The arithmetic mean is the most probable value.
Carl F. Gauss [53], p. 244
It is usually taken for granted that the right method for determining the constants is the method of least squares.
Karl Pearson [146], p. 266
Optimal state estimation requires information about noise and initial values, which is not always available. The requirement of initial values is canceled in the optimal unbiased or ML estimators, which still require accurate noise information. Obviously, these estimators can produce large errors if the noise statistics are inaccurate or the noise is far from Gaussian. In another extreme method of state estimation that gives UFIR estimates corresponding to the LS, the observation mean is tracked only under the zero mean noise assumption. Designed to satisfy only the unbiasedness constraint, such an estimator discards all other requirements and in many cases justifies suboptimality by being more robust. The great thing about the UFIR estimator is that, unlike OFIR, OUFIR, and ML FIR, it only needs an optimal horizon of points to minimize MSE. It is worth noting that determining requires much less effort than noise statistics. Given , the UFIR state estimator, which has no other tuning factors, appears to be the most robust in the family of linear state estimators.
In this chapter, we discuss different kinds of UFIR state estimators, mainly filters and smoothers and, to a lesser extent, predictors.
Let us consider an LTV system represented in discrete‐time state space with the following state and observation equations, respectively,
These equations can be extended on as
using the definitions of vectors and matrices given for (4.7) and (4.14).
The unbiased filtering problem can be solved if our goal is to satisfy only the unbiasedness condition
and then minimize errors by choosing the optimal horizon of points.
For the FIR estimate defined as
and the state model represented with the th row vector in (6.3),
the condition (6.5) gives two unbiasedness constraints,
By multiplying both sides of (6.8) from the right‐hand side by the matrix identity and discarding the nonzero on both sides, we find the fundamental gain of the UFIR filter
where . Then, referring to the forced gain given by (6.9), the a posteriori UFIR filtering estimate can be written as
As can be seen, the gain (6.10a) does not contain any information about noise and initial values, which means that the UFIR filter does not have the inherent disadvantages of the KF and OFIR filter.
The error covariance for the a posteriori UFIR filter (6.11) can be written similarly to the OFIR filter (4.31) as
where the error residual matrices are given by
It now follows that (6.12) has the same structure as (4.55) of the a posteriori OUFIR filter with, however, another gain in (6.13) and (6.14).
Because given by (6.10a) does not require noise covariances, the UFIR filter is not optimal and therefore generally gives larger errors than the OUFIR filter. In practice, however, this flaw can be ignored if a higher robustness is required, as in many applications, especially industrial.
An important indicator of the effectiveness of FIR filtering is the NPG introduced by Trench in [198]. The NPG is the ratio of the output noise variance to the input noise variance , which is akin to the noise figure in wireless communications. For white Gaussian noise, the NPG is equal to the sum of the squared coefficients of the FIR filter impulse response ,
which is the squared norm of .
In state space, the homogeneous gain represents the coefficients of the FIR filter impulse response. Therefore, the product plays the role of a generalized NPG (GNPG) [179]. Referring to (6.10a) and (6.10b), GNPG can be written in the following equivalent forms:
It follows that GNPG is a symmetric square matrix , where the main diagonal components represent the NPGs for the system states, and the remaining components the cross NPGs. The main property of is that its trace decreases with increasing horizon length, which provides effective noise reduction. On the other hand, an increase in causes an increase in bias errors, and therefore must be optimally set by choosing an optimal horizon length.
Recursions for the batch a posteriori UFIR filtering estimate (6.11) can be found by decomposing (6.11) as , where
is the homogeneous estimate and is the forced estimate.
Using (6.16b), the estimate can be transformed to
and, by applying the decomposition
the recursion for the GNPG (6.16b) can be obtained if we transform as
This gives the following forward and backward recursive forms:
Using (6.20) and (6.22) and extracting from (6.10a), we next transform the product as
By combining (6.21) and (6.23), we finally obtain the following recursion for the homogeneous estimate
where is the bias correction gain of the UFIR filter.
To find the recursion for the forced estimate (6.18), we first show that
Then, by extracting from (6.19), can be transformed as
Now, by combining (6.25) and (6.26) and substituting (6.22), the forced estimate (6.18) can be represented with the recursion
which, together with (6.24), gives the recursions for the a priori and a posteriori UFIR filtering estimates, respectively,
The recursive computation of the a posteriori UFIR filtering estimate (6.11) can now be summarized with the following theorem.
Proof The derivation of the recursive forms for (6.11) is provided by (6.17)–(6.29).
Note that the need to compute the initial and in short batch forms arises from the fact that the inverse in (6.16c) does not exist on shorter horizons.
It was shown in [50,195] that the robustness of the UFIR filter can be improved by scaling the GNPG (6.21) with the weight as
The weight is defined by
where , is the integer part of , and is the number of the states. The RMS deviation of the estimate is computed using the innovation residual as
where means the dimension of the target motion. With this adaptation of the GNPG to the operation conditions, the UFIR filter can be approximately twice as accurate as the KF when tracking maneuvering targets [50].
Although the UFIR filter does not require error covariance to obtain an estimate, may be required to evaluate the quality of the estimate on a given horizon. Since the batch form (6.12) is not always suitable for fast estimation, next we present the recursive forms obtained in [226].
Consider given by (6.12) and represent with
where the components are defined as
Using the decomposition , we represent matrix recursively as
Then, referring to (6.20), (6.22), (6.31), and
and substituting with , we represent matrices and recursively with
Similarly, the recursions for and can be obtained as
By combining (6.31), (6.34), (6.35), and (6.36) with (6.30), we finally arrive at the recursive form for the error covariance of the a posteriori UFIR filter
It is worth noting that (6.37) is equivalent to the error covariance of the a posteriori KF (3.77), in which the Kalman gain must be replaced with the bias correction gain of the UFIR filter. The notable difference is that the Kalman gain minimizes the MSE and is thus optimal, whereas the derived from the unbiased constraint (6.17) makes the UFIR estimate truly unbiased, but noisier, since no effort has been made so far to minimize error covariance. The minimum MSE can be reached in the UFIR filter output at an optimal horizon of points, which will be discussed next.
Unlike KF, which has IIR, the UFIR filter operates with data on the horizon of points, which is sometimes called the FIR filter memory. To minimize MSE, should be optimally chosen as . Otherwise, if , noise reduction will be ineffective, and, for , bias errors will prevail as shown in Fig. 6.1.
Since the UFIR filter is not selective, there is no exact relationship between filtering order and memory. Optimizing the memory of a UFIR filter in state space requires finding the derivative of the trace of the error covariance with respect to , which is problematic, especially for LTV systems. In some cases, can be found heuristically for real data. If a reference model or the ground truth is available at the test stage, then can be found by minimizing . For polynomial models, was found in [169] analytically through higher order states, which has explicit limitations. If the ground truth is not available, as in many applications, then can be measured via the derivative of the trace of the measurement residual covariance, as shown in [153]. Next we will discuss several such cases based on the time‐invariant state‐space model
All methods will be separated into two classes depending on the operation conditions: with and without the ground truth.
When state is available through the ground truth measurement, which means that is also available, can be found by referring to Fig. 6.1 and solving on the following optimization problem
There are several approaches to finding using (6.40).
where is prior error covariance of the KF. For the given (6.43), the backward recursion (6.22) can be used to compute the GNPG back in time as
Because an increase in always results in a decrease in the trace of [179], then it follows that can be estimated by calculating the number of steps backward until finally becomes singular. Since is singular when , where is the number of the states, the optimal horizon can be measured as . This approach can serve when or when there are enough data points to find . Its advantage is that can be found even for LTV systems.
The previously considered methods make it possible to obtain by minimizing . However, if the ground truth is unavailable and the noise covariances are not well known, then measured in this ways may be incorrect. If this is the case, then another approach based on the measurement residual rather than the error covariance may yield better results.
