodesics. On a sphere, great-circle arcs are geodesics. On a decoder-induced latent manifold, a geodesic can curve through latent coordinates to avoid high-distortion regions.

This distinction matters for interpolation and planning. If latent coordinates are only a chart, a straight segment between two codes is a coordinate convenience, not a guaranteed meaningful transition. The decoded trajectory can leave high-density regions, violate constraints or take a longer observation-space route than a curved geodesic.

A representation used for planning should specify the metric under which “short path” is meaningful. Otherwise an optimiser is solving a geometry problem that was never defined.

E. Classical MDS turns distances into coordinates through double centring

Classical multidimensional scaling starts from a matrix of squared pairwise distances D². Under Euclidean assumptions, one can recover a centred Gram matrix using

B = -1/2 J D² J,
J = I - (1/n)11ᵀ.

If the distances come exactly from Euclidean points, B is positive semidefinite and equals XXᵀ for centred coordinates X. Eigen-decomposition then yields coordinates from the leading eigenvalues and eigenvectors.

Isomap uses this machinery after replacing ordinary distances with graph shortest-path approximations to geodesic distance. The two-stage structure is important. Negative eigenvalues can indicate that the estimated distance matrix is not exactly Euclidean in the target geometry. A poor final embedding can therefore reflect graph error, intrinsic curvature or insufficient target dimension rather than one undifferentiated “MDS failure”.

F. Isomap consistency depends on sampling density, curvature and avoiding shortcut edges

For graph shortest paths to approximate manifold geodesics, local edges must stay on the same local sheet and the sample must be dense enough that paths can follow the surface. If the neighbourhood radius is larger than the separation between two folded layers, a shortcut edge crosses empty ambient space and destroys the intrinsic distance estimate.

If the radius is too small, the graph can disconnect. In variable-density data, one global radius can be too large in dense regions and too small in sparse regions. k-nearest-neighbour graphs adapt connectivity count but can still bridge geometrically separate structures where sampling is uneven.

A controlled study should report graph connectivity, edge-length distribution, shortcut rates when ground truth is available, and sensitivity across k or radius. The graph is the hidden geometric hypothesis beneath the final coordinates.

G. Hubness is a high-dimensional neighbour pathology worth checking before manifold learning

In high-dimensional spaces, distance concentration can make neighbour structure counterintuitive. Some points become “hubs”, appearing in the nearest-neighbour lists of many other points. This can arise from geometry, density, norm variation or the chosen similarity.

A graph built from such neighbours can over-connect hubs and create artificial global structure. Before attributing a dense central region to the manifold, inspect neighbour occurrence counts, distance distributions and the relationship between vector norm and neighbour frequency.

Alternative metrics, local scaling or normalisation can change hubness dramatically. The choice should be justified by the receiver, not by which setting produces the cleanest plot.

H. Local scaling can repair density imbalance and can also change the scientific question

Neighbour graphs often use Gaussian weights such as exp(-||x_i-x_j||²/σ²). A single bandwidth σ assumes one meaningful distance scale across the dataset. Variable density violates that assumption. Local bandwidths based on neighbour distances can make connectivity more uniform.

But density normalisation is not neutral. If dense regions represent genuinely more common states and that frequency matters, equalising local scale removes part of the distribution. If the job is to recover geometry independent of sampling density, local scaling can be exactly the right move.

Separate geometric occupancy from probability density. A representation can preserve one, the other, or both through separate channels. Do not force a single 2D picture to carry every meaning.

I. LLE’s weights are invariant to some transformations because they encode local affine relationships

For each point, LLE solves a constrained reconstruction problem using its neighbours. The weights sum to one, which gives translation invariance to the local reconstruction. Under global rotation and scaling, the local covariance structure changes in controlled ways, and the same affine relationships can be retained under ideal conditions.

The weight solution can become unstable when neighbours are nearly collinear or when the local covariance matrix is singular. Regularisation is then needed. That regularisation changes the local geometry the method effectively trusts.

A good LLE audit records condition numbers of local covariance matrices and the magnitude of regularisation. Unstable neighbourhoods should be marked rather than silently treated as equally reliable coordinates.

