Young’s inequality
a b ≤ a p p + b q q ab \le \frac{a^p}{p} + \frac{b^q}{q} ab≤pap+qbq
Where a ≥ 0 , b ≥ 0 , p > 1 , q > 1 , 1 p + 1 q = 1 a\ge0,b\ge0,p>1,q>1,\frac{1}{p}+\ frac{1}{q}=1a≥0,b≥0,p>1,q>1,p1+q1=1
if and only ifap = bqa^p=b^qap=bWait when q
Proof:
when a = 0 a=0a=0 orb = 0 b=0b=It is obviously established at 0
当 a > 0 , b > 0 a>0,b>0 a>0,b>0Constant
Jensen's Coefficientln
( app + bqq ) ≥ 1 p ln ap + 1 q ln bq = ln ab \begin{aligned} \ln\left( \frac{a^p}{p} + \; frac{b^q}{q}\right)\ge \frac{1}{p}\ln a^p + frac{1}{q}\ln b^q=\ln ab \end{aligned}ln(pap+qbq)≥p1lnap+q1lnbq=lnab
if and only if ap = bqa^p=b^qap=bWait when q
Hölder's inequality
∑ i = 1 n ∣ x i y i ∣ ≤ ∥ x ∥ p ∥ y ∥ q \sum_{i=1}^{n}\left|x_iy_i\right|\le \|\mathbf{x}\|_p \|\mathbf{y}\|_q i=1∑n∣xiyi∣≤∥x∥p∥y∥q
Where p ≥ 1 , q ≥ 1 , 1 p + 1 q = 1 p\ge 1, q\ge 1,\frac{1}{p}+\frac{1}{q}=1p≥1,q≥1,p1+q1=1
∥ x ∥ p = ( ∑ i = 1 n ∣ x i ∣ p ) 1 p \|\mathbf{x}\|_p=\left(\sum_{i=1}^{n}\left|x_i\right|^p\right)^{\frac{1}{p}} ∥x∥p=(∑i=1n∣xi∣p)p1
If and only if x = 0 \mathbf{x}=\mathbf{0}x=0或y = 0 \mathbf{y}=\mathbf{0}y=0,或 ∃ c 1 , c 2 > 0 \exists c1,c2>0 ∃c1,c2>0,使得 c 1 ∣ x i ∣ p = c 2 ∣ y i ∣ q c_1\left|x_i\right|^p=c_2\left|y_i\right|^q c1∣xi∣p=c2∣yi∣q
Proof:
when x = 0 \mathbf{x}=\mathbf{0}x=0或y = 0 \mathbf{y}=\mathbf{0}y=It is obviously established at 0
当x ≠ 0 , y ≠ 0 \mathbf{x}\q \mathbf{0},\mathbf{y}\q \mathbf{0}x=0,y=0 o'clock
When p = 1 p=1p=1时
∑ i = 1 n ∣ x i y i ∣ ≤ ( max i ∣ y i ∣ ) ∑ i = 1 n ∣ x i ∣ = ∥ x ∥ 1 ∥ y ∥ ∞ \sum_{i=1}^{n} \left|x_iy_i\right|\le \left(\max_i \left|y_i\right|\right)\sum_{i=1}^{n}\left|x_i\right|=\|\mathbf{x}\|_1\|\mathbf{y}\|_\infty ∑i=1n∣xiyi∣≤(maxi∣yi∣)∑i=1n∣xi∣=∥x∥1∥y∥∞
In the same way q = 1 q=1q=1
When p>1 p>1p>1 hour
设 A i = ∣ x i ∣ ∥ x ∥ p , B i = ∣ y i ∣ ∥ y ∥ q A_i = \frac{\left|x_i\right|}{\|\mathbf{x}\|_p}, B_i = \frac{\left|y_i\right|}{\|\mathbf{y}\|_q} Ai=∥x∥p∣xi∣,Bi=∥y∥q∣yi∣
由Young不等式
