# Tag Info

53

1. There is a Difference in terms of optimality criteria Kalman filter is a Linear estimator. It is a linear optimal estimator - i.e. infers model parameters of interest from indirect, inaccurate and uncertain observations. But optimal in what sense? If all noise is Gaussian, the Kalman filter minimizes the mean square error of the estimated parameters. ...

16

I agree that the windowing filter design method is not one of the most important design methods anymore, and it might indeed be the case that it is overrepresented in traditional textbooks, probably due to historical reasons. However, I think that its use can be justified in certain situations. I do not agree that computational complexity is no issue ...

7

You can think of linear-least squares in single dimension. The cost function is something like $a^{2}$. The first derivative (Jacobian) is then $2a$, hence linear in $a$. The second derivative (Hessian) is $2$ - a constant. Since the second derivative is positive, you are dealing with convex cost function. This is eqivalent to positive definite Hessian ...

7

The LMS algorithm is based on the idea of gradient descent to search for the optimal (minimum error) condition, with a cost function equal to the mean squared error at the filter output. However, it doesn't actually calculate the gradient directly, as that would require knowing: $$E(\mathbf{x}[n]e^*[n])$$ where $\mathbf{x}[n]$ is the input vector and $e[... 7 Windowed Sinc filters can be adaptively generated on the fly on processors barely powerful enough to run the associated FIR filter. Windowed Sinc filters can be generated in finite bounded time. The generation of some simple windowed Sinc filters can be completely described (and inspected for malware, etc.) in a few lines of code, versus blind use of some ... 7 Slope from all samples obtained To summarize the question's problem, you want to calculate the slope based on all samples obtained thus far, and as new samples are obtained, update the slope without going through all the samples again. On the page you cite is the equation for calculation of the slope$m_n$that together with$b_n$minimizes the sum of ... 6 The difference is quite huge, since they are two completely different models which can be used to tackle the same problem. Let's do a quick recap. Polynomial regression is a way of function approximation. We have a data set of the form$\lbrace x_i, z_i \rbrace$and wish to determine the functional relationship, which is often expressed by estimating the ... 6 That really depends on context, but generally adaptive implies that the calculations are done on-line / on the fly. In some applications, the filter is updated for a while, then the adaptation is turned off and the last lot of weights are used. 6 To expand on what Peter K. has said, if the signals being used by the filter are stationary, then the filter weights or coefficients can be determined and the filter operates as it was designed without further updates to the filter weights. However, if the signals change, or become quasi-stationary, the filter will adapt continuously. 6 I'll show here one benefit of a windowed design and a trick to get the same benefit from Parks–McClellan. For half-band, quarter-band etc. filters windowing retains the time-domain zeros of the scaled sinc function, which is the prototypical ideal low-pass filter. The zeros end up in the coefficients, reducing the computational cost of the filters. For a ... 5 While I completely agree with Jason R's answer, I would like to add a few things that I consider important. First of all, it is a misunderstanding to believe that for least squares designs the transition band width is smaller than for equi-ripple designs. The width of the transition band depends on many design parameters but it is independent of the ... 5 I'm unsure if this is the answer you are looking for, but why not save and share$|H_i|^2$in addition to the least squares estimates? If you have$w_1^*=\frac 1 {|H_1|^2} H_1^t d_1$,$w_2^*=\frac 1 {|H_2|^2} H_2^t d_2$, and also know$|H_1|^2$and$|H_2|^2$, you get the total least squares estimate with:$w^*=\frac 1 {|H_1|^2+|H_2|^2} (H_1^t d_1 + H_2^t ...

5

Not an expert on kalman filters, however I believe traditional Kalman filtering presumes a linear relationship between the observable data, and data you wish to infer, in contrast to more intricate ones like the Extended Kalman filters that can assume non-linear relationships. With that in mind, I believe that for a traditional Kalman filter, on-line linear ...

5

For anyone who is interested, i coincidentally found a paper describing the method implemented in matlab's filtfilt.m. A link to the paper is attached. At least to my understanding matlab's filtfilt.m doesn't implement the Gustafson algorithm. Sadovsky, P.; Bartusek, K: Optimisation of the Transient Response of a Digital Filter, Radioengineering Vol. 9, No. ...