J. Graph Laplacians approximate differential operators only under assumptions

The graph Laplacian is a discrete operator built from adjacency or weights. In manifold learning, appropriately normalised graph Laplacians can approximate continuous Laplace-type operators on an underlying manifold as sample size increases under suitable density and bandwidth conditions.

Different normalisations answer different questions. The unnormalised Laplacian, random-walk normalisation and symmetric normalisation handle degree and density differently. Spectra can change substantially. A paper saying “we used the Laplacian” should specify which operator and why.

The eigenvectors used for embedding are therefore not magic coordinates. They are modes of variation defined by a particular discrete operator whose continuous interpretation depends on graph construction and normalisation.

K. Diffusion distance integrates many paths instead of trusting one shortest path

Shortest-path distance can be brittle: one erroneous shortcut edge can dramatically reduce distance. Diffusion distance compares transition probabilities of a random walk after t steps. Two points are close when their distributions of possible journeys through the graph are similar.

Spectral decomposition makes this concrete. If a Markov operator has eigenvalues 1=λ₀ ≥ |λ₁| ≥ ... and eigenfunctions ψ, diffusion coordinates weight each component by λ_k^t. Increasing t suppresses fast-decaying modes and emphasises large-scale connectivity.

The time parameter is therefore a scale control. It does not reveal one unique geometry. It defines the geometry relevant to diffusion over that time horizon.

L. t-SNE’s asymmetric KL explains why missing neighbours are punished differently from extra empty space

t-SNE constructs high-dimensional neighbour probabilities p_ij and low-dimensional probabilities q_ij, then minimises KL(P||Q). Because the KL is asymmetric, placing a high-probability high-dimensional neighbour far apart creates a strong penalty, while creating large low-density empty regions between unrelated groups is less directly penalised.

This helps explain why t-SNE excels at revealing local clusters and why global spacing is weakly constrained. The heavy-tailed low-dimensional kernel helps nearby groups spread out to reduce the crowding problem.

Perplexity determines an effective neighbourhood scale by tuning local Gaussian bandwidths. It is not simply a visual smoothing knob. Different perplexities ask the map to preserve different scales of neighbourhood probability.

M. UMAP’s fuzzy graph language still begins with a metric and neighbourhood choice

UMAP constructs local fuzzy membership strengths that represent how strongly points are connected in neighbourhoods, combines these local views into a global fuzzy simplicial structure, and optimises a low-dimensional analogue through a cross-entropy-like objective. The topology-inspired language does not remove dependence on the input metric and k-neighbour graph.

The n_neighbors parameter controls how much local context contributes to the graph. min_dist affects how tightly points can pack in the low-dimensional display. Different values can make clusters look compact, continuous or fragmented without changing the raw observations.

A UMAP figure should therefore be accompanied by stability checks across a reasonable parameter region. If a claimed scientific cluster appears only at one narrow setting, the evidence is weaker than the screenshot suggests.

N. Trustworthiness and continuity measure opposite neighbourhood errors

Trustworthiness penalises points that become close in the embedding even though they were not close in the original space. It asks, “did the map invent neighbours?” A complementary continuity-style metric asks whether original neighbours remain close after embedding: “did the map tear neighbours apart?”

A map can score well on one and poorly on the other. Compressing many regions together can preserve some original neighbours while introducing false neighbours. Spreading groups apart can avoid false neighbours while losing true cross-boundary relations.

Measure both over several k values. Neighbourhood fidelity is a function of scale, not one universal score.

O. Persistent homology asks which topological features survive across scale

A single neighbourhood radius can create arbitrary topological conclusions. Persistent homology addresses this by building a family of complexes over increasing distance scales and tracking when connected components, loops and higher-dimensional holes appear and disappear.

In a Vietoris–Rips construction, points are connected when their pairwise distance falls below a threshold, and higher simplices are filled when all required edges exist. As the threshold grows, topology changes. A long-lived loop is more robust across scale than a loop that appears and vanishes almost immediately.

