Propagators and Boundary Conditions

A differential expression does not determine a quantum Hamiltonian by itself. The same expression

$$-\frac{\hbar^2}{2m}\frac{d^2}{dx^2}$$

can describe a particle on the line, a particle on a half-line, a hard-wall box, or a particle on a ring. The Hilbert space and operator domain distinguish those systems. For a time-independent boundary-value problem, the propagator is

$$U(t_f,t_i)=e^{-iH(t_f-t_i)/\hbar},$$

changing the domain changes the operator $H$, its spectrum, and its kernel.

The practical rule is:

A propagator must preserve the Hamiltonian’s boundary conditions, not merely solve the same differential equation in the interior.

This page is the canonical home for that rule and its main examples. The stationary wave-mechanics meaning of endpoint and matching conditions is reviewed in Boundary Conditions.

The Domain Is Part of the Hamiltonian

Consider

$$H = -\frac{\hbar^2}{2m}\frac{d^2}{dx^2} +V(x)$$

on an interval $a\lt x\lt b$, with real $V$. Integration by parts gives the boundary form

$$\begin{aligned} \langle\phi,H\psi\rangle - \langle H\phi,\psi\rangle &= -\frac{\hbar^2}{2m} \left[ \phi^*\psi' - \phi'^*\psi \right]_a^b. \end{aligned}$$

A symmetric domain makes this form vanish for all $\phi$ and $\psi$ in the domain. A self-adjoint domain is, in addition, maximal in the required domain-sensitive sense: the operator and its adjoint have the same domain. Self-adjointness is what supports the spectral theorem and unitary time evolution.

Common self-adjoint choices include:

Boundary condition	Typical form	Physical reading
Dirichlet	$\psi=0$ at the boundary	hard-wall node
Neumann	$\partial_n\psi=0$	reflecting boundary with zero normal slope
Robin	$\partial_n\psi+\kappa\psi=0$, real $\kappa$	reflecting surface with a length scale
Periodic	endpoint values and derivatives agree	endpoints are identified
Twisted periodic	values and derivatives acquire one common phase	identified endpoints with nontrivial holonomy

Here $\partial_n$ is the outward normal derivative. Dirichlet, Neumann, and Robin conditions are local: they constrain data at each boundary component. Periodic conditions couple data at different endpoints because those endpoints represent the same physical point.

These familiar cases are not the entire mathematical family. For the second derivative on a finite interval, the self-adjoint extensions form a $U(2)$ family and can mix endpoint data. The standard cases above are important subfamilies. The full classification belongs to operator-domain theory; see Symmetric versus Self-Adjoint Operators and the primary references below.

Probability Flux and Unitarity

For a one-dimensional wavefunction, the probability current is

$$j(x,t) = \frac{\hbar}{m} \operatorname{Im} \left( \psi^*\frac{\partial\psi}{\partial x} \right).$$

Dirichlet and Neumann conditions make the normal current vanish at a reflecting endpoint. A Robin condition with a real parameter does the same. For example, on the half-line, if

$$\psi'(0)=\lambda\psi(0), \qquad \lambda\in\mathbb R,$$

then

$$j(0,t) = \frac{\hbar\lambda}{m} \operatorname{Im}\lvert\psi(0,t)\rvert^2 =0.$$

Periodic conditions need not make the current zero. Instead, the current leaving one end re-enters at the identified end. This is why a particle on a ring can carry a persistent probability current without losing norm.

Vanishing or matched boundary flux is the local continuity-equation expression of norm preservation. It is not by itself a complete proof of self-adjointness, but it is a useful diagnostic.

What the Kernel Must Satisfy

Let $\Omega$ be the allowed configuration region and $B$ the boundary operator defining the domain. The kernel must satisfy four linked requirements.

First, it solves the Schrödinger equation in the final variables:

$$i\hbar\frac{\partial}{\partial t_f} K(q_f,t_f;q_i,t_i) = H_{q_f}K(q_f,t_f;q_i,t_i).$$

Second, it obeys the boundary condition in each endpoint variable. Schematically,

$$B_{q_f}K=0, \qquad B_{q_i}^*K=0.$$

For real Dirichlet, Neumann, or Robin data, this means the same boundary condition in $q_f$ and $q_i$.