In cases when the ground truth is not available, can be estimated using the available measurement residual covariance [153]. To justify the approach, we represent as
where the two last terms strongly depend on . Now, three limiting cases can be considered:
The transition between the values can be learned if we take into account (6.46), (6.47), and (6.48). Indeed, when , the bias errors are practically insignificant. They grow with increasing and approach the standard deviation in the estimate at when the filter becomes slightly inconsistent with the system due to process noise.
Since decreases monotonously with increasing due to noise reduction, is also a monotonic function when . It grows monotonously from given by (6.46) to given by (6.47) and passes through with a minimum rate when is minimum. It matters that the rate of around cannot be zero due to increasing bias errors. And this is irrespective of the model, as the UFIR filter reduces white noise variance as the reciprocal of , and the effect of bias is still small when .
Another behavior of can be observed for , when the bias errors grow not equally in different models. Although is always convex on in stable filters, bias errors affecting its ascending right side can grow either monotonously in polynomial models or eventually oscillating in harmonic models. The latter means that can reach its value (6.48) with oscillations, even having a constant average rate.
It follows from the previous that the derivative of with respect to can pass through multiple minima and that the first minimum gives the required value of . The approach developed in [153] suggests that can be determined in the absence of the ground truth by solving the following minimization problem
where minimization should be ensured by increasing , starting from , until the first minimum is reached. To avoid ambiguity when solving the problem (6.49), the number of points should be large enough. Moreover, may require smoothing before applying the derivative.
Let us now look at the properties in more detail and find a stronger justification for (6.49). Since the difference between the a priori and a posteriori UFIR estimates does not affect the dependence of on , we can replace with , with , and with . We also observe that, since , , , and can be replaced by the UFIR estimate (6.29), the last two terms in (6.45) can be represented as
This transforms (6.45) to
Again we see the same picture. By , the bias correction gain in the UFIR filter is close to unity, , and (6.50) gives a value close to zero, as in (6.46). When , and are small enough and (6.50) transforms into (6.47). Finally, for , becomes large and small, which leads to (6.48).
Although we have already discussed many details, it is still necessary to prove that the slope of is always positive up to and around and that this function is convex on . To show this, represent the derivative of as , assuming a unit time‐step for simplicity. For (6.50) at the point we have and, therefore, can be written as
Since the GNPG decreases with increasing , then we have , and it follows that and so is the derivative of ,
Thus, passes through with minimum positive slope. Moreover, since bias errors force to approach as increases, function is reminiscent of . Finally, further minimizing with yields , which can be called the key property of .
Example 6.1 Optimal horizon for two‐state polynomial model [153]. Consider a two‐state polynomial model (6.38) and (6.39) with , , , , , , , , , and .
Functions , , and computed analytically and numerically are shown in Fig. 6.2. As can be seen, is measured either at the minimum of or at the minimum of . For this model, can be calculated using the empirical relationship found in [153] using the noise variances. It also gives
A example of measuring for a two‐state harmonic model can be found in [153].
Example 6.2 Filtering a harmonic model [177]. Given a continuous‐time two‐state harmonic model of a conservative oscillator having the system matrix , where is the fundamental angular frequency, , and is an uncertain factor. An oscillator is observed in discrete time as with . The stationary noise is zero mean white Gaussian with the covariance , where and is the uniform double‐sided PSD of . The measurement noise is .
The continuous‐time state model is converted to , where the system matrix is
and zero mean noise has the covariance
The KF and UFIR filter are run for , , , , and when and otherwise. Filtering errors are shown in Fig. 6.3, where the typical effects inherent to harmonic models are observed. Influenced by the uncertainty acting over a short period of time , KF produces long and oscillating transients, while the UFIR filter returns monotonously to normal filtering after points.
The backward UFIR filter can be derived similarly to the forward UFIR filter if we refer to (5.23) and (5.26) and write the extended model as
for which the definitions of all vectors and matrices can be found after (5.23) and (5.26). Batch and recursive forms for this filter can be obtained as shown next.
The batch backward UFIR estimate can be defined as
where is given by (6.53). The homogenous gain and the forced gain can be determined by satisfying the unbiasedness condition applied to (6.54) and the model
which is the last th row vector in (6.52). This gives two unbiasedness constraints
and the first constraint (6.56) yields the fundamental gain
For obtained by (6.58), the forced gain is defined by (6.57), and we notice that the same result appears when the noise is neglected in the backward OFIR estimate (5.28). The backward a posteriori UFIR filtering estimate is thus given in a batch form by (6.54) with the gains (6.58) and (6.57).
The error covariance for the backward UFIR filter can be written as
where the error residual matrices are specified by
Note that, as in the forward UFIR filter, matrices (6.60) and (6.61) provide an optimal balance between bias and random errors in the backward UFIR filtering estimate if we optimally set the averaging horizon of points, as shown in Fig. 6.1 and Fig. 6.2.
The batch estimate (6.54) can be computed iteratively on if we find recursions for and .
By introducing the backward GNPG
we write the homogeneous estimate as
and represent recursively the inverse of the GNPG (6.62) by
This gives two recursive forms
Referring to taken from (6.63a) and (6.63b), we next represent the product as
By combining (6.65), (6.66), and (6.67), a recursion for can now be written as
where is the bias correction gain of the UFIR filter.
To derive a recursion for the forced gain , we use the decompositions
follow the derivation of the recursive forms for the backward OFIR filter, provide routine transformations, and arrive at
A simple combination of (6.68) and (6.69) finally gives
where the prior estimate is specified by
The backward iterative a posteriori UFIR filtering algorithm can now be generalized with the pseudocode listed as Algorithm 12. The algorithm starts with the initial and computed at in short batch forms (6.62) and (6.54). The iterative computation is performed back in time on the horizon of points so that the filter gives estimates from zero to .
Example 6.3 Forward and backward filtering estimates. Consider a moving vehicle whose trajectory, measured each second, , by a GPS navigation unit, is used as ground truth. To estimate the vehicle state, use the two‐state polynomial model (6.38) and (6.39) with matrices specified in Example 6.1. The scalar noise components and are specified with , and , and the optimal horizon for the UFIR filter is measured as .
Estimates of the vehicle coordinate obtained with the forward and backward KF and UFIR filter in the presence of a temporary uncertainty are shown in Fig. 6.4. As can be seen, beyond the uncertainty, all filters give consistent and hardly distinguishable estimates. The uncertainty that occurs as a step function at affects the estimates in such a way that UFIR filters produce larger excursions and shorter transients, while KFs produce smaller excursions and longer transients.
It is worth noting that the typical differences between KFs and UFIR filters illustrated in Fig. 6.4 are recognized as fundamental [179]. Another thing to mention is that forward and backward filters act in opposite directions, so their responses appear antisymmetric.
The recursive form for the error covariance (6.59) of the backward a posteriori UFIR filter can be found similarly to the forward UFIR filter. To this end, we first represent (6.59) using (6.60) and (6.61) as
We then find recursions for each of the matrix products in the previous relationship, combine the recursive forms obtained, and finally come up with
It can now be shown that there is no significant difference between the error covariances of the forward and backward UFIR filters. Since these filters process the same data, but in opposite directions, it follows that the errors on the given horizon are statistically equal.
In postprocessing and when filtering stationary and quasi stationary signals, smoothing may be the best choice because it provides better noise reduction. Various types of smoothers can be designed using the UFIR approach, although many of them appear to be equivalent as opposed to OFIR smoothing. Here we will first derive the ‐lag FFFM UFIR smoother and then show that all other UFIR smoothing structures of this type are equivalent due to their ability to ignore noise.
Let us look again at the ‐lag FFFM FIR smoothing strategy illustrated in Fig. 5.2. The corresponding UFIR smoother can be designed to satisfy the unbiasedness condition
where the ‐lag estimate can be defined as
and the state model represented by the th row vector in (6.3) as
where is the th row vector in (4.9) and so is in .
