Why PDEs in Finance¶
Partial differential equations arise naturally in quantitative finance whenever we ask: what is the fair price of a derivative today, given uncertainty about the future? The answer turns out to satisfy a deterministic PDE, even though the underlying asset follows a random process. This remarkable fact is the foundation of modern derivatives pricing.
The Core Insight¶
Consider a derivative with payoff \(g(S_T)\) at maturity \(T\). Under the risk-neutral measure, its price at time \(t\) is:
This expectation is a function of the current state \((t, S)\). The Feynman-Kac theorem guarantees that this function satisfies a parabolic PDE:
with terminal condition \(V(T, S) = g(S)\).
The Central Principle
Expected values of diffusion processes satisfy partial differential equations. Every pricing problem for a Markovian underlying reduces to solving a PDE.
From Stochastic Process to PDE¶
The Generator Connection¶
For a general diffusion:
the infinitesimal generator is the differential operator:
The generator encodes the local dynamics of the process. Any expected value \(u(t,x) = \mathbb{E}[g(X_T) \mid X_t = x]\) satisfies:
With discounting at rate \(r(t,x)\), the PDE becomes:
This is the Feynman-Kac equation, and it is the master equation of financial PDE theory.
Why Does a Random Process Lead to a Deterministic Equation?¶
The key is the Markov property. Because the future of \(X_t\) depends only on its current state, the conditional expectation \(u(t,x)\) is a well-defined function of \((t,x)\) alone. The PDE then describes how this function must evolve to remain consistent with the stochastic dynamics.
Intuitively:
- Drift \(\mu\) contributes a first-order (transport) term
- Volatility \(\sigma\) contributes a second-order (diffusion) term
- Discounting \(r\) contributes a zeroth-order (decay) term
Three Approaches to Derivative Pricing¶
Given the price \(V(t,S) = e^{-r(T-t)}\,\mathbb{E}^{\mathbb{Q}}[g(S_T) \mid S_t = S]\), there are three computational strategies.
Approach 1: PDE (Finite Differences)¶
Solve the PDE on a discretized grid in \((t, S)\) space, marching backward from the terminal condition.
Procedure:
- Discretize the domain: grid points \((t_n, S_j)\)
- Apply a finite difference scheme (explicit, implicit, or Crank-Nicolson)
- March backward from \(V(T, S_j) = g(S_j)\) to \(V(0, S_0)\)
Example: For a European call under Black-Scholes, the PDE is:
A Crank-Nicolson scheme with 200 time steps and 400 spatial nodes typically gives accuracy to several decimal places.
Approach 2: Monte Carlo Simulation¶
Simulate paths of \(S_t\) and average the discounted payoff.
Procedure:
- Simulate \(N\) independent paths of \(S_t\) from \(t\) to \(T\)
- Compute the discounted payoff for each path: \(e^{-r(T-t)} g(S_T^{(i)})\)
- Estimate: \(\hat{V} = \frac{1}{N}\sum_{i=1}^N e^{-r(T-t)} g(S_T^{(i)})\)
The standard error is \(O(1/\sqrt{N})\), independent of dimension.
Approach 3: Tree Methods¶
Discretize the stochastic process into a recombining lattice and compute expectations by backward induction.
Procedure:
- Build a binomial or trinomial tree for \(S_t\)
- Assign terminal payoffs at the final nodes
- Discount backward through the tree: \(V_j^n = e^{-r\Delta t}[\hat{p}\,V_{j+1}^{n+1} + (1-\hat{p})\,V_j^{n+1}]\)
Trees are a discrete approximation to the PDE and converge to the continuous solution as the mesh refines.