Third, its equal-time limit is the identity distribution on the allowed domain:

$$\lim_{t_f\to t_i^+} K(q_f,t_f;q_i,t_i) = \delta_\Omega(q_f,q_i).$$

Fourth, composition integrates only over allowed intermediate configurations:

$$\begin{aligned} K(q_f,t_f;q_i,t_i) &= \int_\Omega d\mu(q)\, K(q_f,t_f;q,t)\\ &\qquad\times K(q,t;q_i,t_i). \end{aligned}$$

If $H$ is self-adjoint, unitarity also gives

$$\begin{aligned} &\int_\Omega d\mu(q)\, K(q_f,T;q,0) K^*(q_i,T;q,0)\\ &\qquad= \delta_\Omega(q_f,q_i). \end{aligned}$$

Solving the interior differential equation while failing any one of these conditions does not produce the propagator for the stated boundary-value problem.

Particle on a Half-Line

Take a free particle on $x\gt0$. Write the full-line free kernel as a function of displacement,

$$k_0(z,T) = \left( \frac{m}{2\pi i\hbar T} \right)^{1/2} \exp\left[ \frac{imz^2}{2\hbar T} \right],$$

with $T\gt0$ and the usual convergence prescription.

Dirichlet endpoint

For

$$\psi(0)=0,$$

the half-line kernel is

$$K_{\mathrm D}(x_f,T;x_i,0) = k_0(x_f-x_i,T) - k_0(x_f+x_i,T),$$

for $x_f,x_i\gt0$. The second term is the amplitude from the reflected image point $-x_i$. Because $k_0$ is even in its displacement,

$$K_{\mathrm D}(0,T;x_i,0)=0.$$

The minus sign is the odd-extension rule. A half-line wavefunction satisfying Dirichlet data can be extended oddly to the full line, propagated there, and then restricted back to $x\gt0$.

Neumann endpoint

For

$$\psi'(0)=0,$$

the image has the same sign:

$$K_{\mathrm N}(x_f,T;x_i,0) = k_0(x_f-x_i,T) + k_0(x_f+x_i,T).$$

The derivative of $k_0(z,T)$ is odd in $z$, so

$$\left. \frac{\partial}{\partial x_f} K_{\mathrm N}(x_f,T;x_i,0) \right|_{x_f=0} =0.$$

This is the even-extension rule.

Robin endpoint

The one-parameter family

$$\psi'(0)=\lambda\psi(0), \qquad \lambda\in\mathbb R,$$

interpolates through Neumann at $\lambda=0$ and introduces a boundary length scale $\lvert\lambda\rvert^{-1}$. It generally cannot be represented by a single image with a constant plus or minus sign.

For an incoming continuum mode

$$\psi_k(x) = e^{-ikx}+r_\lambda(k)e^{ikx},$$

the boundary condition gives

$$r_\lambda(k) = \frac{ik+\lambda}{ik-\lambda}.$$

For real $k$ and $\lambda$,

$$\lvert r_\lambda(k)\rvert=1,$$

as expected for a perfectly reflecting self-adjoint boundary. The momentum-dependent reflection phase is why the Robin image rule is more complicated.

When $\lambda\lt0$, the same domain has a boundary-localized bound state:

$$\psi_b(x) = \sqrt{-2\lambda}\,e^{\lambda x},$$

with

$$E_b = -\frac{\hbar^2\lambda^2}{2m}.$$

The Robin propagator must include this discrete spectral contribution in addition to its continuum integral. Thus a boundary condition can create a bound state even though the interior differential expression is free.

Method of Images

The image method constructs a kernel on a simple domain by combining full-space kernels so that unwanted boundary data cancel or reinforce.

For the half-line:

$$K_{\mathrm D}=k_0(\text{direct})-k_0(\text{reflected}),$$ $$K_{\mathrm N}=k_0(\text{direct})+k_0(\text{reflected}).$$

For a Dirichlet interval $0\lt x\lt L$, repeated images give

$$\begin{aligned} K_{\mathrm D}^{[0,L]}(x_f,T;x_i,0) &= \sum_{w\in\mathbb Z} \bigl[ k_0(x_f-x_i+2wL,T)\\ &\qquad- k_0(x_f+x_i+2wL,T) \bigr]. \end{aligned}$$

The two image families enforce nodes at both $x=0$ and $x=L$. Replacing the relative minus sign by a plus sign gives the corresponding Neumann image construction.

