Probability Density via Heat Kernel Expansion¶

The deepest insight into the SABR model comes from recognizing that its dynamics define a diffusion on a Riemannian manifold --- specifically, a surface related to the hyperbolic plane. The transition density of this diffusion is the heat kernel on the manifold, and its short-time asymptotic expansion yields the Hagan implied volatility formula as a direct consequence. This geometric perspective, developed by Henry-Labordere (2008) and Lesniewski (2002), not only provides a rigorous derivation of the Hagan formula but also reveals the mathematical structure that governs the accuracy of the approximation and guides the construction of higher-order corrections.

Learning Objectives

By the end of this section, you will be able to:

Identify the Riemannian metric induced by the SABR diffusion
Compute the geodesic distance on the SABR manifold
Write down the leading-order heat kernel expansion
Explain how the heat kernel leads to the Hagan implied volatility formula
Describe the role of the hyperbolic plane in the SABR geometry

Motivation¶

The standard approach to computing transition densities for SDEs is to solve the Fokker--Planck PDE. For two-dimensional problems like SABR, this requires numerical methods that are computationally expensive and provide limited analytical insight. The geometric approach offers an alternative: by interpreting the diffusion as Brownian motion on a curved surface, we can leverage the rich theory of heat kernels on Riemannian manifolds. The leading-order heat kernel involves only the geodesic distance between two points --- a purely geometric quantity --- and already captures the main features of the SABR smile. Higher-order terms (the Van Vleck determinant and curvature corrections) refine the approximation systematically.

Riemannian Geometry of the SABR Diffusion¶

From SDE to Metric¶

Consider the general SABR model after the coordinate transformation \(y = F^{1-\beta}/(1-\beta)\) (for \(\beta \neq 1\)):

\[ dy_t = -\frac{\beta}{2}\frac{\sigma_t^2}{y_t^{1/(1-\beta)}}\,dt + \sigma_t\,dW_t^{(1)} \]

\[ d\sigma_t = \nu\sigma_t\,dW_t^{(2)} \]

Dropping the drift (which contributes only at higher order in the short-time expansion), the diffusion matrix in the coordinates \(\mathbf{x} = (y, \sigma)\) is:

\[ \mathbf{a}(\mathbf{x}) = \begin{pmatrix} \sigma^2 & \rho\nu\sigma^2 \\ \rho\nu\sigma^2 & \nu^2\sigma^2 \end{pmatrix} \]

The associated Riemannian metric is the inverse of the diffusion matrix:

\[ g_{ij} = (a^{-1})_{ij} = \frac{1}{(1-\rho^2)\sigma^2}\begin{pmatrix} 1 & -\rho/\nu \\ -\rho/\nu & 1/\nu^2 \end{pmatrix} \]

The Line Element¶

The Riemannian distance element is:

\[ ds^2 = g_{ij}\,dx^i\,dx^j = \frac{1}{(1-\rho^2)\sigma^2}\left(dy^2 - \frac{2\rho}{\nu}\,dy\,d\sigma + \frac{1}{\nu^2}\,d\sigma^2\right) \]

For the uncorrelated case (\(\rho = 0\)), this simplifies to:

\[ ds^2 = \frac{dy^2}{\sigma^2} + \frac{d\sigma^2}{\nu^2\sigma^2} \]

Connection to the Hyperbolic Plane¶

The metric for \(\rho = 0\) is recognized as the Poincare upper half-plane metric (up to a rescaling). The Poincare half-plane \(\mathbb{H}^2\) has the metric:

\[ ds_{\mathbb{H}^2}^2 = \frac{dx^2 + dz^2}{z^2} \]

Identifying \(x = y\) and \(z = \sigma/\nu\), the SABR metric becomes:

\[ ds^2 = \frac{dy^2 + (d\sigma/\nu)^2}{\sigma^2/\nu^2} = \frac{dy^2 + dz^2}{z^2} \cdot \frac{1}{\nu^2} \]

which is \(1/\nu^2\) times the hyperbolic metric. The hyperbolic plane is a surface of constant negative Gaussian curvature \(\mathcal{K} = -\nu^2\). This negative curvature is responsible for the "spreading" of geodesics and the characteristic shape of the SABR smile.

