Boundary Conditions

A differential expression does not by itself define a differential operator or a well-posed physical problem. One must also specify where the solution lives and what it does at each boundary. Those requirements are boundary conditions.

In quantum mechanics, boundary conditions have three linked roles:

• they select solutions of a differential equation;
• they help specify the domain of an unbounded operator;
• they control probability flow through the boundary.

The same formal Hamiltonian can therefore have different spectra and different physics on different domains. This page develops the working mathematics. The functional-analytic distinction between symmetric and self-adjoint operators is treated separately in Symmetric versus Self-Adjoint Operators.

Boundary data complete the problem

Consider the regular second-order eigenvalue equation on $a\le x\le b$,

$$-\frac{d}{dx} \left( p(x)\frac{dy}{dx} \right) +q(x)y =\lambda w(x)y,$$

where $p(x)>0$ and $w(x)>0$. Its local solution space is two-dimensional for a fixed value of $\lambda$. A regular scalar problem therefore requires two independent scalar boundary conditions in total. Common choices are one condition at each endpoint, but the two endpoints can also be coupled.

For example,

$$y(a)=0, \qquad y(b)=0$$

is a separated pair, while

$$\begin{aligned} y(b)&=y(a),\\ p(b)y'(b)&=p(a)y'(a) \end{aligned}$$

is a coupled periodic pair.

This counting rule changes with the order of the equation, the number of components, and whether an endpoint is singular. It is a guide, not a substitute for checking the operator domain.

Initial-value and boundary-value problems are different. An initial-value problem usually specifies all data at one point, such as $y(a)$ and $y'(a)$, and evolves away from that point. A boundary-value problem distributes constraints across a region. It can have no solution, one solution, or several solutions; an eigenvalue problem asks which parameter values make nonzero solutions possible. See Ordinary Differential Equations for the corresponding existence and uniqueness framework.

Standard separated conditions

Let $\partial_n$ denote the outward normal derivative. On an interval,

$$\partial_n y(a)=-y'(a), \qquad \partial_n y(b)=y'(b).$$

The sign matters when Robin parameters are assigned geometrically.

Dirichlet conditions fix the value,

$$y\big|_{\partial\Omega}=f.$$

The homogeneous form has $f=0$. In elementary quantum models this can represent a hard wall at which the wavefunction vanishes. In other contexts it means that the boundary value is externally prescribed.

Neumann conditions fix the outward normal derivative,

$$\partial_n y\big|_{\partial\Omega}=g.$$

Homogeneous Neumann data set the normal slope to zero. For a simple real Schrödinger operator they also force the normal probability current to vanish, but zero derivative should not be interpreted as a universal model of every reflecting surface.

Robin conditions combine value and normal derivative:

$$\alpha y +\beta\,\partial_n y =h \qquad \text{on }\partial\Omega.$$

Dirichlet and Neumann conditions are limiting cases. In a homogeneous self-adjoint problem, $\alpha$ and $\beta$ are ordinarily real at each boundary, not both zero. A ratio such as $\alpha/\beta$ introduces a length scale and can encode an effective surface interaction.

Mixed conditions use different types on different boundary components. A one-dimensional example is

$$y(a)=0, \qquad y'(b)=0.$$

Mixed does not mean inconsistent: this is exactly two independent conditions for a second-order equation.

Periodic and twisted conditions

Periodic data identify the two ends of an interval:

$$\begin{aligned} y(b)&=y(a),\\ p(b)y'(b)&=p(a)y'(a). \end{aligned}$$

They model a circle or a periodically repeated cell, not a particle bouncing between two hard walls. The corresponding mode basis is developed in Periodic Functions and Fourier Series.

A phase-twisted, or quasiperiodic, identification is

$$\begin{aligned} y(b)&=e^{i\theta}y(a),\\ p(b)y'(b)&=e^{i\theta}p(a)y'(a), \end{aligned}$$

with real $\theta$. Periodic data have $\theta=0$, and antiperiodic data have $\theta=\pi$. The same phase in both equations makes the endpoint contribution to Green’s identity cancel. Twists occur in ring geometries, Bloch theory, and systems with gauge holonomy. Their model-specific consequences belong in Periodic Boundary Conditions.

Homogeneous data and linear domains

An operator domain must be a vector space. Homogeneous linear conditions have this property: if $u$ and $v$ satisfy them, so does $c_1u+c_2v$. Thus a Dirichlet realization of a differential expression can have a domain schematically written as

$$\mathcal D(L_D) = \left\lbrace y\in\mathcal D_{\max}: y(a)=y(b)=0 \right\rbrace.$$

By contrast, $y(a)=1$ defines an affine set rather than a vector space. Inhomogeneous boundary data are entirely legitimate for a forced boundary-value problem, but they do not directly define the domain of a linear operator. One can often subtract a convenient function with the prescribed boundary values and solve a homogeneous problem for the remainder.

