Optimum Filters

The discrete form of the Wiener filtering problem, is to design a filter to recover a signal $d(n)$ from noisy observations
$$x(n)=d(n)+v(n)$$
Assuming that both $d(n)$ and $v(n)$ are wide-sense stationary random processes, Wiener considered the problem of designing the filter that would produce the minimum mean square error estimate of $d(n)$. Thus, with
$$\xi =E\left\{|e(n){|}^2\right\}=E\left\{|d(n)-\hat{d}(n){|}^2\right\}$$
the problem is to find the filter that minimizes $\xi$.

Depending upon how the signals $x(n)$ and $d(n)$ are related to each other, a number of different and important problems may be cast into a Wiener filtering framework. Some of the problems that will be considered in this chapter include:

• Filtering: estimate $d(n)$ from $x(n)=d(n)+v(n)$.
• Prediction: estimate $x(n+\alpha )$ from $x(n),x(n-1),x(n-2),\dots$
• Deconvolution: estimate $d(n)$ from $x(n)=g(n)\ast d(n)+v(n)$
• Noise cancellation: estimate $v_1(n)$ from $v_2(n)$ and subtract it from $x(n)=d(n)+v_1(n)$

The FIR Wiener Filter

In this section we consider the design of an FIR Wiener filter that produces the minimum mean-square estimate of a given process $d(n)$ by filtering a set of observations of a statistically related process $x(n)$.

Note that we didn’t introduce any assumption about $x(n)$, it can be any process statistically related to $d(n)$.

It is assumed that $x(n)$ and $d(n)$ are jointly wide-sense stationary with known autocorrelations, $r_x(k)$ and $r_d(k)$, and known cross-correlation $r_{dx}(k)$. Denoting the unit sample response of the Wiener filter by $w(n)$, and assuming a $(p-l)$st-order filter, the system function is
$$W(z)=\sum _{n=0}^{p-1}w(n)z^{-n}$$
With $x(n)$ the input to the filter, the output, which we denote by $\hat{d}(n)$, is the convolution of $w(n)$ with $x(n)$,
$$\hat{d}(n)=\sum _{l=0}^{p-1}w(l)x(n-l)\text{\hspace{1em}(FWF.1)}$$
The Wiener filter design problem requires that we find the filter coefficients, $w(k)$, that minimize the mean-square error
$$\xi =E\left\{|e(n){|}^2\right\}=E\left\{|d(n)-\hat{d}(n){|}^2\right\}\text{\hspace{1em}(FWF.2)}$$
In order for a set of filter coefficients to minimize $\xi$ it is necessary and sufficient that the derivative of $\xi$ with respect to $w^{\ast }(k)$ be equal to zero for $k=0,1,\cdots ,p-1$,
∂ ξ ∂ w ∗ ( k ) = ∂ ∂ w ∗ ( k ) E { e ( n ) e ∗ ( n ) } = E { e ( n ) ∂ e ∗ ( n ) ∂ w ∗ ( k ) } = 0 (FWF.3) \frac{\partial \xi}{\partial w^*(k)}=\frac{\partial}{\partial w^*(k)}E\{e(n)e^*(n )\}=E\left\{e(n)\frac{\partial e^*(n)}{\partial w^*(k)}\right\}=0 \tag{FWF.3} ∂w∗(k)∂ξ​=∂w∗(k)∂​E{e(n)e∗(n)}=E{e(n)∂w∗(k)∂e∗(n)​}=0(FWF.3)
With
$$e(n)=d(n)-\sum _{l=0}^{p-1}w(l)x(n-l)\text{\hspace{1em}(FWF.4)}$$
it follows that
$$\frac{\partial e^{\ast }(n)}{\partial w^{\ast }(k)}=-x^{\ast }(n-k)$$
and $(FWF.3)$ becomes
E { e ( n ) x ∗ ( n − k ) } = 0 ; k = 0 , 1 , ⋯   , p − 1 (FWF.5) E\{e(n)x^*(n-k) \}=0;\quad k=0,1,\cdots,p-1\tag{FWF.5} E{e(n)x∗(n−k)}=0;k=0,1,⋯,p−1(FWF.5)
which is known as the orthogonality principle or the projection theorem.