Geometric Intuition

On a flat surface (like the Black--Scholes model), the heat kernel is a simple Gaussian whose width grows as \(\sqrt{t}\). On a negatively curved surface, geodesics diverge faster than on a flat surface, causing the heat kernel to spread more rapidly in certain directions. This enhanced spreading in the SABR geometry produces the heavier tails and more pronounced smile that distinguish SABR from the CEV model.

The Correlated Case¶

When \(\rho \neq 0\), the metric includes off-diagonal terms. The geometry is still a surface of constant negative curvature, but the coordinate system is oblique. The off-diagonal terms can be removed by a rotation:

\[ \tilde{y} = y - \frac{\rho}{\nu}\sigma, \qquad \tilde{\sigma} = \frac{\sqrt{1-\rho^2}}{\nu}\sigma \]

In these coordinates, the metric becomes diagonal and proportional to the hyperbolic metric. The correlation \(\rho\) acts as a shear in the geometry, tilting the geodesics and breaking the left-right symmetry of the smile.

Geodesic Distance¶

Definition¶

The geodesic distance \(d(\mathbf{x}_0, \mathbf{x})\) is the length of the shortest path between two points on the manifold. For the heat kernel expansion, this distance is the key quantity: the leading-order density is proportional to \(\exp(-d^2/(2T))\).

Formula for the SABR Manifold¶

For the SABR model in coordinates \((y, \sigma)\) with \(y = F^{1-\beta}/(1-\beta)\), the geodesic distance between \((y_0, \sigma_0)\) and \((y, \sigma)\) is:

\[ d^2 = \frac{1}{1-\rho^2}\left[\ln^2\!\left(\frac{\sigma}{\sigma_0}\right) - 2\rho\nu\ln\!\left(\frac{\sigma}{\sigma_0}\right)\frac{y - y_0}{\bar{\sigma}} + \nu^2\frac{(y - y_0)^2}{\bar{\sigma}^2}\right] \]

where \(\bar{\sigma}\) is an appropriate average of \(\sigma_0\) and \(\sigma\) along the geodesic. For the uncorrelated case, this simplifies significantly.

Uncorrelated Geodesic Distance (Rho = 0)¶

When \(\rho = 0\), the geodesic distance on the SABR manifold reduces to:

\[ \cosh(\nu\,d) = 1 + \frac{\nu^2(y - y_0)^2 + (\sigma - \sigma_0)^2}{2\sigma_0\sigma} \]

This is the standard hyperbolic distance formula on the Poincare half-plane (with the identification \(z = \sigma/\nu\)). The geodesics are semicircles orthogonal to the \(\sigma = 0\) axis.

Theorem: Geodesic Distance on the Uncorrelated SABR Manifold

For \(\rho = 0\), the geodesic distance between \((y_0, \sigma_0)\) and \((y, \sigma)\) on the SABR manifold is:

\[ d = \frac{1}{\nu}\cosh^{-1}\!\left(1 + \frac{\nu^2(y - y_0)^2 + (\sigma - \sigma_0)^2}{2\sigma_0\sigma}\right) \]

where \(y = F^{1-\beta}/(1-\beta)\) and \(y_0 = F_0^{1-\beta}/(1-\beta)\).

Connection to the z/x(z) Factor¶

The geodesic distance in the Hagan formula appears through the \(z/x(z)\) factor. Recall that:

\[ z = \frac{\nu}{\alpha}(FK)^{(1-\beta)/2}\ln\frac{F}{K} \]

and \(x(z) = \ln((\sqrt{1 - 2\rho z + z^2} + z - \rho)/(1 - \rho))\). The quantity \(x(z)\) is (to leading order) the geodesic distance between the initial and terminal points, evaluated at \(\sigma = \sigma_0 = \alpha\). The ratio \(z/x(z)\) is the Jacobian of the mapping from the flat (Euclidean) parameterization to the geodesic (hyperbolic) parameterization.