Image constructions work because reflection symmetry and a simple geometry let transformed paths be paired. They are not a universal recipe. A generic Robin condition needs a momentum-dependent reflected amplitude, curved boundaries may require more elaborate constructions, and the most general self-adjoint interval domains can couple endpoints nonlocally.

Particle in a Box

For a free particle on $0\lt x\lt L$ with Dirichlet walls,

$$\psi(0)=\psi(L)=0.$$

The normalized eigenfunctions and energies are

$$\phi_n(x) = \sqrt{\frac{2}{L}} \sin\frac{n\pi x}{L}, \qquad n=1,2,\ldots,$$ $$E_n = \frac{n^2\pi^2\hbar^2}{2mL^2}.$$

The spectral kernel is therefore

$$\begin{aligned} K_{\mathrm D}^{[0,L]}(x_f,T;x_i,0) &= \frac{2}{L} \sum_{n=1}^{\infty} \sin\frac{n\pi x_f}{L} \sin\frac{n\pi x_i}{L}\\ &\qquad\times e^{-iE_nT/\hbar}. \end{aligned}$$

This is the same operator as the image sum above. The spectral representation makes energy quantization explicit; the image representation makes repeated reflection explicit.

With Neumann walls,

$$\psi'(0)=\psi'(L)=0,$$

the kernel becomes

$$\begin{aligned} K_{\mathrm N}^{[0,L]}(x_f,T;x_i,0) &= \frac{1}{L} + \frac{2}{L} \sum_{n=1}^{\infty} \cos\frac{n\pi x_f}{L} \cos\frac{n\pi x_i}{L}\\ &\qquad\times \exp\left[ -\frac{i\hbar n^2\pi^2T}{2mL^2} \right]. \end{aligned}$$

The constant term is the zero-energy Neumann mode. It has no Dirichlet counterpart. The interior differential expression is unchanged, yet the spectrum and kernel differ because the domain differs.

The hard-wall spectrum and stationary states belong to Infinite Square Well. Compact exact formulas are collected in Propagator Table.

Periodic Boundary Conditions

Periodic boundary conditions do not describe two reflecting walls. They identify the endpoints of an interval of length $L$:

$$\psi(x+L)=\psi(x), \qquad \psi'(x+L)=\psi'(x).$$

The free eigenfunctions are plane waves with

$$k_n=\frac{2\pi n}{L}, \qquad n\in\mathbb Z.$$

The kernel is

$$\begin{aligned} K_{\mathrm{per}}(x_f,T;x_i,0) &= \frac{1}{L} \sum_{n\in\mathbb Z} \exp\left[ \frac{i2\pi n(x_f-x_i)}{L} \right]\\ &\qquad\times \exp\left[ -\frac{i\hbar(2\pi n)^2T}{2mL^2} \right]. \end{aligned}$$

Equivalently, sum over full-line paths whose lifted endpoints differ by an integer multiple of the circumference:

$$K_{\mathrm{per}}(x_f,T;x_i,0) = \sum_{w\in\mathbb Z} k_0(x_f-x_i+wL,T).$$

The integer $w$ is the winding number in the covering space. Poisson summation relates the winding representation to the discrete momentum representation.

Twisted Periodicity and Ring Geometry

A self-adjoint twisted boundary condition uses one phase $\alpha$ for both the wavefunction and derivative:

$$\psi(x+L)=e^{i\alpha}\psi(x),$$ $$\psi'(x+L)=e^{i\alpha}\psi'(x).$$

The allowed momenta are shifted:

$$k_n = \frac{2\pi n+\alpha}{L}.$$

The corresponding kernel is

$$\begin{aligned} K_\alpha(x_f,T;x_i,0) &= \frac{1}{L} \sum_{n\in\mathbb Z} \exp\left[ \frac{i(2\pi n+\alpha)(x_f-x_i)}{L} \right]\\ &\qquad\times \exp\left[ -\frac{i\hbar(2\pi n+\alpha)^2T}{2mL^2} \right]. \end{aligned}$$

Its winding form is

$$K_\alpha(x_f,T;x_i,0) = \sum_{w\in\mathbb Z} e^{-iw\alpha} k_0(x_f-x_i+wL,T).$$

The phase ensures

$$K_\alpha(x_f+L,T;x_i,0) = e^{i\alpha} K_\alpha(x_f,T;x_i,0).$$

For a particle on a ring, $L=2\pi R$. The untwisted case $\alpha=0$ gives ordinary periodic scalar wavefunctions. A nonzero twist can represent magnetic-flux holonomy after a gauge choice, shifting angular momentum without introducing a reflecting boundary.