A i B i = ∣ x i y i ∣ ∥ x ∥ p ∥ y ∥ q ≤ 1 p ∣ x i ∣ p ∥ x ∥ p p + 1 q ∣ y i ∣ q ∥ y ∥ q q A_iB_i = \frac{\left|x_iy_i\right|}{\|\mathbf{x}\|_p\|\mathbf{y}\|_q}\le \frac{1}{p}\frac{\left|x_i\right|^p}{\|\mathbf{x}\|_p^p}+\frac{1}{q}\frac{\left|y_i\right|^q}{\|\mathbf{y}\|_q^q} AiBi=∥x∥p∥y∥q∣xiyi∣≤p1∥x∥pp∣xi∣p+q1∥y∥qq∣yi∣q
Sum of both sides
\sum_{i=1}^{n}\frac{\left|x_iy_i\right|}{\|\mathbf{x}\|_p\|\mathbf{y}\|_q} &\le\frac{ 1}{p}\frac{ \|\mathbf{x}\|_p^p}{\|\mathbf{x}\|_p^p}+\frac{1}{q}\frac{ \|\ mathbf{y}\|\mathbf{y}\|_q^q}\\ \sum_{i=1}^{n}\frac{\left|x_iy_i\right|}{\ |\mathbf{x}\|_p\|\mathbf{y}\|_q} &\le\fraction{1}{p}+\fraction{1}{q}\\ \sum_{i=1}^ {n}\left|x_iy_i\right| &\le \|\mathbf{x}\|_p \|\mathbf{y}\|_q \end{aligned}i=1∑n∥x∥p∥y∥q∣xiyi∣i=1∑n∥x∥p∥y∥q∣xiyi∣i=1∑n∥x∥p∥y∥q∣xiyi∣i=1∑n∥x∥p∥y∥q∣xiyi∣i=1∑n∣xiyi∣≤i=1∑np1∥x∥pp∣xi∣p+q1∥y∥qq∣yi∣q≤p1∥x∥pp∑i=1n∣xi∣p+q1∥y∥qq∑i=1n∣yi∣q≤p1∥x∥pp∥x∥pp+q1∥y∥qq∥y∥qq≤p1+q1≤∥x∥p∥y∥q
c 1 ∣ x i ∣ p = c 2 ∣ y i ∣ q c 1 ∑ i = 1 n ∣ x i ∣ p = c 2 ∑ i = 1 n ∣ y i ∣ q c 1 c 2 = ∥ y ∥ q q ∥ x ∥ p p \begin{aligned} c_1\left|x_i\right|^p &=c_2\left|y_i\right|^q\\ c_1\sum_{i=1}^{n}\left|x_i\right|^p &= c_2\sum_{i=1}^{n}\left|y_i\right|^q\\ \frac{c_1}{c_2} &=\frac{\|\mathbf{y}\|_q^q}{\|\mathbf{x}\|_p^p} \end{aligned} c1∣xi∣pc1i=1∑n∣xi∣pc2c1=c2∣yi∣q=c2i=1∑n∣yi∣q=∥x∥pp∥y∥qq
c 1 ∣ x i ∣ p = c 2 ∣ y i ∣ q ∣ y i ∣ = c 1 c 2 ∣ x i ∣ p q ∣ y i ∣ = ∥ y ∥ q ∥ x ∥ p p − 1 ∣ x i ∣ p − 1 ∣ x i y i ∣ = ∥ y ∥ q ∥ x ∥ p p − 1 ∣ x i ∣ p ∑ i = 1 n ∣ x i y i ∣ = ∑ i = 1 n ∥ y ∥ q ∥ x ∥ p p − 1 ∣ x i ∣ p ∑ i = 1 n ∣ x i y i ∣ = ∥ x ∥ p ∥ y ∥ q \begin{aligned} c_1\left|x_i\right|^p &=c_2\left|y_i\right|^q\\ \left|y_i\right| &= \sqrt[q]{\frac{c_1}{c_2}\left|x_i\right|^p}\\ \left|y_i\right| &= \frac{\|\mathbf{y}\|_q}{\|\mathbf{x}\|_p^{p-1}}\left|x_i\right|^{p-1}\\ \left|x_iy_i\right| &= \frac{\|\mathbf{y}\|_q}{\|\mathbf{x}\|_p^{p-1}}\left|x_i\right|^{p}\\ \sum_{i=1}^{n}\left|x_iy_i\right| &= \sum_{i=1}^{n}\frac{\|\mathbf{y}\|_q}{\|\mathbf{x}\|_p^{p-1}}\left|x_i\right|^{p}\\ \sum_{i=1}^{n}\left|x_iy_i\right| &= \|\mathbf{x}\|_p \|\mathbf{y}\|_q\\ \end{aligned} c1∣xi∣p∣yi∣∣yi∣∣xiyi∣i=1∑n∣xiyi∣i=1∑n∣xiyi∣=c2∣yi∣q=qc2c1∣xi∣p=∥x∥pp−1∥y∥q∣xi∣p−1=∥x∥pp−1∥y∥q∣xi∣p=i=1∑n∥x∥pp−1∥y∥q∣xi∣p=∥x∥p∥y∥q