The Heat Kernel Expansion¶

General Framework¶

The heat kernel \(p(T, \mathbf{x}_0, \mathbf{x})\) is the fundamental solution of the heat equation on the Riemannian manifold:

\[ \frac{\partial p}{\partial T} = \frac{1}{2}\Delta_g\,p \]

where \(\Delta_g\) is the Laplace--Beltrami operator associated with the metric \(g\). The short-time asymptotic expansion of the heat kernel is:

\[ p(T, \mathbf{x}_0, \mathbf{x}) = \frac{1}{2\pi T}\frac{\Delta_{\text{VV}}(\mathbf{x}_0, \mathbf{x})}{\sqrt{\det g(\mathbf{x})}}\exp\!\left(-\frac{d^2(\mathbf{x}_0, \mathbf{x})}{2T}\right)\left[1 + a_1(\mathbf{x}_0, \mathbf{x})\,T + O(T^2)\right] \]

The components are:

Leading exponential: \(\exp(-d^2/(2T))\) where \(d\) is the geodesic distance. This is the analog of the Gaussian kernel on flat space.

Van Vleck determinant: \(\Delta_{\text{VV}}\) accounts for the focusing or defocusing of geodesics. On the hyperbolic plane, geodesics diverge, so \(\Delta_{\text{VV}} < 1\) (the density is suppressed relative to flat space).

Volume element: \(\sqrt{\det g}\) is the Riemannian volume form, ensuring the density integrates to 1.

Correction terms: The coefficients \(a_1, a_2, \ldots\) involve the Ricci curvature and its derivatives, providing systematic improvements.

Application to SABR¶

For the SABR manifold with constant negative curvature \(\mathcal{K} = -\nu^2\), the heat kernel on the hyperbolic plane has the exact representation (McKean, 1970):

\[ p_{\mathbb{H}^2}(T, d) = \frac{\sqrt{2}}{(2\pi T)^{3/2}}\,e^{-\nu^2 T/8}\int_d^{\infty} \frac{s\,e^{-s^2/(2T)}}{\sqrt{\cosh(\nu s) - \cosh(\nu d)}}\,ds \]

This exact formula involves a one-dimensional integral that can be evaluated numerically. The short-time expansion of this integral recovers the Hagan formula.

From Heat Kernel to Implied Volatility¶

The Density-to-Price-to-IV Pipeline¶

The connection from the heat kernel to the Hagan implied volatility formula follows this sequence:

Transition density: The heat kernel gives \(p(F_T, \sigma_T \mid F_0, \sigma_0; T)\)
Marginal density: Integrate over terminal volatility \(\sigma_T\) to obtain \(p(F_T \mid F_0, \sigma_0; T)\)
Option price: \(C(K, T) = \int_K^{\infty}(F - K)\,p(F, T)\,dF\)
Implied volatility: Invert the Black (or Bachelier) formula to obtain \(\sigma_{\text{impl}}(K)\)

The asymptotic expansion at each step preserves the order of accuracy. The \(O(T)\) Hagan correction arises from the integration over \(\sigma_T\) and the curvature corrections in the heat kernel.

Leading-Order Result¶

At leading order (ignoring the \(O(T)\) correction), the implied volatility is determined entirely by the geodesic distance:

\[ \sigma_B(K) \approx \frac{\alpha}{(FK)^{(1-\beta)/2}} \cdot \frac{z}{x(z)} \]

This shows that the smile shape is a purely geometric quantity, determined by the geodesic structure of the SABR manifold. The \(z/x(z)\) factor is the ratio of the "Euclidean distance" \(z\) to the "geodesic distance" \(x(z)\) in the \((y, \sigma)\) plane.