The canonical system, current, and flux conventions are developed in Particle on a Ring. The interval viewpoint is Periodic Boundary Conditions.

Spectral and Geometric Representations

Boundary-sensitive kernels often have two useful representations:

Representation	What it emphasizes
spectral sum	energies, degeneracies, boundary eigenfunctions
image or winding sum	reflections, topology, path classes

For the box, sine or cosine eigenfunctions display the wall-dependent spectrum, while image sums display repeated reflections. For the ring, angular-momentum modes display quantization, while winding sectors display the topology of $S^1$.

These forms are equal only when their normalization, convergence prescription, and boundary phase agree. The spectral decomposition of the propagator provides the operator-level reason: every self-adjoint domain supplies its own spectral resolution.

Boundaries in Path Integrals

For Dirichlet and Neumann walls in simple geometries, image kernels can be interpreted as sums over reflected path classes with relative signs. On a ring, the path integral separates into winding sectors. These are powerful constructions, but they do not mean that a local classical action automatically determines every quantum boundary condition.

The operator domain remains the primary specification. Some self-adjoint extensions couple distant boundary components and are not captured by a naive local reflection rule. A time-sliced path integral must be checked against the intended short-time kernel and boundary domain, not inferred from the bulk action alone.

Moving boundaries require additional care because the domain itself depends on time. Complex absorbing potentials, outgoing-wave boundaries, and numerical absorbing layers are different again: they intentionally allow probability to leave a computational region and generally define nonunitary effective evolution there.

Boundaries Are Not Finite Barriers

An infinite hard wall is an idealized boundary condition. A finite potential barrier is a potential-energy region in a larger configuration space. Its wavefunction usually penetrates the barrier and may transmit through it.

Accordingly:

• a Dirichlet wall removes the exterior region from the configuration space;
• a finite barrier keeps the exterior region and solves a matching problem;
• a Robin boundary can encode idealized short-distance surface physics;
• an absorbing boundary is an open or effective description, not a reflecting self-adjoint wall.

Using a hard-wall propagator for a finite barrier discards tunneling. Using a full-line free propagator for a hard wall fails to preserve the domain.

Common Mistakes

• Treating the differential expression as the complete Hamiltonian and omitting its domain.
• Using the full-line free kernel on a half-line, interval, or ring.
• Checking the boundary condition in $x_f$ but not in $x_i$.
• Composing a restricted-domain kernel over the full line instead of the allowed region.
• Assuming every reflecting boundary is Dirichlet.
• Treating Neumann and Dirichlet boxes as having the same spectrum and missing the Neumann zero mode.
• Using one constant image coefficient for a generic Robin boundary.
• Forgetting the Robin boundary-bound-state term when $\lambda\lt0$.
• Interpreting periodic endpoints as two walls rather than one identified point.
• Dropping the winding phase in a twisted-periodic or flux-threaded problem.
• Assuming a standard local path integral represents every self-adjoint extension.
• Calling an absorbing numerical boundary unitary closed-system evolution.

Cross-Links

• Propagator Kernel
• Composition Law
• Spectral Decomposition of the Propagator
• Free-Particle Propagator
• Propagators in Multiple Dimensions
• Causality, Support, and Interpretation in Nonrelativistic QM
• Propagator Table
• Boundary Conditions
• Infinite Square Well
• Periodic Boundary Conditions
• Particle on a Ring
• Unbounded Operators
• Symmetric versus Self-Adjoint Operators
• Hermitian vs Self-Adjoint Operators

References

• G. Bonneau, J. Faraut, and G. Valent, “Self-adjoint extensions of operators and the teaching of quantum mechanics,” American Journal of Physics 69, 322–331 (2001), doi:10.1119/1.1328351.
• M. Asorey, A. Ibort, and G. Marmo, “Global theory of quantum boundary conditions and topology change,” International Journal of Modern Physics A 20, 1001–1026 (2005), doi:10.1142/S0217751X05019798.
• M. Asorey, A. Ibort, and G. Marmo, “Path integrals and boundary conditions,” arXiv:quant-ph/0609023 (2006).
• M. Reed and B. Simon, Methods of Modern Mathematical Physics II: Fourier Analysis, Self-Adjointness, Academic Press, 1975.
• S. Albeverio, F. Gesztesy, R. Høegh-Krohn, and H. Holden, Solvable Models in Quantum Mechanics, 2nd ed., AMS Chelsea, 2005.
• L. S. Schulman, Techniques and Applications of Path Integration, Wiley, 1981.
• C. Grosche and F. Steiner, Handbook of Feynman Path Integrals, Springer, 1998.

