Source
Coifman and Lafon (2006), “Diffusion maps”, Applied and Computational Harmonic Analysis 21(1): 5–30, LIT-1246, read in full in NOTE-985. The construction is §3.1. The limit is Theorem 2, proved as Proposition 10 in Appendix B, and the heat kernel is Proposition 3 = Proposition 11. The same family is announced without proof in the 2005 PNAS paper, LIT-981 (NOTE-784).
What was actually shown
Setting. M is a compact C^∞ submanifold of ℝⁿ. The density q > 0 is smooth, and U = −log q. The kernel is k_ε(x, y) = h(‖x − y‖²/ε), with h of exponential decay, scaled so that m₀ = 1 and m₂ = 2. Δ is the Laplace–Beltrami operator, taken positive semi-definite (Δ = −∇² on flat space). Then q_ε = ∫k_ε(·, y)q(y)dy, k^(α) = k_ε/(q_ε^α ⊗ q_ε^α), d^(α) = ∫k^(α)(·, y)q(y)dy, P_{ε,α}φ = ∫k^(α)(·, y)φ(y)q(y)dy / d^(α), and L_{ε,α} = (I − P_{ε,α})/ε.
The source’s Lemma 8 (read, not re-derived). At points farther than ε^γ from the boundary, G_εf = f + ε(ωf − Δf) + O(ε²) uniformly. Here G_ε is the kernel operator normalized by ε^{−d/2}, and ω is a potential set by the embedding.
The interior algebra, re-derived. The constant ε^{−d/2} cancels in every ratio below, so it is dropped.
- q_ε = G_ε q = q + ε(ωq − Δq) + O(ε²), so q_ε^{−α} = q^{−α}(1 + εα(Δq/q − ω)) + O(ε²).
- K^(α)φ = q_ε^{−α} G_ε[φ q q_ε^{−α}] = q_ε^{−α} G_ε[φq^{1−α}(1 + εα(Δq/q − ω))] + O(ε²) = q_ε^{−α} q^{1−α}[φ + ε((1 − α)ωφ + αφΔq/q − Δ(φq^{1−α})/q^{1−α})] + O(ε²).
- Taking φ ≡ 1: d^(α) = q_ε^{−α}q^{1−α}[1 + ε((1 − α)ω + αΔq/q − Δ(q^{1−α})/q^{1−α})] + O(ε²).
- Dividing, the ω and Δq/q terms cancel: P_{ε,α}φ = φ − ε[Δ(φq^{1−α})/q^{1−α} − φΔ(q^{1−α})/q^{1−α}] + O(ε²). So L_{ε,α}φ → Δ(φg)/g − φΔg/g with g = q^{1−α}, which is the source’s Theorem 2.
The drift form (the record’s). With Δ positive semi-definite, the product rule is Δ(φg) = gΔφ + φΔg − 2⟨∇φ, ∇g⟩. So Δ(φg)/g − φΔg/g = Δφ − 2⟨∇φ, ∇log g⟩ = Δφ + 2(1 − α)⟨∇U, ∇φ⟩, since log g = −(1 − α)U. Writing ρ = e^{−2(1−α)U} = q^{2(1−α)}, the operator is −ρ⁻¹∇·(ρ∇φ). It is symmetric and non-negative in L²(ρ dvol), and it annihilates constants. Its negative is the generator of dX = −2(1 − α)∇U(X)dt + √2 dW, reflected at the boundary, whose invariant density is ρ ∝ q^{2(1−α)}. The three cases:
- α = 0: drift −2∇U, invariant density q². This is NOTE-784’s “potential 2U”. The ordinary graph Laplacian over-weights dense regions.
- α = 1/2: drift −∇U, invariant density q. This is the Langevin equation (2) of the source, ẋ = −∇U + √2ẇ. If the samples were drawn from that dynamics’ equilibrium, α = 1/2 recovers the generator of the very dynamics that produced them. Acting on f, it is the backward operator, LIT-981’s wording. Its adjoint, conjugated by q, is the forward Fokker–Planck equation of §3.3.
- α = 1: no drift, uniform invariant density: Brownian motion on M.
The same exponent appears in the discrete stationary measure: by step 3, d^(α)·q ≈ q^{−α}q^{1−α}q = q^{2(1−α)} to leading order.
Numerical check. On the unit circle with q ∝ e^{0.8 cos θ}, the integrals were computed by quadrature on 4000 points with h(s) = e^{−s/4}/√(4π) (m₀ = 1, m₂ = 2 in d = 1). Apply the operator to f = cos 2θ + 0.3 sin θ. The maximum of |L_{ε,α}f − (−f″ + 2(1 − α)U′f′)| over θ, against a scale of about 5, was:
| ε | α = 0 | α = 1/2 | α = 1 |
|---|---|---|---|
| 4 × 10⁻³ | 5.5 × 10⁻² | 2.9 × 10⁻² | 2.3 × 10⁻² |
| 1 × 10⁻³ | 1.4 × 10⁻² | 7.2 × 10⁻³ | 5.9 × 10⁻³ |
So the error is O(ε). The chain’s stationary density matched q^{2(1−α)} to within 4 × 10⁻⁴ at ε = 10⁻³.
What the promotion review left open.
- Lemma 8, the kernel expansion on a submanifold, was read and not re-derived when this entry was filed. It is re-derived for a manifold without boundary in the note of 2026-10-20 below.