Comparison of Methods¶
| Criterion | PDE (Finite Diff.) | Monte Carlo | Trees |
|---|---|---|---|
| Dimension | 1--3 variables | Any dimension | 1--2 variables |
| Convergence | \(O(\Delta x^2 + \Delta t^2)\) | \(O(1/\sqrt{N})\) | \(O(\Delta t)\) |
| Greeks | Direct from grid | Requires bump-and-reprice | Finite differences on tree |
| Early exercise | Free boundary methods | Least-squares MC (Longstaff-Schwartz) | Natural via backward induction |
| Path dependence | Difficult (add state variables) | Natural | Difficult |
| All strikes at once | Yes (full grid) | No (one payoff per run) | Yes (full tree) |
When to Use PDEs
PDEs excel when:
- The problem is low-dimensional (1--3 state variables)
- You need Greeks (sensitivities) — they come directly from the PDE solution
- You need prices for many strikes/maturities simultaneously
- Early exercise or free boundaries are present (American options)
- High accuracy is required with modest computational cost
When to Use Monte Carlo
Monte Carlo excels when:
- The problem is high-dimensional (basket options, CDOs)
- The payoff is path-dependent (Asian, lookback, barrier)
- The model is complex (stochastic volatility + jumps)
- Moderate accuracy suffices
The Black-Scholes PDE as a Canonical Example¶
The most important financial PDE is the Black-Scholes equation. Under geometric Brownian motion:
the price \(V(t, S)\) of any European derivative with payoff \(g(S_T)\) satisfies:
Key features:
- Parabolic: The equation is second-order in \(S\) and first-order in \(t\), with the classification determined by the positive coefficient \(\frac{1}{2}\sigma^2 S^2 > 0\) for \(S > 0\)
- Terminal condition: \(V(T, S) = g(S)\) specifies the payoff at maturity
- Boundary conditions: Depend on the specific derivative (e.g., \(V(t, 0) = 0\) for a call)
- Universal: The same PDE holds for calls, puts, digitals, and any European payoff — only the terminal condition changes
Derivation via Hedging¶
The PDE can also be derived by a replicating portfolio argument (the original Black-Scholes derivation):
- Form a portfolio \(\Pi = V - \Delta S\) where \(\Delta = \partial V / \partial S\)
- By Ito's lemma: \(d\Pi = \left(\frac{\partial V}{\partial t} + \frac{1}{2}\sigma^2 S^2 \frac{\partial^2 V}{\partial S^2}\right)dt\)
- No-arbitrage requires \(d\Pi = r\Pi\,dt\)
- Substituting and rearranging yields the Black-Scholes PDE
This hedging derivation and the Feynman-Kac probabilistic derivation are two sides of the same coin.
Beyond Black-Scholes¶
The PDE framework extends naturally to richer models:
Stochastic Volatility (Heston Model)¶
The option price satisfies a two-dimensional PDE in \((t, S, v)\):
Jump-Diffusion (Merton Model)¶
With jumps of random size \(J\), the PDE becomes a partial integro-differential equation (PIDE):
where \(\lambda\) is the jump intensity and \(\nu\) is the jump size distribution.
Stochastic Interest Rates¶
When the short rate \(r_t\) follows its own diffusion, bond prices satisfy PDEs of the form:
where \(\mathcal{L}_r\) is the generator of the short-rate process.
The PDE Perspective on Greeks¶
A major advantage of the PDE approach is that Greeks are built into the solution. Since \(V(t,S)\) is computed on an entire grid, its derivatives are immediately available:
| Greek | Definition | PDE Interpretation |
|---|---|---|
| Delta (\(\Delta\)) | \(\frac{\partial V}{\partial S}\) | Slope of the solution surface |
| Gamma (\(\Gamma\)) | \(\frac{\partial^2 V}{\partial S^2}\) | Curvature of the solution surface |
| Theta (\(\Theta\)) | \(\frac{\partial V}{\partial t}\) | Time decay, recoverable from the PDE itself |
| Vega | \(\frac{\partial V}{\partial \sigma}\) | Sensitivity to volatility (requires re-solving) |
Theta from the PDE
The Black-Scholes PDE itself gives \(\Theta\) directly:
This is a model-independent relationship (within Black-Scholes) connecting the Greeks.
Historical Context¶
The PDE approach to finance has a rich history:
- 1900: Bachelier models stock prices as Brownian motion and derives option prices via the heat equation — decades before Black-Scholes
- 1973: Black, Scholes, and Merton derive the Black-Scholes PDE via hedging and connect it to risk-neutral expectations
- 1977: Brennan and Schwartz solve American option pricing as a free boundary PDE problem
- 1993: Heston's stochastic volatility model leads to a 2D pricing PDE with a semi-analytical solution
- 2000s: PIDE methods for jump models; ADI schemes for multi-factor models
Summary¶
The PDE approach to finance rests on three pillars:
| Pillar | Statement |
|---|---|
| Feynman-Kac | Expected values of diffusions satisfy parabolic PDEs |
| Risk-neutral pricing | Derivative prices are discounted expectations under \(\mathbb{Q}\) |
| Generator | The drift and volatility of the underlying determine the PDE coefficients |
PDEs provide the analytical backbone of quantitative finance: they yield closed-form solutions when available, efficient numerical schemes when not, and Greeks as a natural byproduct.