Higher-Order Corrections¶

The First Correction Term¶

The \(O(T)\) correction in the Hagan formula:

\[ \varepsilon = \frac{(1-\beta)^2\alpha^2}{24(FK)^{1-\beta}} + \frac{\rho\beta\nu\alpha}{4(FK)^{(1-\beta)/2}} + \frac{2-3\rho^2}{24}\nu^2 \]

arises from three sources:

Drift correction: The first term \(\propto \alpha^2\) comes from the drift that was dropped in the leading-order geometric analysis
Cross-term: The second term \(\propto \rho\nu\alpha\) comes from the correlation-induced asymmetry
Curvature: The third term \(\propto \nu^2\) is a pure curvature correction from the Ricci scalar of the manifold

Systematic Improvement¶

The heat kernel expansion can be extended to \(O(T^2)\) and beyond, producing higher-order corrections to the Hagan formula. Paulot (2009) derived the second-order correction, which is particularly important for:

Long-maturity options (\(T > 5\) years)
High vol-of-vol (\(\nu > 0.6\))
Deep out-of-the-money strikes

Each additional order introduces more terms but maintains the same geometric structure.

Summary¶

The probability density of the SABR model is the heat kernel on a Riemannian manifold of constant negative curvature. The Riemannian metric is determined by the SABR diffusion coefficients, and for \(\rho = 0\) the manifold is isometric to the Poincare hyperbolic plane with curvature \(-\nu^2\). The geodesic distance on this manifold directly yields the \(z/x(z)\) smile factor in the Hagan formula, while the \(O(T)\) correction arises from drift terms, correlation, and curvature. This geometric interpretation provides a rigorous foundation for the Hagan approximation and a systematic framework for computing higher-order corrections. The McKean heat kernel formula gives the exact density for the uncorrelated case, serving as a benchmark for all approximate and numerical methods.

Exercises¶

Exercise 1. The SABR model with \(\rho = 0\) defines a diffusion on a surface isometric to the Poincare hyperbolic plane with curvature \(-\nu^2\). Explain intuitively why the curvature is negative (the "geometry" is hyperbolic rather than Euclidean or spherical). What happens to the curvature as \(\nu \to 0\)? What does this limit correspond to in terms of the SABR model?

Solution to Exercise 1

The curvature of the SABR manifold is negative because the vol-of-vol parameter \(\nu\) causes geodesics to diverge faster than they would on a flat (Euclidean) surface. Intuitively, in the SABR model, the stochastic volatility \(\sigma_t\) acts as a "scale factor" for the forward dynamics: when \(\sigma\) is large, the forward diffuses rapidly, and when \(\sigma\) is small, the forward barely moves. This heterogeneity in local diffusion speed is exactly what the Riemannian metric captures. On the Poincare half-plane, points near the boundary (\(\sigma \to 0\)) are "infinitely far apart" in the geodesic sense --- the metric element \(ds^2 = (dy^2 + dz^2)/z^2\) blows up as \(z = \sigma/\nu \to 0\). This divergence of geodesics is the hallmark of negative curvature (hyperbolic geometry), in contrast to positive curvature (spherical geometry) where geodesics converge.

As \(\nu \to 0\), the curvature \(\mathcal{K} = -\nu^2 \to 0\). In this limit, the manifold becomes flat, and the SABR model reduces to the CEV model (constant elasticity of variance) with no stochastic volatility. In flat geometry, the heat kernel is an ordinary Gaussian, and the implied volatility smile is determined entirely by the CEV backbone \(\alpha/F^{1-\beta}\) with no smile curvature from vol-of-vol. The transition from the hyperbolic plane (\(\nu > 0\)) to flat space (\(\nu = 0\)) corresponds precisely to the transition from a stochastic volatility model with smile curvature to a local volatility model with only skew.

Exercise 2. The geodesic distance on the SABR manifold between the initial point \((F_0, \alpha)\) and the point \((K, \alpha)\) is related to the Hagan smile factor \(z/x(z)\). For \(\rho = 0\), the geodesic distance simplifies to \(d = \frac{1}{\nu}\cosh^{-1}\!\bigl(1 + \frac{\nu^2}{2\alpha^2}\frac{(K-F)^2}{(FK)^{1-\beta}}\bigr)\) (schematic). Explain why a larger geodesic distance (corresponding to a further OTM strike) leads to a higher implied volatility through the heat kernel decay.