dual norm of p norm
Let ∥ ⋅ ∥ \|\cdot \|∥⋅∥ is defined inR n \mathbb{R}^nRNorm of n
Dual norm∥ ⋅ ∥ ∗ \|\cdot \|_*∥⋅∥∗定义为
∥ z ∥ ∗ = sup ∥ x ∥ ≤ 1 z T x \|\mathbf{z}\|_*=\sup\limits_{\|x\|\le 1} \mathbf{z}^T\mathbf{x} ∥z∥∗=∥x∥≤1supzTx
It can also be derived that x T z ≤ ∥ x ∥ ∥ z ∥ ∗ \mathbf{x}^T\mathbf{z}\le \|\mathbf{x}\|\|\mathbf{z}\|_*xTz≤∥x∥∥z∥∗
p norm
∥ z ∥ q = sup ∥ x ∥ p ≤ 1 z T x \|\mathbf{z}\|_q=\sup\limits_{\|x\|_p\le 1} \mathbf{z}^T\mathbf{x} ∥z∥q=∥x∥p≤1supzT x
wherep ≥ 1 , q ≥ 1 , 1 p + 1 q = 1 p\ge 1, q\ge 1, \frac{1}{p}+\frac{1}{q}=1p≥1,q≥1,p1+q1=1
prove:
当z = 0 \mathbf{z}=\mathbf{0}z=0 , it is obviously established
当z ≠ 0 \mathbf{z}\neq \mathbf{0}z=Let 0
hold a function in
z T x ≤ ∑ i = 1 n ∣ zi ∣ ∣ xi ∣ ≤ ∥ x ∥ p ∥ z ∥ q ≤ ∥ z ∥ q \begin{aligned} \mathbf{z}^T\mathbf{x } \le \sum_{i=1}^{n}\left|z_i\right|\left|x_i\right|\le\|\mathbf{x}\|_p\|\mathbf{z}\|_q \le \|\mathbf{z}\|_q \end{aligned}zTx≤i=1∑n∣zi∣∣xi∣≤∥x∥p∥z∥q≤∥z∥q
Let's verify the equality. Let
xi = sign ( zi ) ∣ zi ∣ q − 1 ∥ z ∥ qqp x_i=\frac{\operatorname{sign}\left(z_i\right)\left|z_i\right|^{q -1}}{\|\mathbf{z}\|_q^{\frac{q}{p}}}xi=∥z∥qpqsign(zi)∣zi∣q−1
∥ x ∥ p p = ∑ i = 1 n ∣ sign ( z i ) ∣ z i ∣ q − 1 ∥ z ∥ q q p ∣ p = ∑ i = 1 n ∣ z i ∣ p q − p ∥ z ∥ q q = ∑ i = 1 n ∣ z i ∣ q ∥ z ∥ q q = ∥ z ∥ q q ∥ z ∥ q q = 1 \begin{aligned} \|\mathbf{x}\|_p^p&=\sum_{i=1}^n\left|\frac{\operatorname{sign}\left(z_i\right)\left|z_i\right|^{q-1}}{\|\mathbf{z}\|_q^{\frac{q}{p}}}\right|^p\\ &=\sum_{i=1}^n\frac{\left|z_i\right|^{pq-p}}{\|\mathbf{z}\|_q^q}\\ &=\sum_{i=1}^n\frac{\left|z_i\right|^{q}}{\|\mathbf{z}\|_q^q}\\ &=\frac{\|\mathbf{z}\|_q^q}{\|\mathbf{z}\|_q^q}\\ &=1\\ \end{aligned} ∥x∥pp=i=1∑n
∥z∥qpqsign(zi)∣zi∣q−1
p=i=1∑n∥z∥qq∣zi∣pq − p=i=1∑n∥z∥qq∣zi∣q=∥z∥qq∥z∥qq=1
So
∥ x ∥ p ≤ 1 \|\mathbf{x}\|_p\le 1∥x∥p≤1
因此
