跳到主要内容

京都大学 情報学研究科 知能情報学専攻 2025年8月実施 専門科目 S-4

Author​

itsuitsuki

Description​

大学公表の原題 In the questions below, (⋅)∗(\cdot)^*, (⋅)T(\cdot)^{\mathrm{T}}, and E[⋅]E[\cdot] denote the complex conjugate, the transpose, and the expectation, respectively. R\mathbb{R} and Z\mathbb{Z} denote the set of all real numbers and the set of all integers, respectively.

Q.1​

Let x(n)x(n) be a discrete-time signal with the index of n∈Zn \in \mathbb{Z}, and define the zz-transform of x(n)x(n) as

X(z)=∑n=−∞∞x(n)z−n,(∗)X(z) = \sum_{n=-\infty}^{\infty} x(n)z^{-n}, \quad\quad (*)

where zz is a complex variable. Moreover, define the region of convergence of the zz-transform X(z)X(z) as a set of zz such that the series in the right-hand side of Eq. (∗)(*) is absolutely convergent. Answer the following questions.

(1) Derive the zz-transform and its region of convergence of a discrete-time signal

x1(n)={ann≥00n<0,x_1(n) = \begin{cases} a^n & n \geq 0 \\ 0 & n < 0 \end{cases},

where a∈Ra \in \mathbb{R} and a0=1a^0 = 1.

(2) Derive the discrete-time signal x2(n)x_2(n), whose zz-transform is given by

X2(z)=1(1−14z−1)(1−12z−1),X_2(z) = \frac{1}{\left(1 - \frac{1}{4}z^{-1}\right) \left(1 - \frac{1}{2}z^{-1}\right)},

where the region of convergence is 14<∣z∣<12\frac{1}{4} < |z| < \frac{1}{2}.

(3) Assume that the zz-transform and its region of convergence of a discrete-time signal x3(n)x_3(n) are given by X3(z)X_3(z) and R3\mathcal{R}_3, respectively. Express the zz-transform of a discrete-time signal x3∗(n−k)x_3^*(n - k) for some k∈Zk \in \mathbb{Z} using X3(z)X_3(z). Moreover, answer whether the region of convergence of the zz-transform of x3∗(n−k)x_3^*(n - k) is identical to R3\mathcal{R}_3 or not with reasons.

Q.2​

Consider a filter of NN taps, whose output signal with the index of n∈Zn \in \mathbb{Z} is given by

y(n)=xT(n)h,y(n) = \boldsymbol{x}^{\mathrm{T}}(n)\boldsymbol{h},

where

x(n)=[x(n) x(n−1) … x(n−N+1)]T∈RN,h=[h(0) h(1) … h(N−1)]T∈RN\boldsymbol{x}(n) = [x(n) \ x(n - 1) \ \dots \ x(n - N + 1)]^{\mathrm{T}} \in \mathbb{R}^N,\\ \boldsymbol{h} = [h(0) \ h(1) \ \dots \ h(N - 1)]^{\mathrm{T}} \in \mathbb{R}^N

are the input signal vector and the filter coefficient vector, respectively. Let the input signal x(n)x(n) and the desired signal d(n)d(n) be real-valued wide-sense stationary discrete-time random processes. We assume that E[x(n)xT(n)]E[\boldsymbol{x}(n)\boldsymbol{x}^{\mathrm{T}}(n)], E[d(n)x(n)]E[d(n)\boldsymbol{x}(n)], and E[d2(n)]E[d^2(n)] can be expressed as R=E[x(n)xT(n)]\boldsymbol{R} = E[\boldsymbol{x}(n)\boldsymbol{x}^{\mathrm{T}}(n)], p=E[d(n)x(n)]\boldsymbol{p} = E[d(n)\boldsymbol{x}(n)], and σ2=E[d2(n)]\sigma^2 = E[d^2(n)], respectively.

Moreover, we define the mean-squared error between y(n)y(n) and d(n)d(n) as

J(h)=E[{d(n)−y(n)}2].J(\boldsymbol{h}) = E \left[ \{d(n) - y(n)\}^2 \right].