Condition (6.73) applied to (6.74) and (6.75) gives two unbiasedness constraints,
and after simple manipulations the first one in (6.76) gives a fundamental UFIR smoother gain ,
Referring to , we next transform (6.78) to
where is the homogeneous gain (6.10a) of the UFIR filter, and express the ‐lag homogeneous UFIR smoothing estimate in terms of the filtering estimate as
that does not require recursion; that is, the recursively computed filtering estimate is projected into in one step by the matrix .
For the forced estimate, the recursive form
appears if we first represent the last row vector of the matrix given by (4.9) as
where matrix becomes matrix if we replace by . Also can be written as
and then the subsequent modification of (6.81) gives
Next, using the decomposition , we transform the forced estimate to the form
where the ‐varying product correction term is computed recursively as
By combining (6.80) and (6.82), we finally arrive at the recursion
using which, the recursive ‐lag FFFM a posteriori UFIR smoothing algorithm can be designed as follows. Reorganize (6.83) as
set , assign , and compute for until this recursion gives . Given and , the smoothing estimate is obtained by (6.84). It is worth noting that in the particular case of an autonomous system, , the ‐lag a posteriori UFIR smoothing estimate is computed using a simple projection (6.80).
In batch form, the ‐varying error covariance of the FFFM UFIR smoother is determined by (6.12), although with the renewed matrices,
where the error residual matrices are given by
To find a recursive form for (6.85), we write it in the form
where is the error covariance (6.37) of the UFIR filter, and represent the ‐varying amendment as
It can be seen that naturally becomes zero at due to . Furthermore, the structure of the matrix suggests that and also .
By considering several cases of for ,
and reasoning deductively, we obtain the following recursion for ,
where the matrix is still in batch form as
Similarly, we represent in special cases as
and then replace the sum in the parentheses with
which gives the following recursion for ,
By combining (6.90) and (6.93) with (6.89), we finally obtain the recursive form for the error covariance as
which should be computed starting with for the initial value
using the matrix given by (6.91) and by (6.92). Noticing that the recursive form for the batch matrix given by (6.92) is not available in this procedure, we postpone to “Problems” the alternative derivation of recursive forms for (6.85).
For LTI systems, the matrix , given in batch form as (6.92), can easily be represented recursively as
and the ‐lag FFFM UFIR smoothing algorithm can be modified accordingly. Computing the initial by (6.95) and knowing the matrix , one can update the estimates for as
It should be noted that the computation of the error covariance using (6.99) does not have to be necessarily included in the iterative cycle and can be performed only once after reaches the required lag‐value.
Other types of ‐lag UFIR smoothers can be obtained if we follow the FFBM, BFFM, and BFBM strategies discussed in Chapter . To show the equivalence of these smoothers, we will first look at the FFBM UFIR smoother and then draw an important conclusion.
The ‐lag FFBM UFIR smoother can be obtained similarly to the FFFM UFIR smoother in several steps. Referring to the backward state‐space model in (5.23) and (5.26), we define the state at ) as
and the forward UFIR smoothing estimate as
From (6.100) at the point , we can also obtain
where is the last row vector in (5.41) and so is in .
The unbiasedness condition applied to (6.100) and (6.101) gives
Taking into account that and replacing extracted from (6.102) with , the previous relationship can be split into two unbiasedness constraints
What now follows is that the first constraint (6.103) is exactly the constraint (6.76) for the FFFM UFIR smoother. We then observe that for we have and thus can be transformed as
Finally, we transform the second constraint (6.104) to
and conclude that this is constraint (6.77) for FFFM UFIR smoothing.
Now the following important conclusion can be drawn. Because the FFBM and FFFM UFIR smoothers obey the same unbiasedness constraints in (6.76) and (6.77), then it follows that these smoothers are equivalent. And this is not surprising, since all UFIR structures obeying only the unbiasedness constraint ignore noise. By virtue of that, the forward and backward models become identical at , and thus the FFFM and FMBM UFIR smoothers are equivalent. In addition, since the forward and backward UFIR filters are equivalent for the same reason, then it follows that the BFFM and BFBM UFIR smoothers are also equivalent, and an important finding follows.
Equivalence of UFIR smoothers: Satisfied only the unbiasedness condition, ‐lag FFFM, FFBM, BFFM, and BFBM UFIR smoothers are equivalent.
In view of the previous definition, the ‐lag FFFM UFIR smoother can be used universally as a UFIR smoother in both batch and iterative forms. Other UFIR smoothing structures such as FFBM, BFFM, and BFBM have rather theoretical meaning.
There is a wide class of systems and processes whose states change slowly over time and, thus, can be represented by degree polynomials. Examples can be found in target tracking, networking, and biomedical applications. Signal envelopes in narrowband wireless communication channels, remote wireless control, and remote sensing are also slowly changing.
The theory of UFIR state estimation developed for discrete time‐invariant polynomial models [172] implies that the process can be represented on using the following state‐space equations
where and are zero mean noise vectors. The power of the system matrix has a specific structure
which means that, by , each row in is represented with the descending th degree, , Taylor or Maclaurin series, where is the number of the states.
The UFIR approach applied to the model in (6.106) and (6.107) to satisfy the unbiasedness condition gives a unique th degree polynomial impulse response function for each of the states separately, where is a discrete time shift relative to the current point . Function has many useful properties, which make the UFIR estimate near optimal. Depending on , the following types of can be recognized. When , function is used to obtain UFIR filtering. When , function serves to obtain ‐lag UFIR smoothing filtering. Finally, when , function is used to obtain ‐step UFIR predictive filtering.
Accordingly, the following problems can be solved by applying UFIR structures to a polynomial signal measured as in the presence of zero mean additive noise:
Note that one‐step predictive UFIR filtering, , was originally developed for polynomial models in [71]. In state space, it is known as RH FIR filtering [106] used in state‐feedback control and MPC.
In the previous definitions of the state estimation problems, the functions and are not equal for , but can be transformed into each other. More detail can be found in [181].
Looking at the details of the UFIR strategy, we notice that the approach that ignores zero mean noise allows us to solve universally the filtering problem (6.109), the smoothing filtering problem (6.110), and the predictive filtering problem (6.111) by obtaining a ‐shift estimate [176]. The UFIR approach also assumes that a shift to the past can be achieved at point using data taken from with a positive smoother lag , a shift to the future at point using data taken from with a positive prediction step , and that means filtering.
Thus, the ‐shift UFIR filtering estimate can be defined as
where the components of the gain are diagonal matrices specified by the matrix
whose components, in turn, are the values of the function . The unbiasedness condition applied to (6.114) gives the unbiasedness constraint
where .
For the th system state, the ‐shift UFIR filtering estimate is defined as
and the constraint (6.115) is transformed to
where means the th row in and the remaining th rows are given by
It is worth noting that the linear matrix equation 6.117 can be solved analytically for the th degree polynomial impulse response . This gives the ‐varying function
where , , and the coefficient is defined by
where the determinant of the ‐varying Hankel matrix is specified via the Vandermonde matrix as
and is the minor of . The th component , , of matrix (6.122) is the power series
where is the Bernoulli polynomial.
The coefficients of several low‐degree polynomials , which are most widely used in practice, are given in Table 6.1. Using this table, or (6.121) for higher‐degree systems, one can obtain an analytic form for as a function of , where is for UFIR filtering, for UFIR smoothing filtering, and for UFIR predictive filtering.
Table 6.1 Coefficients of Low‐Degree Functions .
Coefficients | |
---|---|
Uniform: | |
Ramp: | |
Quadratic: | |
The most important properties of the impulse response function are listed in Table 6.2, which also summarizes some of the critical findings [181]. If we set , we can use this table to examine the properties of the UFIR filter. We can also characterize the UFIR filter in terms of system theory. Indeed, since the transfer function of the th degree UFIR filter is equal to unity at zero frequency, , then it follows that this filter is essentially a low‐pass (LP) filter. Moreover, if we analyze other types of FIR structures in a similar way, we can come to the following general conclusion.