Substituting $(FWF.4)$ into $(FWF.5)$ we have
E { d ( n ) x ∗ ( n − k ) } − ∑ l = 0 p − 1 w ( l ) E { x ( n − l ) x ∗ ( n − k ) } = 0 (FWF.6) E\{d(n)x^*(n-k)\}-\sum_{l=0}^{p-1}w(l)E\{x(n-l)x^*(n-k)\}=0\tag{FWF.6} E{d(n)x∗(n−k)}−l=0∑p−1​w(l)E{x(n−l)x∗(n−k)}=0(FWF.6)
Finally, since $x(n)$ and $d(n)$ are jointly WSS then
$$\sum _{l=0}^{p-1}w(l)r_x(k-l)=r_{dx}(k);k=0,1,\cdots ,p-1\text{\hspace{1em}(FWF.7)}$$
In matrix form, using the fact that the autocorrelation sequence is conjugate symmetric, $r_x(k)=r_x^{\ast }(k)$, $(FWF.7)$ becomes
$$\left[\begin{array}{cccc}{r}_x(0)& r_x^{\ast }(1)& \cdots & r_x^{\ast }(p-1)\\ r_x(1)& r_x(0)& \cdots & r_x^{\ast }(p-2)\\ r_x(2)& r_x(1)& \cdots & r_x^{\ast }(p-3)\\ ⋮& ⋮& & ⋮\\ r_x(p-1)& r_x(p-2)& \cdots & r_x(0)\end{array}\right]\left[\begin{array}{c}w(0)\\ w(1)\\ w(2)\\ ⋮\\ w(p-1)\end{array}\right]=\left[\begin{array}{c}{r}_{dx}(0)\\ r_{dx}(1)\\ r_{dx}(2)\\ ⋮\\ r_{dx}(p-1)\end{array}\right]\text{\hspace{1em}(FWF.8)}$$
which is the matrix form of Wiener-Hopf equations. It may be written more concisely as
$${\mathbf{R}}_x\mathbf{w}={\mathbf{r}}_{dx}\text{\hspace{1em}(FWF.9)}$$
where R x = E { x ∗ x T } , x = [ x ( n ) , ⋯   , x ( n − p + 1 ) ] \mathbf R_x=E\{\mathbf x^* \mathbf x^T\},\mathbf x=[x(n),\cdots,x(n-p+1)] Rx​=E{x∗xT},x=[x(n),⋯,x(n−p+1)].

The minimum mean-square error in the estimate of $d(n)$ may be evaluated from $(FWF.2)$ as follows. With
ξ m i n = E { ∣ e ( n ) ∣ 2 } = E { e ( n ) [ d ( n ) − ∑ l = 0 p − 1 w ( l ) x ( n − l ) ] ∗ } = a E { e ( n ) d ∗ ( n ) } = E { [ d ( n ) − ∑ l = 0 p − 1 w ( l ) x ( n − l ) ] d ∗ ( n ) } = r d ( 0 ) − ∑ l = 0 p − 1 w ( l ) r d x ∗ ( l ) (FWF.10) \begin{aligned} \xi_{\mathrm{min}}&=E\left\{|e(n)|^2\right\}=E\left\{e(n)\left[d(n)-\sum_{l=0}^{p-1} w(l)x(n-l) \right]^*\right\}\\ &\stackrel{a}{=}E\{e(n)d^*(n)\}=E\left\{\left[d(n)-\sum_{l=0}^{p-1} w(l)x(n-l) \right]d^*(n)\right\}\\ &=r_d(0)-\sum_{l=0}^{p-1}w(l)r^*_{dx}(l) \end{aligned} \tag{FWF.10} ξmin​​=E{∣e(n)∣2}=E⎩⎨⎧​e(n)[d(n)−l=0∑p−1​w(l)x(n−l)]∗⎭⎬⎫​=aE{e(n)d∗(n)}=E{[d(n)−l=0∑p−1​w(l)x(n−l)]d∗(n)}=rd​(0)−l=0∑p−1​w(l)rdx∗​(l)​(FWF.10)
where $\stackrel{a}{=}$ is due to $(FWF.5)$, i.e., orthogonality theorem. Or using vector notation
$${\xi }_{\mathrm{m}\mathrm{i}\mathrm{n}}=r_d(0)-{\mathbf{r}}_{dx}^H\mathbf{w}=r_d(0)-{\mathbf{r}}_{dx}^H{\mathbf{R}}_x^{-1}{\mathbf{r}}_{dx}\text{\hspace{1em}(FWF.11)}$$

