酉不变范数与对称度规函数
设
∥
⋅
∥
:
C
m
×
n
→
R
+
\lVert\cdot\rVert : \mathbb{C}^{m\times n} \to \mathbb{R}_+
∥⋅∥:Cm×n→R+ 是范数,且
∥
★
∥
=
∥
U
∗
★
V
∥
\lVert \bigstar \rVert = \lVert U^{*} \bigstar V \rVert
∥★∥=∥U∗★V∥ 对所有酉矩阵
U
,
V
U,V
U,V 成立(此时称
∥
⋅
∥
\lVert\cdot\rVert
∥⋅∥酉不变);考虑奇异值分解
A
=
U
Σ
(
A
)
V
∗
A = U \Sigma(A) V^{*}
A=UΣ(A)V∗,其中
Σ
(
A
)
∈
C
m
×
n
\Sigma(A) \in \mathbb{C}^{m\times n}
Σ(A)∈Cm×n的主对角元为
σ
1
(
A
)
≥
⋯
≥
σ
q
(
A
)
≥
0
,
q
=
min
(
m
,
n
)
\sigma_{1}(A) \geq \dots \geq \sigma_{q}(A) \geq 0 ,\ q = \min(m, n)
σ1(A)≥⋯≥σq(A)≥0, q=min(m,n),则
∥
A
∥
=
∥
Σ
(
A
)
∥
\lVert A \rVert = \lVert \Sigma(A) \rVert
∥A∥=∥Σ(A)∥ 只依赖于
sv
(
A
)
=
(
σ
i
(
A
)
)
1
≤
i
≤
q
∈
R
+
q
\operatorname{sv}(A) = (\sigma_{i}(A))_{1\leq i\leq q} \in \mathbb{R}_{+}^{q}
sv(A)=(σi(A))1≤i≤q∈R+q 。
设 g :
F
n
→
R
+
\mathbb{F}^n \to\mathbb{R}_+
Fn→R+ 是范数:
若
g
(
⋅
)
=
g
(
τ
(
⋅
)
)
,
∀
τ
∈
S
n
g(\cdot) = g( \tau(\cdot) ), \ \forall \tau\in S_n
g(⋅)=g(τ(⋅)), ∀τ∈Sn ,则称 g对称或置换不变;
若
g
(
⋅
)
=
g
(
abs
(
⋅
)
)
g(\cdot) = g( \operatorname{abs}(\cdot) )
g(⋅)=g(abs(⋅)),则称 g绝对或度规不变;
若 g 对称且绝对,则称 g 为对称度规函数。
固定正整数 m,n 和
q
=
min
(
m
,
n
)
q = \min(m, n)
q=min(m,n) ,von Neumann发现
若
∥
⋅
∥
:
C
m
×
n
→
R
+
\lVert\cdot\rVert : \mathbb{C}^{m\times n} \to \mathbb{R}_+
∥⋅∥:Cm×n→R+ 是酉不变范数,则
g
∥
⋅
∥
:
(
x
i
)
∈
F
q
↦
∥
X
∥
g_{\lVert\cdot\rVert} : (x_{i})\in\mathbb{F}^{q} \mapsto \lVert X\rVert
g∥⋅∥:(xi)∈Fq↦∥X∥是对称度规函数,其中
X
∈
F
m
×
n
X \in \mathbb{F}^{m\times n}
X∈Fm×n由
diag
(
x
1
,
…
,
x
q
)
\operatorname{diag}(x_{1}, \dots, x_{q})
diag(x1,…,xq)添补零得到;
若
g
:
F
q
→
R
+
g : \mathbb{F}^{q} \to\mathbb{R}_+
g:Fq→R+ 是对称度规函数,则
∥
⋅
∥
g
:
A
∈
C
m
×
n
↦
g
(
sv
(
A
)
)
\lVert\cdot\rVert_{g} : A\in\mathbb{C}^{m\times n} \mapsto g(\operatorname{sv}(A))
∥⋅∥g:A∈Cm×n↦g(sv(A)) 是酉不变范数。
证明:对于范数
∥
⋅
∥
\lVert\cdot\rVert
∥⋅∥,易见
g
∥
⋅
∥
g_{\lVert\cdot\rVert}
g∥⋅∥ 仍是范数;而置换矩阵和
diag
(
exp
(
−
1
θ
1
)
,
…
,
exp
(
−
1
θ
q
)
)