Likewise, a condition that depends on the unknown eigenvalue does not describe one fixed operator domain in the usual way. Such problems can be meaningful, but they are generalized spectral problems and require separate analysis.

The broader principle is developed in Domains of Operators: an unbounded operator is the pair consisting of an action and a domain, not merely a formula.

Green’s identity and the boundary form

Define the Sturm–Liouville expression

$$\tau y = \frac{1}{w} \left[ -(py')'+qy \right]$$

in the weighted inner-product space

$$\langle u,v\rangle_w = \int_a^b u(x)^*v(x)w(x)\,dx.$$

Integration by parts gives Green’s identity,

$$\begin{aligned} \langle u,\tau v\rangle_w -\langle\tau u,v\rangle_w &= \left[ p \left( u'^*v-u^*v' \right) \right]_a^b. \end{aligned}$$

The endpoint expression on the right is the boundary form. Boundary conditions make the realization symmetric when this form vanishes for every pair $u,v$ in the proposed domain.

The standard conditions pass this test:

• Dirichlet data kill every term containing an endpoint value.
• Neumann data kill every term containing an endpoint derivative.
• Real homogeneous Robin data make the two products cancel at each endpoint.
• Periodic or twisted data make the contribution at $b$ cancel the contribution at $a$.

Vanishing of the boundary form proves symmetry, not automatically self-adjointness. One must also show that the adjoint has the same domain. For regular scalar Sturm–Liouville problems, maximal boundary conditions of the standard self-adjoint types give the expected result. Singular endpoints require more care.

Probability-current interpretation

For the one-dimensional Hamiltonian

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

with real $V$, integration by parts yields

$$\begin{aligned} \langle u,Hv\rangle-\langle Hu,v\rangle = \frac{\hbar^2}{2m} \left[ u'^*v-u^*v' \right]_a^b. \end{aligned}$$

The probability current is

$$j(x) = \frac{\hbar}{2mi} \left( \psi^*\psi' -\psi'^*\psi \right).$$

Setting $u=v=\psi$ relates the boundary form to the current:

$$\langle\psi,H\psi\rangle -\langle H\psi,\psi\rangle = -i\hbar \left[ j(b)-j(a) \right].$$

Separated Dirichlet, Neumann, and real Robin conditions set the current to zero at each endpoint. Periodic and twisted conditions can allow nonzero current, but the current leaving one identified end re-enters through the other. In both cases there is no net probability loss from the physical configuration space.

This is a useful diagnostic, not a complete proof of self-adjointness. The continuity equation and applications are developed in Probability Current.

One expression, four spectra

Consider

$$-y''=k^2y \qquad \text{on }[0,L].$$

Different endpoint conditions produce different eigenmodes:

Conditions	Eigenfunctions	Allowed wave numbers
Dirichlet–Dirichlet	$\sin(n\pi x/L)$	$k_n=n\pi/L$, $n=1,2,\ldots$
Neumann–Neumann	$\cos(n\pi x/L)$	$k_n=n\pi/L$, $n=0,1,\ldots$
Dirichlet–Neumann	$\sin((n+\tfrac12)\pi x/L)$	$k_n=(n+\tfrac12)\pi/L$, $n=0,1,\ldots$
Periodic	$e^{2\pi i n x/L}$	$k_n=2\pi n/L$, $n\in\mathbb Z$

The Neumann problem has a constant zero mode. The periodic problem has opposite-traveling modes with the same $k_n^2$ for $n\ne0$. The Dirichlet–Dirichlet problem has neither feature. Nothing about the differential expression $-d^2/dx^2$ alone decides among these spectra.

The hard-wall derivation and its physical interpretation have a canonical home in Infinite Square Well.

A Robin quantization condition

As a simple interpolation, impose

$$\begin{gathered} y(0)=0,\\ y'(L)+h\,y(L)=0, \qquad h\ge0. \end{gathered}$$

For a positive eigenvalue, $y(x)=A\sin(kx)$, and the second condition gives

$$k\cos(kL) +h\sin(kL) =0.$$

Equivalently, away from the zeros of $\cos(kL)$,

$$\tan(kL)=-\frac{k}{h}.$$

The $h=0$ limit is Dirichlet–Neumann, while large positive $h$ approaches a Dirichlet condition at $L$. The roots move continuously as the boundary parameter changes. One must test $k=0$ directly in the differential equation: the trigonometric ansatz degenerates there, and an equation obtained after division can create or discard a spurious root.