---

Filtering

In the filtering problem, a signal $d(n)$ is to be estimated from a noise corrupted observation
$$x(n)=d(n)+v(n)\text{ or }\mathbf{x}=\mathbf{d}+\mathbf{v}$$
where $\mathbf{x}=[x(n),\cdots ,x(n-p+1)]$, $\mathbf{d}=[d(n),\cdots ,d(n-p+1)]$, $\mathbf{v}=[v(n),\cdots ,v(n-p+1)]$.

It will be assumed that the noise $v(n)$ has zero mean and that it is uncorrelated with $d(n)$. Therefore, E { d ( n ) v ∗ ( n − k ) } = 0 E\{d(n)v^*(n - k)\} = 0 E{d(n)v∗(n−k)}=0 and ${\mathbf{R}}_x$ and ${\mathbf{r}}_{dx}$ becomes
R x = E { x ∗ x T } = E { ( d + v ) ∗ ( d + v ) T } = E { d ∗ d T } + E { v ∗ v T } = R d + R v r d x = E { d ( n ) x ∗ } = E { d ( n ) d ∗ } = r d \begin{aligned} \mathbf R_x&=E\{\mathbf x^* \mathbf x^T\}=E\{(\mathbf d+\mathbf v)^* (\mathbf d+\mathbf v)^T\}=E\{\mathbf d^* \mathbf d^T\}+E\{\mathbf v^* \mathbf v^T\}=\mathbf R_d+\mathbf R_v\\ \mathbf r_{dx}&=E\{d(n)\mathbf x^*\}=E\{d(n)\mathbf d^*\}=\mathbf r_d \end{aligned} Rx​rdx​​=E{x∗xT}=E{(d+v)∗(d+v)T}=E{d∗dT}+E{v∗vT}=Rd​+Rv​=E{d(n)x∗}=E{d(n)d∗}=rd​​
The Wiener-Hopf equations then become
$$({\mathbf{R}}_d+{\mathbf{R}}_v)\mathbf{w}={\mathbf{r}}_d\text{\hspace{1em}(FWF.12)}$$

---

Prediction

Noise-free observations

We consider the following data model ($\alpha$-step prediction)
$$\begin{gathered} d(n)=x(n+\alpha ) \\ \hat{x}(n+\alpha )=\sum _{k=0}^{p-1}w(k)x(n-k) \end{gathered}$$
This results in the following expression for ${\mathbf{r}}_{dx}$:
r d x = E { d ( n ) x ∗ } = E { x ( n + α ) x ∗ } ≜ r α \mathbf r_{dx}=E\{d(n)\mathbf x^*\}=E\{x(n+\alpha)\mathbf x^*\}\triangleq \mathbf r_\alpha rdx​=E{d(n)x∗}=E{x(n+α)x∗}≜rα​
The Wiener-Hopf equations then become
$$\underset{{\mathbf{R}}_x}{\underbrace{\left[\begin{array}{cccc}{r}_x(0)& r_x^{\ast }(1)& \cdots & r_x^{\ast }(p-1)\\ r_x(1)& r_x(0)& \cdots & r_x^{\ast }(p-2)\\ ⋮& ⋮& \ddots & ⋮\\ r_x(p-1)& r_x(p-2)& \cdots & r_x(0)\end{array}\right]}}\underset{\mathbf{w}}{\underbrace{\left[\begin{array}{c}w(0)\\ w(1)\\ ⋮\\ w(p-1)\end{array}\right]}}=\underset{{\mathbf{r}}_{\alpha }}{\underbrace{\left[\begin{array}{c}{r}_x(\alpha )\\ r_x(\alpha +1)\\ ⋮\\ r_x(\alpha +p-1)\end{array}\right]}}\text{\hspace{1em}(FWF.13)}$$
For $\alpha =1$, this is similar to the all-pole modeling using Prony’s method, the autocorrelation method, or the Yule-Walker method (the minus sign simply changes the sign of the coefficients).

