Boundary Value Problems¶
A partial differential equation alone does not determine a unique solution. We must also specify boundary conditions -- constraints on the solution at the edges of the domain -- and initial or terminal conditions. The choice of conditions is dictated by the PDE type and the physics (or finance) of the problem. A problem is well-posed when these conditions guarantee existence, uniqueness, and continuous dependence on the data.
The Parabolic Initial-Boundary Value Problem¶
The standard setting for financial PDEs is a parabolic equation on a bounded domain:
with terminal condition:
and boundary conditions on \(\partial\Omega\), where \(\Omega \subset \mathbb{R}^d\) is the spatial domain and:
is the pricing operator (generator minus discounting).
The parabolic boundary consists of the terminal time and the spatial boundary:
Time Direction
In finance, the terminal condition is given at \(t = T\) and the PDE is solved backward. Mathematically this is equivalent (via \(\tau = T - t\)) to a forward initial value problem.
Types of Boundary Conditions¶
Dirichlet Condition (Prescribed Value)¶
The solution is prescribed on the boundary. This is the most common type in finance.
Financial interpretation: The option price is known at the boundary.
Examples:
- European call as \(S \to 0\): \(V(t, 0) = 0\) (worthless if stock is zero)
- European call as \(S \to \infty\): \(V(t, S) \approx S - Ke^{-r(T-t)}\) (deep in-the-money)
- Knock-out barrier option at barrier \(B\): \(V(t, B) = 0\) (option knocked out)
- Killed diffusion: Process absorbed at boundary \(\implies\) \(u = 0\) on \(\partial\Omega\)
Neumann Condition (Prescribed Flux)¶
where \(\frac{\partial}{\partial n}\) is the outward normal derivative. The rate of change of the solution is prescribed, not the value itself.
Financial interpretation: The option's delta is specified at the boundary.
Examples:
- Put option as \(S \to \infty\): \(\frac{\partial V}{\partial S} \to 0\) (delta approaches zero)
- Reflecting boundary: Probability mass is reflected back into the domain
- Symmetry condition: \(\frac{\partial u}{\partial x} = 0\) at a line of symmetry
Robin Condition (Mixed)¶
A linear combination of the solution value and its normal derivative is specified. This interpolates between Dirichlet (\(\beta = 0\)) and Neumann (\(\alpha = 0\)).
Financial interpretation: A partial absorption/reflection condition, or a linear relationship between price and delta at the boundary.
Example: Elastic barriers where partial rebates are paid upon hitting the boundary.
Comparison of Boundary Conditions¶
| Type | Specifies | Probabilistic Meaning | Financial Example |
|---|---|---|---|
| Dirichlet | \(u = h\) | Absorption (killing) at boundary | Barrier option knockout |
| Neumann | \(\partial_n u = h\) | Reflection at boundary | No-flux condition |
| Robin | \(\alpha u + \beta \partial_n u = h\) | Partial absorption/reflection | Elastic barrier |
Well-Posedness in the Sense of Hadamard¶
A problem is well-posed if it satisfies three conditions:
- Existence: A solution exists
- Uniqueness: The solution is unique (in an appropriate function space)
- Continuous dependence: Small changes in the data lead to small changes in the solution
where \(g\) is the terminal data, \(h\) is the boundary data, and \(f\) is the source term.
Ill-Posedness
Not all combinations of PDE type and boundary conditions are well-posed:
- Backward heat equation (\(u_t + \frac{1}{2}u_{xx} = 0\) with initial data at \(t = 0\), solving forward): ill-posed. Arbitrarily small perturbations in data can produce arbitrarily large changes in the solution. This reflects the irreversibility of diffusion.
- Laplace equation with Cauchy data (prescribing both \(u\) and \(\partial_n u\) on part of the boundary): ill-posed (Hadamard's counterexample).
Correct Pairings¶
| PDE Type | Well-Posed Problem | Condition Type |
|---|---|---|
| Elliptic | Dirichlet/Neumann/Robin on entire \(\partial\Omega\) | Boundary only |
| Parabolic | Terminal data + boundary conditions on \(\partial\Omega\) | Initial/terminal + boundary |
| Hyperbolic | Initial data \(u(0,x)\), \(u_t(0,x)\) + boundary | Two initial + boundary |
The Maximum Principle and Uniqueness¶
For parabolic equations, the maximum principle provides both uniqueness and continuous dependence.
