東京大学 情報理工学研究科 2017年8月実施 数学 第1問
Author
Zero, etsurin , 祭音Myyura
Description
次の連立一次方程式を解く問題を考える.
ここで, A ∈ R m × n , b ∈ R m A\in R^{m\times n},b\in R^m A ∈ R m × n , b ∈ R m は与えられた定数の行列とべクトルであり, x ∈ R n x\in R^n x ∈ R n は未知ベクトルである.以下の問いに答えよ.
(1)、
A ˉ = ( A ∣ b ) \bar{A}=(A|b) A ˉ = ( A ∣ b ) のように,行列 A A A の最後の列の後ろに1列追加した m × ( n + 1 ) m\times (n+1) m × ( n + 1 ) 行列を作る.例えば,
A = ( 1 0 − 1 1 1 0 0 1 1 ) , b = ( 2 4 2 ) A=\left (\begin{array}{cccc}
1&0&-1\\
1&1&0\\
0&1&1\\
\end{array}\right),
b=\left (\begin{array}{cccc}
2\\
4\\
2\\
\end{array}\right) A = 1 1 0 0 1 1 − 1 0 1 , b = 2 4 2
の場合には,
A ˉ = ( 1 0 − 1 2 1 1 0 4 0 1 1 2 ) \bar{A}=\left (\begin{array}{cccc}
1&0&-1&2\\
1&1&0&4\\
0&1&1&2\\
\end{array}\right) A ˉ = 1 1 0 0 1 1 − 1 0 1 2 4 2 となる.この例の A ˉ \bar{A} A ˉ の第 i i i 列ベクトルを a i ( i = 1 , 2 , 3 , 4 ) a_{i}(i=1,2,3,4) a i ( i = 1 , 2 , 3 , 4 ) とする.
(i)、a 1 , a 2 , a 3 a_{1},a_{2},a_{3} a 1 , a 2 , a 3 のうち線形独立なベクトルの最大個数を求めよ.
(ii)、a 4 a_{4} a 4 が a 1 , a 2 , a 3 a_{1},a_{2},a_{3} a 1 , a 2 , a 3 の線形和で表されることを, a 4 = x 1 a 1 + x 2 a 2 + a 3 a_{4}=x_{1}a_{1}+x_{2}a_{2}+a_{3} a 4 = x 1 a 1 + x 2 a 2 + a 3 となるスカラー x 1 , x 2 x_{1},x_{2} x 1 , x 2 を求めることで示せ.
(iii)、a 1 , a 2 , a 3 , a 4 a_{1},a_{2},a_{3},a_{4} a 1 , a 2 , a 3 , a 4 のうち線形独立なベクトルの最大個数を求めよ.
(2)、任意の m , n , A , b m,n,A,b m , n , A , b 対して, rank ( A ˉ ) = rank ( A ) \text{rank}(\bar{A})=\text{rank}(A) rank ( A ˉ ) = rank ( A ) のとき連立一次方程式の解が存在することを示せ.
(3)、rank ( A ˉ ) > rank ( A ) \text{rank}(\bar{A})>\text{rank}(A) rank ( A ˉ ) > rank ( A ) ならば解は存在しない.m > n m>n m > n , rank ( A ) = n \text{rank}(A)=n rank ( A ) = n , rank ( A ˉ ) > rank ( A ) \text{rank}(\bar{A})>\text{rank}(A) rank ( A ˉ ) > rank ( A ) のとき, 連立一次方程式の右辺と左辺と差のノルムの2乗 ∥ b − A x ∥ 2 \Vert b-Ax\Vert ^2 ∥ b − A x ∥ 2 を最小にする x x x を求めよ.
(4)、m < n , rank ( A ) = m m<n,\text{rank}(A)=m m < n , rank ( A ) = m のとき,どのような b b b に対しても連立一次方程式を満たす解が複数存在する.解のうちで ∥ x ∥ 2 \Vert x \Vert ^2 ∥ x ∥ 2 を最小にする x x x を,連立一次方程式を制約条件として,ラグランジュ乗数法を用いて求めよ.