Observations with noise

We consider the following data model $(\alpha$ -step prediction)
$$y(n)=x(n)+v(n)\text{ or }\mathbf{y}=\mathbf{x}+\mathbf{v}\text{ and }d(n)=x(n+\alpha )$$
where $\mathbf{x}=[x(n),\cdots ,x(n-p+1)]$, $\mathbf{y}=[y(n),\dots ,y(n-p+1){]}^T$ and $\mathbf{v}=[v(n),\dots ,v(n-p+1){]}^T$
This results in the following expressions for ${\mathbf{R}}_y$ and ${\mathbf{r}}_{dy}$
R y = E { y ∗ y T } = E { ( x + v ) ∗ ( x + v ) T } = E { x ∗ x T } + E { v ∗ v T } = R x + R v r d y = E { d ( n ) y ∗ } = E { x ( n + α ) ( x ∗ + v ∗ ) } = E { x ( n + α ) x ∗ ) = r α \begin{array}{l} \mathbf{R}_{y}=E\left\{\mathbf{y}^* \mathbf{y}^{T}\right\}=E\left\{(\mathbf{x}+\mathbf{v})^* (\mathbf{x}+\mathbf{v})^{T}\right\}=E\left\{\mathbf{x}^* \mathbf{x}^{T}\right\}+E\left\{\mathbf{v}^*\mathbf{v}^{T}\right\}=\mathbf{R}_{x}+\mathbf{R}_{v} \\ \mathbf{r}_{d y}=E\left\{d(n) \mathbf{y}^{*}\right\}=E\left\{x(n+\alpha)\left(\mathbf{x}^{*}+\mathbf{v}^{*}\right)\right\}=E\left\{x(n+\alpha) \mathbf{x}^{*}\right)=\mathbf{r}_{\alpha} \end{array} Ry​=E{y∗yT}=E{(x+v)∗(x+v)T}=E{x∗xT}+E{v∗vT}=Rx​+Rv​rdy​=E{d(n)y∗}=E{x(n+α)(x∗+v∗)}=E{x(n+α)x∗)=rα​​
The Wiener-Hopf equations then become
$$\left({\mathbf{R}}_x+{\mathbf{R}}_v\right)\mathbf{w}={\mathbf{r}}_{\alpha }\text{\hspace{1em}(FWF.14)}$$