Persistence diagrams or barcodes summarise these lifetimes. They do not automatically tell us what a loop means scientifically. They tell us that a topological feature is stable under the chosen metric and filtration. Semantic interpretation still requires domain evidence.

P. Topological distance gives another way to compare representations

If a representation is supposed to preserve topology, compare persistence summaries between source and latent spaces using metrics such as bottleneck or Wasserstein distances on persistence diagrams. This evaluates topological structure rather than only pointwise Euclidean error.

But topology can be intentionally changed. A classifier may benefit from separating intertwined classes that were connected in raw pixel space. An autoencoder intended for faithful reconstruction has a different contract. The correct topology is receiver-relative.

Therefore phrase the goal explicitly: preserve data-manifold topology, simplify topology for classification, or discover robust topological signatures. Those are different jobs.

Q. Intrinsic dimension can vary from one region to another

A dataset can contain a two-dimensional surface joined to a one-dimensional branch, or a motion system where some joints become locked in particular states. There may be no single global intrinsic dimension.

Local dimension estimates can reveal these transitions. Plot estimated dimension across points and neighbourhood sizes. A sudden change can indicate a branch, singularity, boundary, noise regime or simply insufficient sampling.

A variable-dimension structure is not a smooth manifold everywhere. Calling it one manifold can hide the singular points where the local Euclidean model fails. Stratified spaces or unions of manifolds can be better descriptions.

R. Boundaries and singularities deserve explicit representation

A manifold can have a boundary: a line segment differs from a circle because endpoints have neighbourhoods that look like half-lines. Data-generating processes often have hard physical limits, saturation, collisions or forbidden states. Treating boundary points as ordinary interior points can distort local geometry.

Singularities are stronger failures of smoothness. Two branches may cross, or several regimes may meet at one state. A tangent plane can become non-unique. Local PCA may report inflated dimension.

Mark these regions as special rather than forcing smooth interpolation through them. In planning, boundaries can correspond to safety constraints. In scientific data, singularities can be the phenomenon of interest.

S. Hyperbolic geometry can represent hierarchy more economically than Euclidean space

Trees grow exponentially with depth: each level can contain many more nodes than the previous one. Euclidean balls grow polynomially with radius, while hyperbolic space has exponential volume growth. This makes hyperbolic geometry naturally suited to embedding tree-like hierarchies with relatively low distortion.

Poincaré embeddings exploit this property for hierarchical representations. Distance and optimisation use hyperbolic rather than Euclidean geometry.

The broader lesson is that latent geometry should follow relational structure. A hierarchy forced into Euclidean coordinates can require high dimension or large distortion. A curved metric can make the same relationships simpler.

T. Geometry is not probability density, but generative models couple the two

A manifold describes possible or structured states; a probability distribution describes how frequently states occur. A uniform distribution over a curved manifold and a highly concentrated distribution on the same manifold share geometry but differ statistically.

Generative models couple these objects because a decoder maps latent coordinates to observations while a latent prior assigns probability to coordinates. The decoder Jacobian affects both geometry and volume change. Regions where the decoder stretches space can transform density.

When interpreting latent clusters, ask whether the observed spacing comes from geometry, density, prior regularisation or all three. A Euclidean latent prior can impose organisation that is convenient for sampling but not intrinsic to the world.

U. The pullback metric reveals directions the decoder amplifies

For decoder g(z), the local metric G=JᵀJ has eigenvectors and eigenvalues. An eigenvector with large eigenvalue is a latent direction where a small coordinate step causes a large change in observation space. A small eigenvalue indicates a direction the decoder compresses.

The condition number of G measures anisotropy. Near-singular metrics indicate directions in latent space that barely change the output or regions where the decoder loses local invertibility. Such regions can make geodesic computation unstable and can signal redundant latent dimensions.

Plot metric eigenvalues along interpolation paths. A visually smooth latent line can cross a region where observation-space sensitivity spikes, warning that equal coordinate steps are not equal semantic steps.

V. Geodesic planning needs uncertainty, not only a metric

A decoder-induced metric measures local output change for the learned map. It does not guarantee that every region of latent space is supported by data. A geodesic can exploit poorly trained regions where the decoder behaves unpredictably.