(5)、任意 m , n , A m,n,A m , n , A に対して,以下の4つの式を満たす P ∈ R n × m P\in R^{n\times m} P ∈ R n × m が唯一に決まることを示せ.
A P A = A P A P = P ( A P ) T = A P ( P A ) T = P A \begin{aligned}
APA=A \\
PAP=P \\
(AP)^T=AP \\
(PA)^T=PA
\end{aligned} A P A = A P A P = P ( A P ) T = A P ( P A ) T = P A
(6)、(3)て求めた x x x と(4)で求めた x x x が,いずれも x = P b x=Pb x = P b の形で表せることを示せ.
题目描述
考虑线性方程组
其中 A ∈ R m × n A\in\mathbb R^{m\times n} A ∈ R m × n 、b ∈ R m b\in\mathbb R^m b ∈ R m 已知,
x ∈ R n x\in\mathbb R^n x ∈ R n 未知。把 b b b 作为末列接在 A A A 后得到增广矩阵
A ˉ = ( A ∣ b ) \bar A=(A\mid b) A ˉ = ( A ∣ b ) 。回答下列问题。
(1)对
A = ( 1 0 − 1 1 1 0 0 1 1 ) , b = ( 2 4 2 ) , A=\begin{pmatrix}1&0&-1\\1&1&0\\0&1&1\end{pmatrix},
\qquad
b=\begin{pmatrix}2\\4\\2\end{pmatrix}, A = 1 1 0 0 1 1 − 1 0 1 , b = 2 4 2 ,
记 A ˉ \bar A A ˉ 的列为 a 1 , a 2 , a 3 , a 4 a_1,a_2,a_3,a_4 a 1 , a 2 , a 3 , a 4 。
(i)求 a 1 , a 2 , a 3 a_1,a_2,a_3 a 1 , a 2 , a 3 中线性无关向量的最大个数。
(ii)求标量 x 1 , x 2 x_1,x_2 x 1 , x 2 ,使
a 4 = x 1 a 1 + x 2 a 2 + a 3 a_4=x_1a_1+x_2a_2+a_3 a 4 = x 1 a 1 + x 2 a 2 + a 3 ,从而证明 a 4 a_4 a 4 是前三列的线性组合。
(iii)求四个列向量中线性无关向量的最大个数。
(2)对任意 m , n , A , b m,n,A,b m , n , A , b ,证明若
rank ( A ˉ ) = rank ( A ) \operatorname{rank}(\bar A)=\operatorname{rank}(A) rank ( A ˉ ) = rank ( A ) ,则方程组有解。
(3)若
m > n m>n m > n 、
rank ( A ) = n \operatorname{rank}(A)=n rank ( A ) = n 且
rank ( A ˉ ) > rank ( A ) \operatorname{rank}(\bar A)>\operatorname{rank}(A) rank ( A ˉ ) > rank ( A ) ,方程组无精确解。求使
∥ b − A x ∥ 2 \|b-Ax\|^2 ∥ b − A x ∥ 2 最小的 x x x 。
(4)若 m < n m<n m < n 且 rank ( A ) = m \operatorname{rank}(A)=m rank ( A ) = m ,则对任意 b b b 都有多个解。
以 A x = b Ax=b A x = b 为约束,用拉格朗日乘子法求其中使 ∥ x ∥ 2 \|x\|^2 ∥ x ∥ 2 最小的解。
(5)证明对任意 m , n , A m,n,A m , n , A ,满足
A P A = A , P A P = P , ( A P ) T = A P , ( P A ) T = P A APA=A,\quad PAP=P,\quad
(AP)^{\mathsf T}=AP,\quad
(PA)^{\mathsf T}=PA A P A = A , P A P = P , ( A P ) T = A P , ( P A ) T = P A
的 P ∈ R n × m P\in\mathbb R^{n\times m} P ∈ R n × m 唯一确定。
(6)证明第(3)、(4)问所得解均可写成 x = P b x=Pb x = P b 。
Kai
(1)
(i)
( 1 0 − 1 1 1 0 0 1 1 ) → ( 1 0 − 1 0 1 1 0 1 1 ) → ( 1 0 − 1 0 1 1 0 0 0 ) \left (\begin{array}{cccc}
1&0&-1\\
1&1&0\\
0&1&1\\
\end{array}\right) \rightarrow
\left (\begin{array}{cccc}
1&0&-1\\
0&1&1\\
0&1&1\\