z T x = ∑ i = 1 n sign ( z i ) ∣ z i ∣ q − 1 ∥ z ∥ q q p z i = ∑ i = 1 n ∣ z i ∣ q ∥ z ∥ q q p = ∥ z ∥ q q ∥ z ∥ q q p = ∥ z ∥ q \begin{aligned} \mathbf{z}^T\mathbf{x} &=\sum_{i=1}^{n}\frac{\operatorname{sign}\left(z_i\right)\left|z_i\right|^{q-1}}{\|\mathbf{z}\|_q^{\frac{q}{p}}}z_i\\ &=\sum_{i=1}^{n}\frac{\left|z_i\right|^{q}}{\|\mathbf{z}\|_q^{\frac{q}{p}}}\\ &=\frac{\|\mathbf{z}\|_q^q}{\|\mathbf{z}\|_q^{\frac{q}{p}}}\\ &=\|\mathbf{z}\|_q \end{aligned} zTx=i=1∑n∥z∥qpqsign(zi)∣zi∣q−1zi=i=1∑n∥z∥qpq∣zi∣q=∥z∥qpq∥z∥qq=∥z∥q
conjugate
f : R n → R f:\mathbb{R}^n\to \mathbb{R} f:Rn→R
equation
f ∗ ( y ) = max xy T x − f ( x ) f^*\left(\mathbf{y}\right) = \max_{\mathbf{x}}\mathbf{y}^ T\mathbf{x}-f\left(\mathbf{x}\right)f∗(y)=xmaxyTx−f(x)
norm
f ( x ) = ∥ x ∥ f\left(\mathbf{x}\right) = \|\mathbf{x}\| f(x)=∥ x ∥
f ∗ ( y ) = { 0 , ∥ y ∥ ∗ ≤ 1 + ∞ , ∥ y ∥ ∗ > 1 f^*\left(\mathbf{y}\right)=\begin{cases} 0, & \|\mathbf{y}\|_*\le 1\\ +\infty, &\|\mathbf{y}\|_*>1 \end{cases}f∗(y)={
0,+∞,∥y∥∗≤1∥y∥∗>1
Among them ∥ ⋅ ∥ ∗ \|\cdot \|_*∥⋅∥∗is the dual norm
Proof:
When ∥ y ∥ ∗ ≤ 1 \|\mathbf{y}\|_*\le 1∥y∥∗≤1 hour
y T x ≤ ∥ x ∥ ∥ y ∥ ∗ ≤ ∥ the \|\mathbf{x}\|yTx≤∥x∥∥y∥∗≤∥x∥
取 x = 0 \mathbf{x}=\mathbf{0} x=0 ,the function
f ∗ ( y ) = y T x − ∥ x ∥ = 0 f^*\left(\mathbf{y}\right)=\mathbf{y}^T\mathbf{x}-\|\ mathbf{x}\|=0f∗(y)=yTx−∥x∥=0
当∥ y ∥ ∗ > 1 \|\mathbf{y}\|_*> 1∥y∥∗>1 is
defined by the dual norm
∥ y ∥ ∗ = max ∥ x ∥ ≤ 1 x T y > 1 \|\mathbf{y}\|_*=\max_{\|\mathbf{x}\|\le1 } \mathbf{x}^T\mathbf{y}>1∥y∥∗=max∥x∥≤1xTy>1 means that x \mathbf{x}
existsx ,letx ∥ ≤ 1 , x T y > 1 \|\mathbf{x}\|\le 1,\mathbf{x}^T\mathbf{y}>1∥x∥≤1,xTy>1
f ∗ ( y ) = max xy T x − ∥ x ∥ ≥ y T ( tx ) − t ∥ x ∥ = t ( y T x − ∥ x ∥ ) \begin{aligned} f^*\left(\mathbf {y}\right) &=\max_{\mathbf{x}} \mathbf{y}^T\mathbf{x}-\|\mathbf{x}\|\\ &\ge \mathbf{y}^ T\left(t\mathbf{x}\right)-t\|\mathbf{x}\|\\ &=t\left(\mathbf{y}^T\mathbf{x}-\|\mathbf{ x}\|\right)\\\end{aligned}f∗(y)=xmaxyTx−∥x∥≥yT(tx)−t∥x∥=t(yTx−∥x∥)
(The second line is because x \mathbf{x}x is arbitrary, so substitutetxt\mathbf{x}t x will also be established)
t → ∞ t\to \inftyt→∞时,f ∗ ( y ) → ∞ f^*\left(\mathbf{y}\right)\to \inftyf∗(y)→∞