03.02.45 · differential-geometry / manifolds

Brownian motion on a Riemannian manifold

shipped3 tiersLean: none

Anchor (Master): Hsu — Stochastic Analysis on Manifolds (2002) Ch. 3-5; Elworthy — Stochastic Differential Equations on Manifolds (1982); Émery — Stochastic Calculus in Manifolds (1989); Bismut — Large Deviations and the Malliavin Calculus (1984)

Intuition Beginner

A speck of pollen jiggling in water traces a path so jagged it has no velocity at any instant: this is Brownian motion, the random walk taken to its continuous limit. On a flat sheet the recipe is plain. At each tiny step the particle jumps a small random amount in each coordinate direction, the jumps independent, and the coordinates never get in each other's way. You can write down where it goes by adding up independent wiggles, one per axis.

Now bend the sheet into a curved surface — a sphere, a saddle, a crumpled landscape. The trouble is that a curved space has no single set of coordinates that works everywhere, and no fixed "axis directions" that stay parallel as you move around. A frame of arrows you carry from one point to another comes back rotated. So the flat recipe, which leaned on fixed axes, has nothing to stand on.

Visual Beginner

Picture a small frame of arrows — an orthonormal frame — sitting at the particle's current spot, marking out the local directions. The particle takes a random nudge measured against that frame, moves a hair's breadth, and then the frame itself is slid along the new path so it stays as parallel as the curved surface allows. Repeat forever, infinitely fast. The frame is the bookkeeper that lets a flat random recipe run on a bent space.

The picture shows two layers at once. Below, on the surface, is the jagged random path the particle actually walks. Above it rides the frame, turning smoothly as the path curves. The random input is always fed in upstairs, against the frame, where the geometry looks momentarily flat; the curvature enters only through how the frame rotates as it is carried along.

Worked example Beginner

Take the round unit sphere and watch a Brownian particle's distance from the north pole, call it $r$ , which runs from $0$ at the pole to $π$ at the south pole. On a flat plane the distance from a fixed centre tends to grow, because there is more room far out than near in — a random walker drifts outward simply because outward is roomier. The same effect pushes $r$ on the sphere, but only up to the equator.

Past the equator the geometry reverses: circles of latitude start shrinking again as you approach the south pole, so now there is more room inward. The roominess that drove the particle outward now nudges it back. The drift in $r$ is governed by how fast the latitude circles grow, which on the unit sphere is measured by $cot r$ : positive (outward push) for $r$ below $π /2$ , negative (inward push) beyond it.

What this tells us: on a curved space the radial part of Brownian motion is not a plain random walk but a random walk with a built-in drift, and that drift is read straight off the geometry — the rate at which spheres around a point expand. Curvature controls the drift, and so controls whether the particle wanders off to infinity or stays put.

Check your understanding Beginner

Exercise (easy, multiple choice).

Why can the flat-space recipe for Brownian motion (independent wiggles, one per fixed coordinate axis) not be copied directly onto a curved surface?

A. Because random numbers behave differently on a curved surface B. Because a curved space has no fixed axis directions that stay parallel everywhere C. Because the particle moves faster on a curved surface D. Because curved surfaces have no notion of distance

Hint

Think about carrying a frame of arrows around a closed loop on a sphere and what happens to it.

Answer

B. Because a curved space has no fixed axis directions that stay parallel everywhere. A frame carried around a loop on a curved surface comes back rotated, so there is no single global set of axes to feed independent wiggles into. The fix is to carry a frame along the path and feed the randomness against that moving frame.

Formal definition Intermediate+

Let $(M, g)$ be a smooth connected Riemannian manifold of dimension $n$ with Levi-Civita connection $\nabla$ of 03.02.27. Brownian motion on $M$ is the diffusion process whose infinitesimal generator is $\frac{1}{2} Δ$ , where $Δ$ is the Laplace-Beltrami operator $Δ = div grad$ of 03.02.15 (geometer's sign, nonpositive spectrum). On flat $R^{n}$ this is the standard heat-equation generator and the process is ordinary Brownian motion. The construction problem is that there is no canonical chart in which to write independent driving noises; the resolution, due to Eells, Elworthy, and Malliavin, is to lift the equation to a bundle where a canonical frame exists.

Let $π : O (M) \to M$ be the orthonormal frame bundle of 03.05.15: a point $u \in O (M)$ is a pair $(x, e)$ with $x = π (u) \in M$ and $e = (e_{1}, \dots, e_{n})$ a $g$ -orthonormal basis of $T_{x} M$ . Equivalently $u$ is a linear isometry $u : R^{n} \to T_{x} M$ . The Levi-Civita connection splits each tangent space $T_{u} O (M) = H_{u} \oplus V_{u}$ into a horizontal subspace $H_{u}$ and the vertical subspace $V_{u} = ker d π_{u}$ . For each $a \in R^{n}$ the canonical (fundamental) horizontal vector field $H_{a}$ is defined by

H_{a} (u) = the horizontal lift of u (a) \in T_{x} M to T_{u} O (M),

so that $d π_{u} H_{a} (u) = u (a)$ and $H_{a} (u) \in H_{u}$ . Writing $H_{i} = H_{ε_{i}}$ for the standard basis $ε_{1}, \dots, ε_{n}$ of $R^{n}$ , the fields $H_{1}, \dots, H_{n}$ frame the horizontal distribution.

