Note: This blog post is based on my own term paper for MIT 8.06x (Applications of Quantum Mechanics). You can find the original PDF here: Case Study of the Momentum Operator.
Why this post exists
Most quantum mechanics textbooks say that observables correspond to “Hermitian operators”. Charitably, this is a half-truth, and the missing half is exactly the one that gets you into trouble as soon as you try to define a momentum operator on anything more interesting than .
The issue is that “Hermitian” in the physics sense usually means symmetric, while what observables really need to be is self-adjoint, and in infinite dimensions these are not the same thing. The gap between them is invisible in finite-dimensional linear algebra, where every linear operator is bounded, so every densely defined symmetric operator is automatically self-adjoint. In quantum mechanics, where the Hilbert spaces are typically and the operators we care about are unbounded, the distinction cannot be avoided.
This post is a case study. We go through the definitions carefully (operator, adjoint, symmetric, self-adjoint, essentially self-adjoint) and then apply them to the momentum operator in two settings.
- A compact interval with hard-wall (Dirichlet) boundary conditions.
- A circle, i.e. the same interval with periodic boundary conditions.
In short, on the circle the momentum operator is essentially self-adjoint, and there is a unique sensible way to “promote” it to a true observable. On the interval with Dirichlet boundary conditions it is not essentially self-adjoint, and the question of which self-adjoint operator deserves to be called “the momentum on ” has no unique answer. There is a one-parameter family of options, and one has to choose. This is the kind of subtlety that gets lost when is treated as if it were a finite matrix.
I will assume a working knowledge of Hilbert spaces and basic functional analysis (bounded operators, dense subspaces, ), and set throughout.
1. Operators, bounded and unbounded
An operator between two normed spaces and is just a linear map . Nothing special so far, the interesting part is the topology.
In finite dimensions every linear operator is automatically continuous, because all norms on a finite-dimensional vector space are equivalent. This is false in infinite dimensions, and the failure is not a mathematical curiosity. It is the reason quantum mechanics needs a more careful framework than ordinary linear algebra.
Bounded operators
Let be a normed space and a Banach space. A linear operator is bounded if
or equivalently if there is a constant such that for all .
For linear operators, “bounded” and “continuous” mean the same thing (this is a standard lemma). We write for the space of bounded operators .
Unbounded operators
A linear operator that fails the boundedness condition is unbounded. The two most important operators in quantum mechanics, position and momentum, are both unbounded on ,
- Position: .
- Momentum: .
Why are they unbounded? For position, one can find functions concentrated arbitrarily far from the origin, where blows up. For momentum, differentiation amplifies high-frequency components without bound.
A more elementary way to see that the derivative is unbounded is to define
and take . Then but , so the ratio is unbounded.
Important consequence. An unbounded operator cannot be defined on the whole Hilbert space (at least not as a closed linear operator). An everywhere-defined closed operator is automatically bounded (the closed graph theorem), and the special case for symmetric operators is the Hellinger–Toeplitz theorem. So an unbounded operator always lives on a proper subspace, its domain, and specifying the domain is part of specifying the operator. Two operators with the same formal expression but different domains are different operators. If there is one thing to take away from this section, it is this.
2. Adjoints, symmetric and self-adjoint
Densely defined operators
Let be a Hilbert space. A linear operator is densely defined if its domain is dense in , i.e. every vector in can be approximated arbitrarily well by vectors in .
Density is what allows us to define the adjoint at all. Without it, the adjoint is not even well-defined as a single-valued operator.
The adjoint
Let be densely defined. The adjoint is the operator with domain
acting as .
is well-defined. If both and satisfy for all , then for all in a dense set, which forces . (This is exactly where density is used.)
A small but useful proposition is . To see it, for all .
Extensions and an inclusion lemma
We say is an extension of , written , if and for all .
Lemma. If are both densely defined, then .
Proof. Let , so there is an with for all . Since and on , this gives for all . Hence and .
Note that the direction is reversed, i.e. a bigger operator has a smaller adjoint. We will use this repeatedly.
Symmetric vs. self-adjoint
A densely defined operator is symmetric if
It is self-adjoint if , which means that both the actions and the domains agree,
- ,
- for all .