Table 6.2 Main Properties of .
Property | |
---|---|
Region of existence: | |
‐transform at : | (UFIR filter is an LP filter) |
Unit area: | |
Energy (NPG): | |
Value at zero: | |
Zero moments: | , |
, | |
Orthogonality: | , |
, | |
Unbiasedness: | |
State estimator in the transform domain: All optimal, optimal unbiased, and unbiased state estimators are essentially LP structures.
Since the sum of the values of is equal to unity, it follows that the UFIR filter is strictly stable. More specifically, it is a BIBO stable filter due to the FIR. The sum of the squared values of represents the filter NPG, which is equal to the energy of the function . The important thing is that NPG is equal to the function at zero, which, in turn, is equal to the zero‐degree coefficient . This means that the denoising properties of the UFIR filter can be fully explored using the value . It is also worth noting that the family of functions and , , establish an orthogonal basis, and thus high‐degree impulse responses can be computed through low‐degree impulse responses using a recurrence relation.
The fact that all moments of the function are equal to zero means nothing more and nothing less than the UFIR filter is strictly unbiased by design. The unbiasedness of FIR filters can also be checked by equating the area of the transfer function and the area of the squared magnitude frequency response. The same test for unbiasedness in the discrete Fourier transform (DSP) domain is ensured by equating the sum of the DSP values and the sum of the squared magnitude values. At the end of this chapter, when we will consider practical implementations, we will take a closer look at the properties of UFIR filters in the transform domain.
The UFIR filter can be viewed as a special case of the ‐shift UFIR filter when . This transforms the impulse response function (6.120) to
where the coefficient is defined by (6.121) as
The constant (zero‐degree, ) FIR function is used for simple averaging. The ramp (first‐degree, ) FIR function
is applicable for linear signals. The quadratic (second‐degree, ) FIR function
is applicable for quadratically changing signals, etc.
There is also a recurrence relation [131]
that can be used when and to compute of any degree in terms of the lower‐degree functions. Several low‐degree functions computed using (6.128) are shown in Fig. 6.5.
The NPG of a UFIR filter, which is defined as , suggests that the best noise reduction associated with the lowest NPG is obtained by simple averaging (case in Fig. 6.5). An increase in the filter degree leads to an increase in the filter output noise, and the following statements can be made.
Noise reduction with polynomial filters: 1) A zero‐degree filter (simple averaging) is optimal in the sense of minimum produced noise; 2) An increase in the filter degree or, which is the same, the number of states leads to an increase in random errors.
Therefore, because of the better noise reduction, the low‐degree UFIR state estimators are most widely used. In Fig. 6.5, we see an increase in caused by an increase in the degree at .
In [131], the class of the th‐degree polynomial FIR functions , defined by (6.124) and having the previously listed fundamental properties, was tested using the orthogonality condition
where is the Kronecker symbol and . It was found that the set of functions for and is orthogonal on with the square of the weighted norm of given by
where and for is the Pochhammer symbol. The non‐negative weight in (6.129) is the ramp probability density function
An applied significance of this property for signal analysis is that, due to the orthogonality, higher‐degree functions can be computed in terms of lower‐degree functions using a recurrence relation (6.128).
Since the functions have the properties of discrete orthogonal polynomials (DOP), they were named in [61,162] discrete Shmaliy moments (DSM) and investigated in detail. It should be noted that, due to the embedded unbiasedness, DSM belong to the one‐parameter family of DOP, while the classical Meixner and Krawtchouk polynomials belong to the two‐parameter family, and the most general Hahn polynomials belong to the three‐parameter family of DOP [131]. This makes the DSM more suitable for unbiased analysis and reconstruction if signals than the classical DOP and Tchebyshev polynomials. Note that DSM are also generalized by Hanh's polynomials along with the classical DOP.
Both the ‐lag UFIR smoothing filtering problem (6.110) and the UFIR smoothing problem (6.112) can be solved universally using (6.120) and (6.121) if we assign , .
A feature of UFIR smoothing filtering is that the function exists on the horizon , and otherwise it is equal to zero, while the lag is limited to . Zero‐degree UFIR smoothing filtering is still provided by simple averaging. First‐degree UFIR smoothing filtering can be obtained using the ‐varying ramp response function
where is the UFIR filter ramp impulse response (6.126). What can be observed is that better noise reduction is accompanied in (6.132) by loss of stability. Indeed, when approaches unity, the second term in (6.132) grows indefinitely, which also follows from NPG of the first‐degree given by
Approximation (6.134) valid for large and clearly shows that increasing the horizon length leads to a decrease in NPG, which improves noise reduction. Likewise, increasing the lag results in better noise reduction.
The features discussed earlier are inherent to UFIR smoothing filters of any degree, although with the important specifics illustrated in Fig. 6.6. This figure shows that all smoothing filters of degree provide better noise reduction as increases, starting from zero. However, NPG can have multiple minima, and therefore the optimal lag does not necessarily correspond to the middle of the averaging horizon . For odd degrees, can be found exactly in the middle of , while for even degrees at other points. For more information see [174].
The smoothing problem defined by (6.112) and discussed in state space in Chapter can be solved if we introduce a gain as
which exists on with the same major properties as . Indeed, the ramp UFIR smoother can be designed using the FIR function
The NPG for this smoother is defined by
As expected, NPG (6.138) is exactly the same as (6.133) of the UFIR smoothing filter, since noise reduction is provided by both structures with equal efficiency. Note that similar conclusions can be drawn for other UFIR smoothing structures applied to polynomial models.
A special case of the UFIR smoothing filter was originally shown in [158] and is now called the Savitzky‐Golay (SG) filter. The convolution‐based smoothed estimate appears at the output of the SG filter with a lag in the middle of the averaging horizon as
where the set of convolution coefficients is determined by the linear LS method to fit with typically low‐degree polynomial processes. Since the coefficients can be taken directly from the FIR function , the SG filter is a special case of (6.110) with the following restrictions:
It then follows that the UFIR smoothing filter (6.110), developed for arbitrary and lags , generalizes the SG filter (6.139) in the particular case of odd and . Also note that the lag in (6.110) can be optimized for even‐degree polynomials as shown in [174].
The predictive filtering problem (6.111) can be solved directly using the ‐shift FIR function (6.120) if we set . Like filtering and smoothing, zero‐degree UFIR predictive filtering is provided by simple averaging. The first‐degree predictive filter can be designed using a ‐varying ramp function
which makes a difference with smoothing filtering. The NPG of the predictive filter is determined by
and we note an important feature: increasing the prediction step leads to an increase in NPG, and denoising becomes less efficient. Looking at the region around in Fig. 6.5 and referring to (6.138) and (6.142), we come to the obvious conclusion that prediction is less precise than filtering and filtering is less precise than smoothing.
The prediction problem (6.113) can be solved similarly to the smoothing problem by introducing the gain
which exists on with the same main properties as . The ramp UFIR predictor can be designed using the FIR function
and its efficiency can be estimated using the NPG (6.141). Note that similar conclusions can be drawn for other UFIR predictors corresponding to polynomial models.
Example 6.4 Smoothing, filtering, and prediction errors [177]. Consider the dynamics of an object described by the following continuous‐time state model with . Convert this model in discrete time to , where and . The noise has the covariance , where and are the uniform double‐sides PSDs of the second and third state noise, respectively. The object is observed as , where and .
To compare ‐varying estimation errors, we set , , and and run the ‐shift OFIR and UFIR filters. The estimation errors are shown in Fig. 6.7 for and . It can be seen that the minimum errors correspond to the lag and grow at any deviation from this point.