Solution to Exercise 2

The heat kernel on the SABR manifold has the leading-order form

\[ p(T, \mathbf{x}_0, \mathbf{x}) \propto \frac{1}{T}\exp\!\left(-\frac{d^2(\mathbf{x}_0, \mathbf{x})}{2T}\right) \]

where \(d\) is the geodesic distance. The implied volatility \(\sigma_B(K)\) is obtained by matching this density (after integrating over the terminal volatility \(\sigma_T\)) to the density implied by the Black formula. In the Black model with volatility \(\sigma_B\), the density has the analogous form \(\propto \exp(-(\ln(F/K))^2 / (2\sigma_B^2 T))\).

By matching the leading exponential terms, the implied volatility satisfies

\[ \frac{(\ln(F/K))^2}{2\sigma_B^2 T} \sim \frac{d^2}{2T} \]

which gives \(\sigma_B \sim |\ln(F/K)| / d\). When the strike \(K\) is further OTM, the Euclidean-like measure \(|\ln(F/K)|\) grows, but the geodesic distance \(d\) grows more slowly on the hyperbolic plane (because geodesics spread out on a negatively curved surface). Therefore, the ratio \(|\ln(F/K)|/d\) increases for OTM strikes, implying a higher implied volatility in the wings.

Equivalently, the Hagan formula encodes this through the ratio \(z/x(z)\), where \(z\) is a Euclidean-like parameterization and \(x(z)\) is the geodesic distance. Since \(x(z) < z\) for \(z > 0\) on the hyperbolic plane, the ratio \(z/x(z) > 1\), boosting the implied volatility above the backbone level. The further OTM the strike, the larger the gap between \(z\) and \(x(z)\), producing the characteristic SABR smile.

Exercise 3. The short-time heat kernel expansion gives the density as \(p(t, x, y) \sim \frac{1}{t}\exp\bigl(-\frac{d^2(x,y)}{2t}\bigr) \cdot g(x,y)\) where \(d\) is the geodesic distance and \(g\) involves curvature corrections. Explain why the leading \(\exp(-d^2/(2t))\) term becomes more concentrated around \(x = y\) as \(t \to 0\), and how the \(O(t)\) corrections give rise to the skew and convexity terms in the Hagan formula.

Solution to Exercise 3

Concentration as \(t \to 0\): The leading term \(\exp(-d^2(x,y)/(2t))\) is a Gaussian-like function of the geodesic distance \(d\). As \(t \to 0\):

For \(y \neq x\): \(d^2(x,y) > 0\), so \(\exp(-d^2/(2t)) \to 0\) exponentially fast.
For \(y = x\): \(d = 0\), so \(\exp(0) = 1\).

Therefore, the density concentrates at \(y = x\) as a Dirac delta in the limit \(t \to 0\). The prefactor \(1/t\) (in two dimensions; more generally \(1/t^{n/2}\)) provides the correct normalization so that the integral over the manifold remains unity.

Origin of skew and convexity from \(O(t)\) corrections: The full heat kernel expansion is

\[ p(t, x, y) = \frac{1}{2\pi t}\frac{\Delta_{\mathrm{VV}}}{\sqrt{\det g}}\exp\!\left(-\frac{d^2}{2t}\right)\bigl[1 + a_1(x,y)\,t + O(t^2)\bigr] \]

The correction \(a_1(x,y)\) involves the Ricci curvature and the drift terms that were dropped in the leading-order analysis. In the Hagan formula, the \(O(T)\) correction

\[ \varepsilon = \frac{(1-\beta)^2\alpha^2}{24(FK)^{1-\beta}} + \frac{\rho\beta\nu\alpha}{4(FK)^{(1-\beta)/2}} + \frac{2-3\rho^2}{24}\nu^2 \]

enters as \(\sigma_B \approx \sigma_B^{(0)}(1 + \varepsilon T)\).