Augment path cost with uncertainty or density penalties. Penalise regions far from training support, regions with high epistemic uncertainty or states violating physical constraints. The “shortest” useful path is then a constrained path through trustworthy geometry.

This is particularly important for robotics, chemistry and medical trajectory modelling. A smooth decoded transition can still be physically impossible or clinically unsupported.

W. Out-of-sample extension changes a transductive picture into a reusable representation

Methods based on eigenvectors or optimised point coordinates often produce embeddings only for the fitted dataset. New samples require an extension. Nyström methods approximate eigenfunctions at new points. Parametric t-SNE or neural UMAP variants learn explicit mappings. Neighbour interpolation provides simpler local estimates.

Every extension adds assumptions. A new point far from training support can still be assigned coordinates even when those coordinates are meaningless. Out-of-sample systems need a support or uncertainty check.

Evaluate with temporal or geographically held-out data, not only random points from the same cloud. A map used for deployment should prove that its coordinate system survives the arrival of new observations.

X. Nyström approximation is a concrete example of the scale–accuracy trade

Large spectral methods cannot always diagonalise an n×n matrix. Nyström approximation chooses a subset of landmarks, computes the expensive structure on that subset and extends it to the rest. This reduces computation substantially.

Landmark choice becomes part of the geometry. Uniform random landmarks may undersample rare regions. k-means-style landmarks emphasise dense structure. Stratified selection can preserve known subgroups. Approximation error should be measured by region, not only globally.

Scaling a manifold algorithm is therefore not merely an engineering optimisation. It can change which parts of the manifold are represented faithfully.

Y. Stability across random seeds is a geometry claim

Stochastic embeddings such as t-SNE and UMAP can produce different global arrangements across initialisations while preserving similar local neighbourhoods. A stable scientific conclusion should survive the variation relevant to its claim.

If the claim is “these two subgroups form separate local neighbourhoods”, compare k-neighbour overlap or cluster adjacency across seeds. If the claim is “group A lies between B and C on a continuum”, global orientation and ordering must be stable under aligned runs. If not, the narrative exceeds the geometry.

Procrustes alignment can remove arbitrary rotation, translation and scale between embeddings before comparing them. But it cannot repair genuine topological rearrangements. Choose alignment transformations that correspond to equivalences you are willing to treat as the same map.

Z. Bootstrap the dataset to see which geometric conclusions depend on particular samples

Resample observations, refit the graph or embedding, and measure how neighbourhoods, clusters, loops and dimensions change. A structure that disappears whenever a few observations are omitted may be too fragile for a strong claim.

Bootstrap uncertainty can be difficult for embeddings because coordinates are not uniquely aligned across runs. Compare invariant objects instead: pairwise neighbourhood relations, persistent homology, aligned distances or downstream predictions.

The result is a geometry confidence profile rather than one definitive picture.

AA. A visual cluster is not necessarily a natural category

Dimension-reduction objectives can exaggerate gaps to preserve neighbourhoods. Sampling density can create sparse bridges. Labels can bias representation training. A visible cluster therefore requires independent evidence before it is named as a biological subtype, social group, failure mode or semantic class.

Test cluster stability in the original representation, across embeddings, across preprocessing choices and on held-out data. Ask whether the cluster predicts an external variable not used to construct the map. If the cluster disappears under small methodological changes, use exploratory language.

Visualisation is a hypothesis generator. It becomes evidence when the hypothesised structure survives tests outside the picture.

AB. A trajectory in an embedding is not automatically time or causality

Single-cell biology, developmental data and other domains often display embeddings with curves interpreted as progression. A geometric continuum can be consistent with a temporal process without uniquely determining temporal direction or causal sequence.

Direction requires additional evidence: timestamps, lineage, RNA velocity assumptions, interventions or domain constraints. Branch points require validation. Sampling a population at different latent states is not the same as observing one individual move through those states.

This is another manifestation of the world-return principle. Geometry can organise observations; causal and temporal semantics require their own evidence.

