Many quantum devices are limited by noise that fluctuates in time — \(1/f\) charge or flux noise, magnetic-field fluctuations, laser phase noise. Its effect depends on how the noise spectrum overlaps with the “spectrum” of the control. Filter functions make this precise. To leading order, the error of an operation is the integral of the noise power spectral density multiplied by a filter function that depends only on the control. This gives a transparent, frequency-domain way to evaluate, compare and design pulses for time-dependent noise.

The idea

Consider a qubit with weak classical noise, \(H(t) = H_c(t) + \beta(t)\,\sigma_z\), where \(\beta(t)\) is a stationary Gaussian process with power spectral density \(S(\omega)\). To leading order in the noise, the average error of the controlled evolution takes the form

$$1 - \bar F \;\approx\; \frac{1}{2\pi}\int_{-\infty}^{\infty} S(\omega)\,F_c(\omega)\,d\omega ,$$

where the filter function \(F_c(\omega)\) depends only on the control \(H_c(t)\) (normalization conventions differ between papers). A good control makes \(F_c(\omega)\) small wherever \(S(\omega)\) is large.

For pure dephasing under a sequence of instantaneous \(\pi\) pulses, the filter function is the squared Fourier transform of the switching function \(f(t) = \pm 1\), which changes sign at every pulse:

$$F_c(\omega) \;\propto\; \Big|\int_0^T f(t)\,e^{i\omega t}\,dt\Big|^2 .$$

A free evolution passes low frequencies (\(F_c\) peaks at \(\omega = 0\)); a spin echo or CPMG sequence suppresses them and moves the sensitivity to a band near \(\omega \approx \pi/\tau_p\), where \(\tau_p\) is the spacing between pulses. This explains how dynamical decoupling works and why the best sequence depends on the noise spectrum [1]. Kofman and Kurizki gave the general form of this idea for decay into a continuum: the modified decay rate is the overlap of the bath spectrum with the spectrum of the modulation [2].

Filter functions for arbitrary gates

Decoupling sequences ideally do nothing. Gates rotate the qubit, so noise along one axis turns into errors along others. Generalized filter functions handle this. They are built from control matrices that track how each noise axis is transformed by the control, and they cover arbitrary piecewise-constant pulses with noise along all axes [3, 4]. A transfer-function framework shows how arbitrary filter functions are assembled from a finite, computable set of fundamental ones. It also defines a filtering order, which captures features of error suppression that differ from the order of cancellation in the Magnus expansion [5]. The formalism also extends to complete quantum processes of gate sequences, with a simple rule for composing them [6].

Designing pulses with filter functions

Because the error is an explicit functional of the control, filter functions can be used as objectives. Examples:

  • Walsh-synthesized gates. Amplitude and phase modulations built from Walsh functions are optimized to minimize the filter function in chosen bands [7].
  • Experimental noise filtering. Designed controls suppress noise in selected frequency bands, which has been demonstrated in trapped ions [8].
  • Combined robustness. Filter-function objectives for time-dependent noise can be optimized together with sensitivity to static errors; see Robust & ensemble optimal control.

Noise spectroscopy

The relation can also be inverted. Applying sequences with known, narrow-band filter functions and measuring the resulting decay reconstructs the noise spectrum \(S(\omega)\). This has been done in NMR [9], in superconducting flux qubits [10] and, with multi-pulse frequency combs, for non-Gaussian noise [11]. A known spectrum then feeds back into better control design; see System identification for control and Control for sensing & metrology.

Limitations

  • Leading order. The standard formulas assume weak noise. Strong noise needs higher-order terms or full simulation.
  • Classical, Gaussian noise. Most results treat classical, stationary Gaussian noise. Quantum or non-Gaussian environments require extensions [11].
  • Correlations across gates. Noise correlated over many gates breaks the assumption that errors of successive operations are independent. Process-level filter functions address this [6].