The first term (proportional to \(\alpha^2\)) comes from the drift in the transformed coordinate \(y = F^{1-\beta}/(1-\beta)\). This drift was dropped in the leading-order geometric analysis but contributes to the backbone curvature of the smile.
The second term (proportional to \(\rho\nu\alpha\)) produces the skew --- it is linear in \(\rho\) and breaks the left-right symmetry of the smile. Geometrically, it arises from the off-diagonal terms in the metric tensor when \(\rho \neq 0\).
The third term (proportional to \(\nu^2\)) produces additional convexity and is a pure curvature correction from the Ricci scalar of the hyperbolic manifold.

Together, these \(O(T)\) corrections shape the smile by adding skew (tilt) and convexity (curvature) to the leading-order purely geometric smile.

Exercise 4. For the uncorrelated case \(\rho = 0\), the McKean heat kernel provides the exact transition density. Explain why adding correlation (\(\rho \neq 0\)) breaks the simple hyperbolic geometry and requires perturbation methods. How does the Hagan formula handle the correlation: is \(\rho\) included in the geodesic distance, the curvature correction, or both?

Solution to Exercise 4

When \(\rho = 0\), the SABR metric is diagonal:

\[ ds^2 = \frac{dy^2}{\sigma^2} + \frac{d\sigma^2}{\nu^2\sigma^2} \]

This is (up to rescaling) the standard Poincare half-plane metric, which is a space of constant negative curvature \(\mathcal{K} = -\nu^2\). On such a space, the heat kernel has an exact closed-form representation (the McKean formula), and the geodesics are semicircles orthogonal to the boundary \(\sigma = 0\). All geometric quantities --- geodesic distance, Van Vleck determinant, curvature corrections --- can be computed analytically.

When \(\rho \neq 0\), the metric acquires off-diagonal terms:

\[ ds^2 = \frac{1}{(1-\rho^2)\sigma^2}\left(dy^2 - \frac{2\rho}{\nu}\,dy\,d\sigma + \frac{1}{\nu^2}\,d\sigma^2\right) \]

Although the coordinate change \(\tilde{y} = y - (\rho/\nu)\sigma\), \(\tilde{\sigma} = \sqrt{1-\rho^2}\,\sigma/\nu\) diagonalizes the metric, the resulting geometry is still the hyperbolic plane --- the curvature remains constant. However, the drift terms that were dropped in the leading-order analysis now include \(\rho\)-dependent contributions that cannot be absorbed into the geometry. These drift terms break the exact solvability and require perturbation methods.

In the Hagan formula, correlation enters in both places:

Geodesic distance: The \(z/x(z)\) factor involves \(\rho\) through \(x(z) = \ln((\sqrt{1-2\rho z + z^2} + z - \rho)/(1-\rho))\). This captures the leading-order effect of correlation on the smile shape --- the shearing of geodesics that breaks the left-right symmetry.
Curvature correction: The \(O(T)\) correction includes the term \(\rho\beta\nu\alpha/(4(FK)^{(1-\beta)/2})\), which is a first-order perturbation in \(\rho\) arising from the drift terms.

Thus, \(\rho\) affects both the geometry (geodesic distance) and the perturbative corrections, and the Hagan formula treats each contribution at the appropriate order.

Exercise 5. The geometric approach yields higher-order corrections to the Hagan formula (Paulot 2009). The second-order correction improves accuracy from \(O(\nu^2 T^2)\) to \(O(\nu^2 T^3)\). For a 5-year swaption with \(\nu = 0.5\), estimate the improvement factor \(T\) in the error bound. Is the second-order correction worth implementing for practical applications?

