学习教材主要是Boyd的《Convex Optimization》
Chapter 2 convex sets,主要是convex sets的一些数学定义,以及一些常见的convex sets。
2.1 Affine and convex sets
-
lines: y = θ x 1 + ( 1 − θ ) x 2 = x 2 + θ ( x 1 − x 2 ) y=\theta x_1+(1-\theta)x_2=x_2+\theta (x_1-x_2) y=θx1+(1−θ)x2=x2+θ(x1−x2), θ ∈ R \theta\in\mathbb{R} θ∈R, x 1 , x 2 ∈ R n , x 1 ≠ x 2 x_1,x_2\in\mathbb{R^n},x_1\neq x_2 x1,x2∈Rn,x1=x2
-
affine sets: x 1 , x 2 ∈ C x_1,x_2\in C x1,x2∈C, θ ∈ R \theta\in\mathbb{R} θ∈R, θ x 1 + ( 1 − θ ) x 2 ∈ C \theta x_1+(1-\theta)x_2\in C θx1+(1−θ)x2∈C
- affine hull: a f f C = { θ 1 x 1 + ⋯ + θ k x k ∣ x 1 , … , x k ∈ C , θ 1 + ⋯ + θ k = 1 } \mathbf{aff} C=\{\theta_1 x_1+\dots+\theta_k x_k|x_1,\ldots,x_k\in C, \theta_1+\dots+\theta_k=1\} affC={θ1x1+⋯+θkxk∣x1,…,xk∈C,θ1+⋯+θk=1}, affine dimension of C C C is the dimension of its affine hull
- relative interior: r e l i n t C = { x ∈ C ∣ B ( x , r ) ∩ a f f C ⊆ C for some r > 0 } \mathbf{relint}C=\{x\in C|B(x,r)\cap\mathbf{aff}C\subseteq C\;\text{for some}\;r>0\} relintC={x∈C∣B(x,r)∩affC⊆Cfor somer>0}, B ( x , r ) = { y ∣ ∥ y − x ∥ ≤ r } B(x,r)=\{y|\Vert y-x\Vert\leq r\} B(x,r)={y∣∥y−x∥≤r}
- relative boundary: c l C \ r e l i n t C \mathbf{cl}C\backslash \mathbf{relint}C clC\relintC,
-
convex sets: x 1 , x 2 ∈ C x_1,x_2\in C x1,x2∈C, 0 ≤ θ ≤ 1 0\leq\theta\leq 1 0≤θ≤1, θ x 1 + ( 1 − θ ) x 2 ∈ C \theta x_1+(1-\theta)x_2\in C θx1+(1−θ)x2∈C
- convex hull: c o n v C = { θ 1 x 1 + ⋯ + θ k x k ∣ x i ∈ C , θ i ≥ 0 , i = 1 , … , k , θ 1 + ⋯ + θ k = 1 } \mathbf{conv}C=\{\theta_1x_1+\dots+\theta_kx_k|x_i\in C, \theta_i\geq 0, i=1,\ldots,k,\theta_1+\dots+\theta_k=1\} convC={θ1x1+⋯+θkxk∣xi∈C,θi≥0,i=1,…,k,θ1+⋯+θk=1}
- p : R n → R p:\mathbb{R}^n\rightarrow\mathbb{R} p:Rn→R, p ( x ) ≥ 0 p(x)\geq 0 p(x)≥0, x ∈ C x\in C x∈C, C ⊆ R n C\subseteq\mathbb{R}^n C⊆Rn is convex, ∫ C p ( x ) d x = 1 \int_Cp(x)\mathrm{d}x=1 ∫Cp(x)dx=1, ∫ C p ( x ) x d x ∈ C \int_Cp(x)x\mathrm{d}x\in C ∫Cp(x)xdx∈C
-
cones: x ∈ C x\in C x∈C, θ ≥ 0 \theta\geq 0 θ≥0, θ x ∈ C \theta x\in C θx∈C
- convex cone: x 1 , x 2 ∈ C x_1,x_2\in C x1,x2∈C, θ 1 , θ 2 ≥ 0 \theta_1,\theta_2\geq 0 θ1,θ2≥0, θ 1 x 1 + θ 2 x 2 ∈ C \theta_1 x_1+\theta_2 x_2\in C θ1x1+θ2x2∈C
- conic hull: { θ 1 x 1 + ⋯ + θ k x k ∣ x i ∈ C , θ i ≥ 0 , i = 1 , … , k } \{\theta_1x_1+\dots+\theta_kx_k |x_i\in C, \theta_i\geq 0, i=1,\ldots,k\} {θ1x1+⋯+θkxk∣xi∈C,θi≥0,i=1,…,k}
affine → \rightarrow →convex