Theorem (Parabolic Maximum Principle): If \(u\) satisfies \(\partial_t u + \mathcal{L}u \leq 0\) in \((0,T) \times \Omega\) (with \(r \geq 0\) in \(\mathcal{L}\)), then:
where \(\Gamma\) is the parabolic boundary.
Corollary (Uniqueness): If \(u_1\) and \(u_2\) both solve the same parabolic IBVP, then \(w = u_1 - u_2\) satisfies the homogeneous problem with zero data on \(\Gamma\). By the maximum principle, \(w \leq 0\) and \(-w \leq 0\), so \(w = 0\). \(\square\)
Corollary (Stability): The continuous dependence estimate:
follows directly from the maximum principle.
Boundary Conditions in the Black-Scholes Framework¶
For a European derivative with payoff \(g(S)\) under Black-Scholes dynamics, the PDE domain is typically \((t, S) \in [0, T] \times [0, \infty)\).
Terminal Condition¶
This is universal -- it encodes the derivative's payoff structure.
Boundary at \(S = 0\)¶
At \(S = 0\), the Black-Scholes operator degenerates: \(\frac{1}{2}\sigma^2 S^2 \partial_{SS} \to 0\). The PDE reduces to:
For a call (\(g(0) = 0\)): \(V(t, 0) = 0\). For a put (\(g(0) = K\)): \(V(t, 0) = Ke^{-r(T-t)}\).
Why No Boundary Condition Is Needed at \(S = 0\)
The degeneracy of the PDE at \(S = 0\) means \(S = 0\) is an absorbing boundary for geometric Brownian motion: once \(S_t = 0\), it remains there forever. Probabilistically, \(\mathbb{P}(S_t = 0 \text{ for some } t) = 0\), so the boundary is never reached. The PDE determines its own boundary behavior there.
Boundary as \(S \to \infty\)¶
An artificial truncation at \(S = S_\text{max}\) requires a far-field boundary condition. Common choices:
| Derivative | Far-Field Condition | Justification |
|---|---|---|
| European call | \(V \approx S - Ke^{-r(T-t)}\) | Deep ITM asymptotics |
| European put | \(V \approx 0\) | Worthless for large \(S\) |
| General | \(\frac{\partial^2 V}{\partial S^2} \approx 0\) | Gamma vanishes far from strike |
Barrier Options and Domain Truncation¶
Barrier options provide the cleanest financial example of Dirichlet boundary conditions.
Down-and-Out Call¶
Domain: \((t, S) \in [0, T] \times (B, \infty)\) where \(B < K\) is the barrier.
The Dirichlet condition \(V(t, B) = 0\) encodes the knockout feature: the option becomes worthless upon hitting the barrier.
Up-and-Out Put¶
Domain: \((t, S) \in [0, T] \times (0, B)\) where \(B > K\) is the upper barrier.
Rebate Barrier Option¶
If a rebate \(R\) is paid upon knockout:
Free Boundary Problems and American Options¶
American options introduce a fundamentally new type of boundary condition: the free boundary (also called the optimal exercise boundary).
The holder can exercise at any time \(t \leq T\), so the price satisfies the linear complementarity problem:
This decomposes the domain into:
- Continuation region: \(V > g\) and \(\frac{\partial V}{\partial t} + \mathcal{L}V = 0\) (standard PDE)
- Exercise region: \(V = g\) (option exercised immediately)
- Free boundary \(S^*(t)\): The curve separating the two regions
At the free boundary, two conditions hold simultaneously:
Smooth Pasting
The smooth-pasting condition states that not only the price but also the delta is continuous across the free boundary. This is a necessary condition for the optimal exercise policy and emerges naturally from the variational inequality formulation.
Existence and Regularity Theory¶
For the parabolic IBVP with Dirichlet conditions:
Theorem (Classical Existence): If \(\Omega\) has smooth boundary, coefficients are smooth, \(a_{ij}\) is uniformly elliptic, and data \((f, g, h)\) are compatible and smooth, then there exists a unique classical solution \(u \in C^{2,1}(\overline{Q}_T)\).
Theorem (Weak/Viscosity Solutions): Under weaker conditions (bounded measurable coefficients, continuous data), the problem admits a unique viscosity solution. This is the appropriate framework for degenerate equations and non-smooth payoffs encountered in finance.
Summary¶
| Aspect | Description |
|---|---|
| Dirichlet | Prescribes \(u\) on boundary; absorbing/knockout |
| Neumann | Prescribes \(\partial_n u\); reflecting/no-flux |
| Robin | Linear combination; partial absorption |
| Well-posedness | Existence + uniqueness + stability |
| Maximum principle | Ensures uniqueness and stability for parabolic PDEs |
| Free boundary | American options; exercise boundary determined as part of solution |
The choice of boundary conditions reflects the financial structure of the derivative: knockout barriers impose Dirichlet conditions, no-arbitrage constraints determine far-field behavior, and early exercise creates free boundaries.