\operatorname{diag}\left( \exp(\sqrt{-1}\theta_{1}), \dots, \exp(\sqrt{-1}\theta_{q}) \right)
diag(exp(−1θ1),…,exp(−1θq))都是酉矩阵,所以酉不变范数确实诱导了对称度规函数。反之,由
sv
(
U
∗
A
V
)
=
sv
(
A
)
\operatorname{sv}(U^{*}AV) = \operatorname{sv}(A)
sv(U∗AV)=sv(A) 可知
∥
⋅
∥
g
\lVert\cdot\rVert_{g}
∥⋅∥g酉不变,而正定性和齐次性容易验证,只需说明三角不等式成立。任取
A
,
B
∈
C
m
×
n
A, B \in \mathbb{C}^{m\times n}
A,B∈Cm×n,奇异值的樊畿不等式断言
(
σ
i
(
A
+
B
)
)
1
≤
i
≤
q
≺
w
(
σ
i
(
A
)
+
σ
i
(
B
)
)
1
≤
i
≤
q
(\sigma_{i}(A+B))_{1 \leq i \leq q} \prec_\mathrm{w} ( \sigma_{i}(A) + \sigma_{i}(B) )_{1 \leq i \leq q}
(σi(A+B))1≤i≤q≺w(σi(A)+σi(B))1≤i≤q,弱受控表明存在双随机矩阵 D 适合
D
(
sv
(
A
)
+
sv
(
B
)
)
≥
sv
(
A
+
B
)
D( \operatorname{sv}(A)+ \operatorname{sv}(B) ) \geq \operatorname{sv}(A+B)
D(sv(A)+sv(B))≥sv(A+B),而双随机矩阵是置换矩阵的凸组合,应用下述引理通过简单推导即证。
■
\blacksquare
■
[引理] 设
g
:
F
n
→
R
+
是范数,若
g
(
v
)
≥
g
(
w
)
,
∀
v
,
w
∈
F
n
,
abs
(
v
)
≥
abs
(
w
)
g : \mathbb{F}^n \to\mathbb{R}_+ 是范数,若 g(v)\geq g(w),\ \forall v,w\in\mathbb{F}^n,\, \operatorname{abs}(v)\geq \operatorname{abs}(w)
g:Fn→R+是范数,若g(v)≥g(w), ∀v,w∈Fn,abs(v)≥abs(w) ,则称 g单调;范数 g 单调当且仅当 g 绝对。
证明:若 g 单调,则
g
(
⋅
)
≥
g
(
abs
(
⋅
)
)
≥
g
(
⋅
)
g(\cdot) \geq g\left( \operatorname{abs}(\cdot) \right) \geq g(\cdot)
g(⋅)≥g(abs(⋅))≥g(⋅) 。若 g 绝对,设
∣
α
∣
≤
1
|\alpha| \leq 1
∣α∣≤1 ,则
g
(
z
1
,
…
,
z
k
−
1
,
α
z
k
,
z
k
+
1
,
…
,
z
n
)
=
g
(
z
1
,
…
,
∣
α
∣
z
k
,
…
,
z
n
)
≤
1
+
∣
α
∣
2
g
(
z
1
,
…
,
z
k
,
…
,
z
n
)
+
1
−
∣
α
∣
2
g
(
z
1
,
…
,
−
z
k
,
…
,
z
n
)
=
g
(
z
1
,
…
,
z
n
)
,
1
≤
k
≤
n
.
\begin{align} &\ \ g(z_1, \dots, z_{k-1}, \alpha z_k , z_{k+1}, \dots, z_n) = g(z_1, \dots, |\alpha| z_k , \dots, z_n) \\ &\leq \frac{1+|\alpha|}{2}g(z_1, \dots, z_k , \dots, z_n) + \frac{1-|\alpha|}{2}g(z_1, \dots, -z_k , \dots, z_n) \\ &= g(z_1, \dots, z_n), \quad 1\leq k\leq n. \end{align}
g(z1,…,zk−1,αzk,zk+1,…,zn)=g(z1,…,∣α∣zk,…,zn)≤21+∣α∣g(z1,…,zk,…,zn)+21−∣α∣g(z1,…,−zk,…,zn)=g(z1,…,zn),1≤k≤n.