AC. Manifold geometry inside neural networks can be layer-specific

Representations change through a network. Early layers may preserve local sensory variation. Middle layers can separate task factors. Later layers may collapse within-class variability while increasing margins between classes. There is no reason to expect one manifold geometry across every layer.

Analyse neighbourhoods, intrinsic dimension, class geometry and topology layer by layer. A decrease in dimension can indicate useful abstraction or destructive collapse. Increased separation can improve classification while reducing information needed for another task.

Layer geometry should therefore be interpreted relative to the computation that follows it, not as an intrinsic quality score.

AD. Behavioural interventions can test whether representational directions matter to the model

Correlating a latent direction with behaviour does not prove the model uses that direction causally. One stronger test intervenes in representation space: move activation along a direction or manifold coordinate while holding other aspects fixed as far as the architecture permits, then observe model output.

Recent manifold-steering work explores this relationship between representation geometry and behaviour. Such interventions require caution. Moving activations can take the network off its natural activation manifold, producing behaviour that is hard to interpret.

A trustworthy intervention checks whether edited states remain in-distribution, uses control directions, measures dose–response and compares multiple layers. Geometry becomes mechanistic evidence only when interventions respect the system’s valid state space.

AE. Representation curvature can change optimisation even when topology stays the same

Two spaces can be topologically identical yet metrically very different. A stretched, curved surface and a flat sheet can share topology while distances and geodesics differ. Optimisation, interpolation and nearest-neighbour retrieval depend on metric geometry, not topology alone.

Conversely, two representations can preserve local distances reasonably while differ in topology because a loop has been cut. This is why topology metrics and metric-distortion metrics answer different questions.

Do not use “structure preserved” as a single umbrella phrase. Name whether the preserved object is neighbourhood, distance, angle, density, topology, order, hierarchy or downstream function.

AF. Geometry-aware batching matters when local neighbourhoods define the objective

Neural manifold objectives often estimate neighbours or contrastive relationships within minibatches. A random batch can omit true neighbours and replace them with distant points. Large batches improve coverage but increase compute. Memory banks or approximate neighbour indices provide alternatives.

If batch composition changes the local graph, it changes the effective geometry used for learning. Record batch size and sampling policy as part of the representation method. Stratified batching can introduce its own bias.

For datasets with temporal or spatial correlation, random batching may also leak near-duplicates across train and validation. Geometry evaluation should respect independence at the level relevant to deployment.

AG. The geometry evidence card

• Objects: what observations or latent states are points?
• Units: what preprocessing defines coordinate scale?
• Metric: what does distance mean?
• Neighbour rule: k, radius, weighting and approximate search.
• Scale: what neighbourhood or diffusion horizon is claimed?
• Dimension: global, local and estimator uncertainty.
• Topology: expected components, loops, boundaries or singularities.
• Density: preserved, normalised or separately represented.
• Algorithmic contract: variance, geodesic distance, local weights, diffusion, probabilities or fuzzy graph.
• Out-of-sample rule: how new points enter.
• Stability: seeds, hyperparameters and bootstraps.
• Metric distortion: local and global diagnostics.
• Topological distortion: persistence or connectivity diagnostics.
• Receiver: visualisation, retrieval, planning, science or control.
• Refusal rule: which visual conclusions are explicitly unsupported.

AH. Worked laboratory: double-centre known distances and recover a plane

Generate points in a two-dimensional plane, calculate exact pairwise distances and form the squared distance matrix. Apply the double-centring formula and eigen-decompose B. The two positive leading eigenvalues should recover the plane up to translation, rotation and reflection. Extra eigenvalues should be numerical noise.

Now replace exact distances with deliberately corrupted graph distances containing one shortcut. Repeat MDS. Observe how a single bad global relation can create extra distortion. This separates the embedding machinery from the distance-estimation machinery.

AI. Worked laboratory: compare shortest-path and diffusion robustness to one shortcut

Construct a ring graph with many nodes. Add one artificial edge connecting opposite sides. Shortest-path distance across the ring changes drastically for pairs using the shortcut. A short-time diffusion distance may be less dominated because probability spreads across multiple routes rather than following one shortest path.