---

Deconvolution

We consider a noisy convolutive model, with an FIR filter $g(n)$ of order $L:$
$$x(n)=g(n)\ast d(n)+v(n)\text{ or }\mathbf{x}=\mathbf{G}\mathbf{d}+\mathbf{v}$$
where ${\mathbf{d}}_{(p+L)\times 1}=[d(n),\dots ,d(n-p+1),\dots ,d(n-L-p+1){]}^T,{\mathbf{v}}_{p\times 1}=[v(n),\dots ,v(n-p+1){]}^T$ and
$${\mathrm{G}}_{p\times (p+L)}=\left[\begin{array}{ccccc}g(0)& \cdots & g(L)& \cdots & 0\\ ⋮& \ddots & \ddots & \ddots & ⋮\\ 0& \cdots & g(0)& \cdots & g(L)\end{array}\right].$$
This results in the following expressions for ${\mathbf{R}}_x$ and ${\mathrm{r}}_{dx}$ :
R x = E { x ∗ x T } = G ∗ E { d ∗ d T } G T + E { v ∗ v T } = G ∗ R d G T + R v r d x = E { d ( n ) x ∗ } = G ∗ E { d ( n ) d ∗ } = G ∗ r d \begin{array}{l} \mathbf{R}_{x}=E\left\{\mathbf{x}^{*} \mathbf{x}^{T}\right\}=\mathbf{G}^{*} E\left\{\mathbf{d}^{*} \mathbf{d}^{T}\right\} \mathbf{G}^{T}+E\left\{\mathbf{v}^{*} \mathbf{v}^{T}\right\}=\mathbf{G}^{*} \mathbf{R}_{d} \mathbf{G}^{T}+\mathbf{R}_{v} \\ \mathbf{r}_{d x}=E\left\{d(n) \mathbf{x}^{*}\right\}=\mathbf{G}^{*} E\left\{d(n) \mathbf{d}^{*}\right\}=\mathbf{G}^{*} \mathbf{r}_{d} \end{array} Rx​=E{x∗xT}=G∗E{d∗dT}GT+E{v∗vT}=G∗Rd​GT+Rv​rdx​=E{d(n)x∗}=G∗E{d(n)d∗}=G∗rd​​
The Wiener-Hopf equations then become
$$\left({\mathbf{G}}^{\ast }{\mathbf{R}}_d{\mathbf{G}}^T+{\mathbf{R}}_v\right)\mathbf{w}={\mathbf{G}}^{\ast }{\mathbf{r}}_d\text{\hspace{1em}(FWF.15)}$$

---

Noise cancellation

We consider the same data model as for filtering:
$$x(n)=d(n)+v_1(n)\text{ or }\mathbf{x}=\mathbf{d}+{\mathbf{v}}_1$$
where $\mathrm{d}=[d(n),\dots ,d(n-p+1){]}^T$ and ${\mathrm{v}}_1={\left[v_1(n),\dots ,v_1(n-p+1)\right]}^T$
This time we estimate $v_1(n)$ from a correlated noise source $v_2(n),$ and estimate $d(n)$ as
$$\hat{d}(n)=x(n)-{\hat{v}}_1(n)\text{ with }{\hat{v}}_1(n)={\mathbf{w}}^T{\mathbf{v}}_2$$
where ${\mathbf{v}}_2={\left[v_2(n),\dots ,v_2(n-p+1)\right]}^T$

To estimate $v_1(n)$ from $v_2(n),$ we start from the Wiener-Hopf equations
$${\mathbf{R}}_{v_2}\mathbf{w}={\mathbf{r}}_{v_1{v}_2}$$
since ${\mathbf{r}}_{v1v2}$ is generally not known, we can rewrite this as
r v 1 v 2 = E { v 1 ( n ) v 2 ∗ } = E { ( d ( n ) + v 1 ( n ) ) v 2 ∗ } = E { x ( n ) v 2 ∗ } = r x v 2 \mathbf{r}_{v_{1} v_{2}}=E\left\{v_{1}(n) \mathbf{v}_{2}^{*}\right\}=E\left\{\left(d(n)+v_{1}(n)\right) \mathbf{v}_{2}^{*}\right\}=E\left\{x(n) \mathbf{v}_{2}^{*}\right\}=\mathbf{r}_{x v_{2}} rv1​v2​​=E{v1​(n)v2∗​}=E{(d(n)+v1​(n))v2∗​}=E{x(n)v2∗​}=rxv2​​
and thus the Wiener-Hopf equations can be written as
$${\mathbf{R}}_{v_2}\mathbf{w}={\mathbf{r}}_{x{v}_2}\text{\hspace{1em}(FWF.16)}$$
As already mentioned, $d(n)$ is then estimated as
$$\hat{d}(n)=x(n)-{\hat{v}}_1(n)\text{ with }{\hat{v}}_1(n)={\mathbf{w}}^T{\mathbf{v}}_2$$

Discrete Kalman Filter

In section [The FIR Wiener Filter](# The FIR Wiener Filter) we considered the problem of designing a casual Wiener filter to estimate a process $d(n)$ from a set of noisy observations $x(n)=d(n)+v(n)$. The primary limitation with the solution that was derived is that it requires that $d(n)$ and $x(n)$ be jointly wide-sense stationary processes. Since most processes encountered in practice are nonstationary, this constraint limits the usefulness of the Wiener filter. Therefore, in this section we re-examine this estimation problem within the context of nonstationary processes and derive what is known as the discrete Kalman filter.