Answer the following questions.

(1) Express J(h)J(\boldsymbol{h}) using R,p,σ2\boldsymbol{R}, \boldsymbol{p}, \sigma^2, and h\boldsymbol{h}.

(2) Derive the equation that the filter coefficient vector h\boldsymbol{h} satisfies to minimize J(h)J(\boldsymbol{h}) (Wiener-Hopf equation).

题目描述​

以下 (⋅)∗(\cdot)^*、(⋅)T(\cdot)^{\mathrm T} 与 E[⋅]E[\cdot] 分别表示复共轭、转置和期望;R\mathbb R 与 Z\mathbb Z 分别表示实数集与整数集。

  1. 设 x(n)x(n) 为下标 n∈Zn\in\mathbb Z 的离散时间信号,其 zz 变换定义为

    X(z)=∑n=−∞∞x(n)z−n,(*)X(z)=\sum_{n=-\infty}^{\infty}x(n)z^{-n},\tag{*}

    其中 zz 是复变量。X(z)X(z) 的收敛域定义为使式 (∗)(*) 右端级数绝对收敛的所有 zz 构成的集合。

    (1)求离散时间信号

    x1(n)={an(n≥0),0(n<0),x_1(n)= \begin{cases} a^n & (n\geq0),\\ 0 & (n<0), \end{cases}

    的 zz 变换及其收敛域,其中 a∈Ra\in\mathbb R 且 a0=1a^0=1。

    (2)已知

    X2(z)=1(1−14z−1)(1−12z−1)X_2(z)= \frac{1} {\left(1-\frac14z^{-1}\right) \left(1-\frac12z^{-1}\right)}

    的收敛域为

    14<∣z∣<12,\frac14<|z|<\frac12,

    求其对应的离散时间信号 x2(n)x_2(n)。

    (3)设离散时间信号 x3(n)x_3(n) 的 zz 变换及其收敛域分别为 X3(z)X_3(z) 和 R3\mathcal R_3。对任意 k∈Zk\in\mathbb Z,用 X3(z)X_3(z) 表示 x3∗(n−k)x_3^*(n-k) 的 zz 变换;并说明该变换的收敛域是否与 R3\mathcal R_3 相同及其理由。

  2. 考虑一个 NN 抽头滤波器,其输出为

    y(n)=xT(n)h,n∈Z,y(n)=\boldsymbol{x}^{\mathrm T}(n)\boldsymbol h,\qquad n\in\mathbb Z,

    其中输入信号向量和滤波器系数向量分别为

    x(n)=[x(n)  x(n−1)  ⋯  x(n−N+1)]T∈RN,\boldsymbol{x}(n) =[x(n)\ \ x(n-1)\ \ \cdots\ \ x(n-N+1)]^{\mathrm T}\in\mathbb R^N,
    h=[h(0)  h(1)  ⋯  h(N−1)]T∈RN.\boldsymbol h =[h(0)\ \ h(1)\ \ \cdots\ \ h(N-1)]^{\mathrm T}\in\mathbb R^N.

    设输入信号 x(n)x(n) 与期望信号 d(n)d(n) 是实值广义平稳离散时间随机过程,并记

    R=E[x(n)xT(n)],p=E[d(n)x(n)],σ2=E[d2(n)].\boldsymbol R =E[\boldsymbol{x}(n)\boldsymbol{x}^{\mathrm T}(n)],\qquad \boldsymbol p =E[d(n)\boldsymbol{x}(n)],\qquad \sigma^2=E[d^2(n)].

    输出 y(n)y(n) 与期望信号 d(n)d(n) 之间的均方误差定义为

    J(h)=E ⁣[{d(n)−y(n)}2].J(\boldsymbol h) =E\!\left[\{d(n)-y(n)\}^2\right].