Vary diffusion time. At long times the random walk approaches stationary behaviour and fine distinctions disappear. The experiment demonstrates that robustness and scale are connected: no one diffusion time is universally correct.

AJ. Worked laboratory: persistent loop under increasing noise

Sample points from a circle and add increasing Gaussian noise. Build Vietoris–Rips filtrations and track the dominant one-dimensional persistence interval. At low noise the main loop should persist over a substantial range. As noise grows or sampling becomes sparse, the interval shortens and spurious small loops appear.

This gives a quantitative way to ask when “the data contain a loop” remains justified. The answer is not a picture of a ring; it is a feature that remains stable across scale and perturbation under the chosen metric.

AK. Worked laboratory: latent Euclidean line versus decoder geodesic

Use a two-dimensional latent space with an analytic decoder that stretches one region strongly. Choose two endpoints. Compute the straight Euclidean segment and its observation-space length using the pullback metric. Then numerically optimise a curve with lower integrated metric length.

Decode both paths and compare intermediate states. The experiment makes the central point tangible: coordinates draw the map, while the metric determines travel cost.

AL. Worked laboratory: one hierarchy in Euclidean and hyperbolic coordinates

Generate a balanced tree. Fit low-dimensional Euclidean embeddings and hyperbolic embeddings under comparable objectives. Measure distortion of graph distances across depth. The hyperbolic model should often represent exponential branching more compactly.

Then add cross-links that make the graph less tree-like. Observe where the hyperbolic advantage changes. Geometry is a prior matched to structure, not a universal ranking of spaces.

AM. Manifold learning and causal representation meet at valid trajectories

A causal state space contains more than geometric proximity. Two states may be near in observation space but require very different interventions. Conversely, a small causal intervention can create a large sensory change. Geometry can help organise state while causality defines which paths are possible under actions.

A world model used for planning should therefore learn or preserve action-conditioned transitions, not merely an attractive manifold. Geodesics can suggest smooth state changes; dynamics determine executable state changes. The intersection of the two is the valid planning corridor.

Evaluate by rollouts: does a short latent path correspond to a sequence of feasible actions? If not, the geometric metric is not the control metric.

AN. Manifold learning and equivariance meet when a transformation traces structured paths

A continuous group action can trace an orbit through representation space. Rotating an object through angle θ forms a one-dimensional closed trajectory if a full turn returns to the same state. A good equivariant representation can organise this orbit according to the group’s geometry.

This gives a direct diagnostic: sample a transformation trajectory, embed the representations and compare its geometry with the expected group orbit. Does a rotation produce a loop? Are equal angle increments represented consistently? Does the orbit collapse under an invariant head only where intended?

Symmetry supplies known geometry that can calibrate manifold analysis.

AO. Final audit: when does a manifold claim deserve release?

1. The point objects and preprocessing are defined.
2. The metric has a semantic justification.
3. Neighbourhood construction and scale are reported.
4. Connectivity and shortcut failure are audited.
5. The target geometric property is named.
6. Intrinsic dimension is treated as an estimate with scale dependence.
7. Boundaries, branches and singularities are considered.
8. Density and geometry are not conflated.
9. Topology is tested when topological claims are made.
10. Random-seed and hyperparameter stability match the public interpretation.
11. Out-of-sample behaviour is evaluated if the map will be reused.
12. Latent interpolation is checked against decoder geometry and data support.
13. Visual clusters receive independent validation before semantic naming.
14. Temporal or causal narratives use temporal or causal evidence.
15. The article explicitly states which global distances or areas should not be read literally.

The mature view of manifold representation learning is not that high-dimensional data secretly contain one perfect low-dimensional picture. It is that many datasets contain structured relations that become easier to work with when the representation preserves the right geometry at the right scale. Sometimes that geometry is Euclidean. Sometimes it is graph-based, circular, spherical, hyperbolic, topological or decoder-induced. The map is successful when it makes the receiver’s operation simpler without disguising the distortions needed to make that simplification possible.