See Also¶
- Why PDEs in Finance -- motivation for the PDE approach
- Classification of Second-Order PDEs -- which boundary conditions are appropriate for which PDE type
- Maximum Principle -- the tool for uniqueness and stability
- Kolmogorov Backward Equation -- boundary conditions for diffusion expectations
Exercises¶
Exercise 1. For a European put option under Black-Scholes dynamics, state the terminal condition \(V(T, S)\) and the boundary conditions at \(S = 0\) and as \(S \to \infty\). Verify that the boundary condition at \(S = 0\) is consistent with the degenerate PDE \(\partial_t V = rV\) at that point.
Solution to Exercise 1
The European put has payoff \(g(S) = (K - S)^+\), so the terminal condition is:
Boundary at \(S = 0\): At \(S = 0\), the Black-Scholes PDE degenerates since \(\frac{1}{2}\sigma^2 S^2 \to 0\), leaving \(\frac{\partial V}{\partial t} = rV\). The payoff at \(S = 0\) is \(g(0) = K\), so the solution of \(\partial_t V = rV\) with terminal value \(K\) is:
This is the present value of receiving \(K\) at maturity, which is consistent: if the stock is at zero, it stays at zero (absorbing state for GBM), and the put pays \(K\) with certainty.
Boundary as \(S \to \infty\): For large \(S\), the put is deep out-of-the-money and becomes worthless:
Verification: At \(S = 0\) the PDE reduces to \(\partial_t V = rV\). Substituting \(V(t, 0) = Ke^{-r(T-t)}\):
confirming consistency with the degenerate PDE.
Exercise 2. Consider a down-and-out call option with barrier \(B < K\) and strike \(K\). Write the complete initial-boundary value problem: the Black-Scholes PDE on the domain \(S \in (B, \infty)\), the terminal condition, and the Dirichlet boundary condition at \(S = B\). What is the probabilistic interpretation of the Dirichlet condition?
Solution to Exercise 2
The complete initial-boundary value problem for the down-and-out call with barrier \(B < K\) and strike \(K\) is:
PDE on the domain \(S \in (B, \infty)\), \(t \in [0, T]\):
Terminal condition at \(t = T\):
Dirichlet boundary condition at \(S = B\):
Far-field condition as \(S \to \infty\):
Probabilistic interpretation of the Dirichlet condition: The condition \(V(t, B) = 0\) corresponds to the option being killed (knocked out) when the stock price hits the barrier \(B\). In probabilistic terms, if \(\tau_B = \inf\{s \geq t : S_s = B\}\) is the first hitting time of the barrier, then the option pays zero if \(\tau_B \leq T\). The Dirichlet condition encodes this absorbing boundary: the diffusion is killed upon reaching \(B\), and the discounted payoff from that point onward is zero.
Exercise 3. The Hadamard well-posedness conditions require existence, uniqueness, and continuous dependence on data. Using the maximum principle, prove the continuous dependence estimate
for two solutions \(u_1\), \(u_2\) of the parabolic IBVP with different terminal data \(g_1\), \(g_2\) and boundary data \(h_1\), \(h_2\).
Solution to Exercise 3
Let \(u_1\) and \(u_2\) solve the parabolic IBVP with the same operator but different data:
with terminal data \(u_i(T, x) = g_i(x)\) and boundary data \(u_i(t, x) = h_i(t, x)\) on \(\partial\Omega\).
Define \(w = u_1 - u_2\). Then \(w\) satisfies:
with terminal data \(w(T, x) = g_1(x) - g_2(x)\) and boundary data \(w = h_1 - h_2\) on \(\partial\Omega\).
By the parabolic maximum principle, the maximum of \(w\) over \(\overline{Q}_T = [0,T] \times \overline{\Omega}\) is attained on the parabolic boundary \(\Gamma = \{T\} \times \Omega \cup [0,T] \times \partial\Omega\):
Applying the same argument to \(-w\) (which also satisfies the homogeneous PDE):
Combining these two inequalities:
This is the continuous dependence estimate, confirming well-posedness in the sense of Hadamard. \(\square\)
Exercise 4. Explain why the backward heat equation \(\partial_t u + \frac{1}{2}\partial_{xx} u = 0\) with initial data \(u(0, x) = f(x)\) solved forward in time is ill-posed. Construct a sequence of initial data \(f_n(x) = \frac{1}{n}\sin(nx)\) such that \(\|f_n\|_\infty \to 0\) but the solution at any \(t > 0\) blows up. What does this say about the irreversibility of diffusion?
