From individual samples to local geometry

When the neighbors share the slope

A sharp score can pull toward one remembered point. A smoother score can combine nearby points. The curve decides how safely we can make that average.

Smoothing / training-cell spacingσmin / ℓ 0.25

Geometric bias scale / spacingσmin² / (τℓ) 0.005

These ratios relate the controls to observation scale and curvature. The second is the geometric term in the manifold-error analysis; it is not a certified error bound by itself.

Try

Arclength means distance along the curve. Each cell groups two observations around position kℓ, at ℓ(k ± 1/6); consecutive gaps alternate between ℓ/3 and 2ℓ/3. The broad cutoff stays fixed at σmax = 3.

Reach τ = 10Reach is the radius of a neighborhood around the curve in which every location has one nearest point on the curve.

The local pull toward the data

Training pointQuery xWeighted average μ̄Fixed start

Drag the orange query, or move it with the arrow keys. Every one of the 82 observations contributes, including observations outside this local view. Score arrows show direction; their lengths are normalized.

Who determines this score?

Largest center weight—
Effective contributors—

Refinement step 0 / 1200

Playback follows visible movement and compresses nearly stationary refinement. Step labels remain the actual update numbers; Show outcomes displays the full numerical endpoint.

Different starts, different outcomes

Refining the fixed starts
Vertical range — · normal zoom 1×

Near a training point 0Between samples, near the curve 0Away from the curve 0Still refining 15
Distinct resolved cells —Median training distance / ℓ —Median normal distance / ℓ —Largest remaining pull / ℓ —

The upper view shows the reference curve. This view separates position along the curve from normal displacement, both measured in units of ℓ. The remaining pull is the distance from a point to its current weighted average: how far its update still wants to move. The same fixed starts are used after every parameter change. Cell counts describe this illustrative ensemble, rather than population coverage probabilities.

Novelty, fidelity, and diversity are separate. An outcome between training points is useful only if it remains close to the manifold, and multiple regions remain populated. The plot displays these diagnostics separately; it does not certify population generalization.

The model and the resolution of the diagnostics

The reference curve is the parabola μ(u) = (u, κu²/2). The control κ is its curvature at the vertex. Away from the vertex its curvature is κ/(1 + κ²u²)3/2. At κ = 0 the curve is flat, an affine line. Training pairs are placed at equal arclength-cell spacing, so the spacing control retains its geometric meaning as the curve bends. Reach is the radius of a neighborhood in which each location has a unique nearest point on the curve; it describes the geometric room available for local averaging.

The manifold analysis bounds the endpoint tube radius by ε₀ = Cloc mσmin² / (2τ) under its posterior-localization, contraction, and landscape assumptions. Thus the second live ratio is only the geometric factor in ε₀ / ℓ = (Clocm/2) · σmin²/(τℓ). The local constant and theorem assumptions have not been certified by these sliders; a small ratio alone supplies no guarantee. The numerical preview instead reports the trajectories actually computed for this finite cloud.

We evaluate the exact σ²-weighted Gaussian mixture in three ambient dimensions, with 32 logarithmically spaced scales and σmax = 3. The plot and starts lie in the invariant plane x₃ = 0. These illustrative initializations are not samples from the full three-dimensional Gaussian initialization used in distributional theorems.

The displayed contributions are the drift weights qj(x) = ∑i w̃ij(x), with w̃ij ∝ 𝒩3(x; μj, σi²I). They satisfy s(x) = β(x)(∑j qj(x)μj − x). Effective contributors are Neff = 1/∑jqj²; all weights enter this statistic.

Ascent uses the normalized update x ← x + 0.35(μ̄(x) − x), until the barycentric residual is at most 10−6ℓ or 1200 updates are reached. The update has an exact-mixture ascent guarantee; it is not a time-dependent diffusion sampler. Moving the query inspects the score without changing the ensemble starts.

For a resolved numerical outcome, near a training point means training distance < ℓ/12. Otherwise between samples, near the curve means normal distance ≤ ℓ/10. A larger normal distance is marked away from the curve. Unresolved trajectories remain still refining. These thresholds are visible-resolution proxies, not definitions or proofs of exact memorization or generalization. A finite empirical model can still send trajectories to isolated destinations; outputs between training observations do not by themselves recover a continuous population. More curvature can increase the normal bias of an average; larger reach gives more room for smoothing under the paper's other assumptions.