See Also¶
- The SDE-PDE Bridge -- the mathematical connection between diffusions and PDEs
- Classification of Second-Order PDEs -- elliptic, parabolic, hyperbolic
- Boundary Value Problems -- Dirichlet, Neumann, Robin conditions
- Feynman-Kac Formula -- the rigorous statement and proof
- Black-Scholes PDE -- detailed PDE structure
Exercises¶
Exercise 1. A stock follows geometric Brownian motion \(dS_t = rS_t\,dt + \sigma S_t\,dW_t^{\mathbb{Q}}\) with \(r = 0.05\) and \(\sigma = 0.25\). Write down the infinitesimal generator \(\mathcal{L}\) in terms of \(S\), and verify that the Black-Scholes PDE \(\partial_t V + \mathcal{L}V = rV\) matches the standard form \(\partial_t V + rS\,\partial_S V + \frac{1}{2}\sigma^2 S^2\,\partial_{SS}V = rV\).
Solution to Exercise 1
For geometric Brownian motion \(dS_t = rS_t\,dt + \sigma S_t\,dW_t^{\mathbb{Q}}\), the drift is \(\mu(S) = rS\) and the diffusion coefficient is \(\sigma(S) = \sigma S\).
The infinitesimal generator is:
The Feynman-Kac equation with constant discounting rate \(r\) is \(\partial_t V + \mathcal{L}V - rV = 0\), which gives:
Rearranging:
With \(r = 0.05\) and \(\sigma = 0.25\), this becomes:
This matches the standard Black-Scholes PDE form.
Exercise 2. Using the Black-Scholes PDE relationship \(\Theta = rV - rS\Delta - \frac{1}{2}\sigma^2 S^2 \Gamma\), compute Theta for a European call option with \(S = 100\), \(K = 100\), \(T - t = 0.5\), \(r = 0.05\), \(\sigma = 0.20\), given that \(\Delta = 0.5987\), \(\Gamma = 0.0261\), and \(V = 7.97\). Compare your result with the interpretation that Theta represents time decay.
Solution to Exercise 2
Given \(S = 100\), \(K = 100\), \(T - t = 0.5\), \(r = 0.05\), \(\sigma = 0.20\), \(\Delta = 0.5987\), \(\Gamma = 0.0261\), and \(V = 7.97\).
Using the Black-Scholes PDE relationship:
Substituting:
So \(\Theta \approx -7.82\) per year. Converting to daily theta (dividing by 365): \(\Theta_{\text{daily}} \approx -0.0214\) per day.
Interpretation: The option loses approximately \(\$0.02\) per day due to time decay. This is consistent with the fact that an at-the-money option with six months to expiry experiences significant time decay. The three components of theta reveal the balance: the positive \(rV\) term represents the return earned on the option's value, while the \(-rS\Delta\) and \(-\frac{1}{2}\sigma^2 S^2\Gamma\) terms represent the cost of carrying the hedging position. The net effect is negative, confirming that a long call option position decays in value as time passes.
Exercise 3. A Crank-Nicolson scheme has convergence \(O(\Delta x^2 + \Delta t^2)\), while a Monte Carlo estimator with \(N\) paths has error \(O(1/\sqrt{N})\). Suppose each finite-difference step costs \(O(J)\) operations (where \(J\) is the number of spatial grid points), and each Monte Carlo path costs \(O(M)\) operations (where \(M\) is the number of time steps). For a one-dimensional pricing problem requiring four decimal places of accuracy, estimate which method is more efficient and explain your reasoning.
Solution to Exercise 3
Finite difference (Crank-Nicolson): With \(J\) spatial points and \(N_t\) time steps, the total cost is \(O(J \cdot N_t)\) (tridiagonal solves at each time step cost \(O(J)\)). For convergence \(O(\Delta x^2 + \Delta t^2)\), to achieve accuracy \(\epsilon = 10^{-4}\):
- Need \(\Delta x \sim \epsilon^{1/2} = 10^{-2}\), so \(J \sim 1/\Delta x = 100\) spatial points
- Need \(\Delta t \sim \epsilon^{1/2} = 10^{-2}\), so \(N_t \sim 1/\Delta t = 100\) time steps
- Total cost: \(O(J \cdot N_t) = O(10^4)\) operations
Monte Carlo: With \(N\) paths and \(M\) time steps per path, the total cost is \(O(N \cdot M)\). The statistical error is \(O(\sigma_g / \sqrt{N})\) where \(\sigma_g\) is the standard deviation of the discounted payoff. To achieve accuracy \(\epsilon = 10^{-4}\):
- Need \(N \sim \sigma_g^2 / \epsilon^2\). For typical option pricing, \(\sigma_g \sim O(1)\), so \(N \sim 10^8\) paths
- Each path needs \(M \sim 100\) time steps for discretization accuracy
- Total cost: \(O(N \cdot M) = O(10^{10})\) operations
Comparison: For this one-dimensional problem, the PDE method is vastly more efficient: \(O(10^4)\) vs. \(O(10^{10})\) operations. The PDE method wins by roughly six orders of magnitude.