Consider the following nonstationary state space model:
$$\begin{array}{l}\mathbf{x}(n)=\mathbf{A}(n-1)\mathbf{x}(n-1)+\mathbf{w}(n)\\ \mathbf{y}(n)=\mathbf{C}(n)\mathbf{x}(n)+\mathbf{v}(n)\end{array}\text{\hspace{1em}(DKF.1)}$$
$\mathbf{x}(n)$: the $p\times 1$ state vector

$\mathbf{A}(n-1)$: the $p\times p$ state transition matrix

$\mathbf{w}(n)$: the state noise with $E\left\{\mathbf{w}(n){\mathbf{w}}^H(n)\right\}={\mathbf{Q}}_w(n)\delta (n-k)$

$\mathbf{y}(n)$: the $q\times 1$ observation vector

$\mathbf{C}(n)$: the $q\times p$ observation matrix

$\mathbf{v}(n)$: the observation noise with $E\left\{\mathbf{v}(n){\mathbf{v}}^H(k)\right\}={\mathbf{Q}}_v(n)\delta (n-k)$, independent of the observation noise

It is assumed that $\mathbf{A}(n),\mathbf{C}(n),{\mathbf{Q}}_w(n),$ and ${\mathbf{Q}}_v(n)$ are known.

We are going to show that the optimum linear estimate of $\mathbf{x}(n)$ to be expressed in the form
$$\hat{\mathbf{x}}(n)=\mathbf{A}(n-1)\hat{\mathbf{x}}(n-1)+\mathbf{K}(n)[\mathbf{y}(n)-\mathbf{C}(n)\mathbf{A}(n-1)\hat{\mathbf{x}}(n-1)]\text{\hspace{1em}(DKF.2)}$$
With the appropriate Kalman gain matrix $\mathbf{K}(n)$, this recursion corresponds to the discrete Kalman filter.

Let us define $\hat{\mathbf{x}}(n|n-1)$ and $\hat{\mathbf{x}}(n|n)$ as the best linear estimate of $\mathbf{x}(n)$ given the observations $\mathbf{y}(n)$ up to time $n-1$ and $n,$ respectively.
Let us denote the corresponding errors as
$$\begin{array}{rl}\mathbf{e}(n|n-1)& =\mathbf{x}(n)-\hat{\mathbf{x}}(n|n-1)\\ \mathbf{e}(n|n)& =\mathbf{x}(n)-\hat{\mathbf{x}}(n|n)\end{array}\text{\hspace{1em}(DKF.3)}$$
with covariance matrices
$$\begin{array}{rl}\mathbf{P}(n|n-1)& =E\left\{\mathbf{e}(n|n-1){\mathbf{e}}^H(n|n-1)\right\}\\ \mathbf{P}(n|n)& =E\left\{\mathbf{e}(n|n){\mathbf{e}}^H(n|n)\right\}\end{array}\text{\hspace{1em}(DKF.4)}$$
For each $n>0$, given $\hat{\mathbf{x}}(n-1|n-1)$ and $\mathbf{P}(n-1|n-1)$, when a new observation, $y(n)$, becomes available, the problem is to find the minimum mean-square estimate $\hat{\mathbf{x}}(n|n)$ for the state vector $\mathbf{x}(n)$, i.e., to solve
$$\underset{\hat{\mathbf{x}}(n|n)}{\min }\mathrm{t}\mathrm{r}(\mathbf{P}(n|n))\text{\hspace{1em}(DKF.5)}$$
The solution to this problem will be derived in two steps.