归纳即知 g 单调。
■
\blacksquare
■
樊畿控制定理
前述证明其实突出了樊畿范数的重要性(弱受控),整理可得如下命题。
设
g
:
F
n
→
R
+
g : \mathbb{F}^n \to\mathbb{R}_+
g:Fn→R+是对称度规函数,若
∥
v
∥
[
k
]
≤
∥
w
∥
[
k
]
,
1
≤
k
≤
n
\lVert v\rVert_{[k]} \leq \lVert w\rVert_{[k]}, \ 1\leq k\leq n
∥v∥[k]≤∥w∥[k], 1≤k≤n,则
g
(
v
)
≤
g
(
w
)
g(v) \leq g(w)
g(v)≤g(w) 。
设
∥
⋅
∥
:
C
m
×
n
→
R
+
\lVert\cdot\rVert : \mathbb{C}^{m\times n} \to \mathbb{R}_+
∥⋅∥:Cm×n→R+是酉不变范数,若
∥
A
∥
[
k
]
≤
∥
B
∥
[
k
]
,
1
≤
k
≤
min
(
m
,
n
)
\lVert A\rVert_{[k]} \leq \lVert B\rVert_{[k]}, \ 1\leq k\leq \min(m,n)
∥A∥[k]≤∥B∥[k], 1≤k≤min(m,n) ,则
∥
A
∥
≤
∥
B
∥
\lVert A\rVert \leq \lVert B\rVert
∥A∥≤∥B∥ 。
至此,矩阵的酉不变范数有了较清晰的图景。特别地,如果对称度规函数取为向量的 l_p 范数,则相应的矩阵酉不变范数即为Schatten范数,满足次可加性。
酉不变范数的次可乘性质
设
∥
⋅
∥
:
C
n
×
n
→
R
+
\lVert\cdot\rVert : \mathbb{C}^{n\times n} \to \mathbb{R}_+
∥⋅∥:Cn×n→R+ 是酉不变范数,则对
P
,
A
,
Q
∈
C
n
×
n
P, A, Q \in \mathbb{C}^{n\times n}
P,A,Q∈Cn×n 有
∥
P
A
Q
∥
≤
σ
max
(
P
)
∥
A
∥
σ
max
(
Q
)
\lVert PAQ \rVert \leq \sigma_{\max}(P) \lVert A \rVert \sigma_{\max}(Q)
∥PAQ∥≤σmax(P)∥A∥σmax(Q)。
证明:在
∥
★
∥
=
g
∥
⋅
∥
(
sv
(
★
)
)
\lVert\bigstar\rVert = g_{\lVert\cdot\rVert}(\operatorname{sv}(\bigstar))
∥★∥=g∥⋅∥(sv(★)) 中利用
g
∥
⋅
∥
g_{\lVert\cdot\rVert}
g∥⋅∥的单调性,只需说明
σ
max
(
P
)
σ
max
(
Q
)
sv
(
A
)
≥
sv
(
P
A
Q
)
\sigma_{\max}(P)\sigma_{\max}(Q)\operatorname{sv}(A) \geq \operatorname{sv}(PAQ)
σmax(P)σmax(Q)sv(A)≥sv(PAQ),这通过下述引理易得。
■
\blacksquare
■
[引理] 将奇异值(计入 0 )排列为
σ
1
(
⋅
)
≥
⋯
≥
σ
n
(
⋅
)
\sigma_{1}(\cdot) \geq \dots \geq \sigma_{n}(\cdot)
σ1(⋅)≥⋯≥σn(⋅),则对
A
,
B
∈
C
n
×
n
A, B \in \mathbb{C}^{n\times n}
A,B∈Cn×n和
1
≤
i
≤
n
有
σ
i
(
A
B
)
≤
σ
max
(
A
)
σ
i
(
B
)
和
σ
i
(
A
B
)
≤
σ
max
(
B
)
σ
i
(
A
)
1\leq i\leq n 有 \sigma_{i}(AB) \leq \sigma_{\max}(A)\sigma_{i}(B) 和 \sigma_{i}(AB) \leq \sigma_{\max}(B)\sigma_{i}(A)
1≤i≤n有σi(AB)≤σmax(A)σi(B)和σi(AB)≤σmax(B)σi(A) 。
证明:在奇异值的变分刻画
σ
i
(
★
)
=
sup
dim
(
W
)
=
i
inf
v
∈
W
:
∥
v
∥
2
=
1
∥
★
v
∥
2
\sigma_{i}(\bigstar) = \sup_{\dim(W)=i} \inf_{v \in W : \lVert v \rVert_{2} = 1} \lVert \bigstar v \rVert_{2}