Solution to Exercise 4
Consider the backward heat equation \(\partial_t u + \frac{1}{2}\partial_{xx} u = 0\) with initial data \(u(0, x) = f(x)\), solved forward in time (i.e., for \(t > 0\)).
Why it is ill-posed: Under the change \(\tau = -t\), this becomes the standard heat equation solved backward in time. Equivalently, by separation of variables or Fourier analysis, an initial mode \(e^{ikx}\) evolves as \(e^{ikx + \frac{1}{2}k^2 t}\), which grows exponentially in \(t\). High-frequency components are amplified rather than damped.
Explicit construction: Take \(f_n(x) = \frac{1}{n}\sin(nx)\). Then \(\|f_n\|_\infty = \frac{1}{n} \to 0\) as \(n \to \infty\).
The solution of \(\partial_t u + \frac{1}{2}\partial_{xx}u = 0\) with \(u(0,x) = \frac{1}{n}\sin(nx)\) is:
One can verify: \(\partial_t u_n = \frac{n}{2}\sin(nx)\,e^{\frac{1}{2}n^2 t}\) and \(\frac{1}{2}\partial_{xx}u_n = -\frac{n}{2}\sin(nx)\,e^{\frac{1}{2}n^2 t}\), so \(\partial_t u_n + \frac{1}{2}\partial_{xx}u_n = 0\).
At any fixed \(t > 0\):
So the initial data converges to zero uniformly, but the solution blows up for any \(t > 0\). This violates continuous dependence on data.
Connection to irreversibility of diffusion: The heat equation (forward in time) smooths and loses information -- high-frequency details are exponentially damped. Reversing this process requires reconstructing those lost details, which amplifies any noise exponentially. This reflects the fundamental thermodynamic irreversibility of diffusion: you cannot "unmix" a diffused substance.
Exercise 5. For an American put option, the free boundary \(S^*(t)\) separates the continuation region from the exercise region. State the smooth-pasting condition at \(S^*(t)\) and explain why both value matching and smooth pasting must hold simultaneously. What would happen to the hedging strategy if the delta were discontinuous across the exercise boundary?
Solution to Exercise 5
For an American put with payoff \(g(S) = (K - S)^+\), the free boundary \(S^*(t)\) is the critical stock price below which immediate exercise is optimal.
Value matching at the free boundary:
This states that the option price equals the exercise value at the boundary -- continuity of the price across the exercise boundary.
Smooth pasting at the free boundary:
This states that the delta of the option equals the delta of the exercise payoff at the boundary -- the first derivative is also continuous.
Why both must hold simultaneously: Value matching alone is insufficient because it only ensures price continuity. The smooth-pasting condition is a necessary optimality condition that arises from the variational inequality formulation. If only value matching held but \(\partial_S V \neq g'\) at \(S^*(t)\), then the exercise boundary could be shifted slightly to improve the option value, contradicting optimality.
Consequences of discontinuous delta: If the delta were discontinuous across the exercise boundary, the hedging portfolio \(\Pi = V - \Delta S\) would require a discrete, instantaneous rebalancing at the moment the stock crosses \(S^*(t)\). This would create:
- A jump in the hedge ratio, leading to a non-self-financing portfolio
- Infinite transaction costs in the presence of market frictions
- An inability to perfectly replicate the option payoff
The smooth-pasting condition ensures that the hedge ratio transitions continuously, making delta hedging implementable in practice.
Exercise 6. At \(S = 0\), the Black-Scholes PDE degenerates because \(\frac{1}{2}\sigma^2 S^2 \to 0\). Explain why this means no boundary condition needs to be imposed at \(S = 0\). Connect this to the probabilistic fact that geometric Brownian motion can never reach zero: \(\mathbb{P}(S_t = 0 \text{ for some } t > 0) = 0\).
Solution to Exercise 6
The Black-Scholes PDE for \(V(t, S)\) is:
At \(S = 0\), the coefficient of the second-order term vanishes: \(\frac{1}{2}\sigma^2 S^2 \big|_{S=0} = 0\). Likewise, the first-order drift term vanishes: \(rS\big|_{S=0} = 0\). The PDE degenerates to the ODE \(\frac{\partial V}{\partial t} = rV\), which has the unique solution \(V(t, 0) = e^{-r(T-t)}g(0)\) given the terminal condition \(V(T, 0) = g(0)\).