Solution to Exercise 5

The error in the Hagan formula (first-order heat kernel expansion) is \(O(\nu^2 T^2)\), while the Paulot second-order correction improves this to \(O(\nu^2 T^3)\). For a 5-year swaption with \(\nu = 0.5\):

First-order error: \(\nu^2 T^2 = 0.25 \times 25 = 6.25\)
Second-order error: \(\nu^2 T^3 = 0.25 \times 125 = 31.25\)

However, these are the scaling of the error terms, not the actual errors (which include small numerical prefactors). The relevant comparison is the ratio of the errors:

\[ \frac{\text{second-order error}}{\text{first-order error}} = \frac{O(\nu^2 T^3)}{O(\nu^2 T^2)} = O(T) = 5 \]

This means the improvement factor is approximately \(T = 5\): the second-order correction reduces the relative error by a factor of about 5. In absolute terms, if the first-order Hagan formula has an implied volatility error of, say, 50 basis points for a deep OTM 5-year swaption with \(\nu = 0.5\), the second-order correction would reduce this to roughly 10 basis points.

Practical assessment: For \(T = 5\) and \(\nu = 0.5\), the product \(\nu^2 T = 1.25\) is well outside the "small-parameter" regime (\(\nu^2 T \ll 1\)) where the first-order expansion is reliable. The second-order correction is therefore not just worthwhile but necessary for practical accuracy. In fact, for such parameters, even the second-order expansion may be insufficient, and practitioners often supplement the Hagan formula with numerical PDE or Monte Carlo methods for long-dated options with high vol-of-vol.

Exercise 6. Compare the geometric (heat kernel) derivation of the Hagan formula with the direct PDE perturbation approach. Both yield the same leading-order formula. What additional insight does the geometric approach provide? Specifically, explain how the Riemannian metric interpretation helps determine when the approximation will fail (e.g., for strikes far from ATM on the manifold).

Solution to Exercise 6

Both the geometric (heat kernel) and the direct PDE perturbation approaches yield the same Hagan formula at leading order, but the geometric approach provides several additional insights:

1. Structural understanding of the smile: The geometric approach reveals that the smile shape is fundamentally a geodesic phenomenon. The \(z/x(z)\) factor is the ratio of Euclidean to geodesic distances on the SABR manifold. This makes it clear that the smile arises from the curvature of the parameter space, not from any particular algebraic manipulation of the PDE.

2. Systematic higher-order corrections: The heat kernel expansion provides a natural hierarchy of corrections organized by powers of \(T\). Each order involves well-defined geometric quantities (Van Vleck determinant, Ricci curvature, etc.), making the expansion systematic and principled. The PDE perturbation approach, while producing the same results, is more ad hoc in its expansion structure.

3. Criterion for approximation failure: The Riemannian metric interpretation provides a clear criterion for when the approximation breaks down. The heat kernel expansion is valid when the geodesic distance \(d\) satisfies \(d^2 / T \gg 1\) (the density is well-approximated by its saddle-point expansion) and when \(T\) is small enough that higher-order corrections remain perturbative.

The approximation fails when:

The geodesic distance is large (deep OTM strikes): On the hyperbolic plane, the geodesic distance grows as \(\cosh^{-1}(\cdot)\), which is logarithmic for large arguments. Very deep OTM strikes correspond to points far from the initial point on the manifold, where the short-time asymptotics become unreliable because multiple geodesics (or geodesics that approach the boundary \(\sigma = 0\)) contribute to the heat kernel.
The maturity is long (\(\nu^2 T\) not small): The heat kernel expansion is a short-time expansion. When \(\nu^2 T \gtrsim 1\), the higher-order terms in the expansion become comparable to the leading term, and the approximation loses accuracy.
The curvature is large (\(\nu\) large): High curvature \(|\mathcal{K}| = \nu^2\) means the manifold deviates strongly from flat space, making the Gaussian approximation to the heat kernel less accurate at any given maturity.

The PDE perturbation approach does not provide these geometric failure criteria as naturally. It can identify that the expansion breaks down when \(\nu^2 T\) is large, but it does not explain why in terms of the underlying geometry. The Riemannian perspective makes it clear that the failure is due to the mismatch between the flat-space Gaussian and the true heat kernel on a curved manifold --- a mismatch that grows with curvature, time, and distance.