Chapter 23 · Connections to PDEs

The Feynman-Kac Theorem

The bridge between probability (random paths) and engineering (heat flow equations) — two languages, one answer.
Probability meets engineering

Everything in Chapters 18–22 priced derivatives by asking: "What is the average payoff across all possible random paths?" This is the probabilistic view — compute an expectation. Chapter 23 asks a completely different question: "How does the option price flow and spread across time and stock price levels, like heat spreading through a room?" This is the engineering view — solve a partial differential equation. The stunning result: both answers are identical.

Two views of the same problem — random paths averaged vs heat equation solved Left panel shows many random stock paths with a payoff being averaged at expiry — the Monte Carlo view. Right panel shows a smooth option price surface spreading across time and price — the PDE view. Probabilistic View (Monte Carlo) PDE View (Heat Equation) Average 10,000 path payoffs → price Solve PDE across time-price grid → price =

Same problem, two solution methods. The Feynman-Kac Theorem proves they always give the identical answer.

1 · SDEs — the recipe for randomness

Before we can build the bridge, we need to understand the starting point: a Stochastic Differential Equation (SDE). An SDE is simply the recipe that describes how a price moves over each tiny instant of time. Every SDE has exactly two ingredients.

🚗 The Drift β(t, X)

β(t,X) dt

The predictable part. Like a car's cruise control — a steady, deterministic push in a specific direction. In GBM, the drift is r·X (the risk-neutral growth at the risk-free rate).

🧊 The Diffusion γ(t, X)

γ(t,X) dW

The random part. Like a car hitting patches of ice — unpredictable shaking caused by Brownian Motion. In GBM, the diffusion is σ·X (volatility scaled by price).

dX(t) = β(t, X(t)) dt + γ(t, X(t)) dW(t)
The SDE is the complete description of how X evolves.
The Markov Property ensures: the future depends only on X now — not on how we got here.

The Markov Property is what makes SDEs useful for pricing. Because tomorrow's price only depends on today's price (not the entire history), we can build a PDE that tracks the option value across a two-dimensional grid of (time, price) — without needing to store the full path history. This is the computational efficiency that makes PDE methods orders of magnitude faster than Monte Carlo for standard European options.

2 · The Feynman-Kac theorem — the magic bridge

This is the heart of Chapter 23. The Feynman-Kac Theorem states that the solution to a PDE — a calculus problem about how prices flow across a grid — is identical to the solution of a risk-neutral expectation — a probability problem about averaging random paths. The two approaches are mathematically equivalent.

⚡ The Feynman-Kac Theorem ⚡
Probabilistic (Expectation)
V(t,x) = 𝔼̃ₓ[e^(−r(T−t)) h(X(T))]
Average the discounted payoff h(X(T)) over all risk-neutral paths starting from X(t) = x.
Calculus (PDE)
Vₜ + βVₓ + ½γ²Vₓₓ = rV
Solve this PDE with terminal condition V(T,x) = h(x). The PDE describes how V flows backward through time.
Feynman-Kac as two routes to the same mountain top — Monte Carlo averaging and PDE solving both reach the same option price A mountain diagram showing two paths up the same mountain. The left path is labeled Monte Carlo averaging and goes up through many random steps. The right path is labeled PDE solving and goes up through a smooth grid. Both reach the same peak labeled option price. Option Price Monte Carlo Average random paths PDE Solver Solve on time-price grid Two routes up the same mountain — always the same answer at the top

Monte Carlo averages paths from the bottom up. PDE solves backwards from the expiry payoff. Both reach the same option price.

The room-temperature analogy makes this intuitive: to find the temperature at a point in a room, you could either release 1,000 tiny thermometer-sensors and average their readings (Monte Carlo) — or you could write down the heat equation and solve it mathematically (PDE). In finance, the payoff at expiry is the "heat source," and the option price today is the temperature at an earlier time. Both methods are correct. You choose based on speed and convenience.

3 · The three-step recipe for any PDE

For any market model driven by an SDE, Shreve provides a universal three-step process to convert the stochastic problem into a PDE. This recipe works for stocks, interest rates, exchange rates — any Markovian process.

1
Find the Martingale — discounted value process
Under the risk-neutral measure ℙ̃, the discounted derivative price e^(−rt)V(t, X(t)) must be a Martingale. This is Girsanov's gift: under ℙ̃, discounted prices drift nowhere.
D(t)V(t, X(t)) is a ℙ̃-Martingale
2
Take the differential — apply Itô-Doeblin
Apply the Itô-Doeblin formula to d(D(t)V(t,X(t))). This generates a dt term (the drift) and a dW̃ term (the noise). Since the result is a Martingale, the dt term must equal zero.
d(DV) = [dt terms] dt + [dW̃ terms] dW̃
3
Kill the drift — set the dt coefficient to zero
Setting the dt coefficient to zero — because Martingales have no drift — produces a PDE that V(t,x) must satisfy. The terminal condition V(T,x) = h(x) (the payoff at expiry) closes the system.
Set [dt coefficient] = 0 → the PDE for V
Result: The Black-Scholes-Merton PDE
Vₜ + rxVₓ + ½σ²x²Vₓₓ = rV
Vₜ
Theta
How price changes as time passes. The "time decay" of the option value.
rxVₓ
Delta
How price changes as the stock moves. The slope of V in the x direction.
½σ²x²Vₓₓ
Gamma
Curvature profit from volatility. The Itô correction — always surviving.