Definition (Eells-Elworthy-Malliavin Brownian motion). Let $B = (B^{1}, \dots, B^{n})$ be a standard $R^{n}$ -valued Brownian motion on a filtered probability space, and let $U_{t}$ solve the Stratonovich stochastic differential equation on $O (M)$

d U_{t} = i = 1 \sum n H_{i} (U_{t}) \circ d B_{t}^{i}, U_{0} = u_{0} \in π^{- 1} (x_{0}),

up to its (possibly finite) explosion time $ζ$ . Then the projection $X_{t} = π (U_{t})$ is a Brownian motion on $M$ started at $x_{0}$ : a diffusion with generator $\frac{1}{2} Δ$ , defined for $t < ζ$ . The lift $U_{t}$ is the horizontal Brownian motion, and the linear isometry $u_{t} \circ u_{0}^{- 1} : T_{x_{0}} M \to T_{X_{t}} M$ is stochastic parallel transport along the path $X$ .

The Stratonovich convention is forced. Stratonovich integrals obey the ordinary chain rule, so the equation transforms as a genuine geometric object under change of coordinates on $O (M)$ ; the Itô form would acquire chart-dependent second-order correction terms and fail to define a connection-invariant process. This invariance is exactly why the construction is stated on the frame bundle with Stratonovich noise rather than in a single chart with Itô noise (see 02.15.05).

Counterexamples to common slips

Writing $\sum_{i} H_{i} \circ d B^{i}$ with Itô rather than Stratonovich differentials does not give a $\frac{1}{2} Δ$ -diffusion: the Itô version is not coordinate-invariant on $O (M)$ , and the conversion term changes the generator. The Stratonovich-to-Itô correction is precisely what restores $\frac{1}{2} \sum_{i} H_{i}^{2}$ on the bundle.
The horizontal sum $\sum_{i} H_{i}^{2}$ on $O (M)$ is Bochner's horizontal Laplacian $Δ_{O (M)}$ , not the Laplace-Beltrami operator of $O (M)$ for any metric. Its defining property is that it is $π$ -related to $Δ$ on $M$ : $Δ_{O (M)} (f \circ π) = (Δ f) \circ π$ . Mistaking it for an intrinsic Laplacian of the bundle loses this relation.
Brownian motion need not exist for all time. On a manifold that is geodesically complete the process can still explode in finite time if the volume of balls grows too fast; geodesic completeness and stochastic completeness are different conditions.

Key theorem with proof Intermediate+

Theorem (Eells-Elworthy-Malliavin; the projected horizontal diffusion has generator $\frac{1}{2} Δ$ ). Let $U_{t}$ solve $d U_{t} = \sum_{i} H_{i} (U_{t}) \circ d B_{t}^{i}$ on $O (M)$ and set $X_{t} = π (U_{t})$ . Then for every $f \in C^{\infty} (M)$ and $t < ζ$ , $$ df(X_t) = \sum_{i=1}^n (H_i \widetilde f)(U_t), dB_t^i + \tfrac12 (\Delta f)(X_t), dt, \qquad \widetilde f := f\circ\pi, $$ so $X_{t}$ is a diffusion with generator $\frac{1}{2} Δ$ ; that is, $f (X_{t}) - \frac{1}{2} \int_{0}^{t} (Δ f) (X_{s}) d s$ is a local martingale.

Proof. Work with the lifted function $f = f \circ π \in C^{\infty} (O (M))$ . Stratonovich calculus obeys the chain rule, so along the solution $U_{t}$ , $$ d\widetilde f(U_t) = \sum_{i=1}^n (H_i \widetilde f)(U_t)\circ dB_t^i . $$ Convert each Stratonovich differential to Itô form. For a Stratonovich integral $\int Y \circ d B^{i}$ one has $\int Y \circ d B^{i} = \int Y d B^{i} + \frac{1}{2} ⟨ Y, B^{i} ⟩$ , and here $Y = H_{i} f (U_{\cdot})$ has Itô differential $d Y = \sum_{j} (H_{j} H_{i} f) d B^{j} + (\dots) d t$ by the same calculus applied once more. Since $⟨ B^{j}, B^{i} ⟩ = δ_{ij} t$ , the cross-variation contributes $\frac{1}{2} \sum_{i} (H_{i} H_{i} f) d t$ . Therefore $$ d\widetilde f(U_t) = \sum_{i=1}^n (H_i \widetilde f)(U_t), dB_t^i + \tfrac12\sum_{i=1}^n (H_i^2 \widetilde f)(U_t), dt . $$ The drift coefficient is the horizontal Laplacian $Δ_{O (M)} = \sum_{i} H_{i}^{2}$ applied to $f$ . Bochner's identity for the canonical horizontal fields states that $Δ_{O (M)}$ is $π$ -related to the Laplace-Beltrami operator: for any $f \in C^{\infty} (M)$ , $$ \sum_{i=1}^n H_i^2 (f\circ\pi) = (\Delta f)\circ\pi . $$

