Smoothness and Regularity Issues¶
Regularity of \(V\) and Greeks depends on: - regularity of payoff \(\Phi\), - ellipticity/smoothness of coefficients, - time distance to maturity.
Diffusion smoothing: precise statement¶
Parabolic equations smooth initial/terminal data for \(t<T\). Even if \(\Phi\) is not \(C^1\), \(V(t,\cdot)\) is often smooth for \(t<T\) (away from degeneracies such as \(S=0\)).
Schauder estimates. For the Black–Scholes equation with bounded, measurable terminal data \(\Phi\), the solution satisfies:
More precisely, for any \(\alpha \in (0,1)\) and compact \(K \subset (0,T) \times (0,\infty)\), we have interior Hölder estimates:
where the parabolic Hölder space \(C^{2+\alpha, 1+\alpha/2}\) captures two spatial derivatives and one time derivative with Hölder continuity.
Heat kernel interpretation. The fundamental solution (heat kernel) of the log-transformed Black–Scholes equation is Gaussian:
Convolution with any \(L^1\) terminal data produces a \(C^\infty\) function for \(t < T\).
Kinks and gamma concentration¶
For \(\Phi(S)=(S-K)^+\), the second derivative at maturity is distributional:
For \(t<T\), diffusion replaces this by a narrow bump of width \(\mathcal{O}(\sqrt{T-t})\).
Quantitative smoothing. The gamma at time \(t < T\) satisfies:
The maximum gamma occurs at \(S = K e^{-(r-\frac12\sigma^2)\tau} \approx K\) and equals:
This shows how the Dirac mass "regularizes" into a smooth bump with: - Height: \(\mathcal{O}(\tau^{-1/2})\) - Width: \(\mathcal{O}(\sqrt{\tau})\) - Area (integral of \(\Gamma\)): \(\mathcal{O}(1)\), independent of \(\tau\)
Degenerate coefficients¶
The Black–Scholes operator
degenerates at \(S = 0\) (the coefficient of \(\partial_{SS}\) vanishes). This creates:
- Boundary layer effects near \(S = 0\)
- Reduced regularity compared to uniformly elliptic equations
- Need for weighted Sobolev spaces for rigorous analysis
The standard approach is to work in log-coordinates \(x = \ln S\), where the operator becomes uniformly elliptic:
American options and free boundaries¶
Free boundaries reduce regularity; viscosity solutions provide the right weak framework.
Free boundary regularity. For the American put, the optimal exercise boundary \(S^*(t)\) satisfies: - \(S^* \in C^{0,1/2}\) (Lipschitz in \(\sqrt{T-t}\)) near maturity - \(S^*\) is smooth away from maturity
Greeks across the boundary. At the exercise boundary: - \(V\) is \(C^1\) (smooth pasting): \(\Delta\) is continuous - \(V\) is generally not \(C^2\): \(\Gamma\) has a jump discontinuity - The jump in \(\Gamma\) relates to the local time of \(S\) at the boundary
Function space classification¶
| Domain | Price regularity | Greek regularity |
|---|---|---|
| \(t < T\), European | \(C^\infty\) in \((t,S)\) | All Greeks smooth |
| \(t = T\), kinked payoff | \(C^0\) only | \(\Delta\) discontinuous, \(\Gamma\) distributional |
| American, continuation region | \(C^\infty\) | All Greeks smooth |
| American, at exercise boundary | \(C^1\) | \(\Delta\) continuous, \(\Gamma\) may jump |
What to remember¶
- Smoothness improves for \(t<T\) but deteriorates as \(t\uparrow T\).
- Interior regularity follows from Schauder estimates; \(V \in C^\infty\) for \(t < T\).
- Payoff kinks create large gamma near maturity: height \(\sim \tau^{-1/2}\), width \(\sim \sqrt{\tau}\).
- The Black–Scholes operator degenerates at \(S=0\); log-transform restores uniform ellipticity.
- Early exercise introduces weaker regularity and free boundary effects; \(\Gamma\) may be discontinuous.
Exercises¶
Exercise 1. The heat kernel for the log-transformed Black--Scholes equation is Gaussian. Explain why convolution of any bounded measurable terminal data with this kernel produces a \(C^\infty\) function for \(t < T\), and relate this to the Schauder estimate \(\|V\|_{C^{2+\alpha, 1+\alpha/2}(K)} \leq C_K \|\Phi\|_{L^\infty}\).
Solution to Exercise 1
In log-coordinates \(x = \ln S\), the Black--Scholes PDE becomes the constant-coefficient parabolic equation:
The fundamental solution (Green's function) is the Gaussian kernel:
For any bounded measurable terminal data \(\Phi\), the solution is:
This is \(C^\infty\) for \(t < T\) because:
- The Gaussian \(G\) is \(C^\infty\) in \((t,x)\) for \(t < T\), since the exponential of a polynomial is smooth and \(T - t > 0\) prevents any singularity.