Exercises

1. Verify that the Dirichlet and Neumann half-line image kernels obey their stated boundary conditions.

Solution

Because

$$k_0(-z,T)=k_0(z,T),$$

the Dirichlet kernel at the boundary is

$$\begin{aligned} K_{\mathrm D}(0,T;x_i,0) &= k_0(-x_i,T)-k_0(x_i,T)\\ &=0. \end{aligned}$$

The derivative $\partial_zk_0(z,T)$ is odd. Therefore

$$\begin{aligned} \left. \partial_{x_f}K_{\mathrm N} \right|_{x_f=0} &= \partial_zk_0(-x_i,T) + \partial_zk_0(x_i,T)\\ &=0. \end{aligned}$$

Each term separately solves the free Schrödinger equation, so each linear combination does as well.

2. For the Robin half-line condition, show that reflection is unitary and derive the boundary bound state when it exists.

Solution

For real $k$ and $\lambda$,

$$r_\lambda(k) = \frac{ik+\lambda}{ik-\lambda}.$$

Its modulus squared is

$$\lvert r_\lambda(k)\rvert^2 = \frac{k^2+\lambda^2}{k^2+\lambda^2} =1.$$

For a bound state, set

$$\psi_b(x)=Ae^{-\kappa x}, \qquad \kappa\gt0.$$

The boundary condition gives

$$-\kappa=\lambda.$$

Thus a normalizable state exists only for $\lambda\lt0$, with $\kappa=-\lambda$. Normalization gives

$$A=\sqrt{2\kappa}=\sqrt{-2\lambda},$$

and the energy is

$$E_b = -\frac{\hbar^2\kappa^2}{2m} = -\frac{\hbar^2\lambda^2}{2m}.$$

3. Explain why the Neumann box has a zero-energy mode but the Dirichlet box does not.

Solution

The constant function

$$\phi_0(x)=\frac{1}{\sqrt L}$$

satisfies

$$\phi_0'(0)=\phi_0'(L)=0$$

and

$$-\frac{\hbar^2}{2m}\phi_0''=0.$$

It is therefore a normalized Neumann eigenfunction with zero energy. A nonzero constant cannot satisfy

$$\phi(0)=\phi(L)=0,$$

so the Dirichlet spectrum begins with the $n=1$ sine mode at positive energy.

4. Verify the twisted boundary law of the winding kernel.

Solution

Start from

$$K_\alpha(x_f,T;x_i,0) = \sum_{w\in\mathbb Z} e^{-iw\alpha} k_0(x_f-x_i+wL,T).$$

After replacing $x_f$ by $x_f+L$,

$$\begin{aligned} K_\alpha(x_f+L,T;x_i,0) &= \sum_w e^{-iw\alpha} k_0(x_f-x_i+(w+1)L,T). \end{aligned}$$

Set $w'=w+1$. Then

$$e^{-i(w'-1)\alpha} = e^{i\alpha}e^{-iw'\alpha},$$

so

$$K_\alpha(x_f+L,T;x_i,0) = e^{i\alpha}K_\alpha(x_f,T;x_i,0).$$

5. Why must half-line kernel composition use an integral over $x\gt0$ rather than the full line?

Solution

The half-line Hilbert space is $L^2(0,\infty)$. Its position resolution of identity is

$$I = \int_0^\infty dx\, \lvert x\rangle\langle x\rvert.$$

Inserting that identity at an intermediate time gives

$$\begin{aligned} K(x_f,t_f;x_i,t_i) &= \int_0^\infty dx\, K(x_f,t_f;x,t)\\ &\qquad\times K(x,t;x_i,t_i). \end{aligned}$$

Integrating over the full line would insert states outside the physical configuration space and would not represent the identity on $L^2(0,\infty)$.