• Given $\hat{\mathbf{x}}(n-1|n-1)$ we will find $\hat{\mathbf{x}}(n|n-1)$, which is the best estimate of $\mathbf{x}(n)$ without the observation $\mathbf{y}(n)$.
• Given $\mathbf{y}(n)$ and $\hat{\mathbf{x}}(n|n-1)$ we will estimate $\hat{\mathbf{x}}(n|n)$.

In the first step, since no new measurements are used to estimate $\mathbf{x}(n)$, all that is known is that $\mathbf{x}(n)$ evolves according to the state equation
$$\mathbf{x}(n)=\mathbf{A}(n-1)\mathbf{x}(n-1)+\mathbf{w}(n)$$
since $\mathbf{w}(n)$ is a zero mean white noise process (and the values of $\mathbf{w}(n)$ are unknown), then we may predict $\mathbf{x}(n)$ as follows,
$$\hat{\mathbf{x}}(n|n-1)=\mathbf{A}(n-1)\hat{\mathbf{x}}(n-1|n-1)\text{\hspace{1em}(DKF.6)}$$
which has an estimation error given by
$$\begin{array}{rl}\mathbf{e}(n|n-1)& =\mathbf{x}(n)-\hat{\mathbf{x}}(n|n-1)\\ & =\mathbf{A}(n-1)\mathbf{x}(n-1)+\mathbf{w}(n)-\mathbf{A}(n-1)\hat{\mathbf{x}}(n-1|n-1)\\ & =\mathbf{A}(n-1)\mathbf{e}(n-1|n-1)+\mathbf{w}(n)\end{array}\text{\hspace{1em}(DKF.7)}$$
Since the estimation error $\mathbf{e}(n-1|n-1)$ is uncorrelated with $\mathbf{w}(n)$ (a consequence of the fact that $\mathbf{w}(n)$ is a white noise sequence), then
$$\mathbf{P}(n|n-1)=\mathbf{A}(n-1)\mathbf{P}(n-1|n-1){\mathbf{A}}^H(n-1)+{\mathbf{Q}}_w(n)\text{\hspace{1em}(DKF.8)}$$
In the second step, we incorporate the new measurement $\mathbf{y}(n)$ into the estimate $\hat{\mathbf{x}}(n|n-1)$.

A linear estimate of $\mathbf{x}(n)$ that is based on $\hat{\mathbf{x}}(n|n-1)$ and $\mathbf{y}(n)$ is of the form
$$\hat{\mathbf{x}}(n|n)={\mathbf{K}}^{\prime }(n)\hat{\mathbf{x}}(n|n-1)+\mathbf{K}(n)\mathbf{y}(n)\text{\hspace{1em}(DKF.9)}$$
where $\mathbf{K}(n)$ and ${\mathbf{K}}^{\prime }(n)$ are matrices to be specified.

• Requirement 1: $\hat{\mathbf{x}}(n|n),\hat{\mathbf{x}}(n|n-1)$ are unbiased
• Requirement 2: $\hat{\mathbf{x}}(n|n)$ minimize the mean-square error, $E\Vert \mathbf{e}(n|n){\Vert }^2=\mathrm{t}\mathrm{r}(\mathbf{P}(n|n))$