A word of warning about language. In physics, “Hermitian” is used inconsistently, sometimes as a synonym for symmetric and sometimes as a synonym for self-adjoint. Statements like “observables correspond to Hermitian operators” gloss over a real distinction that only disappears in finite dimensions, so I will avoid the word entirely.
If is symmetric, then . (For , set . Symmetry gives for all , so with .) The whole question is then whether this inclusion is an equality.
A maximality property
Self-adjoint operators have no proper symmetric extensions. Concretely, if is self-adjoint, is symmetric and , then .
Proof. From the lemma gives . Symmetry of means , so . Together with this gives .
This is why self-adjointness is the right condition for observables. The spectral theorem guarantees that self-adjoint operators have real spectra and a sensible functional calculus, which symmetric operators in general do not. The maximality property says that a self-adjoint operator is “as big as it can be”, i.e. there is no room to extend it further and keep it symmetric.
3. Closability, closure, essential self-adjointness
We need one more layer. Even if is not self-adjoint, it may have a unique self-adjoint extension hiding inside , and this is what essential self-adjointness captures.
Closures
A densely defined operator is closable if its adjoint is itself densely defined. Equivalently (and more usefully), has a smallest closed extension, the closure , which satisfies .
Symmetric operators are always closable. If is symmetric, then implies . Since is dense, so is , and is closable. So for symmetric operators we always have available.
For symmetric , the closure sits between and ,
Essentially self-adjoint operators
A symmetric operator is essentially self-adjoint (e.s.a.) if its closure is self-adjoint.
This is weaker than self-adjointness. A self-adjoint operator is automatically e.s.a., since gives, after taking adjoints, , so and is self-adjoint.
Theorem. If is essentially self-adjoint, then is the unique self-adjoint extension of .
Proof. is a self-adjoint extension by hypothesis. For uniqueness, let be any other self-adjoint extension, . Self-adjoint operators are closed (since and adjoints are always closed), so , as the closure is the smallest closed extension. But is itself self-adjoint, and a self-adjoint operator has no proper symmetric extensions, so .
The practical upshot. To turn a symmetric operator into a genuine observable, we do not need it to be self-adjoint already. It is enough that it is essentially self-adjoint. The closure is then the canonical observable, and it is unique.
When essential self-adjointness fails, life gets harder. The symmetric operator may then have many self-adjoint extensions (or none at all), and choosing one becomes a piece of physical input. The tool for counting and parametrizing these extensions is von Neumann’s deficiency index theory. It is beyond the scope of this post, but it comes up explicitly in the interval case below.
4. The momentum operator: setup
Now to the main part. We define the momentum operator on the -th coordinate as
where the domain still has to be specified. We work on in one dimension and consider two natural-looking domains,
- Interval (Dirichlet): .
- Circle (periodic): .
The Dirichlet condition is strictly stronger than the periodic one (forcing both endpoints to vanish implies periodicity, but not the other way round), so as operators , i.e. the circle operator is an extension of the interval operator.
We also need a couple of facts about function spaces. The chain of inclusions
relates classically differentiable functions () to absolutely continuous functions () and to the Sobolev space , where
- if there is a Lebesgue-integrable such that , in which case almost everywhere,
- .
The point of is that it is the natural domain for ” has an derivative”. It lets us apply and stay in , without requiring classical differentiability everywhere.
5. Momentum on the interval
Take with domain .
Step 1. is symmetric
For we integrate by parts,
The boundary term vanishes because , and the remaining integral is
So is symmetric. ✓
Step 2. The adjoint has a bigger domain
Now we ask for which there is an with
The strategy is to pick any with a.e. (such an exists, since every function on a compact interval has an absolutely continuous antiderivative). Substituting and integrating the right-hand side by parts,
where the boundary term vanishes because vanishes at the endpoints. So becomes
or, equivalently, in .
Next we identify this orthogonal complement. The set consists of exactly those continuous functions on with , i.e. , where is the constant function. (In one direction, if with , then . In the other, given such a , the antiderivative vanishes at both endpoints and lies in .)
So the closure of is , and its orthogonal complement is , the constant functions. Hence
Since , this gives , with no boundary conditions at all. The requirement then forces , i.e. . Conversely, every function arises this way, so
The adjoint is the same differential expression on a much larger domain, functions with no boundary conditions, compared to functions that vanish at both endpoints.