\end{array}\right) \rightarrow
\left (\begin{array}{cccc}
1&0&-1\\
0&1&1\\
0&0&0\\
\end{array}\right) 1 1 0 0 1 1 − 1 0 1 → 1 0 0 0 1 1 − 1 1 1 → 1 0 0 0 1 0 − 1 1 0
There are 2 linearly independent vectors in a 1 , a 2 , a 3 a_{1},a_{2},a_{3} a 1 , a 2 , a 3
(ii)
a 4 = 3 a 1 + a 2 + a 3 a_{4}=3a_{1}+a_{2}+a_{3} a 4 = 3 a 1 + a 2 + a 3
x 1 = 3 , x 2 = 1 x_{1}=3,x_{2}=1 x 1 = 3 , x 2 = 1
(iii)
a 4 = 2 a 1 + 2 a 2 , rank ( A ‾ ) = 2 a_4 = 2a_1 + 2a_2, \ \text{rank}(\overline{A}) = 2 a 4 = 2 a 1 + 2 a 2 , rank ( A ) = 2
(2)
Assuming that rank ( A ‾ ) = rank ( A ) = r \text{rank}(\overline{A}) = \text{rank}(A)=r rank ( A ) = rank ( A ) = r and there is no solution with A x = b Ax=b A x = b .
Hence the vector b b b , i.e. a n + 1 a_{n+1} a n + 1 , cannot be represented as a linear combination of ( a 1 , a 2 , … , a n ) (a_{1},a_{2},\ldots,a_{n}) ( a 1 , a 2 , … , a n ) .
Hence,
rank ( A ‾ ) = r + 1 > rank ( A ) \text{rank}(\overline{A}) =r+1> \text{rank}(A) rank ( A ) = r + 1 > rank ( A )
which is contradictory to the fact that rank ( A ˉ ) = rank ( A ) \text{rank}(\bar{A}) = \text{rank}(A) rank ( A ˉ ) = rank ( A ) .
Therefore, for any m , n , A , b m,n,A,b m , n , A , b , when rank ( A ‾ ) = rank ( A ) \text{rank}(\overline{A}) = \text{rank}(A) rank ( A ) = rank ( A ) the equation A x = b Ax=b A x = b has a solution.
(3)
L = ∥ A x − b ∥ 2 = ( b − A x ) T ( b − A x ) = b T b − x T A T b − b T A x + x T A T A x \begin{aligned}
\mathcal{L} &= \| Ax - b \|^2 = (b-Ax)^T(b-Ax) \\
&= b^Tb - x^TA^Tb - b^TAx + x^T A^T A x
\end{aligned} L = ∥ A x − b ∥ 2 = ( b − A x ) T ( b − A x ) = b T b − x T A T b − b T A x + x T A T A x
∂ L ∂ x = − A T b − A T b + ( A T A + ( A T A ) T ) x = 2 A T A x − 2 A T b = 0 \begin{aligned}
\frac{\partial \mathcal{L}}{\partial x} &= -A^Tb - A^Tb + (A^TA + (A^TA)^T)x \\
&= 2A^T Ax - 2A^Tb \\
&= 0
\end{aligned} ∂ x ∂ L = − A T b − A T b + ( A T A + ( A T A ) T ) x = 2 A T A x − 2 A T b = 0
Therefore,
x = ( A T A ) − 1 A T b . x=(A^TA)^{-1}A^Tb. x = ( A T A ) − 1 A T b .
Since A A A has full column rank, A T A A^TA A T A is positive definite. The objective has positive-definite Hessian 2 A T A 2A^TA 2 A T A , so this stationary point is the unique global minimum.