However, this advantage is specific to low dimensions. For a \(d\)-dimensional PDE, the grid cost becomes \(O(J^d \cdot N_t) = O(100^d \cdot 100)\), which grows exponentially in \(d\). Monte Carlo cost remains \(O(10^{10})\) regardless of \(d\) (the \(O(1/\sqrt{N})\) rate is dimension-independent). The crossover occurs around \(d \approx 3\)--\(4\), beyond which Monte Carlo becomes more practical.
Exercise 4. Derive the Black-Scholes PDE via the hedging argument. Starting from a portfolio \(\Pi = V - \Delta S\) with \(\Delta = \partial V/\partial S\), apply Ito's lemma to \(dV\), substitute, and impose \(d\Pi = r\Pi\,dt\) to obtain the PDE. Identify where the no-arbitrage condition enters and explain why the physical drift \(\mu\) drops out.
Solution to Exercise 4
Step 1: Assume the stock follows \(dS_t = \mu S_t\,dt + \sigma S_t\,dW_t\) under the physical measure, and \(V(t, S)\) is a twice-differentiable function. By Ito's lemma:
Since \((dS)^2 = \sigma^2 S^2\,dt\):
Step 2: Form the hedged portfolio \(\Pi = V - \Delta S\) with \(\Delta = \frac{\partial V}{\partial S}\):
The \(dW_t\) terms cancel exactly because \(\Delta = \partial V/\partial S\). The portfolio is locally riskless.
Step 3: No-arbitrage requires that a riskless portfolio earns the risk-free rate:
Step 4: Equating the two expressions for \(d\Pi\):
Rearranging:
Where no-arbitrage enters: The condition \(d\Pi = r\Pi\,dt\) is the no-arbitrage requirement. If \(d\Pi > r\Pi\,dt\), one could borrow at rate \(r\) and invest in \(\Pi\) for a riskless profit; if \(d\Pi < r\Pi\,dt\), reverse the trade.
Why \(\mu\) drops out: The physical drift \(\mu\) appears in \(dS\) but cancels when forming \(d\Pi = dV - \Delta\,dS\) because the hedge ratio \(\Delta = \partial V/\partial S\) eliminates all exposure to the stock's random movements. Since the portfolio is riskless, its return cannot depend on \(\mu\) -- it must equal \(r\) regardless of the stock's expected return. This is the essence of risk-neutral pricing.
Exercise 5. The Heston model produces a two-dimensional PDE in \((t, S, v)\) for the option price \(V\). Explain why PDE methods become impractical for basket options on \(d \geq 4\) assets, even though they provide excellent accuracy in low dimensions. Relate your answer to the "curse of dimensionality" and contrast with the dimension-independence of Monte Carlo convergence rates.
Solution to Exercise 5
For a basket option on \(d\) assets, the pricing PDE lives in \(d + 1\) dimensions (one time variable plus \(d\) asset prices):
The curse of dimensionality: A finite-difference grid with \(J\) points per spatial dimension requires \(J^d\) total grid points. With \(J = 100\) and \(d = 4\): \(100^4 = 10^8\) grid points per time step. For \(d = 10\): \(100^{10} = 10^{20}\) grid points -- far beyond any computer's memory or processing capacity.
The computational cost scales as \(O(J^d \cdot N_t)\), growing exponentially in \(d\). For \(d \geq 4\), the memory required to store the grid alone becomes prohibitive, and iterative solvers for the resulting linear systems become impractical.
Monte Carlo dimension-independence: The Monte Carlo estimator \(\hat{V} = \frac{1}{N}\sum_{i=1}^N e^{-rT}g(S_T^{(i)})\) has standard error \(\sigma_g / \sqrt{N}\), where \(\sigma_g\) is the payoff standard deviation. Crucially, this convergence rate \(O(1/\sqrt{N})\) does not depend on \(d\). Simulating a path of \(d\) correlated assets costs \(O(d)\) per step (via Cholesky decomposition of the correlation matrix), so total cost is \(O(N \cdot M \cdot d)\) -- polynomial in \(d\), not exponential.