Why no boundary condition is needed: Since the PDE itself determines \(V(t, 0)\) uniquely from the terminal condition, imposing an additional boundary condition at \(S = 0\) would overdetermine the problem. The degeneracy of the second-order term means the PDE transitions from a parabolic equation (requiring boundary data) to an ODE (self-contained) at \(S = 0\).
Probabilistic connection: Geometric Brownian motion \(dS_t = rS_t\,dt + \sigma S_t\,dW_t\) has the explicit solution \(S_t = S_0 \exp\left((r - \frac{1}{2}\sigma^2)t + \sigma W_t\right)\). Since the exponential function is strictly positive for all finite values of its argument, \(S_t > 0\) for all \(t \geq 0\) whenever \(S_0 > 0\). Therefore:
The boundary \(S = 0\) is never reached by the diffusion. In the Feller classification, \(S = 0\) is an entrance boundary (or natural boundary, depending on parameters): the process cannot reach it from the interior. Since no sample paths touch this boundary, no boundary condition is needed to determine the conditional expectation \(\mathbb{E}^{\mathbb{Q}}[e^{-r(T-t)}g(S_T) \mid S_t = S]\).
In contrast, a finite barrier \(B > 0\) is reachable by the diffusion with positive probability, so a Dirichlet condition at \(S = B\) carries genuine information about what happens upon arrival.
Exercise 7. Compare Dirichlet, Neumann, and Robin boundary conditions in terms of their probabilistic meaning for a diffusion process. For a particle undergoing Brownian motion in the interval \((a, b)\), describe the behavior at the boundary under each condition: absorption (killing), reflection, and partial absorption with probability \(\alpha\).
Solution to Exercise 7
Consider a particle undergoing standard Brownian motion \(X_t\) in the interval \((a, b)\). The three boundary conditions correspond to distinct physical behaviors at the endpoints:
Dirichlet condition \(u(t, a) = 0\), \(u(t, b) = 0\): This corresponds to absorption (killing). When the Brownian particle reaches the boundary \(a\) or \(b\), it is immediately removed from the system. In the PDE context, \(u(t, x) = \mathbb{E}_x[g(X_T)\,\mathbf{1}_{\{\tau > T\}}]\) where \(\tau = \inf\{t : X_t \notin (a,b)\}\) is the first exit time. The solution accounts only for particles that survive to time \(T\) without hitting the boundary.
- Probabilistic meaning: The particle is killed upon reaching the boundary
- Financial analogy: Knock-out barrier options, where the contract terminates at the barrier
Neumann condition \(\frac{\partial u}{\partial n}(t, a) = 0\), \(\frac{\partial u}{\partial n}(t, b) = 0\): This corresponds to reflection. When the particle reaches the boundary, it is instantaneously pushed back into the interior. The resulting process is reflected Brownian motion, which satisfies \(X_t \in [a, b]\) for all \(t\) and accumulates local time at the boundary. The zero-flux condition \(\partial_n u = 0\) ensures no probability mass leaks out of the domain.
- Probabilistic meaning: The particle is reflected at the boundary
- Financial analogy: Models where the state variable is constrained to a range (e.g., interest rates with a floor)
Robin condition \(\alpha u(t, a) + \beta \frac{\partial u}{\partial n}(t, a) = 0\): This corresponds to partial absorption (elastic boundary). When the particle reaches the boundary, it is killed with probability proportional to \(\alpha / (\alpha + \beta)\) and reflected with the complementary probability. This interpolates between pure absorption (\(\beta = 0\), Dirichlet) and pure reflection (\(\alpha = 0\), Neumann).
- Probabilistic meaning: The particle is killed with probability \(p\) and reflected with probability \(1 - p\) at each boundary encounter, where \(p\) depends on the ratio \(\alpha/\beta\)
- Financial analogy: Elastic barrier options where a partial rebate is paid upon hitting the boundary, and the option has a probability of surviving the barrier hit
In summary:
| Condition | Boundary Behavior | Probability Mass |
|---|---|---|
| Dirichlet (\(u = 0\)) | Absorption/killing | Mass destroyed at boundary |
| Neumann (\(\partial_n u = 0\)) | Reflection | Mass conserved, pushed back |
| Robin (\(\alpha u + \beta \partial_n u = 0\)) | Partial absorption | Fraction destroyed, rest reflected |