Like KF, the UFIR filter can also be generalized for Gauss‐Markov colored noise, if we take into account that unbiased averaging ignores zero mean noise. Accordingly, if we convert a model with colored noise to another with white noise and then ignore the white noise sources, the UFIR filter can be used directly. In Chapter, we generalized KF for CMN and CPN. In what follows, we will look at the appropriate modifications to the UFIR filter and focus on what makes it better in the first place: the ability to filter out more realistic nonwhite noise. Although such modifications require tuning factors, and therefore the filter may be more vulnerable and less robust, the main effect is usually positive.
We consider the following state‐space model with Gauss‐Markov CMN, which was used in Chapter to design the GKF,
where and have the covariances and . The coloredness factor is chosen such that the Gauss‐Markov noise is always stationary, as required. Using measurement differencing as , we write a new observation as
where is the new observation matrix, the auxiliary matrices are defined as and , and the noise
is zero mean white Gaussian with the properties , , and , where .
It can be seen that the modified state‐space model in (6.145) and (6.148) contains white and time‐correlated noise sources and . Unlike KF, the UFIR filter does not require any information about noise, except for the zero mean assumption. Therefore, both and can be ignored, and thus the UFIR filter is unique for both correlated and de‐correlated and . However, the UFIR filter cannot ignore CMN , which is biased on a finite horizon .
The pseudocode of the a posteriori UFIR filtering algorithm for CMN is listed as Algorithm 13. To initialize iterations avoiding singularities, the algorithm requires a short measurement vector and an auxiliary block matrix
It can be seen that for this algorithm becomes the standard UFIR filtering algorithm. More details about the UFIR filter developed for CMN can be found in [183].
The error covariance for the UFIR filter is given by [179]
where the matrices , , and are defined earlier and the reader should remember that the GNPG matrix is symmetric.
Typical RMSEs produced by the KF and UFIR algorithms versus are shown in Fig. 6.8 [183], where we recognize several basic features. It can be seen that the KF and UFIR filter modified for CMN give fewer errors than the original ones. It should also be noted that GKF performs better when , and that the general UFIR filter is more accurate when . This means that a more robust UFIR filter may be a better choice when measurement noise is heavily colored.
Unlike the CMN, which always needs to be filtered out, the CPN, or at least its slow spectral components, can be tracked to avoid losing information about the process behavior. Let us show how to deal with CMN based on the following the state‐space model
where matrices , , and are nonsingular, , and is the Gauss‐Markov CPN. Noise vectors and are mutually uncorrelated with the covariances and . The coloredness factor matrix is chosen such that the CPN is always stationary.
Using state differencing (3.175a), a new state equation can be written as
where is white Gaussian, , , and is defined by solving for initial the NARE , where .
Using (3.188), we write the modified observation equation as
where can be substituted with the available past estimate .
The pseudocode of the UFIR algorithm developed for CPN is listed as Algorithm 14. To initialize iterations, Algorithm 14 employs a short data vector and an auxiliary matrix
Note that, by setting and , Algorithm 14 becomes the standard iterative UFIR filtering algorithm.
The error covariance of the UFIR filter modified for CPN can be found if we notice that , where . Since this estimate is subject to the constraint [182], the error for can be transformed to
and the corresponding error covariance found to be
This finally gives
Typical RMSEs produced by the modified and original filters for a two‐state polynomial model with CPN are shown in Fig. 6.9 as functions of the scalar coloredness factor [182]. The filtering effect here is reminiscent of the effect shown in Fig. 6.6 for CMN, and we notice that the accuracy of the original filters has been improved. However, a significant improvement in performance is recognized only when the coloredness factor is relatively large, . Otherwise, the discrepancies between the filter outputs are not significant.
Considering the previous modifications of the UFIR filter, we finally conclude that the filtering effect in the presence of CMN and/or CPN is noticeable only with strong coloration.
Representation of physical processes and approximation of systems using linear models does not always fit with practical needs. Looking at the nonlinear model in (3.226) and (3.227) and analyzing the Taylor series approach that results in the extended KF algorithms, we conclude that UFIR filtering can also be adapted to nonlinear behaviors [178], as will be shown next.
Given a nonlinear state‐space model
where and are mutually uncorrelated, zero mean, and not obligatorily Gaussian additive noise vectors. In Chapter, when derived EKF, it was shown that (6.156) and (6.157) can be approximated using the second‐order Taylor series expansion as
where is the modified observation vector and and represent the components resulting from the linearization,
in which , , and and are Cartesian basis vectors with the th and th components unity, and all others are zeros. The nonlinear functions are represented by
where and are Jacobian matrices and and are Hessian matrices.
Based on (6.158) and (6.159), we can now develop the first‐ and second‐order extended UFIR filtering algorithms. Similarly to the EKF‐1 and EKF‐2 algorithms, we will refer to the extended UFIR algorithms as EFIR‐1 and EFIR‐2.
Let us look at the model in (6.158) and (6.159) again and assume that the mutually uncorrelated and are zero mean and white Gaussian. The EFIR‐1 (first‐order) filtering algorithm can be designed using the following recursions [178]
The pseudocode of the EFIR‐1 filtering algorithm is listed as Algorithm 15. As in the EKF‐1 algorithm, here the prior estimate is obtained using a nonlinear projection , and then is projected onto the observation as . Also, the system and observation matrices and are Jacobian. The error covariance for this filter can be computed using (6.37) and the Jacobian matrices and .
The EFIR‐1 filtering algorithm developed in this way turns out to be simple and in most cases truly efficient. However, it should be noted that the initial state , computed linearly in line 5 of Algorithm 15, may be too rough when the nonlinearity is strong. If so, then can be computed using multiple projections as , , ..., . Otherwise, an auxiliary EKF‐1 algorithm can be used to obtain .
The derivation of the second‐order EFIR‐2 filter is more complex, and its application may not be very useful because it loses the ability to ignore noise covariances. To come up with the EFIR‐2 algorithm, we will mainly follow the results obtained in [178].
We first define the prior estimate by
for which, by the cyclic property of the trace operator, the expectation of can be found as
We can show that for given by (6.164), the expectation of the prior error becomes identically zero,
Averaging the nonlinear function gives
where the expectation of can be found as
that allows us to obtain the error covariance as will be shown next.
Using (6.158) and (6.164) and taking into account that, for zero mean Gaussian noise, the vector and other similar vectors are equal to zero, the a priori error covariance can be transformed as
where matrix is specialized via its th component
Reasoning similarly, the a posteriori error covariance can be transformed using extended nonlinearities if we first represent it as
take into account that and , provide the averaging, and obtain
Due to the symmetry of the matrices , , and , the following expectations can be transformed as
Then, taking into account that the expectations of other products are matrices with zero components, we finally write the covariance in the form
where, the th component of matrix is defined by
and the th component of matrix is
It has to be remarked now that when developing the second‐order extended algorithms, the authors use two forms of . The complete form in (6.170) can be found in [12,157]. On the contrary, in [72,185] only first‐order components are preserved.
The pseudocode of the EFIR‐2 filtering algorithm is listed as Algorithm 16. For the given , , and , the set of auxiliary matrices is computed and updated at each . Then all matrices and vectors are updated iteratively, and the last updated and go to the output when . Although the second‐order EFIR‐2 algorithm has been designed to improve accuracy, empirical studies show that its usefulness is questionable due to the following drawbacks:
Overall, it can be said that the EFIR‐2 and EKF‐2 filtering algorithms do not demonstrate essential advantages over the EFIR‐1 and EKF‐1 algorithms. Moreover, these algorithms are more computationally complex and slower in operation.