This identity is the computational heart. Fix $u_{0} = (x_{0}, e)$ and evaluate at $u_{0}$ . Because $H_{i}$ is horizontal, $H_{i} f (u_{0}) = d f_{x_{0}} (e_{i})$ , the directional derivative of $f$ along the orthonormal vector $e_{i}$ . Applying $H_{j}$ again and using that the horizontal lift of $e_{i}$ has, at $u_{0}$ , vanishing Christoffel correction in normal coordinates centred at $x_{0}$ adapted to the frame $e$ , one gets $H_{j} H_{i} f (u_{0}) = Hess f (e_{j}, e_{i}) + (terms killed by the trace)$ . Summing the diagonal $i = j$ over the orthonormal frame yields $\sum_{i} Hess f (e_{i}, e_{i}) = tr_{g} Hess f = Δ f (x_{0})$ . The off-diagonal connection terms cancel because $e$ is orthonormal and parallel-transported, which is exactly the horizontality of $H_{i}$ . Substituting back, $$ d f(X_t) = d\widetilde f(U_t) = \sum_{i=1}^n (H_i\widetilde f)(U_t), dB_t^i + \tfrac12 (\Delta f)(X_t), dt, $$ the stated decomposition. The first term is a local martingale and the second exhibits $\frac{1}{2} Δ$ as the generator. $□$

Bridge. This identity builds toward the entire probabilistic theory of the heat semigroup: it says the random path $X_{t}$ carries the operator $\frac{1}{2} Δ$ on its back, so the heat semigroup $e^{t Δ/2}$ is realised as the expectation $E_{x} [f (X_{t})]$ , and this is exactly the Feynman-Kac representation that appears again in the probabilistic heat kernel of 03.09.29. The foundational reason the construction is intrinsic is that the horizontal Laplacian $\sum_{i} H_{i}^{2}$ is $π$ -related to $Δ$ regardless of which frame $u_{0}$ we start from; the frame-bundle lift trades the missing canonical chart on $M$ for the canonical horizontal frame on $O (M)$ , and the central insight is that this trade is exact — no curvature correction leaks into the generator, only into the rotation of the frame. The same Stratonovich-on-a-bundle device is dual to the deterministic horizontal lift of 03.05.15: deterministic parallel transport solves $\overset{u}{˙} = H_{\overset{γ}{˙}} (u)$ along a smooth curve, and stochastic parallel transport solves the very same equation driven by $\circ d B$ . Putting these together, Itô's formula on $M$ , the heat semigroup, and stochastic parallel transport are three faces of one lifted SDE, and that SDE generalises the flat-space recipe by replacing fixed axes with the moving orthonormal frame.

Exercises Intermediate+

Exercise 1 (easy, short-answer).

Show that on flat $R^{n}$ with the standard frame, the Eells-Elworthy-Malliavin SDE reduces to $d X_{t} = d B_{t}$ , ordinary Brownian motion.

Hint

On flat space the connection vanishes: the horizontal lift of $ε_{i}$ is just $\partial_{i}$ , and the frame never rotates, so $O (R^{n}) = R^{n} \times O (n)$ and the frame part is constant.

Answer

On $R^{n}$ with the Euclidean metric the Christoffel symbols vanish, so the horizontal subspace at $(x, e)$ is spanned by $\partial_{x^{1}}, \dots, \partial_{x^{n}}$ and $H_{i} (x, e) = \sum_{j} e_{i}^{j} \partial_{x^{j}}$ . Taking $e = Id$ at $t = 0$ , the frame stays constant ( $d e = 0$ since the connection form vanishes on horizontal vectors), so $H_{i} = \partial_{x^{i}}$ throughout. The Stratonovich SDE becomes $d X_{t} = \sum_{i} \partial_{x^{i}} \circ d B_{t}^{i}$ , i.e. $d X_{t}^{j} = d B_{t}^{j}$ . Stratonovich and Itô agree because the coefficients are constant. Hence $X_{t} = x_{0} + B_{t}$ . Rubric: full credit for vanishing connection, constant frame, and the constant-coefficient Stratonovich=Itô reduction.

Exercise 2 (medium, symbolic).

Verify Bochner's relation $\sum_{i} H_{i}^{2} (f \circ π) = (Δ f) \circ π$ in dimension $1$ : take $M = R$ with a metric $g = φ (x)^{2} d x^{2}$ , compute $Δ f$ , and confirm it matches the horizontal Laplacian.

Hint

In one dimension $O (M)$ has a single orthonormal frame field $e_{1} = φ^{- 1} \partial_{x}$ up to sign, and $Δ f = φ^{- 1} \partial_{x} (φ^{- 1} \partial_{x} f) + (\partial_{x} lo g φ) φ^{- 2} \partial_{x} f$ . Use $Δ f = \frac{1}{∣ g ∣} \partial_{x} (∣ g ∣ g^{xx} \partial_{x} f)$ .

Answer

Here $∣ g ∣ = φ$ and $g^{xx} = φ^{- 2}$ , so $Δ f = φ^{- 1} \partial_{x} (φ \cdot φ^{- 2} \partial_{x} f) = φ^{- 1} \partial_{x} (φ^{- 1} \partial_{x} f)$ . The single horizontal field acts as the orthonormal derivative $H_{1} = φ^{- 1} \partial_{x}$ followed by horizontal transport of the frame; carrying the frame parallel along $e_{1}$ adds the connection term, giving $H_{1}^{2} f = φ^{- 1} \partial_{x} (φ^{- 1} \partial_{x} f)$ after the frame correction $Γ = \partial_{x} lo g φ$ is included. Both equal $φ^{- 1} \partial_{x} (φ^{- 1} \partial_{x} f)$ , confirming $H_{1}^{2} (f \circ π) = (Δ f) \circ π$ . Rubric: full credit for the metric Laplacian formula and matching the horizontal second derivative with its connection correction.

