Tracking Topologies

A topology specifies the tests that an approximation must pass. Tracking it through a proof means checking whether those tests still control the next operation: a supremum over inputs, a long-time limit, a sharp spectral cutoff, or a measurement probability. The most revealing question is often which quantities are fixed while the approximation parameter changes.

Required background. Topology and Convergence Minimum defines the principal vector and operator topologies. Helpful background. Fidelity and Trace Distance relates state overlap to probability errors.

Expand the quantifiers before using the limit

A statement $B_n\to B$ strongly on a common Hilbert space means that each fixed vector eventually has a small error. The index needed for an error tolerance can depend on that vector. An operator-norm statement supplies one index that works for every vector in the unit ball.

This difference matters whenever the proposed use chooses an input depending on $n$. A sequence of difficult test states can evade every fixed-state convergence statement. Similarly, convergence for each fixed time does not automatically control times that grow with the approximation parameter.

When the Hilbert spaces themselves change with a cutoff, specify the embeddings or comparison maps before naming an operator topology. Matrices of different sizes cannot be subtracted until they have been related to a common space. An implicit zero extension can introduce its own spectral features.

A small Hamiltonian change can accumulate at long times

Consider a qubit with $H_0=0$ and

$$H_\varepsilon=\hbar\varepsilon Z, \qquad Z=\begin{pmatrix}1&0\\0&-1\end{pmatrix}.$$

Here $\varepsilon$ has units of inverse time, and $U_\varepsilon(t)=e^{-i\varepsilon tZ}$. Direct diagonalization gives

$$\|U_\varepsilon(t)-I\| =2|\sin(\varepsilon t/2)| \leq|\varepsilon t|.$$

Therefore the evolution converges in norm uniformly for $|t|\leq T$ with fixed $T$. It does not converge uniformly over all times. For $\varepsilon>0$, at $t_\varepsilon=\pi/(2\varepsilon)$ the operator error is $\sqrt2$.

This is observable, not just a global phase issue. The initial state $|+\rangle=(|0\rangle+|1\rangle)/\sqrt2$ evolves at that time to $-i|-\rangle$, orthogonal to $|+\rangle$. A detector for the original preparation changes its probability from one to zero.

The correct approximation statement must therefore name its time window. Finite-time convergence and long-time accuracy are distinct claims even in two dimensions, where the usual matrix topologies otherwise agree.

A discontinuous spectral question can defeat norm convergence

Fix an energy unit $E_*>0$ and let $H_n=(E_*/n)I$ on a nonzero finite-dimensional space. Then $H_n\to0$ in operator norm. Nevertheless,

$$E_{H_n}((-\infty,0])=0, \qquad E_0((-\infty,0])=I.$$

The threshold question is discontinuous exactly at the limiting eigenvalue. Even this strong form of operator convergence does not justify applying an arbitrary discontinuous spectral function.

For comparison, a continuous function $f$ has $f(E_*/n)I\to f(0)I$. If the threshold is moved away from zero, the corresponding projections also eventually agree in this example. In general spectral-convergence theorems, conditions on boundary spectral mass play this role.

When reading a limit of sharp measurement events, inspect the endpoints and the limiting spectral projections. When reading a limit of smooth detector responses, inspect the continuity and boundedness assumptions on the response function. Those are different mathematical tests.

Convert a state cutoff into a measurement-error statement

Let $P_N$ be a finite-rank orthogonal projection and let $\psi$ be normalized. Write its omitted weight as

$$\eta_N=\|(I-P_N)\psi\|^2<1, \qquad \phi_N=\frac{P_N\psi}{\sqrt{1-\eta_N}}.$$

The overlap is $|\langle\psi,\phi_N\rangle|^2=1-\eta_N$. The pure-state trace-distance identity therefore gives

$$\frac12\bigl\||\psi\rangle\langle\psi| -|\phi_N\rangle\langle\phi_N|\bigr\|_1 =\sqrt{\eta_N}.$$

Consequently every effect has a probability error at most $\sqrt{\eta_N}$. This conclusion translates a vector cutoff into a uniform statement over measurements on those two states. It is stronger and more useful than inspecting a few selected matrix entries.

For number-basis truncation to $0,\ldots,N$, a mean-number bound $\langle\mathsf N\rangle\leq M$ implies $\eta_N\leq M/(N+1)$ by bounding $\mathsf N$ below by $N+1$ on the omitted sector. The resulting estimate supplies a rate for that energy-constrained class. It is not uniform over unrestricted states, which can place all their weight beyond the cutoff.

Check continuity of the operation you actually need

After identifying convergence, list the next mathematical operation. Products of strongly convergent bounded-operator sequences behave differently from products of weakly convergent sequences. Unbounded moments need additional integrability or energy control. Resolvent convergence is designed to compare unbounded self-adjoint operators, but spectral conclusions still have their own hypotheses.

The convergence primer supplies the relevant counterexamples and definitions. The practical reading rule is to retain the topology and parameter range until the desired prediction has been obtained; deleting them midway can change the assertion being proved.

Exercise: read a cutoff requirement

A family of pure states has mean occupation at most $M=3$. Using only the bound above, how large must $N$ be to guarantee that the normalized cutoff changes every effect probability by at most $0.1$?

Solution

It suffices that $\sqrt{3/(N+1)}\leq0.1$, or $N+1\geq300$. Thus $N\geq299$ is sufficient. This is a conservative guarantee based only on a first-moment bound; additional decay information can substantially improve it. The guarantee concerns bounded effect probabilities, not convergence of every unbounded observable moment.

References

• G. Teschl, Mathematical Methods in Quantum Mechanics: With Applications to Schrödinger Operators, 2nd ed., Graduate Studies in Mathematics 157, American Mathematical Society, 2014, Chapters 1 and 6 — operator topologies and resolvent convergence.
• J. Watrous, The Theory of Quantum Information, Cambridge University Press, 2018, Chapter 3 — trace distance, pure-state overlap, and measurement distinguishability.