    (1)用 R,p,σ2,h\boldsymbol R,\boldsymbol p,\sigma^2,\boldsymbol h 表示 J(h)J(\boldsymbol h)。

    (2)推导使 J(h)J(\boldsymbol h) 最小的滤波器系数向量 h\boldsymbol h 所满足的方程,即 Wiener–Hopf 方程。

Kai​

Q.1​

(1) 对 a≠0a\ne0,几何级数给出

X1(z)=∑n=0∞(a/z)n=11−az−1,∣z∣>∣a∣.X_1(z)=\sum_{n=0}^{\infty}(a/z)^n=\frac1{1-az^{-1}},\qquad |z|>|a|.

当 a=0a=0 时,x1(n)=δ[n]x_1(n)=\delta[n],变换为常数 11。它在所有非零复数处收敛,并可延拓至 z=0z=0;若定义允许有限 Laurent 多项式在零点取值,则此时零点也属于收敛域。

(2) 部分分式分解为

X2(z)=−11−14z−1+21−12z−1.X_2(z)=-\frac1{1-\frac14z^{-1}}+\frac2{1-\frac12z^{-1}}.

第一项须取右边序列,第二项须取左边序列,因为

11−az−1=−∑m=1∞(z/a)m(∣z∣<∣a∣).\frac1{1-az^{-1}}=-\sum_{m=1}^{\infty}(z/a)^m\quad(|z|<|a|).

因此

x2(n)=−(14)nu[n]−2(12)nu[−n−1],\boxed{x_2(n)=-\left(\frac14\right)^n u[n]-2\left(\frac12\right)^n u[-n-1]},

其中 u[n]=1u[n]=1 当 n≥0n\geq0,否则为 00。两个收敛域相交正是 1/4<∣z∣<1/21/4<|z|<1/2。

(3) 换元 m=n−km=n-k,得

∑nx3∗(n−k)z−n=z−k[X3(z∗)]∗.\sum_n x_3^*(n-k)z^{-n}=z^{-k}\left[X_3(z^*)\right]^*.

对任何 0<∣z∣<∞0<|z|<\infty,

∑n∣x3∗(n−k)z−n∣=∣z∣−k∑m∣x3(m)∣∣z∣−m.\sum_n|x_3^*(n-k)z^{-n}|=|z|^{-k}\sum_m|x_3(m)||z|^{-m}.

故非零有限复数处的收敛性完全相同,收敛环及其边界圆上的收敛性均不变。若收敛域包括 00(或扩充复平面中的 ∞\infty),移位可能改变这些点的归属。例如 δ[n]\delta[n] 的变换是 11,而延迟后的 δ[n−1]\delta[n-1] 的变换是 z−1z^{-1},在零点有极点。

Q.2​

(1) 展开平方并取期望,得

J(h)=σ2−2pTh+hTRh.\boxed{J(\boldsymbol h)=\sigma^2-2\boldsymbol p^{\mathrm T}\boldsymbol h+\boldsymbol h^{\mathrm T}\boldsymbol R\boldsymbol h}.

(2) R\boldsymbol R 对称且半正定,故

∇J=2Rh−2p=0,\nabla J=2\boldsymbol R\boldsymbol h-2\boldsymbol p=0,

给出 Wiener–Hopf 方程

Rh=p.\boxed{\boldsymbol R\boldsymbol h=\boldsymbol p}.

若 R\boldsymbol R 可逆,唯一解为 h=R−1p\boldsymbol h=\boldsymbol R^{-1}\boldsymbol p。若 R\boldsymbol R 奇异,方程仍有解:v∈ker⁡R\boldsymbol v\in\ker\boldsymbol R 意味着 E[(vTx)2]=0E[(\boldsymbol v^{\mathrm T}\boldsymbol x)^2]=0,故 vTp=0\boldsymbol v^{\mathrm T}\boldsymbol p=0,即 p∈im⁡R\boldsymbol p\in\operatorname{im}\boldsymbol R。所有解均为全局最小点,可写作 R+p+ker⁡R\boldsymbol R^+\boldsymbol p+\ker\boldsymbol R。