Exercise 3 (medium, short-answer).

Explain why the Stratonovich convention, not Itô, is the correct one for defining Brownian motion via the frame-bundle SDE.

Hint

Compare how each convention transforms under a change of coordinates on $O (M)$ . Which one obeys the ordinary chain rule?

Answer

Stratonovich integrals satisfy the ordinary (first-order) chain rule, so an SDE written as $\sum_{i} V_{i} \circ d B^{i}$ for vector fields $V_{i}$ is invariant under smooth change of coordinates: the same geometric object is described in any chart. The Itô integral instead obeys a second-order chain rule, so $\sum_{i} V_{i} d B^{i}$ acquires a chart-dependent drift correction $\frac{1}{2} \sum_{i} \partial V_{i} \cdot V_{i}$ under coordinate change, which is not a connection-invariant vector field. Defining the process with Stratonovich noise on $O (M)$ makes $X_{t} = π (U_{t})$ depend only on the connection and metric, not on coordinates; the Stratonovich-to-Itô correction is precisely the term that produces the geometrically correct generator $\frac{1}{2} Δ$ . Rubric: full credit for the chain-rule/invariance contrast and the connection to the generator.

Exercise 4 (medium, short-answer).

State the Itô formula for $f (X_{t})$ on a manifold and identify its martingale and drift parts in terms of $grad f$ and $Δ f$ .

Hint

Use the theorem's decomposition $df (X_{t}) = \sum_{i} (H_{i} f) (U_{t}) d B_{t}^{i} + \frac{1}{2} (Δ f) (X_{t}) d t$ and read $H_{i} f (U_{t}) = ⟨ grad f (X_{t}), u_{t} ε_{i} ⟩$ .

Answer

Since $H_{i} f (u) = df (u ε_{i}) = ⟨ grad f, u ε_{i} ⟩$ , the martingale term is $\sum_{i} ⟨ grad f (X_{t}), u_{t} ε_{i} ⟩ d B_{t}^{i}$ , the anti-development of $grad f$ against the driving noise; its quadratic variation is $∣ grad f (X_{t}) ∣^{2} d t$ . The drift is $\frac{1}{2} (Δ f) (X_{t}) d t$ . So $f (X_{t}) = f (X_{0}) + M_{t} + \frac{1}{2} \int_{0}^{t} Δ f (X_{s}) d s$ with $M$ a local martingale of quadratic variation $\int_{0}^{t} ∣ grad f ∣^{2} d s$ . This is the manifold Itô formula. Rubric: full credit for both parts and the quadratic-variation identification.

Exercise 5 (hard, short-answer).

On the unit sphere $S^{n}$ , derive the stochastic differential equation for the radial process $r_{t} = d (o, X_{t})$ (distance from a fixed point $o$ ), away from $o$ and its antipode.

Hint

Apply the manifold Itô formula to $f = r$ (the distance function). Use $∣ grad r ∣ = 1$ and $Δ r = (n - 1) cot r$ on the sphere, away from the cut locus.

Answer

Away from $o$ and its antipode the distance function $r$ is smooth with $∣ grad r ∣ = 1$ , and the mean curvature of geodesic spheres gives $Δ r = (n - 1) cot r$ on $S^{n}$ . Itô's formula yields $$ dr_t = \beta_t + \tfrac12 \Delta r,(X_t),dt = d\beta_t + \tfrac{n-1}{2}\cot(r_t),dt, $$ where $β_{t} = \int_{0}^{t} ⟨ grad r, u_{s} d B_{s} ⟩$ is a one-dimensional Brownian motion because its quadratic variation is $\int_{0}^{t} ∣ grad r ∣^{2} d s = t$ (Lévy's characterisation). So the radial process is a one-dimensional diffusion with drift $\frac{n - 1}{2} cot r$ : positive (outward) for $r < π /2$ , negative beyond. Rubric: full credit for the Itô computation, the Lévy identification of $β$ , and the explicit $cot$ drift.

Exercise 6 (hard, short-answer).

Show that stochastic parallel transport $/ /_{t} = u_{t} u_{0}^{- 1} : T_{x_{0}} M \to T_{X_{t}} M$ is, for each $t$ , a linear isometry, and explain the sense in which it solves the same equation as deterministic parallel transport.

Hint

Each $u_{t} \in O (M)$ is by construction a linear isometry $R^{n} \to T_{X_{t}} M$ ; the horizontal SDE keeps $U_{t}$ inside $O (M)$ . Compare $d U_{t} = \sum_{i} H_{i} (U_{t}) \circ d B_{t}^{i}$ with the deterministic horizontal lift equation $\overset{u}{˙} = H_{\overset{γ}{˙}} (u)$ .