From Requirement 1,
$$E(\mathbf{x}(n))={\mathbf{K}}^{\prime }(n)E(\mathbf{x}(n))+\mathbf{K}(n)(\mathbf{C}(n)E(\mathbf{x}(n))+E(\mathbf{v}(n)))$$
Since $E(\mathbf{v}(n))=0$, we have
$${\mathbf{K}}^{\prime }(n)=\mathbf{I}-\mathbf{K}(n)\mathbf{C}(n)\text{\hspace{1em}(DKF.10)}$$
Substituting $(DKF.10)$ into $(DKF.9)$, we have
$$\hat{\mathbf{x}}(n|n)=\hat{\mathbf{x}}(n|n-1)+\mathbf{K}(n)[\mathbf{y}(n)-\mathbf{C}(n)\hat{\mathbf{x}}(n|n-1)]\text{\hspace{1em}(DKF.11)}$$
and the error
$$\begin{array}{rl}\mathbf{e}(n|n)& ={\mathbf{K}}^{\prime }(n)\mathbf{e}(n|n-1)-\mathbf{K}(n)\mathbf{v}(n)\\ & =[\mathbf{I}-\mathbf{K}(n)\mathbf{C}(n)]\mathbf{e}(n|n-1)-\mathbf{K}(n)\mathbf{v}(n)\end{array}\text{\hspace{1em}(DKF.12)}$$
Thus, the error covariance matrix for $\mathbf{e}(n\mid n)$ is
$$\begin{array}{rl}\mathbf{P}(n|n)& =E\left\{\mathbf{e}(n|n){\mathbf{e}}^H(n|n)\right\}\\ & =[\mathbf{I}-\mathbf{K}(n)\mathbf{C}(n)]\mathbf{P}(n|n-1)[\mathbf{I}-\mathbf{K}(n)\mathbf{C}(n){]}^H+\mathbf{K}(n){\mathbf{Q}}_v(n){\mathbf{K}}^H(n)\end{array}\text{\hspace{1em}(DKF.13)}$$
Next, we must find the value for the Kalman gain $\mathbf{K}(n)$ that minimizes the mean-square error
ξ ( n ) = tr ⁡ { P ( n ∣ n ) } \xi(n)=\operatorname{tr}\{\mathbf{P}(n | n)\} ξ(n)=tr{P(n∣n)}
Differentiating $\xi (n)$ with respect to $\mathbf{K}(n),$ setting the derivative to zero,
d d K tr ⁡ { P ( n ∣ n ) } = − 2 [ I − K ( n ) C ( n ) ] P ( n ∣ n − 1 ) C H ( n ) + 2 K ( n ) Q v ( n ) = 0 (DKF.14) \frac{d}{d \mathbf{K}} \operatorname{tr}\{\mathbf{P}(n |n)\}=-2[\mathbf{I}-\mathbf{K}(n) \mathbf{C}(n)] \mathbf{P}(n | n-1) \mathbf{C}^{H}(n)+2 \mathbf{K}(n) \mathbf{Q}_{v}(n)=0 \tag{DKF.14} dKd​tr{P(n∣n)}=−2[I−K(n)C(n)]P(n∣n−1)CH(n)+2K(n)Qv​(n)=0(DKF.14)
Solving for $\mathbf{K}(n)$ gives the desired expression for the Kalman gain,
$$\mathbf{K}(n)=\mathbf{P}(n|n-1){\mathbf{C}}^H(n){\left[\mathbf{C}(n)\mathbf{P}(n|n-1){\mathbf{C}}^H(n)+{\mathbf{Q}}_v(n)\right]}^{-1}\text{\hspace{1em}(DKF.15)}$$
Having found the Kalman gain vector, we may simplify the expression given in $(DKF.13)$ for the error covariance. First, we rewrite the expression for $\mathbf{P}(n|n)$ as follows,
$$\begin{array}{rl}\mathbf{P}(n|n)=& [\mathbf{I}-\mathbf{K}(n)\mathbf{C}(n)]\mathbf{P}(n|n-1)\\ & -\left\{[\mathbf{I}-\mathbf{K}(n)\mathbf{C}(n)]\mathbf{P}(n|n-1){\mathbf{C}}^H(n)+\mathbf{K}(n){\mathbf{Q}}_v(n)\right\}{\mathbf{K}}^H(n)\end{array}$$
From $(DKF.14)$ , however, it follows that the second term is equal to zero, which leads to the desired expression for the error covariance matrix
$$\mathbf{P}(n|n)=[\mathbf{I}-\mathbf{K}(n)\mathbf{C}(n)]\mathbf{P}(n|n-1)\text{\hspace{1em}(DKF.16)}$$
Note: another derivation method can be found in the Slides 7c_Kalman