Practice dictates that having a good estimator is not enough if the operating conditions are not satisfied from various perspectives, especially in unknown environments. In such a case, the required estimator must also be robust. In Chapter, we introduced the fundamentals of robustness, which is seen as the ability of an estimator not to respond to undesirable factors such as errors in noise statistics, model errors, and temporary uncertainties. We are now interested in investigating the trade‐off in robustness between the UFIR estimator and some other available state estimators. To obtain reliable results, we will examine along the KF, which is optimal and not robust, and the robust game theory recursive filter (3.307)–(3.310), which was developed in [185]. We will use the following state‐space model
where the noise covariances and in the noise vectors and are not necessarily known, as required by KF. To apply the filter, and will be thought of as norm‐bounded with symmetric and positive definite matrices and .
In our scenario, there is not enough information about and to optimally tune KF, and the maximum values of the matrices and are also not known to make the filter robust. Therefore, we organize a scaling test for robustness by representing any uncertain component in (6.173) and (6.174) as , where the increment is due to a positive‐valued scaling factor such that means undisturbed . This test can be applied with the following substitutions: , , , , , and , where the set of scaling factors can either change matrices or not when . For , the product matrix becomes
We can learn about the robustness of recursive estimators by examining the sensitivity of the bias correction gain to interfering factors and its immunity to these factors. Taken from the UFIR Algorithm 11, the alternate KF (3.92)–(3.96) [184], and the filter (3.307)–(3.310) [185], the bias corrections gains can be written as, respectively,
where , , and matrix is constrained by the positive definite matrix . Note that the tuning factor is required by the game theory filter in order to outperform KF.
The influence of on (6.176)–(6.178) can be learned if we use the following rule: if the error factor causes a decrease in , then the bias errors will grow; otherwise, random errors will dominate. The increments caused by the error factors in the bias correction gains can now be represented for each of the filters as
and we notice that lower means higher robustness. There is one more important preliminary remark to be made. While more tuning factors make an estimator potentially more accurate and precise, it also makes an estimator less robust due to possible tuning errors (Fig. 1.3). In this sense, the two‐parameter UFIR filter appears to be more robust, but less accurate, while the four‐parameter KF and five‐parameter game theory filter are less robust but potentially more accurate.
The additional errors produced by estimators applied to stochastic models and caused by and occur due to the practical inability to collect accurate estimates of noise covariances and norm‐bounded weights, especially for LTV systems.
The UFIR filter does not require any information about zero mean noise. Thus, this filter is insensitive (robust) to variations in and , unlike the KF and game theory filter.
For and , errors in the KF can be learned in stationary mode if we allow . Accordingly, the error covariance can be represented using the DARE
The solution to (6.179) does not exist in simple form. But if we accept , which is true for low measurement noise, , then (6.179) can be replaced by the discrete Lyapunov equation [184] (A.34)
which is solvable (A.35). Although approximation (6.180) can be rough when is not small, it helps to compare the KF and filter for robustness, as will be shown next.
Lemma 6.1 Given and , where , is the spectral radius of , and , the solution to (6.180) can be represented for and with negative definiteness in the form [180]
Proof Write the solution to (6.180) as an infinite sum [83],
and transform the ‐norm of (6.182) as
Choose the gain factor so that , go to (6.181), and complete the proof.
When , we can assume that KF is tracking state with high precision and let that gives . This gives and , and we conclude that . Note that this relationship is consistent with the equality in (6.183), which was obtained for the isolated cases of and .
Referring to lemma 6.1, the bias correction gain (6.177) can now be approximated for by the inequality
which does not hold true for . It is worth noting that and act in opposite directions, and therefore the worst case will be when and or vice versa.
Reasoning similarly for equal weighting errors obtained with , the bias correction gain (6.178) for the recursive game theory filter can be approximated by the inequality [180]
It can be seen that makes no difference with the KF gain (6.184) for Gaussian models implying and . Otherwise, we can expect KF to be less accurate and the filter to perform better, provided that both and are properly maximized. Indeed, if we assume that and , then we can try to find some small such that the effect of and is compensated. However, if is not set properly, the filter can cause even more errors than the KF. Moreover, even if is set properly, any deviation of from unity will additionally chance the gain through the product of . More information and examples on this topic can be found in [180].
What follows behind the approximations (6.184) and (6.185) is that, if properly tuned with , the filter will be more accurate than the KF in the presence of errors in noise covariances. In practice, however, it can be difficult to find and set the correct , especially for LTV systems that imply a new at each point in time. If so, then the filter may give more errors and even diverge. Thus, the relationship between the filters in terms of robustness to errors in noise statistics should be practically as follows:
where holds true for properly maximized and and properly set . Otherwise, the KF may perform better. In turn, the UFIR filter that is ‐invariant may be a better choice, especially with large errors in noise covariances.
Permanent estimation errors arise from incorrect modeling and persistent model uncertainties, known as mismodeling, when the model is not fully consistent with the process over all time [69]. If no improvement is expected in the model, then the best choice may be to use robust estimators, in which case testing becomes an important subject of confirmation of the estimator accuracy and stability.
In UFIR filtering, errors caused by are taken into account only on the horizon . Accordingly, the bias correction gain can be expressed as
and approximated using the following lemma.
Lemma 6.2 Given model (6.173) and (6.174) with and , where , is the spectral radius of and . Then the GNPG (6.16c) of the UFIR filter is represented at with
where and can be singular.
Proof Let and . Represent on with the finite sum as
Apply the triangle inequality to the norms, use the properties of the norms, and transform the ‐norm of (6.189) as
Now transform (6.190) using the geometrical sum if , replace the result by if , arrive at (6.188), and complete the proof.
The gain can finally be approximated with
It is seen that the factor affects directly and as a reciprocal. However, the same cannot be said about the factor raised to the power . Indeed, for , the gain can significantly grow. Otherwise, when , it decreases. Thus, we conclude that the effect of errors in the system matrix on the bias correction gain of the UFIR filter is more complex than the effect of errors in the observation matrix.
When examining Kalman filter errors associated with mismodeling, and , it should be kept in mind that such errors are usually observed along with errors in noise covariances. If we ignore errors in the description of noise, the picture will be incomplete due to multiplicative effects. Therefore, we assume that and , and analyze the errors similarly to the UFIR filter. To do this, we start with the discrete Lyapunov equation 6.180, the solution of which can be approximated using the following lemmas.
Lemma 6.3 Given , , and , where , , and is the spectral radius of . Then the prior error covariance of the KF can be approximated for with
Proof The proof is postponed to “Problems” and can be provided similarly to lemma 6.1.
Lemma 6.4 Given , , and , where , , and is the spectral radius of . Then the prior error covariance of the KF can be represented for with
Proof The proof can be found in [180] and is postponed to “Problems.”
Referring to lemma 6.3 and lemma 6.4, the bias correction gain of the KF, which is given by (6.177), can be approximated as
and we notice that inequality (6.196) is inapplicable to some models.
Analyzing (6.195), it can be concluded that the mismodeling errors, , and errors in noise covariance, , cause multiplicative effects in the Kalman gain. In the best case, these effects can compensate for each another, although not completely. In the worst case, which is of practical interest, errors can grow essentially. Again, we see that and act in opposite directions, as in the case of .
The game theory filter obtained in [185] does not take into account errors in the system and observation matrices. Therefore, it is interesting to know its robustness to such errors in comparison with the UFIR filter and KF. Referring to (6.195) and (6.196), the bias correction gain of the filter, which is given by (6.178), can be approximated for and by
These inequalities more clearly demonstrate the key advantage of the filter: a properly set can dramatically improve the estimation accuracy. Indeed, if the tuning factor constrained by is chosen such that the relationships in parentheses become zero, then it follows that the effects of are reduced to zero, and errors in the observation matrix will remain the main source. Furthermore, by adjusting , the effect of can be fully compensated.
Unfortunately, the previous remedy is effective only if the correct is available at any given time. Otherwise, the filter may produce more errors than the KF, and can even diverge. For more information, the reader can read [180], where a complete analysis for robustness is presented with typical examples.