- Differentiation passes under the integral sign by dominated convergence: for any multi-index of derivatives, \(|\partial_t^j \partial_x^k G(t,x;T,y)| \leq C_{j,k}(T-t)^{-(j+k/2)}G_\epsilon(t,x;T,y)\) for a slightly wider Gaussian \(G_\epsilon\), which is integrable against bounded \(\Phi\).
- Each derivative produces another smooth function, so \(u \in C^\infty((0,T) \times \mathbb{R})\).
The Schauder estimate \(\|V\|_{C^{2+\alpha,1+\alpha/2}(K)} \leq C_K\|\Phi\|_{L^\infty}\) quantifies this smoothing. It states that on any compact set \(K\) strictly inside the domain (bounded away from \(t = T\)), the \(C^{2+\alpha}\) norm of the solution is controlled solely by the \(L^\infty\) norm of the terminal data. This is a hallmark of parabolic regularity: even discontinuous or merely bounded initial data produces solutions with two spatial derivatives in Holder spaces, with the constant \(C_K\) depending on the distance from \(K\) to the boundary \(t = T\).
Exercise 2. For the call payoff \(\Phi(S) = (S-K)^+\), the gamma at time \(t < T\) has height \(\mathcal{O}(\tau^{-1/2})\), width \(\mathcal{O}(\sqrt{\tau})\), and area \(\mathcal{O}(1)\). Verify the area claim by computing \(\int_0^\infty \Gamma(t,S) \, dS\) for the Black--Scholes gamma formula and showing it equals \(1/S\) (or a constant independent of \(\tau\)).
Solution to Exercise 2
The Black--Scholes gamma for a European call is:
where \(N'(d_1) = \frac{1}{\sqrt{2\pi}}e^{-d_1^2/2}\) and \(d_1 = \frac{\ln(S/K) + (r + \frac{1}{2}\sigma^2)\tau}{\sigma\sqrt{\tau}}\).
Compute the integral:
Substitute \(u = d_1\), which gives \(du = \frac{1}{S\sigma\sqrt{\tau}}\,dS\) (since \(d_1 = \frac{\ln(S/K) + (r+\frac{1}{2}\sigma^2)\tau}{\sigma\sqrt{\tau}}\) and \(\frac{\partial d_1}{\partial S} = \frac{1}{S\sigma\sqrt{\tau}}\)). As \(S\) ranges from \(0\) to \(\infty\), \(d_1\) ranges from \(-\infty\) to \(+\infty\). Therefore:
The area under the gamma curve equals 1, independent of \(\tau\), \(\sigma\), \(r\), or \(K\). This is consistent with the fact that \(\int_0^\infty \Gamma\,dS = \int_0^\infty V_{SS}\,dS = V_S\big|_0^\infty = \Delta(\infty) - \Delta(0) = 1 - 0 = 1\) for a call option. As \(\tau \to 0\), the height grows as \(\tau^{-1/2}\) and the width shrinks as \(\sqrt{\tau}\), but their product (the area) remains constant at 1, consistent with the distributional limit \(\Gamma(T,S) = \delta(S-K)\) which also has unit integral.
Exercise 3. The Black--Scholes operator degenerates at \(S = 0\). Explain why the change of variable \(x = \ln S\) transforms the operator into a uniformly elliptic one. Write down the transformed PDE explicitly.
Solution to Exercise 3
The Black--Scholes operator in the original \(S\) variable is:
The coefficient of \(\frac{\partial^2}{\partial S^2}\) is \(\frac{1}{2}\sigma^2 S^2\), which vanishes at \(S = 0\). This means the operator is degenerate elliptic at \(S = 0\): the second-order term that provides diffusion (and thus smoothing) disappears. Standard elliptic regularity theory requires uniform ellipticity, meaning the coefficient of the second-order term is bounded away from zero.
Under \(x = \ln S\) (so \(S = e^x\), \(x \in (-\infty, \infty)\)), we use:
Substituting into the PDE \(\frac{\partial V}{\partial t} + \frac{1}{2}\sigma^2 S^2 V_{SS} + rSV_S - rV = 0\):
Simplifying:
The coefficient of \(\frac{\partial^2 V}{\partial x^2}\) is now \(\frac{1}{2}\sigma^2\), a positive constant independent of \(x\). This makes the operator uniformly elliptic on all of \(\mathbb{R}\), so standard parabolic regularity theory (Schauder estimates, maximum principles) applies without any special treatment of boundary layers.
Exercise 4. For an American put, the exercise boundary \(S^*(t)\) creates a jump in gamma. Using the smooth-pasting condition (\(V\) is \(C^1\) across the boundary), explain why \(\Delta\) is continuous but \(\Gamma\) has a discontinuity. What is the sign of the gamma jump?
Solution to Exercise 4
For an American put with exercise boundary \(S^*(t)\), the smooth-pasting condition states that the option price \(V(t,S)\) is \(C^1\) across the boundary: both \(V\) and \(V_S = \Delta\) are continuous at \(S = S^*(t)\).