Summary: PDE methods excel in 1--3 dimensions due to their superior convergence rate (\(O(h^2)\) vs. \(O(1/\sqrt{N})\)) and direct access to Greeks. For \(d \geq 4\), the exponential growth of the grid makes PDEs impractical, while Monte Carlo remains viable due to its dimension-independent convergence.
Exercise 6. In the Merton jump-diffusion model, the pricing equation is a partial integro-differential equation (PIDE) with an integral term \(\lambda\int_{\mathbb{R}}[V(t, Se^z) - V(t, S)]\,\nu(dz)\). Explain physically what this integral term represents, why it is nonlocal, and why a standard finite difference scheme for the Black-Scholes PDE cannot be applied directly without modification.
Solution to Exercise 6
The integral term in the Merton PIDE is:
Physical meaning: This term represents the expected change in option value due to jumps. At rate \(\lambda\) (the jump intensity), the stock price jumps from \(S\) to \(Se^z\), where \(z\) is the log-jump size drawn from the distribution \(\nu\). The integral averages \(V(t, Se^z) - V(t, S)\) over all possible jump sizes, weighted by \(\nu(dz)\).
- \(V(t, Se^z)\): option value after a jump of size \(z\)
- \(V(t, S)\): option value before the jump
- The difference captures the instantaneous P\&L from a jump event
Why it is nonlocal: The integral term evaluates \(V\) at the point \(Se^z\) for all possible \(z\). When \(z\) is large, \(Se^z\) can be far from \(S\). This means the PIDE at the point \((t, S)\) depends on the values of \(V\) at distant points in the spatial domain. This is fundamentally different from a standard PDE, where the equation at \((t, S)\) depends only on \(V\) and its derivatives at \((t, S)\) -- purely local information.
Why standard finite differences fail: A standard finite difference scheme for the Black-Scholes PDE approximates \(\partial_S V\) and \(\partial_{SS} V\) using values of \(V\) at neighboring grid points (e.g., \(V_{j-1}\), \(V_j\), \(V_{j+1}\)). This works because the PDE is local. The integral term, however, requires evaluating \(V(t, Se^z)\) at points that may span the entire grid. To handle this:
- The integral must be discretized separately, typically by numerical quadrature over the jump-size distribution
- If \(\nu\) has fat tails (e.g., Gaussian jumps), the integral couples distant grid points, producing a dense matrix rather than the tridiagonal structure of the diffusion part
- Implicit time-stepping for the integral part leads to a full linear system (not tridiagonal), dramatically increasing computational cost
- Common approaches split the operator: treat the diffusion part implicitly (tridiagonal) and the integral part explicitly, or use FFT-based methods to evaluate the convolution efficiently
Exercise 7. The text states that tree methods are "a discrete approximation to the PDE and converge to the continuous solution as the mesh refines." For a one-step binomial tree with up factor \(u = e^{\sigma\sqrt{\Delta t}}\), down factor \(d = 1/u\), and risk-neutral probability \(\hat{p} = (e^{r\Delta t} - d)/(u - d)\), show that the backward pricing equation \(V = e^{-r\Delta t}[\hat{p}V_u + (1-\hat{p})V_d]\) is consistent with the Black-Scholes PDE in the limit \(\Delta t \to 0\) by performing a Taylor expansion to second order in \(\sqrt{\Delta t}\).
Solution to Exercise 7
The one-step binomial pricing equation is:
with \(u = e^{\sigma\sqrt{\Delta t}}\), \(d = e^{-\sigma\sqrt{\Delta t}}\), and \(\hat{p} = \frac{e^{r\Delta t} - d}{u - d}\).
Taylor expansions (letting \(h = \sqrt{\Delta t}\)):
The risk-neutral probability:
Now expand \(V\) at the up and down nodes. Let \(V_S = \partial V/\partial S\), \(V_{SS} = \partial^2 V/\partial S^2\), and \(V_t = \partial V/\partial t\):
Computing the weighted average:
Substituting \(2\hat{p} - 1 = \frac{(r - \frac{1}{2}\sigma^2)h}{\sigma} + O(h^2)\):
Multiplying by \(e^{-r\Delta t} = 1 - rh^2 + O(h^4)\):
For the binomial equation \(V = e^{-r\Delta t}[\hat{p}\,V_u + (1-\hat{p})\,V_d]\) to hold to order \(h^2\), we need:
This is exactly the Black-Scholes PDE, confirming that the binomial tree is consistent with it in the limit \(\Delta t \to 0\).