Answer

The horizontal vector fields $H_{i}$ are tangent to $O (M)$ , so the solution $U_{t}$ stays in the frame bundle: each $u_{t}$ is a linear isometry $R^{n} \to T_{X_{t}} M$ preserving the inner products by definition of $O (M)$ . Hence $/ /_{t} = u_{t} u_{0}^{- 1}$ is a composition of isometries, itself an isometry $T_{x_{0}} M \to T_{X_{t}} M$ . Deterministic parallel transport along a smooth curve $γ$ solves $\overset{u}{˙}_{t} = H_{\overset{γ}{˙} (t)} (u_{t})$ , the horizontal lift of $\overset{γ}{˙}$ ; the stochastic transport solves the identical equation with the smooth driver $\overset{γ}{˙} d t$ replaced by the Stratonovich differential $\sum_{i} u_{t} ε_{i} \circ d B_{t}^{i}$ of the random path. Because Stratonovich calculus obeys the chain rule, the two are the same geometric ODE, one driven smoothly and one driven by white noise. Rubric: full credit for the isometry argument (staying in $O (M)$ ) and the same-equation comparison via Stratonovich.

Advanced results Master

The radial process governs whether Brownian motion can run for all time. Fix $o \in M$ , let $r (x) = d (o, x)$ , and away from the cut locus apply the manifold Itô formula to $r$ : as in Exercise 5, $r_{t} = r_{0} + β_{t} + \frac{1}{2} \int_{0}^{t} Δ r (X_{s}) d s$ with $β$ a one-dimensional Brownian motion and $Δ r$ the mean curvature of the geodesic sphere through $X_{s}$ . The Laplacian comparison theorem of 03.02.05 bounds this drift: if $Ric \geq (n - 1) κ g$ , then $Δ r \leq (n - 1) \frac{sn _{κ}^{'} ( r )}{sn _{κ} ( r )}$ in the barrier sense, where $sn_{κ}$ is the model sine function ( $sn_{κ} (r) = sin (κ r) / κ$ for $κ > 0$ , $r$ for $κ = 0$ , $sinh (- κ r) / - κ$ for $κ < 0$ ). A Ricci lower bound therefore caps the outward drift of $r_{t}$ , and comparison with the dominating one-dimensional diffusion controls the radial process from above.

The payoff is a curvature criterion for stochastic completeness. A manifold is stochastically complete when Brownian motion has infinite lifetime, $ζ = \infty$ almost surely, equivalently when the heat semigroup conserves probability, $\int_{M} p_{t} (x, y) d vol (y) = 1$ for all $t$ . Comparison with the model radial diffusion gives the Yau-Dodziuk theorem: a complete Riemannian manifold with Ricci curvature bounded below, $Ric \geq - (n - 1) K$ for a constant $K \geq 0$ , is stochastically complete. The probabilistic mechanism is that the drift $\frac{n - 1}{2} \frac{sn _{K}^{'} ( r )}{sn _{K} ( r )} \sim \frac{n - 1}{2} K$ grows at most linearly, so $r_{t}$ cannot reach infinity in finite time any faster than the explicitly solvable model process, which is itself conservative. Stochastic completeness is strictly weaker than geodesic completeness: there are geodesically complete manifolds, with volume growing faster than $e^{c r^{2}}$ , on which Brownian motion explodes — the Grigor'yan volume test $\int^{\infty} \frac{r d r}{l o g vol B ( o , r )} = \infty$ is the sharp criterion, and it is implied by but does not imply a Ricci lower bound.

Stochastic parallel transport upgrades Itô's formula to tensors and yields the probabilistic side of the Bochner method of 03.02.15. For a one-form $α$ , the process $/ /_{t}^{- 1} α (X_{t})$ is an $R^{n}$ -valued semimartingale whose drift is $- \frac{1}{2} / /_{t}^{- 1} (Δ_{H} α) (X_{t})$ , where $Δ_{H}$ is the Hodge-de Rham Laplacian and the Weitzenböck identity $Δ_{H} = \nabla^{*} \nabla + Ric$ inserts the Ricci tensor exactly where the Laplacian comparison inserted it for the radial process. This is the source of probabilistic gradient estimates: $\nabla (e^{t Δ/2} f) = E [/ /_{t}^{- 1} \nabla f (X_{t}) Q_{t}]$ with $Q_{t}$ solving the matrix ODE $\dot{Q}_{t} = - \frac{1}{2} Ric (X_{t}) Q_{t}$ , the Bismut formula for the derivative of the heat semigroup, which makes a Ricci lower bound into an exponential contraction estimate on $∣\nabla e^{t Δ/2} f ∣$ .

The horizontal Laplacian $Δ_{O (M)} = \sum_{i} H_{i}^{2}$ is the operator the whole construction is organised around. It is Hörmander's sum-of-squares form for $\frac{1}{2} Δ$ lifted to $O (M)$ : the canonical horizontal fields $H_{1}, \dots, H_{n}$ together with the fundamental vertical fields of $o (n)$ satisfy Hörmander's bracket condition, so $Δ_{O (M)}$ is hypoelliptic and its diffusion has a smooth transition density. Bismut's lift of the heat semigroup to $O (M)$ runs the entire heat-kernel asymptotic analysis upstairs, where the geometry is encoded in the brackets $[H_{i}, H_{j}]$ — which are vertical and measure curvature — rather than in coordinate Christoffel symbols.