(4)
L ( x , λ ) = x T x − λ T ( A x − b ) \mathcal{L}(x,\lambda)=x^Tx-\lambda^T(Ax-b) L ( x , λ ) = x T x − λ T ( A x − b )
∂ L ( x , λ ) ∂ x = 2 x − A T λ = 0 \begin{aligned}
\frac{\partial L(x,\lambda)}{\partial x} &= 2x-A^T\lambda = 0 \\
\end{aligned} ∂ x ∂ L ( x , λ ) = 2 x − A T λ = 0
∂ L ( x , λ ) ∂ λ = − ( A x − b ) = 0 \begin{aligned}
\frac{\partial L(x,\lambda)}{\partial \lambda} &= -(Ax-b) =0
\end{aligned} ∂ λ ∂ L ( x , λ ) = − ( A x − b ) = 0
∴ x = A T λ 2 A A T λ = 2 b \therefore x = \frac{A^T \lambda}{2} \qquad AA^T \lambda = 2b ∴ x = 2 A T λ A A T λ = 2 b
Hence
λ = 2 ( A A T ) − 1 b \lambda = 2(AA^T)^{-1}b λ = 2 ( A A T ) − 1 b
Finally
x = A T ( A A T ) − 1 b . x=A^T(AA^T)^{-1}b. x = A T ( A A T ) − 1 b .
The matrix A A T AA^T A A T is positive definite because A A A has full row rank. Every other solution is x + z x+z x + z with A z = 0 Az=0 A z = 0 . As x x x lies in the range of A T A^T A T , x T z = 0 x^Tz=0 x T z = 0 , and ∥ x + z ∥ 2 = ∥ x ∥ 2 + ∥ z ∥ 2 \|x+z\|^2=\|x\|^2+\|z\|^2 ∥ x + z ∥ 2 = ∥ x ∥ 2 + ∥ z ∥ 2 . Thus this is the unique minimum-norm solution.
(5)
Let A = U Σ V T A=U\Sigma V^T A = U Σ V T be a singular value decomposition and define
P = V Σ + U T , P=V\Sigma^+U^T, P = V Σ + U T ,
where Σ + ∈ R n × m \Sigma^+\in\mathbb R^{n\times m} Σ + ∈ R n × m is the transposed rectangular diagonal matrix with every nonzero singular value replaced by its reciprocal and all other entries zero. Direct substitution gives all four equations, so such a matrix exists.
For uniqueness, let both P P P and Q Q Q satisfy the equations. The matrices A P AP A P and A Q AQ A Q are symmetric idempotents, and
range ( A P ) = range ( A Q ) = range ( A ) . \operatorname{range}(AP)=\operatorname{range}(AQ)=\operatorname{range}(A). range ( A P ) = range ( A Q ) = range ( A ) .
Thus they are the same orthogonal projector, so A P = A Q AP=AQ A P = A Q . Similarly, P A PA P A and Q A QA Q A are symmetric idempotents with
ker ( P A ) = ker ( Q A ) = ker ( A ) , \ker(PA)=\ker(QA)=\ker(A), ker ( P A ) = ker ( Q A ) = ker ( A ) ,
so P A = Q A PA=QA P A = Q A . Hence
P = P A P = P ( A Q ) = ( P A ) Q = ( Q A ) Q = Q A Q = Q . P=PAP=P(AQ)=(PA)Q=(QA)Q=QAQ=Q. P = P A P = P ( A Q ) = ( P A ) Q = ( Q A ) Q = Q A Q = Q .
Therefore, P P P is unique.
(6)
For (3), rank ( A ) = n \text{rank}(A)=n rank ( A ) = n and we have x = ( A T A ) − 1 A T b x=(A^TA)^{-1}A^Tb x = ( A T A ) − 1 A T b , hence
P = ( A T A ) − 1 A T P=(A^TA)^{-1}A^T P = ( A T A ) − 1 A T
which satisfies the four equations in (5).
For (4), rank ( A ) = m \text{rank}(A)=m rank ( A ) = m and we have x = A T ( A A T ) − 1 b x=A^T(AA^T)^{-1}b x = A T ( A A T ) − 1 b , hence
P = A T ( A A T ) − 1 P=A^T(AA^T)^{-1} P = A T ( A A T ) − 1
which also satisfies the four equations in (5).