In particular , so is not self-adjoint.
Step 3. The closure is also too small
Maybe the closure saves us? We compute directly.
means that there is a with for all . Since and both act as , we have , and the question is whether the boundary terms vanish.
For (which we already know is necessary, since ) and , integrating by parts gives
This has to vanish for all , and functions can take arbitrary, independent values at and , so and . Therefore
which is strictly smaller than . Hence , and is not essentially self-adjoint.
What is really going on
This is not a defect of our choice of starting domain. It reflects a real physical ambiguity. There is in fact a one-parameter family of self-adjoint extensions of , labelled by , with domains
Different values of correspond to physically different systems (for instance a particle on an interval threaded by a magnetic flux, or a “twisted” boundary condition). The deficiency-index analysis confirms that this is the whole story. The deficiency indices of are , which gives exactly a worth of self-adjoint extensions.
The lesson is that when an operator is symmetric but not essentially self-adjoint, the formal expression alone does not tell you which observable to use. You have to supply a boundary condition, and that choice is part of defining the system.
6. Momentum on the circle
Now we repeat the analysis with the milder boundary condition , without requiring the common value to be zero.
Step 1. Symmetric
For with and , the boundary term from integrating by parts is
where we used both boundary conditions. So is symmetric. ✓
Step 2. The adjoint
We can shorten the computation using the inclusion . By the inclusion lemma, , so , and the action is still .
To fix the boundary conditions, requires the boundary term to vanish for all with ,
Since is arbitrary, this forces , so
i.e. the adjoint inherits the same periodic boundary condition, only on a bigger function space.
Step 3. Not self-adjoint, but…
Comparing the two,
so is not self-adjoint either, but only because of the regularity gap , and not because of any mismatch in boundary conditions.
Step 4. Essentially self-adjoint!
Now the closure. For and , the boundary term is
(using ). For it to vanish for all such we need , but nothing more, since is the only free quantity. So
and therefore , i.e. is essentially self-adjoint.
Conclusion
The momentum operator on the circle has a unique self-adjoint realization. Taking the closure of the naive operator on with periodic BC gives the operator with domain .
Its eigenfunctions are with eigenvalues , which is exactly the spectrum of momentum on one computes formally in any QM textbook. The point of all the machinery is to justify that calculation, by showing that the operator we are diagonalizing actually exists as a genuine observable.
7. The takeaway
| Setting | Symmetric? | Self-adjoint? | Essentially self-adjoint? | Self-adjoint extensions |
|---|---|---|---|---|
| Interval , | yes | no | no | one-parameter family |
| Circle, | yes | no (regularity only) | yes | unique: the closure |
The interval case is an example of something that comes up again and again in quantum mechanics. A perfectly innocent-looking differential operator can fail to define a unique observable, because the boundary conditions carry physical information that cannot be recovered from the formal expression alone. The circle is the happier case where everything works out, and it works out because the topology of the configuration space already fixes the boundary condition for us.
The general method for the harder cases is von Neumann’s. One counts the deficiency indices , and gets a -family of self-adjoint extensions when , or no extensions at all when . This is the right tool for asking systematically whether a symmetric operator is a genuine observable, and if not, which data is missing.
We have also said nothing about why momentum is in the first place. That is the content of the Stone–von Neumann theorem, which recovers the Schrödinger representation up to unitary equivalence from the canonical commutation relation (plus an irreducibility assumption). The classical analogue is the Poisson bracket . Quantization replaces Poisson brackets by commutators, and Stone–von Neumann says that, up to some mathematical niceness conditions, this prescription has essentially one realization. I will save that for some other day.
References
- B. C. Hall, Quantum Theory for Mathematicians, Springer (2013) — Chapters 7–10.
- M. Reed and B. Simon, Methods of Modern Mathematical Physics, Volume I: Functional Analysis, Academic Press (1980) — especially Chapters VI–VIII.
- F. Schuller, Lectures on Quantum Theory, University of Erlangen-Nürnberg, YouTube playlist (Lectures 6–9 cover this material).
- P. Szekeres, A Course in Modern Mathematical Physics: Groups, Hilbert Space and Differential Geometry, Cambridge University Press (2004) — Chapters 13–14 for background reading.