Convex State Space

Density operators form a convex, trace-norm closed subset of the trace-class operators. In finite dimension the set is compact and has a nonempty relative interior; both conclusions fail for an infinite-dimensional Hilbert space. Compactness can be recovered under an energy constraint when the Hamiltonian confines probability to finite-dimensional low-energy sectors.

Helpful background. Density operators defines the state set; trace and operator norms explains the topology used here.

Convexity and trace-norm closedness

Let

$$\mathcal D(\mathcal H) =\{\rho\in\mathcal S_1:\rho\geq0,\ \operatorname{Tr}\rho=1\}.$$

For states $\rho,\sigma$ and $0\leq t\leq1$, the operator $t\rho+(1-t)\sigma$ is positive and has trace one. Operationally this is random selection between preparation procedures, with the selection label discarded.

If $\rho_n\to\rho$ in trace norm, then convergence also holds in operator norm. Every vector therefore has a nonnegative limiting quadratic value, so $\rho\geq0$. Trace continuity gives $\operatorname{Tr}\rho=1$. Hence $\mathcal D$ is closed in the real Banach space of self-adjoint trace-class operators.

Every state has norm one, so the set is bounded. Infinite-dimensional closed bounded sets need not be compact. The normal state space is a particularly physical example of this basic functional-analytic distinction.

Orthogonal states defeat compactness

For an orthonormal sequence $(e_n)$, define $\rho_n=|e_n\rangle\langle e_n|$. If $n\ne m$, the difference has eigenvalues $1,-1$, and

$$\|\rho_n-\rho_m\|_1=2.$$

There is no trace-norm convergent subsequence. Thus the full state set is not trace-norm compact.

This does not contradict weak-* compactness of the state space of the unital algebra $\mathcal B(\mathcal H)$. That larger state space lives in $\mathcal B(\mathcal H)^*$ and includes nonnormal states. A weak-* cluster point of the vector states above can vanish on every finite-rank projection while assigning one to $I$. It then lies outside the density-operator state space. Always specify both the ambient dual space and the topology before invoking compactness.

The other natural duality, $\mathcal S_1=\mathcal K(\mathcal H)^*$, has yet another effect: against compact tests, $\rho_n$ converges to zero. The trace-one condition is not closed in that weak-* topology because $I$ is not a compact test operator.

Why faithful states are not interior points

In dimension $d<\infty$, a positive definite state has a smallest eigenvalue $p_{\min}>0$. Sufficiently small self-adjoint trace-zero perturbations preserve positivity, so it is interior relative to the trace-one affine hyperplane.

In infinite dimension, a density operator has no strictly positive uniform lower bound. For any $\epsilon>0$, choose a unit vector $v$ with $\langle v,\rho v\rangle<\epsilon$ and a unit vector $w\perp v$. Then

$$T=\rho+\epsilon|w\rangle\langle w| -\epsilon|v\rangle\langle v|$$

has trace one, but $\langle v,Tv\rangle<0$. Its trace-norm distance from $\rho$ is $2\epsilon$. Every relative neighborhood therefore contains a nonstate. Even a faithful state, whose kernel is zero, lies on the boundary. Faithfulness is an algebraic nondegeneracy property, not a uniform stability margin.

Energy constraints restore compactness

Let $H\geq0$ be self-adjoint with eigenvalues tending to infinity and finite-dimensional spectral subspaces below every finite cutoff. For finite $E$, consider

$$\mathcal D_E=\{\rho\in\mathcal D:\operatorname{Tr}(\rho H)\leq E\},$$

where energy is the nonnegative spectral expectation. If $P_L=E_H([0,L])$ and $Q_L=I-P_L$, then

$$\operatorname{Tr}(\rho Q_L)\leq E/L.$$

This is a uniform tail bound. Factor products through $\rho^{1/2}$ and use Hilbert–Schmidt Cauchy–Schwarz to obtain

$$\|\rho-P_L\rho P_L\|_1 \leq2\sqrt{\operatorname{Tr}(\rho Q_L)} \leq2\sqrt{E/L}.$$

The compressed operators lie in a bounded set of a finite-dimensional matrix space, hence in a compact set. Increasing $L$ therefore approximates the entire energy-constrained family uniformly by finite-dimensional compact sets: $\mathcal D_E$ is totally bounded. It is closed because energy is the supremum of continuous bounded-cutoff expectations. Completeness then gives compactness.

This proof needs confinement. If $H=0$ on an infinite-dimensional subspace, the energy constraint admits infinitely many mutually orthogonal states and cannot yield compactness. Holevo discusses such constraints in infinite-system information theory.

Exercises

1. For the harmonic-oscillator number Hamiltonian $H=\hbar\omega N$, bound the probability above occupation $M$ in any state of mean energy at most $E$.

Solution

On the subspace $n>M$, $H\geq\hbar\omega(M+1)I$. Hence its probability is at most $E/[\hbar\omega(M+1)]$. This makes a uniform finite-subspace approximation possible for density-operator calculations with a fixed energy budget.

2. Is the convex hull of finitely many fixed pure states enough to approximate every state in trace norm?

Solution

No. Their vectors span a finite-dimensional subspace. A pure state supported in its orthogonal complement has trace distance one from every mixture in that convex hull. A single fixed finite collection cannot capture an infinite-dimensional state space.

References

• Bratteli, Ola, and Derek W. Robinson, Operator Algebras and Quantum Statistical Mechanics 1, 2nd ed., Springer, 1987 — convex state spaces and weak-* topology. Publisher.
• Holevo, Alexander S., Quantum Systems, Channels, Information: A Mathematical Introduction, 2nd revised and expanded ed., De Gruyter, 2019, Chapter 11 — energy constraints and compactness in infinite systems. Publisher.