σi(★)=supdim(W)=iinfv∈W:∥v∥2=1∥★v∥2中,利用
∥
A
B
v
∥
2
≤
∥
∣
A
∣
∥
2
∥
B
v
∥
2
=
σ
max
(
A
)
∥
B
v
∥
2
\lVert AB v \rVert_{2} \leq \lVert\lvert A \rvert\rVert_{2} \lVert Bv \rVert_{2} = \sigma_{\max}(A) \lVert Bv \rVert_{2}
∥ABv∥2≤∥∣A∣∥2∥Bv∥2=σmax(A)∥Bv∥2即得
σ
i
(
A
B
)
≤
σ
max
(
A
)
σ
i
(
B
)
\sigma_{i}(AB) \leq \sigma_{\max}(A)\sigma_{i}(B)
σi(AB)≤σmax(A)σi(B) 。另一方面,我们有
sv
(
★
)
=
sv
(
★
∗
)
。
■
\operatorname{sv}(\bigstar) = \operatorname{sv}(\bigstar^{*}) 。 \blacksquare
sv(★)=sv(★∗)。■
p次对称度规函数
设 g : F n → R + g : \mathbb{F}^n \to\mathbb{R}_+ g:Fn→R+ 是对称度规函数,任取 p ∈ [ 1 , ∞ ) p\in[1,\infty) p∈[1,∞),则 g ( p ) : ( u 1 , … , u n ) ∈ F n ↦ [ g ( ∣ u 1 ∣ p , … , ∣ u n ∣ p ) ] 1 / p g^{(p)} : (u_1,\dots,u_n) \in \mathbb{F}^{n} \mapsto \left[ g(|u_1|^p, \dots, |u_n|^p) \right]^{1/p} g(p):(u1,…,un)∈Fn↦[g(∣u1∣p,…,∣un∣p)]1/p 也是对称度规函数——只需验证三角不等式。
[Hölder不等式]
g
(
u
∘
v
)
≤
g
(
p
)
(
u
)
g
(
p
′
)
(
v
)
,
1
p
+
1
p
′
=
1
g(u\circ v) \leq g^{(p)}(u) g^{(p')}(v) ,\ \tfrac{1}{p}+\tfrac{1}{p'}=1
g(u∘v)≤g(p)(u)g(p′)(v), p1+p′1=1;
[Minkowski不等式]
g
(
p
)
(
u
+
v
)
≤
g
(
p
)
(
u
)
+
g
(
p
)
(
v
)
g^{(p)}(u+v) \leq g^{(p)}(u) + g^{(p)}(v)
g(p)(u+v)≤g(p)(u)+g(p)(v) 。
证明:由Young不等式可知
α
β
=
min
t
>
0
{
(
α
/
t
)
p
p
+
(
t
β
)
p
′
p
′
}
,
∀
α
,
β
∈
R
+
\alpha\beta = \min_{t>0} \left\{ \frac{(\alpha/t)^p}{p} + \frac{(t\beta)^{p'}}{p'} \right\} ,\ \forall \alpha,\beta \in \mathbb{R}_{+}
αβ=mint>0{p(α/t)p+p′(tβ)p′}, ∀α,β∈R+ ,进而利用 g 的单调性、次可加性和齐次性易证Holder不等式。不妨设
u
,
v
∈
R
+
n
和
p
>
1
u,v \in \mathbb{R}_{+}^{n} 和 p>1
u,v∈R+n和p>1,由
(
u
+
v
)
p
=
u
∘
(
u
+
v
)
p
−
1
+
v
∘
(
u
+
v
)
p
−
1
和
p
′
(
p
−
1
)
=
p
(u+v)^{p} = u\circ(u+v)^{p-1} + v\circ(u+v)^{p-1} 和 p'(p-1) = p
(u+v)p=u∘(u+v)p−1+v∘(u+v)p−1和p′(p−1)=p 可得
g
(
(
u
+
v
)
p
)
≤
g
(
u
∘
(
u
+
v
)
p
−
1
)
+
g
(
v
∘
(
u
+
v
)
p
−
1
)
≤
g
(
p
)
(
u
)
[
g
(
(
u
+
v
)
p
)
]
1
/
p
′
+
g
(
p
)
(
v
)
[
g
(
(
u
+
v
)
p
)
]
1
/
p
′
=
[
g
(
p
)
(
u
)
+
g
(
p
)
(
v
)
]
⋅
[
g
(
(
u
+
v
)
p
)
]
1
/
p
′