Synthesis. The frame-bundle construction is the foundational reason that a single lifted Stratonovich SDE simultaneously produces the diffusion, its generator, and its parallel transport: the canonical horizontal fields $H_{i}$ carry the connection, their squares sum to the horizontal Laplacian, and that operator is $π$ -related to $Δ$ — this is the central insight that converts a geometric structure into a probabilistic one without loss. Putting these together, the radial process, the Laplacian comparison theorem, and the Weitzenböck identity are three appearances of one fact, that curvature controls the second-order behaviour of the flow: in the radial process curvature enters as the drift of $r_{t}$ , in the Bochner method it enters as the Ricci term of the drift of $/ /_{t}^{- 1} α$ , and in the Bismut formula it enters as the matrix ODE damping the transported gradient. This is exactly the pattern that recurs in the probabilistic heat kernel of 03.09.29, where the short-time asymptotics of $p_{t} (x, y)$ are computed by Brownian bridges whose concentration is governed by the same comparison geometry; the Ricci lower bound is dual to stochastic completeness on one side and to the gradient estimate on the other, and the horizontal Laplacian is the bridge that makes both computable upstairs on $O (M)$ . The deterministic horizontal lift of 03.05.15 generalises to the stochastic one by the single substitution $\overset{γ}{˙} d t \mapsto \circ d B$ , and that substitution is the entire content of doing analysis on paths rather than on curves.

Full proof set Master

The central decomposition and the Bochner $π$ -relation are proved in the Key theorem section. The remaining Master claims are recorded here.

Proposition (the projected process is independent of the lift $u_{0}$ in law). Let $u_{0}, u_{0}^{'}$ be two frames over the same point $x_{0}$ , related by $u_{0}^{'} = u_{0} a$ for some $a \in O (n)$ . Let $U_{t}, U_{t}^{'}$ solve the horizontal SDE from $u_{0}, u_{0}^{'}$ driven by Brownian motions $B, B^{'}$ with $B^{'} = a^{- 1} B$ . Then $π (U_{t}) = π (U_{t}^{'})$ for all $t$ , and in particular the law of $X_{t} = π (U_{t})$ does not depend on the choice of initial frame.

Proof. The canonical horizontal fields are $O (n)$ -equivariant: for the right action $R_{a}$ of $a \in O (n)$ on $O (M)$ one has $(R_{a})_{*} H_{i} = \sum_{j} (a^{- 1})_{j i} H_{j}$ , because $H_{a} (u) =$ horizontal lift of $u (a)$ and $R_{a} u = u \circ a$ . Set $V_{t} = R_{a} U_{t} = U_{t} a$ . By equivariance and the Stratonovich chain rule, $$ dV_t = (R_a)* \sum_i H_i(U_t)\circ dB^i_t = \sum{i,j} (a^{-1})_{ji} H_j(V_t)\circ dB^i_t = \sum_j H_j(V_t)\circ d(a^{-1}B)^j_t . $$ Thus $V_{t}$ solves the horizontal SDE from $V_{0} = u_{0} a = u_{0}^{'}$ driven by $a^{- 1} B = B^{'}$ , i.e. $V_{t} = U_{t}^{'}$ . Since $R_{a}$ acts in the fibre direction, $π (V_{t}) = π (U_{t})$ , so $π (U_{t}^{'}) = π (U_{t})$ . Because $B^{'} = a^{- 1} B$ has the same law as $B$ ( $a^{- 1} \in O (n)$ is orthogonal and Brownian motion is rotation-invariant), the law of $X_{t}$ is unchanged. $□$

Proposition (Laplacian comparison drift bound on the radial process). Let $Ric \geq (n - 1) κ g$ on a complete $M$ , and let $r_{t} = d (o, X_{t})$ . Then, in the sense of the comparison with the model diffusion, the drift of $r_{t}$ satisfies $\frac{1}{2} Δ r (X_{t}) \leq \frac{n - 1}{2} \frac{sn _{κ}^{'} ( r _{t} )}{sn _{κ} ( r _{t} )}$ for $X_{t}$ outside the cut locus of $o$ , and $r_{t}$ is stochastically dominated by the one-dimensional diffusion $d ρ_{t} = d β_{t} + \frac{n - 1}{2} \frac{sn _{κ}^{'} ( ρ _{t} )}{sn _{κ} ( ρ _{t} )} d t$ .

Proof. Outside the cut locus $r$ is smooth and $Δ r$ equals the trace of the second fundamental form of the geodesic sphere through $X_{t}$ , i.e. the mean curvature. The Riccati comparison underlying the Laplacian comparison theorem of 03.02.05 states that under $Ric \geq (n - 1) κ g$ this mean curvature is at most that of the model sphere of radius $r$ in the constant-curvature space form, which is $(n - 1) sn_{κ}^{'} (r) / sn_{κ} (r)$ ; the inequality holds globally in the barrier (distributional) sense by Calabi's trick at the cut locus, where $r$ acquires a nonpositive singular contribution that only decreases the drift. Apply Itô's formula to $r$ : the martingale part $β_{t}$ has quadratic variation $\int_{0}^{t} ∣ grad r ∣^{2} d s = t$ , hence is a standard Brownian motion by Lévy's characterisation, and the drift is $\frac{1}{2} Δ r (X_{s}) d s$ , bounded above by the model drift. A one-dimensional comparison theorem for SDEs with the same driving martingale then gives the pathwise domination $r_{t} \leq ρ_{t}$ for the model diffusion $ρ$ . $□$

Proposition (Yau-Dodziuk stochastic completeness). A complete Riemannian manifold with $Ric \geq - (n - 1) K$ , $K \geq 0$ , is stochastically complete: Brownian motion has infinite lifetime almost surely.