The BSM PDE is the direct output of applying the three-step recipe to GBM (β = rx, γ = σx). Any other SDE produces a different PDE with the same structure but different coefficients. The recipe is universal — only the ingredients change. The Hull-White interest rate model and the CIR model (below) are both products of this exact same three-step process.

4 · Real-world models — interest rate PDEs

Applying the three-step recipe to interest rate models produces PDEs for bond prices. Two famous models illustrate the trade-offs between mathematical tractability and physical realism.

Hull-White Model

dR = (a − bR)dt + σ dW

Mean-reverting drift (rubber band toward a/b). Constant volatility σ. The resulting PDE has a closed-form solution — fast and analytically tractable. Used widely for interest rate derivatives.

⚠ Flaw: Normal distribution → negative rates possible. A model flaw that became reality in Europe post-2008.

Cox-Ingersoll-Ross (CIR)

dR = (a − bR)dt + σ√R dW

Same rubber-band drift. But volatility scales as σ√R — shrinks toward zero as R → 0. The PDE's volatility term vanishes at the boundary, preventing negative rates mathematically.

✓ Advantage: Rates stay positive. More realistic for bonds and interest rate options.
Hull-White vs CIR near zero — Hull-White crosses into negative rates while CIR bounces off the floor Two interest rate paths shown. Hull-White in blue dashes crosses below zero. CIR in gold remains above zero, bouncing off the floor as its volatility shrinks near zero. R = 0% Target rate Hull-White — can go negative ⚠ CIR — bounces off zero ✓

Hull-White (blue dashed) crosses into negative rates. CIR (gold) bounces off the zero floor as σ√R → 0 — the model self-corrects.

5 · The volatility smile and surface

The BSM PDE assumes σ is a constant. But in real markets, options traders observe that different strike prices and maturities imply different volatilities. If BSM were perfectly correct, all implied volatilities would be the same. They are not — and the pattern they form reveals important market structure.

The Volatility Smile (one maturity, varying strike)

Deep in-the-money and out-of-the-money options consistently show higher implied volatility than at-the-money — the "smile" effect. BSM cannot explain it with a constant σ.

When you track the smile across multiple maturities, you get a Volatility Surface — a 3D landscape where σ is a function σ(t, x) that varies with both time to expiry and the strike price. The CEV model captures part of this by letting σ depend on the current stock level.

Standard BSM — constant σ

σ is a single number. Fast and clean. But implies a flat volatility surface — which no real market ever shows. Systematically misprices out-of-the-money options.

PDE: Vₜ + rxVₓ + ½σ²x²Vₓₓ = rV

Local Volatility — σ(t, x)

σ is a function of time and stock price. Matches any observed smile perfectly by construction. Produces a more complex PDE but still solvable numerically.

PDE: Vₜ + rxVₓ + ½σ(t,x)²x²Vₓₓ = rV

The Dupire formula (1994) is the classic result: given observed market option prices at all strikes and maturities, you can extract σ(t,x) exactly — the unique local volatility surface that makes BSM consistent with every observed price simultaneously. This is the PDE approach running in reverse: instead of solving the PDE to find prices, you use observed prices to recover the PDE coefficient σ(t,x).

6 · Monte Carlo vs PDE — choosing your tool

The Feynman-Kac Theorem proves both methods give the same answer. The choice between them is entirely a question of computational efficiency for your specific problem.

FeatureMonte CarloPDE (Finite Difference)
How it works Simulate thousands of paths. Average the payoffs. Solve backward on a time-price grid.
Best for Path-dependent options (Asian, lookback, barrier) European and American options in 1–2 dimensions
Speed in 1D Slow (needs many paths for accuracy) Very fast — microseconds for standard options
Speed in high dimensions Scales well — "curse of dimensionality" avoided Exponential blowup — impractical above 3D
Handles stochastic vol? Yes — just add a second random driver Yes — but adds a dimension to the PDE grid
Error convergence 1/√N — slow (need 4× paths to halve error) O(Δt², Δx²) — fast with fine grid
Strategy type Complex path-dependent backtests Real-time options pricing engine
Try it — Feynman-Kac in action

Compare the Monte Carlo price (averaging random paths) with the BSM formula price (the closed-form PDE solution). Both implement the same Feynman-Kac theorem. Watch the Monte Carlo price converge to the exact BSM answer as you increase the number of paths — and observe how much faster the PDE answer is.

$100
$100
20%
5,000
BSM (exact PDE)
Monte Carlo avg
MC error
95% CI width

Gold dashed line = exact BSM price (PDE solution)  |  Blue bars = Monte Carlo path payoff distribution  |  As paths increase, the MC average converges to the BSM price

Chapter 23 — the Feynman-Kac bridge in five ideas:

What comes next: The multidimensional extension of these PDE methods — pricing options on multiple correlated stocks, interest rate term structure models, and the computational machinery (finite difference grids, ADI schemes) that makes PDE pricing practical at scale.