\begin{align} g((u+v)^{p}) &\leq g(u\circ(u+v)^{p-1}) + g(v\circ(u+v)^{p-1}) \\ &\leq g^{(p)}(u) [g((u+v)^{p})]^{1/p'} + g^{(p)}(v) [g((u+v)^{p})]^{1/p'} \\ &= [g^{(p)}(u) + g^{(p)}(v)] \cdot [g((u+v)^{p})]^{1/p'} \end{align}
g((u+v)p)≤g(u∘(u+v)p−1)+g(v∘(u+v)p−1)≤g(p)(u)[g((u+v)p)]1/p′+g(p)(v)[g((u+v)p)]1/p′=[g(p)(u)+g(p)(v)]⋅[g((u+v)p)]1/p′
,两端除以
[
g
(
(
u
+
v
)
p
)
]
1
/
p
′
[g((u+v)^{p})]^{1/p'}
[g((u+v)p)]1/p′即证。
■
\blacksquare
■
设 ∥ ⋅ ∥ : C n × n → R + \lVert\cdot\rVert : \mathbb{C}^{n\times n} \to \mathbb{R}_+ ∥⋅∥:Cn×n→R+是酉不变范数,则对 A , B ∈ C n × n A, B \in \mathbb{C}^{n\times n} A,B∈Cn×n 有Hölder不等式 ∥ A B ∥ ≤ ∥ ( A A ∗ ) p ∥ 1 / p ∥ ( B B ∗ ) p ′ ∥ 1 / p ′ , 1 p + 1 p ′ = 1 \lVert AB \rVert \leq \lVert (\sqrt{AA^{*}})^{p} \rVert^{1/p} \lVert (\sqrt{BB^{*}})^{p'} \rVert^{1/p'} ,\ \tfrac{1}{p}+\tfrac{1}{p'}=1 ∥AB∥≤∥(AA∗)p∥1/p∥(BB∗)p′∥1/p′, p1+p′1=1 。文章来源:https://www.toymoban.com/news/detail-674391.html
证明:将奇异值(计入 0 )排列为 σ 1 ( ⋅ ) ≥ ⋯ ≥ σ n ( ⋅ ) \sigma_{1}(\cdot) \geq \dots \geq \sigma_{n}(\cdot) σ1(⋅)≥⋯≥σn(⋅),由奇异值的乘积型受控 ( σ i ( A B ) ) 1 ≤ i ≤ n ≺ w ( σ i ( A ) σ i ( B ) ) 1 ≤ i ≤ n 可得 g ∥ ⋅ ∥ ( sv ( A B ) ) ≤ g ∥ ⋅ ∥ ( sv ( A ) ∘ sv ( B ) ) ≤ g ∥ ⋅ ∥ ( p ) ( sv ( A ) ) g ∥ ⋅ ∥ ( p ′ ) ( sv ( B ) ) (\sigma_{i}(AB))_{1\leq i\leq n} \prec_{\mathrm{w}} (\sigma_{i}(A)\sigma_{i}(B))_{1\leq i\leq n} 可得 g_{\lVert\cdot\rVert}(\operatorname{sv}(AB)) \leq g_{\lVert\cdot\rVert}(\operatorname{sv}(A)\circ\operatorname{sv}(B)) \leq g_{\lVert\cdot\rVert}^{(p)}(\operatorname{sv}(A)) g_{\lVert\cdot\rVert}^{(p')}(\operatorname{sv}(B)) (σi(AB))1≤i≤n≺w(σi(A)σi(B))1≤i≤n可得g∥⋅∥(sv(AB))≤g∥⋅∥(sv(A)∘sv(B))≤g∥⋅∥(p)(sv(A))g∥⋅∥(p′)(sv(B)) ,所以注意到 σ i ( A ) p = σ i ( ( A A ∗ ) p ) , 1 ≤ i ≤ n \sigma_{i}(A)^{p} = \sigma_{i}((\sqrt{AA^{*}})^{p}) , \ 1\leq i\leq n σi(A)p=σi((AA∗)p), 1≤i≤n即可。 ■ \blacksquare ■文章来源地址https://www.toymoban.com/news/detail-674391.html
到了这里,关于高等数学:矩阵的酉不变范数,樊畿控制定理,次可乘性质,p次对称度规函数的文章就介绍完了。如果您还想了解更多内容,请在右上角搜索TOY模板网以前的文章或继续浏览下面的相关文章,希望大家以后多多支持TOY模板网!