Proof. The explosion time is $ζ = lim_{m \to \infty} τ_{m}$ where $τ_{m} = in f {t : r_{t} \geq m}$ are the exit times of geodesic balls; explosion means $r_{t} \to \infty$ in finite time. By the previous proposition with $κ = - K$ , $r_{t}$ is dominated by the model diffusion $ρ_{t}$ with drift $\frac{n - 1}{2} K coth (K ρ_{t}) \leq \frac{n - 1}{2} K + \frac{n - 1}{2 ρ _{t}}$ , which is bounded for $ρ$ away from $0$ and grows at most linearly. A one-dimensional diffusion with at-most-linear drift and unit diffusion coefficient is non-explosive: by Feller's test, $\int^{\infty} exp (- \int^{ρ} 2 b (s) d s) d ρ = \infty$ because the linear drift makes the integrand decay only like $ρ^{- (n - 1)} e^{- c ρ}$ paired against the volume factor, and the Khasminskii non-explosion criterion applies. Hence $ρ_{t}$ , and therefore $r_{t} \leq ρ_{t}$ , stays finite for all finite $t$ , so $ζ = \infty$ almost surely. $□$

Connections Master

Sectional, Ricci, and scalar curvature 03.02.05 enter the moment a Brownian path is asked to live for all time. The Laplacian comparison theorem proved there from a Ricci lower bound is exactly the bound on the drift of the radial process $r_{t}$ used in the stochastic-completeness criterion; the Ricci tensor that appears in the Weitzenböck identity is the same tensor that damps the Bismut gradient formula. Curvature is not decoration here — it is the second-order coefficient of the diffusion, and every analytic estimate on Brownian motion is a curvature estimate in disguise.

The Levi-Civita connection and exponential map 03.02.27 provide the horizontal lift that the entire construction rests on. The horizontal subspace $H_{u} \subset T_{u} O (M)$ is defined by that connection; the canonical horizontal fields $H_{i}$ are horizontal lifts of frame vectors; and stochastic parallel transport solves the stochastic version of the parallel-transport ODE of that unit. The exponential map supplies the normal coordinates in which the Bochner $π$ -relation $\sum_{i} H_{i}^{2} (f \circ π) = (Δ f) \circ π$ is computed cleanly, with the Christoffel terms vanishing at the centre.

The Bochner technique 03.02.15 is the analytic twin of this unit's probabilistic machinery. The Laplace-Beltrami and Hodge-de Rham Laplacians studied there are the generators whose semigroups Brownian motion represents; the Weitzenböck decomposition $Δ_{H} = \nabla^{*} \nabla + Ric$ is what turns Itô's formula for one-forms into a vanishing theorem, and the horizontal Laplacian $\sum_{i} H_{i}^{2}$ is the frame-bundle avatar of the Bochner Laplacian $\nabla^{*} \nabla$ . Where that unit integrates against the manifold, this one integrates against the random path.

Linear connections, the frame bundle, and the soldering form 03.05.15 are the geometric stage. The orthonormal frame bundle $O (M)$ , its horizontal distribution, and the canonical horizontal fields built from the soldering form are precisely the objects on which the driving SDE is written. Deterministic horizontal lift there becomes stochastic horizontal lift here by replacing the smooth driver with a Brownian one, and the curvature read off from the bracket $[H_{i}, H_{j}]$ being vertical is the same curvature that obstructs flat parallel transport.

Stochastic differential equations and Itô's formula 02.15.03 and the Stratonovich integral 02.15.05 are the analytic prerequisites realised on a manifold by this unit. The frame-bundle SDE is a Stratonovich SDE in the exact sense of 02.15.05, chosen because only the Stratonovich convention transforms as a geometric object under change of chart on $O (M)$ ; the manifold Itô formula is the chart-free upgrade of the Itô formula of 02.15.03, with the second-order term organised by the connection rather than by coordinate second derivatives.

The probabilistic heat kernel 03.09.29 is the principal downstream consumer. Brownian motion on $M$ represents the heat semigroup $e^{t Δ/2}$ through $E_{x} [f (X_{t})]$ , so the transition density $p_{t} (x, y)$ of this process is the heat kernel, and its short-time asymptotics are computed by the concentration of Brownian bridges along the minimal geodesic — the geometric foundation that unit builds its analysis on.

Historical & philosophical context Master

The idea that the heat operator on a curved space should have a stochastic counterpart goes back to the 1950s, but the intrinsic construction was crystallised by James Eells and K. David Elworthy and, independently with a different emphasis, by Paul Malliavin, in the late 1960s and 1970s. The decisive move — solving a stochastic differential equation for an orthonormal frame on the bundle $O (M)$ and projecting down — appears in Elworthy's lectures and in his 1982 Stochastic Differential Equations on Manifolds ^{[Elworthy 1982]}, and it is the lift that converts the coordinate-bound flat recipe into a connection-invariant one. Malliavin's parallel programme, developed to give a probabilistic proof of Hörmander's hypoellipticity theorem, used the same horizontal lift and grew into the Malliavin calculus; Jean-Michel Bismut's Large Deviations and the Malliavin Calculus ^{[Bismut 1984]} carried the lifted heat semigroup on $O (M)$ to a probabilistic proof of the Atiyah-Singer index theorem.