The general conclusion that can be drawn from the analysis of the impact of mismodeling errors is that, under and , the KF performance can degrade dramatically and that the can improve it if we set a proper tuning factor at each time instant. Since this is hardly possible for many practical applications, the UFIR filter is still preferred, and the relationship between the filters in terms of robustness to model errors will generally be the same (6.186), where holds true for properly maximized errors in and if and are properly maximized and is set properly. Otherwise, the filter may give large errors and go to divergence, while the KF may perform better.
Certainly persistent model errors can seriously degrade the estimator performance, causing uncontrollable bias errors. But their effect can be mitigated by improving the model, at least theoretically. What cannot be done efficiently is the elimination of errors caused by temporary uncertainties such as jumps in phase, frequency, and velocity. Such errors are not easy to deal with due to the unpredictability and short duration of the impacts. The problem is complicated by the fact that temporary uncertainties exist in different forms, and it is hard to imagine that they have a universal model.
One way to investigate the associated estimation errors is to assume that the model has no past errors up to the time index and that the uncertainty caused by occurs at . The estimator that is less sensitive to such an impact caused by at can be said to be the most robust.
Suppose the model is affected by over all time and that is known at . At the next step, at , the system experiences an unpredictable impact caused by and . Provided that is also known at , the bias correction gain of the UFIR filter (6.176) can be transformed to
As can be seen, and act in the same way in (6.199) as in the case of model errors represented by (6.192). However, the effect in (6.199) turns out to be stronger because of the square of , in contrast to (6.192). This means that temporary data errors have a stronger effect on the UFIR filter than permanent model errors. The conclusion just made should not be unexpected, given the fact that rapid changes cause transients, the values of which usually exceed steady‐state errors.
The Kalman gain can be approximated similarly to the bias correction gain of the UFIR filter. Substituting the prior estimate into (6.177), where the matrices are scaled by the error factors and taking into account that , we obtain the approximation of the Kalman gain in the form
Analyzing (6.199) and (6.200), we draw attention to an important feature. Namely, the effect of the first is partially compensated by the second in the parentheses, which is not observed in the UFIR filter. Because of this, the KF responds to temporary uncertainty with lower values than the UFIR filter. In other words, the UFIR filter is more sensitive to temporary uncertainties. On the other hand, all transients in the UFIR filter are limited by points, while the KF filter generates long transients due to IIR. Both these effects, which are always seen in practice, represent one of the fundamental differences between the UFIR filter and the KF [179].
Transients in the UFIR filter and KF: Responding to the stepwise temporary uncertainty, the UFIR filter produces points transient with a larger overshoot, and the KF produces a longer transient with a shorter overshoot.
Examples of transients in the UFIR filter and KF are given in Fig. 4.2 and Fig. 6.3, and we notice that the difference between the overshoots is usually .
Providing similar transformations as for (6.200), we approximate the bias correction gain of the filter (6.178) as
and notice an important difference with (6.200). In (6.201), there is an additional term containing , which can be chosen such that the effect of is fully compensated. This is considered to be the main advantage of the game theory filter over KF. The drawback is that the tuning factor is an unknown analytical function of the set , which contains uncertain components. Therefore, if cannot be set properly at every point of time, the filter performance may significantly degrade, and the filter may even demonstrate the divergence [180]. Analyzing (6.201), it can also be noted that the problem is complicated by the higher sensitivity of to .
All that follows from the previous analysis of estimation errors caused by temporary uncertainties is that the UFIR filter can still be considered the most robust solution. The game theory filter can improve the performance of KF by properly setting the tuning factor . However, since the exact is not available at every point of time, this filter can become unstable and diverge.
Hardware implementation of digital filters is usually not required in state space. In contrast, scalar input‐to‐output structures have found much more applications. Researchers often prefer to use UFIR polynomial structures, whenever suboptimal filtering is required. Examples can be found in tracking, control, positioning, timekeeping, etc. If such a UFIR filter matches the model, then the group delay reaches a minimum. Otherwise, the group delay grows at a lower rate than in IIR filters. This makes UFIR structures very useful in engineering practice of suboptimal filtering. Next we will present and analyze the polynomial UFIR structures in the domain [29] and discrete Fourier transform (DFT) domain [30].
The transfer function of the th degree polynomial UFIR filter is specialized by the ‐transform applied to the FIR function given by (6.120) with ; that is,
where , is an angular frequency, is the sampling time, is given by (6.121), and conventionally we will also use in as the imaginary sign . The following properties of can be listed in addition to the inherent ‐periodicity, symmetry of , and antisymmetry of .
Because is always positive‐valued, , and for all , the counterclockwise circular integration in (6.204a) always produces a positive imaginary value, which gives
and it follows that the energy (or squared norm) of the FIR function in the domain is equal to the value of at .
Theorem 6.2 Given a digital FIR filter with an th degree transfer function and an input polynomial signal specified on . Then the FIR filter will be unbiased if it satisfies the following unbiasedness condition in the domain:
Proof The proof appears instantly when compared to (6.204b) and (6.207c).
It is worth noting that using to the properties listed, it becomes possible to design and optimize polynomial FIR filter structures in the domain with maximum efficiency for real‐time operation.
Although the inner sum in (6.202b) does not have a closed form, the properties of the FIR filter in the ‐domain allow one to represent as
By assigning the following subtransfer functions
the generalized block diagram of the th degree UFIR filter can be shown as in Fig. 6.10. The low‐degree coefficients , , and are listed for this structure in Table 6.3 [29].
Table 6.3 Coefficients , , and of Low‐Degree UFIR Filters.
0 | 1 | 2 | 3 | |
---|---|---|---|---|
0 | − | − | − | |
0 | 0 | |||
0 | 0 | 0 | − | |
− | − | |||
0 | − | − | ||
0 | 0 | − | ||
0 | 0 | 0 | − | |
−1 | −2 | −3 | −4 | |
0 | 1 | 3 | 6 | |
0 | 0 | −1 | −4 | |
0 | 0 | 0 | 1 |
Example 6.5 Transfer function of the first‐degree UFIR filter in domain [29]. If we use the coefficients given in Table 6.3, then the transfer function of the first‐degree polynomial UFIR filter having a ramp impulse response function comes from (6.210) as
Simple analysis shows that the region of convergence (ROC) for (6.213) is all and that the filter is both stable and causal. A block diagram corresponding to (6.213) is exhibited in Fig. 6.11, and it follows that the filter utilizes six multipliers, four adders, and three time‐delay subfilters. It can be shown that this structure can be further optimized to have three multipliers, five adders, and three time‐delay subfilters, as shown in [27]. Note that block diagrams of other low‐degree polynomial UFIR filters can be found in [29].
The magnitude response functions of low‐degree polynomial UFIR filters are shown in Fig. 6.12. A distinctive feature of LP filters of this type is a negative slop of 20 dB per decade in the Bode plot (Fig. 6.12b) in the transient region. Another feature is that increasing the degree of the polynomial filter expands the bandwidth. It can also be seen that the transfer function has multiple intensive side lobes that are related to the shape of the impulse response (Fig. 6.5). The later property is discouraged in the design of standard LP filters. However, it is precisely this shape of that guarantees the filter unbiasedness.
Figure 6.12a assures that bias is eliminated at the filter output by shifting and lifting the side lobes of a zero‐degree uniform FIR filter, which provides simple averaging. Accordingly, the ‐degree filter passes the spectral content close to zero without change, magnifies the components falling into the first lobe, and attenuates the power of the higher‐frequency components with 10 dB per decade (Fig. 6.12b).
The phase response functions of low‐degree polynomial UFIR filters are shown in Fig. 6.13a, and the group delay functions obtained by are shown in Fig. 6.13b. The phase response function of this filter is linear on average (Fig. 6.13a). However, the phase response changes following variations in the magnitude response, which, in turn, leads to changes in the group delay around a small constant value (Fig. 6.13b). From the standpoint of the design of basic LP filters, the presence of periodic variations in the phase response is definitely a disadvantage. But it is also a key condition that cannot be violated without making the filter output biased.
