Trace Norm and Operator Norm

The operator norm controls the largest action on one unit vector, whereas the trace norm sums all singular-value weights. For quantum states, trace norm controls prediction errors uniformly over bounded observables. Infinite-dimensional operator-norm convergence alone can lose all normalization, even when every approximating operator is a valid state.

Helpful background. Trace-class operators introduces $\|\cdot\|_1$ and the ideal bound; density operators supplies trace-one normalization.

Three norms and their different questions

For trace-class $T$ with singular values $s_n$,

$$\|T\|=\sup_ns_n,\qquad \|T\|_2=\left(\sum_ns_n^2\right)^{1/2},\qquad \|T\|_1=\sum_ns_n.$$

Consequently $\|T\|\leq\|T\|_2\leq\|T\|_1$. When $\operatorname{rank}T\leq r$, Cauchy–Schwarz and the maximum bound give

$$\|T\|_1\leq\sqrt r\,\|T\|_2\leq r\|T\|.$$

The constants depend on rank. There is no dimension-independent inequality bounding trace norm by a constant times operator norm on all trace-class operators. All norms are equivalent in a fixed finite-dimensional vector space, but the constants need not survive a growing cutoff.

The distinction matters when a numerical calculation reports a small matrix norm. A small largest matrix error can coexist with a large total error distributed over many levels. State errors should be interpreted using the norm tied to the desired observables.

Trace duality and bounded predictions

The dual of the Banach space $\mathcal S_1(\mathcal H)$ is $\mathcal B(\mathcal H)$ under the trace pairing. In particular,

$$\|T\|_1 =\sup_{\|A\|\leq1}|\operatorname{Tr}(AT)|.$$

The ideal estimate proves the upper bound. For the lower bound, write $T=V|T|$ by polar decomposition and choose $A=V^*$. Since $V^*V$ is the support projection of $|T|$, $\operatorname{Tr}(V^*T)=\operatorname{Tr}|T|$. The partial isometry $V^*$ is a contraction, so it is an admissible test operator.

It follows that

$$|\operatorname{Tr}((\rho-\sigma)A)| \leq\|\rho-\sigma\|_1\|A\|$$

for every bounded observable $A$. For an effect, the sharper bound is $\tfrac12\|\rho-\sigma\|_1$ because the difference is self-adjoint and has trace zero. The proof using its positive and negative parts appears in fidelity and trace distance.

This is uniform control over all bounded tests with the prescribed norm. Convergence of a few selected matrix elements gives less information. Even trace-norm convergence does not uniformly control unbounded energy, position, or particle number without an additional moment condition.

Probability mass escaping to higher dimensions

Let $P_N$ project onto the first $N$ basis vectors in $\ell^2(\mathbb N)$ and set $\rho_N=P_N/N$. Then

$$\operatorname{Tr}\rho_N=1,\qquad \|\rho_N\|=\frac1N,\qquad \|\rho_N\|_2=\frac1{\sqrt N}.$$

Thus $\rho_N\to0$ in both operator and Hilbert–Schmidt norms, while $\|\rho_N\|_1=1$. The limiting operator has trace zero. This is not a contradiction: the trace is continuous in trace norm, not in operator norm on an unbounded-rank family.

For every fixed finite-rank projection $P$, $\operatorname{Tr}(\rho_NP)\to0$, while the identity observable always has expectation one. Fixed finite-dimensional probes miss the probability spread across ever more states.

A related example is $\sigma_n=(1-\epsilon_n)P_0+\epsilon_nP_n$, with orthogonal rank-one projections and $\epsilon_n\to0$. Here $\|\sigma_n-P_0\|_1=2\epsilon_n\to0$. If $HP_n=E_nP_n$, the energy excess is $\epsilon_n(E_n-E_0)$ and can diverge when $E_n$ grows sufficiently fast. Bounded-observable convergence and energy convergence are distinct promises.

Strong convergence under a trace

A useful positive result restores many limit interchanges. If $A_n\to A$ strongly and $\sup_n\|A_n\|<\infty$, then

$$\operatorname{Tr}(TA_n)\longrightarrow\operatorname{Tr}(TA) \qquad(T\in\mathcal S_1).$$

Approximate $T$ by a finite-rank $F$ in trace norm. The finite-rank pairing converges by strong convergence on finitely many vectors. The remaining error is bounded uniformly by $\|T-F\|_1(\sup_n\|A_n\|+\|A\|)$. First choose $F$, then take $n$ large. This order of limits is the rigorous reason a fixed normal state can evaluate strongly convergent bounded cutoff observables.

Simon develops the ideal duality behind these arguments; Holevo uses the resulting trace topology for state approximation.

Exercises

1. Take $\rho_N=P_N/N$ and $\rho_{2N}=P_{2N}/(2N)$. Compute their trace distance.

Solution

The difference has $N$ eigenvalues $1/(2N)$ and $N$ eigenvalues $-1/(2N)$. Its trace norm is one, so its trace distance is $1/2$. These states are not trace-norm Cauchy, despite their common operator-norm limit.

2. For a finite-dimensional density-matrix calculation with a rank-$r$ error $\Delta$, the operator-norm error is at most $\delta$. What trace-distance guarantee follows?

Solution

One obtains $\tfrac12\|\Delta\|_1\leq r\delta/2$. The bound can be poor when $r$ grows. Reporting $\delta$ without the rank or another summability estimate does not provide a uniform probability-error guarantee.

References

• Holevo, Alexander S., Quantum Systems, Channels, Information: A Mathematical Introduction, 2nd revised and expanded ed., De Gruyter, 2019 — trace topology and infinite-dimensional state approximation. Publisher.
• Simon, Barry, Trace Ideals and Their Applications, 2nd ed., Mathematical Surveys and Monographs 120, American Mathematical Society, 2005 — duality, norm inequalities, and finite-rank approximation. Publisher.