The use of the Stratonovich rather than the Itô integral is the technical hinge on which intrinsic meaning turns: only the Stratonovich differential obeys the chain rule and so survives change of coordinates as a geometric object, a point Kiyosi Itô himself and later Michel Émery in Stochastic Calculus in Manifolds ^{[Émery 1989]} made into the organising principle of manifold-valued stochastic calculus. The curvature criteria for stochastic completeness were settled by Shing-Tung Yau and Józef Dodziuk via the Laplacian comparison method, and sharpened to a volume-growth test by Alexander Grigor'yan; Elton Hsu's Stochastic Analysis on Manifolds ^{[Hsu 2002]} is the modern synthesis of the whole subject, from the frame-bundle construction through the Bismut gradient formula to the comparison geometry of the radial process.

Bibliography Master

@book{hsu2002,
  author    = {Hsu, Elton P.},
  title     = {Stochastic Analysis on Manifolds},
  series    = {Graduate Studies in Mathematics},
  volume    = {38},
  publisher = {American Mathematical Society},
  year      = {2002}
}

@book{elworthy1982,
  author    = {Elworthy, K. David},
  title     = {Stochastic Differential Equations on Manifolds},
  series    = {London Mathematical Society Lecture Note Series},
  volume    = {70},
  publisher = {Cambridge University Press},
  year      = {1982}
}

@book{ikedawatanabe1989,
  author    = {Ikeda, Nobuyuki and Watanabe, Shinzo},
  title     = {Stochastic Differential Equations and Diffusion Processes},
  edition   = {2nd},
  publisher = {North-Holland / Kodansha},
  year      = {1989}
}

@book{emery1989,
  author    = {\'Emery, Michel},
  title     = {Stochastic Calculus in Manifolds},
  series    = {Universitext},
  publisher = {Springer-Verlag, Berlin},
  year      = {1989},
  note      = {With an appendix by P.-A. Meyer}
}

@book{bismut1984,
  author    = {Bismut, Jean-Michel},
  title     = {Large Deviations and the Malliavin Calculus},
  series    = {Progress in Mathematics},
  volume    = {45},
  publisher = {Birkh\"auser, Boston},
  year      = {1984}
}

@article{dodziuk1983,
  author  = {Dodziuk, J\'ozef},
  title   = {Maximum principle for parabolic inequalities and the heat flow on open manifolds},
  journal = {Indiana University Mathematics Journal},
  volume  = {32},
  number  = {5},
  pages   = {703--716},
  year    = {1983}
}

@article{grigoryan1999,
  author  = {Grigor'yan, Alexander},
  title   = {Analytic and geometric background of recurrence and non-explosion of the Brownian motion on Riemannian manifolds},
  journal = {Bulletin of the American Mathematical Society},
  volume  = {36},
  number  = {2},
  pages   = {135--249},
  year    = {1999}
}

Prerequisites

03.02.05
03.02.27
03.02.15
03.05.15

Tier anchors

beginner: Hsu — Stochastic Analysis on Manifolds (2002) Ch. 1-2 (informal); Stroock — An Introduction to the Analysis of Paths on a Riemannian Manifold (2000) Ch. 1
intermediate: Hsu — Stochastic Analysis on Manifolds (2002) Ch. 2-3 (Eells-Elworthy-Malliavin construction); Ikeda-Watanabe — Stochastic Differential Equations and Diffusion Processes (1989) Ch. V
master: Hsu — Stochastic Analysis on Manifolds (2002) Ch. 3-5; Elworthy — Stochastic Differential Equations on Manifolds (1982); Émery — Stochastic Calculus in Manifolds (1989); Bismut — Large Deviations and the Malliavin Calculus (1984)

References

Hsu — Stochastic Analysis on Manifolds (Graduate Studies in Mathematics 38, AMS, 2002) · Ch. 2 (horizontal lift, Eells-Elworthy-Malliavin), Ch. 3 (stochastic parallel transport, Itô formula), Ch. 4-5 (radial process, stochastic completeness)
Elworthy — Stochastic Differential Equations on Manifolds (London Math. Soc. Lecture Note Series 70, 1982) · §§II-III (SDEs on the frame bundle, horizontal vector fields, the projected diffusion)
Ikeda — Watanabe, Stochastic Differential Equations and Diffusion Processes (North-Holland/Kodansha, 1989) · Ch. V (Stratonovich SDEs on manifolds, the horizontal lift of Brownian motion)
Émery — Stochastic Calculus in Manifolds (Springer Universitext, 1989) · Ch. 5-7 (intrinsic Itô formula, martingales in a manifold, the role of the connection)
Bismut — Large Deviations and the Malliavin Calculus (Progress in Math. 45, Birkhäuser, 1984) · Ch. 1-2 (horizontal Laplacian, the lift of the heat semigroup to the frame bundle)

Estimated time

beginner: 18m
intermediate: 45m
master: 85m