5

The Jacobian is not computed numerically but analytically and then just evaluated. The frequency response of the IIR filter is $$H(e^{j\omega})=\frac{b_0+b_1e^{-j\omega}+\ldots+b_Me^{-jM\omega}}{1+a_1e^{-j\omega}+\ldots+a_Ne^{-jN\omega}}=\frac{B(e^{j\omega})}{A(e^{j\omega})}\tag{1}$$ Now you need the derivative with respect to the filter coefficients: $$\... 4 First of all, the minimum norm least square solution is A^+b, where A^+ is the pseudoinverse. Only when the left inverse A_L^{-1} (or right inverse A_R^{-1}) exists, you have A^+=A_L^{-1}=(A^TA)^{-1}A^T (or A^+=A_R^{-1}=A^T(AA^T)^{-1}). Likewise, the projection matrix onto the column space is P=AA^+, or P=AA_R^{-1}=AA^T(AA^T)^{-1}=I if A_R^{... 4 Kalman filtering gives multiple predictions for the next state, where an extrapolation of a regression would not. The Kalman filters are also focused on including noise factors (based on Gaussian distributions). 4 Here's the way I think about a discrete Wiener Filter Consider a sequence of observations \mathbf{y} \in \Re^n  Form a matrix from the input \mathbf{x} \in \Re^{n+r-1} by shifting columns one sample each:$$ X= \begin{bmatrix} x_1 & x_2 & ... & x_r \\ x_2 & x_3 & & x_{r+1} \\ x_3 & x_4 & & x_{r+2} \\ ... & &...

4

Kalman filter is the best linear estimator regardless of stationarity or Gaussianity. Also in the Gaussian case it does not require stationarity (unlike Wiener filter). In the linear Gaussian case Kalman filter is also a MMSE estimator or the conditional mean.

3

One main difference is the cost function used in the two design methods: Equiripple filters seek to minimize the maximum error between the desired filter response and the designed approximation. Least-squares filters seek to minimize the total squared error betwen the desired filter response and the designed approximation. These different strategies lead ...

3

If you want to solve for single value of $\lambda$ in the model: $$\arg \min_{x} \frac{1}{2} {\left\| A x - b \right\|}_{2}^{2} + \lambda {\left\| x \right\|}_{1}$$ Then you can use Coordinate Descent method which is the fastest and simplest and doesn't require any matrix inversion. I have a MATLAB code for in my ${L}_{1}$ Regularized Least Squares ...

3

The least-squares re-synthesis procedure is very similar to the overlap-add (OLA) procedure. Let w_n be the discrete window vector and y_n(:,k) be the time domain vector computed using the IFFT on a column k of the 2D matrix. Both w_n and y_n(:,k) are the same length. Then, using Matlab syntax, we compute the point-by-point multiplication with the window: ...

3

A lot has been said already, allow me to add some comments: Kalman filters are an application of Bayesian probability theory, which means that "a priori information" or "prior uncertainty" can (and must) be specified. As I understand, this is not the case with traditional least-squares fitting. While observations (data) can be weighted with probabilities in ...

3

From here: $$(AB)(B^{-1}A^{-1}) = A (BB^{-1}) A^{-1} = A A^{-1} = I$$ So $(AB)^{-1} = B^{-1} A^{-1}$ provided $A$ and $B$ are invertible.

3

I don't understand the subscript $n$ notation, however, in the least squares problem that is given by: $${\bf{y}}={\bf{H}}{\theta}+\bf{n},$$ where ${\bf{n}}\sim\mathcal{N}(\bf{0}, \sigma^2I_N)$ is a zero mean additive white Gaussian noise and $I_N$ is the $N \times N$ identity matrix, the maximum likelihood and the least squares ...

3

The idea is to represent all operation sing Matrices. Once it is done, it is easy to solve the problems as a Least Squares problems. The way to represent Convolution Operation using a Matrix is by Toeplitz Matrix. For 1D it is pretty straight forward to do (Just pay attention to boundary). So let's take the simple model in the comment: $$g = f \circ h + ... 3 Since this is a linear model if you add noise which isn't centered (Non zero mean noise) your estimation will be good up to a bias term. The easy way to do so is to remove the bias from  y  and solve. In case you know the mean of the added noise, just remove it from your measurement and your model becomes the classic Linear Least Squares. If you don't ... 3 Our system has the impulse response (why we changed to imaginary instead of real? :) ):$$h(t)=\mathbb{Im}(Ae(iwt)e(-bt))=Ae^{-bt}sin(wt)$$With the following structure (ref.) (discretized):$$H(s)=\frac{Kw}{(s+b)^2+w^2}$$Under matlab, you only need to use procest: procest(data,'P2U'); Which models the same structure under the form:$$H(s)=\frac{K}{(1+2\...

3

The equation you're trying to solve is $$\mathbf{y}=\mathbf{X}\mathbf{h},$$ where $\mathbf{h}$ is your unknown. The matrix $\mathbf{X}$ is going to have a time-shifted structure that reflects the convolution operator. If we assume that the $\mathbf{y}$ vector starts with y(3) i.e. ignores the first two zeroed out elements of y, then the corresponding \$\...

Only top voted, non community-wiki answers of a minimum length are eligible