The DFT of the th degree UFIR filter impulse response is obtained as
where . In addition to the inherent ‐periodicity, the symmetry of , and the antisymmetry of , the following properties are important for the implementation of UFIR filters in the DFT domain.
Since holds for all , the sum of the DFT coefficients is real and positive. It then follows that
Theorem 6.3 Given a digital FIR filter with a th degree transfer function and an input polynomial signal specified on , then the FIR filter will be unbiased if the following unbiasedness condition is satisfied in the DFT domain:
Proof. To prove (6.220), use the following fundamental condition for optimal filtering: the order of the optimal and/or unbiased filter must be the same as that of the system. Then represent by (6.124),
where the coefficient is given by (6.125). Recall that function (6.124) has the following main properties, given in Table 6.2: the sum of its coefficients is equal to unity, and the moments zero; that is,
Use Parseval's theorem and obtain the following relationships,
Now observe that the IDFT applied to at gives that finally leads to (6.220) and completes the proof.
Note that the previously discussed properties of polynomial UFIR filters in the DFT domain are preserved for any polynomial degree, but in practice only low‐degree digital UFIR filters, , are usually found. The reason for using low‐degree filters is explained earlier: increasing the degree leads to larger random errors at the filter output.
Using the properties listed, the transfer function of the th degree UFIR filter can be represented in the DFT domain with the sum
in which the inner sum has no closed‐form solution for an arbitrary . However, solutions can be found for low‐degree polynomials as shown in [30]. Next we will illustrate such a solution for a first‐degree UFIR filter and postpone to “Problems” the search for solutions associated with higher degrees.
Example 6.6 First‐degree UFIR filter in DFT domain [30]. The first‐degree (ramp) UFIR filter is associated with signals, which are linear on . The impulse response of this filter that is specialized with the coefficients and has the DFT
where . At and , the function becomes unity. Note that for , can be found in [30].
The magnitude response , Bode plot, and phase response functions of the th degree, , polynomial UFIR filter are shown in Fig. 6.14. For the 1st degree, the functions are computed by (6.226b) and for the next two low degrees the transfer functions can be found in [30]. In addition to the inherent properties of periodicity with a repetition period of points, symmetry of , and antisymmetry of , the transfer function has several other features that are of practical importance.
Finally, we come to an important practical conclusion. Since the transfer function of the polynomial UFIR filter is monotonic and does not contain the periodic variations observed in the ‐domain (Fig. 6.13), it follows that the filter can be easily implemented with all of the advantages of suboptimal unbiased filtering.
By abandoning the requirement for initial values and ignoring zero mean noise, the UFIR state estimator appears to be the most robust among other linear estimators such as OFIR, OUFIR, and ML state estimators. Moreover, its iterative algorithm is more robust than the recursive KF algorithm. The only tuning factor required for the UFIR filter is the optimal horizon, which can be determined much more easily than noise statistics. Furthermore, at given averaging horizons, the UFIR state estimator becomes blind, which is very much appreciated in practice. In general, it is only within a narrow range of error factors caused by various types of uncertainties that optimal and ML estimators are superior to the UFIR estimator. For the most part, this explains the better performance of the UFIR state estimator in many real‐world applications.
Like other FIR filters, the UFIR filter operates on the averaging horizon of points, from to . Its discrete convolution‐based batch resembles the LS estimator. But, the latter is not a state estimator. The main performance characteristic of a scalar input‐to‐output UFIR filter is NPG. If the UFIR filter is designed to work in state space, then its noise reduction properties are characterized by GNPG.
The recursive forms used to iteratively compute the batch UFIR estimate are not Kalman recursions. An important feature is that UFIR filter recursions serve any zero mean noise, while Kalman recursions are optimal only for white Gaussian noise. It is worth noting that the error covariance of the UFIR and Kalman filters are represented by the same Riccati equation. The difference lies in the different bias correction gains. The optimal bias correction gain of the KF is called the Kalman gain , while the bias correction gain of the UFIR filter is computed in terms of GNPG as .
Since the UFIR filter does not involve noise covariances to the algorithm, it minimizes the MSE on the optimal horizon of points (Fig. 6.1). The significance of is that random errors increase if the horizon length is chosen such that . Otherwise, bias errors grow if . Using (6.49), the optimal horizon can be estimated through the measurement residual without using ground truth. In response to stepwise temporary uncertainties, the UFIR filter generates finite time transients at points with larger overshoots, while the KF generates longer transients with lower overshoots. This property represents the main difference between the transients in both filters.
It turns out that, since zero mean noise is ignored by UFIR state estimators, the FFFM, FFBM, BFFM, and BFBM ‐lag UFIR smoothing algorithms are equivalent. It should also be kept in mind that all unbiased, optimal, and optimal unbiased state estimators are essentially LP filters. Moreover, the zero‐degree (simple average) UFIR filter is the best in the sense of the minimum produces noise. This is because an increase in the filter degree or the number of the states leads to an increase in random errors at the estimator output. It also follows that the set of degree polynomial impulse responses of the UFIR filter establishes a class of discrete orthogonal polynomials that are suitable for signal analysis and restoration due to the built‐in unbiasedness.
Although the UFIR filter ignores zero mean noise, it is not suitable for colored noise that is biased on short horizons. If the colored noise is well‐approximated by the Gauss‐Markov process, then the general UFIR filter becomes universal for time‐correlated and uncorrelated driving white noise sources. Like GKF, a general UFIR filter can be designed to work with CMN and CPN. It can also be applied to state‐space models with smooth nonlinearities using the first‐ or second‐order Taylor series expansions. However, it turns out that the second‐order EFIR‐2 filter has no practical advantage over the first‐order EFIR‐1 filter.
The implementation of polynomial UFIR filters can be most efficiently obtained in the DFT domain due to the following critical property: the DFT transfer function is smooth and does not exhibit periodic variations inherent to the ‐domain.
Since holds for , write as
substitute to the previous relation, and arrive at the recursion (6.21) for .
Similarly, represent the product as
combine with , obtain , substitute , and come up with
From (6.21) find , substitute into the previous relation, and arrive at the recursion (6.29) for the UFIR estimate.
Compare this relation to the error covariance of the KF and explain the difference between the bias correction gain of the UFIR filter and the Kalman gain . Why can the Kalman gain not be less than the optimal value? Why does the bias correction gain of the UFIR decrease with increasing the horizon length?
Why does the OFIR filter not need an optimal horizon? Why does the error in the OFIR filter decrease with increasing the horizon length, while the UFIR filter minimizes the MSE by only ? Give simple and intuitive explanations.
where , , is a deterministic weight, which can be either unity or zero. The time‐stamped data indicate a value of for which the weight is unity and the other weights are zero. Transform the model to another without delay and develop a UFIR filter.
Give an explanation to this fact and discuss the problems that may arise in the practical implementations of this method.
where and , , and are state variables. Write the state equation with additive white Gaussian noise.
Represent GNPG as and specify .
Reasoning similarly as for the forward UFIR filter, provide the derivation of (6.72). Analyze this relationship and find out what will happen to if the noise covariance is set equal to zero.
Find a recursive form for (6.92) or, otherwise, develop a different approach for recursive computation of (6.85).
Show other possible ways to obtain recursions for (6.90) and (6.93).
where is a constant angular measure, and and are zero mean noise sources. Using the derivation of the UFIR filter for polynomial models, derive the scalar FIR functions for the first state and the second state to satisfy the unbiasedness condition .
where and are given constants. Using the proof of lemma 6.1, give a corresponding proof of the lemma 6.3.
Obtain the corresponding constraint for the general UFIR filter matrix gain case .
Obtain the corresponding constraint for the UFIR filter matrix gain .