- The boundary step bounds L_{ε,α}φ by O(1) on a strip of measure O(ε^γ), and concludes R_ε = O(ε^γ) in L². The L² norm of that piece is O(ε^{γ/2}). The convergence survives, but the rate as printed does not.
- Proposition 11’s semigroup step replaces (I − εΔ)^{−1}(I − εΔ − εR_ε) by I − εR_ε. That needs R_ε to commute with Δ. The step also drops ‖(I − εΔ)^{−t/ε}‖, which grows like e^{tν_K²}, and it does not show that P_{ε,1} keeps E_K invariant. So the heat-kernel limit P_{ε,1}^{t/ε} → e^{−tΔ} is not proved by the source as written.
What could have failed: the curvature potential ω could have survived into the limit, which would make the α = 1 operator depend on the embedding and not only on the intrinsic geometry. It cancels in step 4.
What this does not say
- Not that finite samples behave this way at a given n. The finite-sample error is O(m^{−1/2}ε^{−d/4−1/2}) by the bounds the source cites, which needs m ≫ ε^{−d/2−1}. The source’s Criterion 4 states the square root of this.
- Not that α = 1/2 recovers any dynamics from any data. It recovers the reversible gradient (Langevin) dynamics whose equilibrium density is the sampling density, with isotropic unit diffusion. Non-equilibrium samples, anisotropic noise, or non-gradient drift are not covered.
- Not that the eigenfunctions converge. The statement is about the operator on band-limited functions. Spectral convergence of the finite-sample operator is a further result.
- Not the source’s §3.3 display. That display, Δφ − (‖∇U‖²/4 − ΔU/2)φ, uses the ∇² convention against the paper’s own convention, and it does not annihilate √q. The correct Schrödinger form is Δφ + (‖∇U‖²/4 + ΔU/2)φ with Δ positive, which is −∇²φ + (|∇U|²/4 − ∇²U/2)φ.
Connections
- LIT-981 and NOTE-784. This confirms and makes explicit their α-family takeaways (“backward Fokker–Planck” at α = 1/2, potential 2U at α = 0), now with the source of the proof.
- QUESTION-217. An operator-induced order depends on the normalization. The same samples give different slow modes at different α, so a “spectral order” of data is relative to a choice of how density enters.
- QUESTION-222. At α = 1/2 the slow eigenfunctions of the data’s operator are those of the generator of the dynamics that produced the data. That is the route by which Nadler et al. treat them as reaction coordinates, the Markovian slow variables of a reduced model. Whether those coordinates close without memory is not addressed here.
- CLAIM-331. For a reversible Langevin system, the slow modes this operator recovers are the system’s slow variables, which is one source of the timescale hierarchy that the claim’s “characteristically” runs together.
- QUESTION-231. The boundary rate and the semigroup limit, which the title no longer claims.
Note of 2026-10-20: the kernel expansion, re-derived
The promotion audit (batch 060) re-derived the source’s Lemma 8 at interior points of a compact manifold M without boundary, embedded in ℝⁿ. Take m₀ = ∫h(|u|²)du = 1 and m₂ = ∫u₁²h(|u|²)du = 2, with G_εf(x) = ε^{−d/2}∫_M h(‖x − y‖²/ε)f(y)dvol(y).
- Localization. M is compact and embedded, so ‖x − y‖ is bounded below by a multiple of the geodesic distance near x, and by a constant away from it. Points beyond geodesic distance ε^{1/2−κ} contribute O(ε^∞) uniformly, by the exponential decay of h. So work in normal coordinates y = exp_x(s).
- Expansions. ‖x − exp_x(s)‖² = |s|² − |II(s, s)|²/12 + O(|s|⁵), with II the second fundamental form. dvol = (1 − Ric(s, s)/6 + O(|s|³))ds. f(exp_x s) = f + s·∇f + ½sᵀHs + O(|s|³), where the trace of the coordinate Hessian H is −Δf at x.
- Scaling s = √ε u removes the factor ε^{−d/2}. The integrand is [h − εh′|II(u, u)|²/12][1 − εRic(u, u)/6][f + √ε u·∇f + (ε/2)uᵀHu], up to terms of order ε^{3/2} and ε².
- Parity. h(|u|²), |II(u, u)|² and Ric(u, u) are even in u. The √ε term and every ε^{3/2} term carry exactly one odd factor, so they vanish, and the remainder is O(ε²) uniformly in x.
- The order-ε term. ½∫h uᵀHu du = ½m₂ tr H = −Δf. The terms in f give εωf, with ω = −(1/12)∫h′|II(u, u)|²du − (1/6)∫h Ric(u, u)du, a smooth function set by the embedding.
So G_εf = f + ε(ωf − Δf) + O(ε²) uniformly, with no first-derivative term. That is what the interior algebra above needs: ω multiplies the function, and so it cancels in step 4 whatever its value. The function φq q_ε^{−α} in step 2 depends on ε, but smoothly and with bounded derivatives, so the O(ε²) holds for it too, and L_{ε,α}φ converges uniformly at rate O(ε). The sign is now in the title. With Δ ≥ 0, Δφ + 2(1 − α)⟨∇U, ∇φ⟩ is minus the generator ∇²φ − 2(1 − α)∇U·∇φ = ρ⁻¹∇·(ρ∇φ), with ρ = q