Interfaces and matching conditions

An internal interface behaves like a boundary shared by two regions. For

$$-(py')'+qy=\lambda wy,$$

integrating across an ordinary finite jump at $x_0$ gives continuity of the flux variable:

$$\left[ py' \right]_{x_0^-}^{x_0^+} =0.$$

In the absence of a distributional source, $y$ is also normally continuous. For the constant-mass Schrödinger equation with a finite potential step, this reduces to continuity of $\psi$ and $\psi'$.

A delta interaction changes the derivative condition. For

$$V(x)=g\,\delta(x-x_0),$$

the wavefunction is continuous and

$$\psi'(x_0^+)-\psi'(x_0^-) = \frac{2mg}{\hbar^2}\psi(x_0).$$

The coefficient follows by integrating the Schrödinger equation through the interface. It should not be guessed from the finite-step rule. The complete model calculation belongs in Delta-Function Potential.

More generally, matching conditions must preserve the appropriate boundary form. Continuity of probability current is a quick physical check.

Infinite and singular endpoints

At an infinite endpoint, a bound-state wavefunction must be square-integrable, but the slogan that it must simply vanish at infinity can hide important details. Decay rate, differentiability, and the action of the operator all matter. Scattering states are not square-integrable and instead satisfy incoming, outgoing, or standing-wave asymptotic conditions; see Scattering States and Boundary Conditions.

At a singular endpoint, square-integrability may select one local behavior or may leave a family of admissible conditions. This is the distinction behind limit-point and limit-circle classification. Its systematic home is Sturm–Liouville Theory.

The radial origin is a common source of mistakes. If

$$\psi(\mathbf r) = \frac{u_\ell(r)}{r} Y_{\ell m}(\theta,\phi),$$

regularity of the ordinary three-dimensional wavefunction generally selects

$$u_\ell(0)=0.$$

This does not mean that the unreduced radial factor must always vanish. Singular potentials and point interactions can alter the domain analysis, so the origin should not be treated as an ordinary endpoint by reflex.

Spatial boundaries in time-dependent problems

The time-dependent Schrödinger equation requires an initial state in time and boundary conditions in space. These play distinct roles:

$$\psi(\mathbf x,t_0)=\psi_0(\mathbf x)$$

sets the initial state, while a condition such as

$$\psi(\mathbf x,t)=0 \qquad \text{for }\mathbf x\in\partial\Omega$$

specifies the spatial operator domain for every time. A valid initial state must lie in the Hilbert space, and stronger regularity is required if one wants it in the Hamiltonian domain. Time evolution preserves the boundary condition when generated by the corresponding self-adjoint Hamiltonian.

For partial differential equations, Dirichlet, Neumann, and Robin data are imposed on boundary surfaces, and $\partial_n$ is the geometric normal derivative. Corners, nonsmooth boundaries, and unbounded regions require additional analysis.

Numerical practice

Boundary conditions must be built into the discretization, not appended after diagonalization.

1. Specify the continuum operator and domain before choosing a grid.
2. Eliminate fixed Dirichlet degrees of freedom or encode them consistently in the matrix.
3. Use one-sided, ghost-point, finite-element, or spectral constructions that retain the intended order of accuracy for Neumann and Robin data.
4. Wrap both the function and derivative stencil for periodic data.
5. Check that the discrete Hamiltonian is Hermitian in the discrete inner product.
6. Vary the grid spacing and artificial box size separately. Agreement under only one refinement does not establish convergence to the intended unbounded-domain problem.

A boundary implementation that produces a visibly non-Hermitian Hamiltonian, complex eigenvalues for a closed system, or probability drift should be treated as a defect until explained.

A practical decision sequence

When setting up a new quantum boundary-value problem:

1. Write the differential expression and the Hilbert-space measure.
2. Identify regular endpoints, singular endpoints, interfaces, and infinity.
3. Derive interface jumps by integrating the equation.
4. State a homogeneous linear domain for the operator.
5. Evaluate the boundary form for two arbitrary domain functions.
6. Check probability current and dimensions.
7. Only then solve the spectral equation or discretize it.

This sequence separates physical assumptions from algebraic consequences and makes it easier to compare two apparently similar models.

Common mistakes

• Treating a differential formula as a complete Hamiltonian.
• Imposing two conditions at each endpoint of a second-order scalar problem.
• Using inhomogeneous data as though they formed a linear operator domain.
• Calling periodic conditions a hard-wall approximation.
• Assuming every reflecting boundary requires $\psi=0$.
• Matching $\psi'$ across a delta interaction as if the potential were finite.
• Forgetting the coefficient $p$ when the equation is in divergence form.
• Using a complex Robin parameter without checking the boundary form and flux.
• Requiring a scattering state to be square-integrable.
• Setting the reduced radial function and the unreduced radial factor to the same endpoint value.
• Proving symmetry by integration by parts and calling it self-adjointness.
• Trusting a discretized spectrum without varying both grid and domain size.