2.2 Some important examples
-
hyperplane: { x ∣ a T x = b } \{x|a^\mathrm{T}x=b\} {x∣aTx=b}, a ∈ R a\in\mathbb{R} a∈R, a ≠ 0 a\neq 0 a=0, b ∈ R b\in\mathbb{R} b∈R, affine set
-
halfspace: { x ∣ a T x ≤ b } \{x|a^\mathrm{T}x\leq b\} {x∣aTx≤b}, convex
-
Euclidean ball: B ( x c , r ) = { x ∣ ∥ x − x c ∥ 2 ≤ r } = { x c + r u ∣ ∥ u ∥ 2 ≤ 1 } B(x_c,r)=\{x|\Vert x-x_c\Vert_2\leq r\}=\{x_c+ru|\Vert u\Vert_2\leq 1\} B(xc,r)={x∣∥x−xc∥2≤r}={xc+ru∣∥u∥2≤1}, convex
- ellipsoid: E = { x ∣ ( x − x c ) T P − 1 ( x − x c ) ≤ 1 } = { x c + A u ∣ ∥ u ∥ 2 ≤ 1 } \mathcal{E}=\{x|(x-x_c)^\mathrm{T}P^{-1}(x-x_c)\leq 1\}=\{x_c+Au|\Vert u\Vert_2\leq 1\} E={x∣(x−xc)TP−1(x−xc)≤1}={xc+Au∣∥u∥2≤1} , convex
-
norm ball: { x ∣ ∥ x − x c ∥ ≤ r } \{x|\Vert x-x_c\Vert\leq r\} {x∣∥x−xc∥≤r}, convex; norm cone C = { ( x , t ) ∣ ∥ x ∥ ≤ t } ⊆ R n + 1 C=\{(x,t)|\Vert x\Vert\leq t\}\subseteq\mathbb{R}^{n+1} C={(x,t)∣∥x∥≤t}⊆Rn+1, convex
- second-order cone:
C = { ( x , t ) ∈ R n + 1 ∣ ∥ x ∥ 2 ≤ t } = { [ x t ] ∣ [ x t ] T [ I 0 0 − 1 ] [ x t ] ≤ 0 , t ≥ 0 } \begin{aligned} C&=\{(x,t)\in\mathbb{R}^{n+1}|\Vert x\Vert_2\leq t\}\\ &=\left\{\begin{bmatrix} x \\ t \end{bmatrix}\left |\begin{bmatrix} x \\ t \end{bmatrix}^\mathrm{T}\begin{bmatrix} I & 0 \\ 0 &-1 \end{bmatrix}\begin{bmatrix} x \\ t \end{bmatrix}\leq 0,\quad t\geq 0\right.\right\} \end{aligned} C={(x,t)∈Rn+1∣∥x∥2≤t}={[xt]∣∣∣∣∣[xt]T[I00−1][xt]≤0,t≥0}
- second-order cone:
-
polyhedra: P = { x ∣ a j T x ≤ b j , j = 1 , … , m , c j T x = d j , j = 1 , … , p } = { x ∣ A x ⪯ b , C x = d } \mathcal{P}=\{x|a_j^\mathrm{T}x\leq b_j,\;j=1,\ldots,m,\;c_j^\mathrm{T}x=d_j,\;j=1,\ldots,p\}=\{x|Ax\preceq b,Cx=d\} P={x∣ajTx≤bj,j=1,…,m,cjTx=dj,j=1,…,p}={x∣Ax⪯b,Cx=d}, convex
- k k k-dimensional simplexes: C = c o n v { v 0 , … , v k } = { θ 0 v 0 + ⋯ + θ k v k ∣ θ ⪰ 0 , 1 T θ = 1 } C=\mathbf{conv}\{v_0,\ldots,v_k\}=\{\theta_0v_0+\cdots+\theta_kv_k|\theta\succeq0,\mathbf{1}^\mathrm{T}\theta=1\} C=conv{v0,…,vk}={θ0v0+⋯+θkvk∣θ⪰0,1Tθ=1}, v 1 − v 0 , … , v k − v 0 v_1-v_0,\ldots,v_k-v_0 v1−v0,…,vk−v0 linearly independent
-
positive semidefinite cone: S + n = { X ∈ S n ∣ X ⪰ 0 } \mathbf{S}_+^n=\{X\in\mathbf{S}^n|X\succeq 0\} S+n={X∈Sn∣X⪰0}, the set of symmetric positive semidefinite matrices, where S n = { X ∈ R n × n ∣ X = X T } \mathbf{S}^n=\{X\in\mathbb{R}^{n\times n}|X=X^\mathrm{T}\} Sn={X∈Rn×n∣X=XT}.