Why delta is continuous. In the exercise region (\(S \leq S^*\)), \(V(t,S) = K - S\), so \(\Delta = -1\). In the continuation region (\(S > S^*\)), \(\Delta\) is determined by the PDE solution. Smooth pasting requires that \(\Delta\) approaches \(-1\) continuously as \(S \downarrow S^*\) from the continuation region. This is not automatic -- it is a condition that determines the free boundary \(S^*\) itself.
Why gamma is discontinuous. In the exercise region, \(V = K - S\) is linear, so \(\Gamma = V_{SS} = 0\). In the continuation region, the PDE is active and \(V\) has non-trivial curvature. At the boundary, we can compute:
The jump in gamma is:
The sign is positive because the put price in the continuation region is convex (curving upward away from the linear intrinsic value). Mathematically, \(V \geq K - S\) everywhere with equality at the boundary, so \(V - (K-S)\) is non-negative and vanishes at \(S^*\), implying \((V-(K-S))_{SS} \geq 0\) at the boundary, which gives \(\Gamma^+ \geq 0\). Strict inequality holds because the diffusion term is active in the continuation region.
This gamma discontinuity is a fundamental feature of American options and creates practical challenges for hedging near the exercise boundary.
Exercise 5. Consider a digital call with payoff \(\Phi(S) = \mathbf{1}_{S > K}\), which is discontinuous at \(S = K\). Describe the regularity of the digital call price \(V(t,S)\) for \(t < T\). Is delta smooth? How does the delta of a digital call behave as \(\tau \to 0\)?
Solution to Exercise 5
The digital call payoff \(\Phi(S) = \mathbf{1}_{S > K}\) is discontinuous at \(S = K\) with a jump of size 1.
Regularity for \(t < T\). Despite the discontinuous terminal data, the digital call price:
is \(C^\infty\) in both \(t\) and \(S\) for any \(t < T\). This follows from the parabolic smoothing property: convolution of \(\mathbf{1}_{S>K}\) with the smooth heat kernel produces a smooth function. The Schauder estimate guarantees \(V \in C^{2+\alpha,1+\alpha/2}\) on compact subsets of \((0,T) \times (0,\infty)\).
Delta is smooth for \(t < T\). The delta of the digital call is:
This is a smooth, positive, bell-shaped function of \(S\) centered near \(S = K\), identical in shape to the gamma of a vanilla call (up to a factor of \(e^{-r\tau}\)).
Behavior as \(\tau \to 0\). As \(\tau \to 0\), the delta of the digital call concentrates:
- At \(S = K\) (ATM): \(\Delta_{\text{ATM}} = \frac{e^{-r\tau}}{K\sigma\sqrt{2\pi\tau}} \sim \tau^{-1/2} \to \infty\)
- For \(S \neq K\): \(\Delta \to 0\) exponentially fast
The delta converges to a Dirac delta \(\delta(S - K)\) in the distributional sense, reflecting the fact that \(\Phi'(S) = \delta(S-K)\). The digital delta becomes unhedgeable near expiry: it requires infinite rebalancing at the strike. In practice, this means digital options near expiry cannot be delta-hedged with the underlying alone, motivating the use of call spreads as approximate replication.
Exercise 6. A finite-difference scheme for the Black--Scholes PDE must handle the terminal condition \(V(T,S) = (S-K)^+\), which has a kink at \(S = K\). Explain why using the exact payoff as initial data can introduce oscillations in the numerical gamma, and propose two remedies.
Solution to Exercise 6
Why oscillations arise. The kink in \((S-K)^+\) at \(S = K\) means that the numerical second difference:
is approximately \(1/\Delta S\) at the grid point nearest \(K\) and zero elsewhere. This is a crude approximation to \(\delta(S-K)\) that creates a numerical gamma with large spurious oscillations. When this discontinuous initial gamma is propagated one time step backward, the finite-difference scheme (especially explicit schemes) can produce alternating positive and negative values -- the classic Gibbs phenomenon for parabolic equations with non-smooth data.
The problem is particularly severe for gamma because the exact solution transitions smoothly from \(\Gamma = 0\) (far from strike) to \(\Gamma_{\max}\) (at strike), but the numerical approximation inherits the grid-scale oscillations from the non-smooth initial condition.
Remedy 1: Rannacher time-stepping. Use a few (typically 2--4) implicit Euler steps at the start of the backward time-marching (near \(t = T\)), then switch to the more accurate Crank--Nicolson scheme. The fully implicit steps are strongly damping and eliminate the high-frequency oscillations introduced by the kink, at the cost of only first-order accuracy for those few steps. The subsequent Crank--Nicolson steps restore second-order accuracy once the solution has been smoothed.
Remedy 2: Payoff smoothing (cell averaging). Replace the pointwise payoff values with cell-averaged values near the kink:
For grid points away from \(K\), this equals \((S_i - K)^+\). For the grid point nearest \(K\), the integral smooths the kink over one cell width. This produces a consistent numerical gamma that is \(\mathcal{O}(1/\Delta S)\) at the kink point (matching the true Dirac scaling) without oscillations in neighboring cells.