Use control pulses to select temporal noise

Consider longitudinal coupling with angular-frequency shift $\gamma B(s)$, where B is a field and $\gamma$ is the coupling in radians per second per field unit. Assume ideal instantaneous pi pulses about a transverse axis. In the toggling frame, each pulse flips a sign $y(s)$ between plus and minus one. The accumulated phase is the time integral of $\gamma y(s)B(s)$. Free induction has y equal to one throughout. A Hahn echo has one sign in the first half and the opposite sign in the second.

For a static field, the two equal halves of an ideal echo cancel. This protects a memory from a constant offset but also removes sensitivity to a desired constant signal with the same coupling. For a field that reverses sign at the pulse, the same echo can make both contributions add. Control therefore discriminates temporal structure rather than simply reducing every unwanted effect. Finite pulse duration, pulse-axis errors, and additional Hamiltonian terms can invalidate the simple sign model and must be included when relevant.

Now let B be a stationary, zero-mean classical Gaussian field. Define its autocorrelation by $C_B(u)=\langle B(s+u)B(s)\rangle$ and a two-sided angular-frequency spectrum through $C_B(u)=\int[d\omega/(2\pi)]S_B(\omega)e^{-i\omega u}$. In this convention the spectrum has field-squared times seconds as its unit. Gaussian phase averaging gives normalized coherence $W=\langle e^{i\phi}\rangle=\exp(-\chi)$, where $\chi$ is half the phase variance. Substitution of the correlation representation yields the filter overlap.

Factors of two and two pi depend on whether a paper uses one-sided or two-sided spectra and ordinary or angular frequency. They are conventions to translate, not competing physical predictions. For non-Gaussian classical noise, higher cumulants can contribute, while a quantum bath may require an operator correlation treatment. An observed exponential decay alone does not establish a Gaussian spectrum model. Before fitting data, state the noise and pulse assumptions, the normalization of measured coherence, and the frequency coordinate used by the reconstruction.

A sequence usually measures a weighted integral of the spectrum, not its value at one exact frequency. Repeated periodic modulation can produce narrow filter peaks and harmonics, motivating approximations that associate a decay with selected frequency bands. Their widths depend on total sequence duration and their side peaks need accounting. Varying pulse separation moves the weighting, while varying duration changes resolution and signal-to-noise tradeoffs. Calling a filter a perfect delta function hides leakage and can bias the inferred spectrum.

Recovering a spectrum from several decays is an inverse problem. Distinct spectral shapes may fit a finite set of overlaps within uncertainty, so parameterization, accessible bandwidth, and reconstruction stability should be reported. Additional pulse sequences can test a proposed noise model rather than merely adding fitted points. A numerical Fourier illustration is useful for visualizing interference, but it is not evidence that an instrument implements the required control sequence or observes stochastic bath fluctuations. Keep the coherent signal calculation separate from the ensemble-averaged decoherence model.