2.3 Operations that preserve convexity
- intersection: if S 1 S_1 S1 and S 2 S_2 S2 are convex, then S 1 ⋂ S 2 S_1\bigcap S_2 S1⋂S2 is convex
- affine functions: if S ⊆ R n S\subseteq\mathbb{R}^n S⊆Rn is convex, f ( x ) = A x + b f(x)=Ax+b f(x)=Ax+b, then f ( S ) = { f ( x ) ∣ x ∈ S } f(S)=\{f(x)|x\in S\} f(S)={f(x)∣x∈S} and f − 1 ( S ) = { x ∣ f ( x ) ∈ S } f^{-1}(S)=\{x|f(x)\in S\} f−1(S)={x∣f(x)∈S} are convex
- perspective function: P ( z , t ) = z / t P(z,t)=z/t P(z,t)=z/t, t > 0 t>0 t>0. If C C C is convex, then P ( C ) = { P ( x ) ∣ x ∈ C } P(C)=\{P(x)|x\in C\} P(C)={P(x)∣x∈C} and P − 1 ( C ) = { ( x , t ) ∈ R n + 1 ∣ x / t ∈ C , t > 0 } P^{-1}(C)=\{(x,t)\in\mathbb{R}^{n+1}|x/t\in C, t>0\} P−1(C)={(x,t)∈Rn+1∣x/t∈C,t>0} are convex
- linear-fractional function: f ( x ) = ( A x + b ) / ( c T x + d ) f(x)=(Ax+b)/(c^\mathrm{T}x+d) f(x)=(Ax+b)/(cTx+d), c T x + d > 0 c^\mathrm{T}x+d>0 cTx+d>0. If C C C is convex, then f ( C ) f(C) f(C) and f − 1 ( C ) f^{-1}(C) f−1(C) are convex
2.4 Generalized inequalities
- proper cone: a cone
K
⊆
R
n
K\subseteq\mathbb{R}^n
K⊆Rn is convex, closed, solid and pointed.
- x ⪯ K y ⇔ y − x ∈ K x\preceq_Ky\Leftrightarrow y-x\in K x⪯Ky⇔y−x∈K; x ≺ K y ⇔ y − x ∈ i n t K x\prec_Ky\Leftrightarrow y-x\in\mathbf{int}K x≺Ky⇔y−x∈intK;
- minimum element: S ⊆ x + K S\subseteq x+K S⊆x+K; minimal element: ( x − K ) ⋂ S = { x } (x-K)\bigcap S=\{x\} (x−K)⋂S={x}
2.5 Separating and supporting hyperplanes
- separating hyperplane theorem: Suppose C C C and D D D are nonempty disjoint convex sets, i.e., C ⋂ D = ∅ C\bigcap D=\emptyset C⋂D=∅. Then there exist a ≠ 0 a\neq 0 a=0 and b b b such that a T x ≤ b a^\mathrm{T}x\leq b aTx≤b for all x ∈ C x\in C x∈C and a T x ≥ b a^\mathrm{T}x\geq b aTx≥b for all x ∈ D x\in D x∈D.
- supporting hyperplane: C ⊆ R n C\subseteq\mathbb{R}^n C⊆Rn, x 0 ∈ b d C = c l C \ i n t C x_0\in\mathbf{bd}C=\mathbf{cl}C\backslash\mathbf{int}C x0∈bdC=clC\intC, if a ≠ 0 a\neq 0 a=0 satisfies a T x ≤ a T x 0 a^\mathrm{T}x\leq a^\mathrm{T}x_0 aTx≤aTx0 for all x ∈ C x\in C x∈C, then the hyperplane { x ∣ a T x = a T x 0 } \{x|a^\mathrm{T}x=a^\mathrm{T}x_0\} {x∣aTx=aTx0} is called a supporting hyperplane to C C C at the point x 0 x_0 x0.
- supporting hyperplane theorem: for any nonempty convex set C C C and any x 0 ∈ b d C x_0\in\mathbf{bd}C x0∈bdC, there exists a supporting hyperplane to C C C at x 0 x_0 x0.
2.6 Dual cones and generalized inequalities
- dual cone: Let K K K be a cone, the set K ∗ = { y ∣ x T y ≥ 0 f o r a l l x ∈ K } K^*=\{y|x^\mathrm{T}y\geq 0\;{\rm for\;all}\;x\in K\} K∗={y∣xTy≥0forallx∈K} is called the dual cone of K K K.
- minimum element: x x x is the minimum element of S S S, with respect to the generalized inequality ⪯ K \preceq_K ⪯K, if and only if for all λ ≻ K ∗ 0 \lambda\succ_{K^*}0 λ≻K∗0, x x x is the unique minimizer of λ T z \lambda^\mathrm{T}z λTz over z ∈ S z\in S z∈S.
- minimal element: if
λ
≻
K
∗
0
\lambda\succ_{K^*}0
λ≻K∗0 and
x
x
x minimizes
λ
T
z
\lambda^\mathrm{T}z
λTz over
z
∈
S
z\in S
z∈S, then
x
x
x is minimal.
- If S S S is convex, for any minimal element x x x there exists a nonzero λ ⪰ K ∗ 0 \lambda\succeq_{K^*}0 λ⪰K∗0 such that x x x minimizes λ T z \lambda^\mathrm{T}z λTz over z ∈ S z\in S z∈S.