On algebraic differentiators
Algebraic differentiators have been derived and discussed in the systems and control theory community. The initial works based on differential-algebraic methods have been developed by Mboup, Join, and Fliess in [2]. These numerical, non-asymptotic approximation approaches for higher-order derivatives of noisy signals are well suited for real-time embedded systems. A historical overview and a detailed discussion of these differentiators and their time-domain and frequency-domain properties are given in the survey [1].
The approximation-theoretic derivation recalled in the survey [1] permits the interpretation of the estimation process by the following three steps illustrated in the figure below stemming from [1]:
Projection: At time \(t\), the sough \(n\)-th order time derivative \(y^{(n)}\) over the interval \(I_T(t)\) is projected onto the space of polynomials of degree \(N\). This yields the polynomial \(p_N\) depicted in the left and middle part of Figure 2.
Evaluation: The polynomial \(p_N\) is evaluated at \(t-\delta_t\), which gives an estimate \(\hat{y}^{(n)}(t)=p_N(t-\delta_t)\) for the derivative \(y^{(n)}\) as depicted in the central part of Figure 2. Choosing the delay to be the largest root of a special Jacobi polynomial increases the approximation order by 1 with a minimal delay. Alternatively, a delay-free estimation or even a prediction of the future derivative might be selected, at the cost of a reduced accuracy.
Repetition: The first two steps are repeated at each discrete time instant while keeping the parameters of the differentiator constant. This yields the estimate \(\hat{y}^{(n)}\) depicted in the right part of the Figure 2.
Algebraic differentiators can be interpreted as linear time-invariant filters with a finite-duration impulse response. Figure 3 visualizes the online estimation process of the first derivative of a noisy signal. The filter window, the buffered signal, and the filter kernel can be clearly seen.
These filters can be approximated as lowpass filters with a known cutoff frequency and a stopband slope. Figure 4 presents the amplitude and phase spectra of two exemplary filters. The lowpass approximation is also shown.
See [1], [3], [4], and [5] for more details on the parametrization of these differentiators.