Exercises

1. Let $L=-d^2/dx^2$ on $[0,L_0]$. Show directly from the boundary form that the separated Robin conditions

   $$\begin{aligned} u'(0)&=h_0u(0),\\ u'(L_0)&=-h_1u(L_0) \end{aligned}$$

   make $L$ symmetric when $h_0,h_1\in\mathbb R$.

Solution

For two domain functions $u$ and $v$, the boundary form is

$$\left[ u'^*v-u^*v' \right]_0^{L_0}.$$

At the left endpoint, substitute the conditions term by term:

$$\begin{aligned} u'(0)^*v(0) &=h_0u(0)^*v(0),\\ u(0)^*v'(0) &=h_0u(0)^*v(0). \end{aligned}$$

Their difference is zero, where reality of $h_0$ was used. At the right endpoint, both derivatives carry the same factor $-h_1$, so the two terms cancel in the same way. The complete boundary form vanishes. This establishes symmetry on the proposed domain; a separate maximal-domain argument establishes self-adjointness.

2. Solve $-y''=k^2y$ on $[0,L]$ with $y(0)=0$ and $y'(L)=0$. Normalize the eigenfunctions in $L^2([0,L])$.

Solution

For $k>0$,

$$y(x)=A\sin(kx)+B\cos(kx).$$

The condition at zero gives $B=0$. The condition at $L$ gives

$$A k\cos(kL)=0.$$

A nonzero eigenfunction therefore requires

$$k_n = \frac{(n+\tfrac12)\pi}{L}, \qquad n=0,1,2,\ldots.$$

Since

$$\int_0^L \sin^2(k_nx)\,dx = \frac{L}{2},$$

the normalized modes are

$$y_n(x) = \sqrt{\frac{2}{L}} \sin\left( \frac{(n+\tfrac12)\pi x}{L} \right).$$

The $k=0$ equation gives $y=Ax+B$; the two boundary conditions force $A=B=0$, so there is no zero mode.

3. Verify that twisted conditions preserve both the boundary form and the probability current for $p=1$:

   $$\begin{aligned} \psi(L)&=e^{i\theta}\psi(0),\\ \psi'(L)&=e^{i\theta}\psi'(0). \end{aligned}$$

Solution

For arbitrary domain functions $u$ and $v$, the endpoint expression at $L$ is

$$\begin{aligned} u'(L)^*v(L) &= e^{-i\theta}u'(0)^* e^{i\theta}v(0)\\ &=u'(0)^*v(0),\\ u(L)^*v'(L) &= e^{-i\theta}u(0)^* e^{i\theta}v'(0)\\ &=u(0)^*v'(0). \end{aligned}$$

The values at the two ends are equal, so their difference in the boundary form vanishes. For $u=v=\psi$, the same cancellation gives

$$j(L)=j(0).$$

The current need not vanish. It circulates through the identified endpoint rather than leaking from the configuration space.

4. A particle obeys

   $$-\frac{\hbar^2}{2m}\psi'' +g\delta(x)\psi =E\psi.$$

   Derive the derivative jump and show that a real $g$ preserves current across the interface when $\psi$ is continuous.

Solution

Integrate from $-\epsilon$ to $\epsilon$. The energy term vanishes in the limit because its integral is $O(\epsilon)$, while the kinetic and delta terms give

$$-\frac{\hbar^2}{2m} \left[ \psi'(0^+)-\psi'(0^-) \right] +g\psi(0) =0.$$

Hence

$$\psi'(0^+)-\psi'(0^-) = \frac{2mg}{\hbar^2}\psi(0).$$

Let $\Delta\psi'=\psi'(0^+)-\psi'(0^-)$. Then the current difference is

$$\begin{aligned} j(0^+)-j(0^-) &= \frac{\hbar}{m} \operatorname{Im} \left\lbrace \psi(0)^*\Delta\psi' \right\rbrace\\ &= \frac{2g}{\hbar} \operatorname{Im}\lvert\psi(0)\rvert^2\\ &=0. \end{aligned}$$

Reality of $g$ is essential in the last step. A complex coupling would model gain or loss rather than a closed self-adjoint interaction.

References

• E. A. Coddington and N. Levinson, Theory of Ordinary Differential Equations, McGraw–Hill, 1955.
• A. Zettl, Sturm–Liouville Theory, American Mathematical Society, 2005.
• G. Teschl, Mathematical Methods in Quantum Mechanics, 2nd ed., American Mathematical Society, 2014.
• M. Reed and B. Simon, Methods of Modern Mathematical Physics, Vol. II: Fourier Analysis, Self-Adjointness, Academic Press, 1975.
• B. C. Hall, Quantum Theory for Mathematicians, Springer, 2013.
• D. J. Griffiths and D. F. Schroeter, Introduction to Quantum Mechanics, 3rd ed., Cambridge University Press, 2018.