【CLP】Gaussian Projection Reduction & Randomized Binary Reduction

minuxAE發表於2021-01-05

Gaussian Projection Reduction

A randomized procedure to produce an approximate SDP solution with a desired low rank d d d. Again, let X ∗ \mathbf{X}^* X be an optimal solution of SDP with rank r > d r>d r>d and we factorize X ∗ \mathbf{X}^* X as
X ∗ = ( V ∗ ) T V ∗ , V ∗ ∈ E r × n \mathbf{X}^*=(\mathbf{V}^*)^T\mathbf{V}^*,\quad \mathbf{V}^*\in \mathbb{E}^{r\times n} X=(V)TV,VEr×n
We then generate i.i.d Gaussian random variables ξ i j \xi_i^j ξij with mean 0 and variance 1 / d 1/d 1/d, i = 1 , … , r ; j = 1 , … , d i=1,\dots, r; j=1,\dots, d i=1,,r;j=1,,d, and we define
X ^ = ( V ∗ ) T [ ∑ j = 1 d ξ j ( ξ j ) T ] V ∗ //rank is  d \hat{\mathbf{X}}=(\mathbf{V}^*)^T\bigg[\sum_{j=1}^d\xi^j(\xi^j)^T\bigg]\mathbf{V}^*\quad\text{//rank is $d$} X^=(V)T[j=1dξj(ξj)T]V//rank is d
Note that the rank of X ^ \hat{\mathbf{X}} X^ is d d d and
E ( X ^ ) = ( V ∗ ) T I V ∗ = X ∗ //approximate \mathbb{E}(\hat{\mathbf{X}})=(\mathbf{V}^*)^T\mathbf{I}\mathbf{V}^*=\mathbf{X}^*\quad\text{//approximate} E(X^)=(V)TIV=X//approximate
We can further show that X ^ \hat{\mathbf{X}} X^ would be a good rank- d d d approximate SDP solution in many cases.

Randomized Binary Reduction

As QP optimization, we want to produce a vector x \mathbf{x} x where each entry is either − 1 -1 1 or 1 1 1. A procedure to achieve this is as follows. Let X ∗ \mathbf{X}^* X be any optimal solution of SDP and we factorize X ∗ \mathbf{X}^* X as
X ∗ = ( V ∗ ) T V ∗ , V ∗ ∈ E n × n \mathbf{X}^*=(\mathbf{V}^*)^T\mathbf{V}^*,\quad \mathbf{V}^*\in\mathbb{E}^{n\times n} X=(V)TV,VEn×n
Then, we generate a random n n n-dimensional vector ξ \xi ξ where each entry is a i.i.d Gaussian random variable with mean 0 and variance 1. Define
x ^ = s g n ( ( V ∗ ) T ξ ) \hat{\mathbf{x}}=sgn((\mathbf{V}^*)^T\xi) x^=sgn((V)Tξ)
It was proved by Sheppard:
E [ x ^ i x ^ j ] = 2 π arcsin ⁡ ( X i j ∗ ) , i , j = 1 , 2 , … , n \mathbb{E}[\hat{x}_i\hat{x}_j]=\frac{2}{\pi}\arcsin(\mathbf{X}^*_{ij}),i,j=1,2,\dots, n E[x^ix^j]=π2arcsin(Xij),i,j=1,2,,n
Consider the primal homogeneous binary quadratic maximization problem
z ∗ : = max ⁡ x T Q x s . t . x j = { 1 , − 1 } , j = 1 , … , n z^*:=\max \mathbf{x}^T\mathbf{Q}\mathbf{x}\\ s.t.\quad x_j=\{1, -1\}, j=1,\dots, n z:=maxxTQxs.t.xj={1,1},j=1,,n
where we assume Q \mathbf{Q} Q is positive semidefinite. Then, the SDP relaxation would be
z S D P : = max ⁡ Q ⋅ X s . t . I j ⋅ X = 1 , j = 1 , … , n X ∈ S + n z^{SDP}:=\max \mathbf{Q}\cdot \mathbf{X}\\ s.t.\quad \mathbf{I}_j\cdot \mathbf{X}=1, j=1,\dots, n\\ \mathbf{X}\in \mathcal{S}_+^n zSDP:=maxQXs.t.IjX=1,j=1,,nXS+n
and let X ∗ \mathbf{X}^* X be any optimal solution, from which we produced a random binary vector x ^ \hat{\mathbf{x}} x^. And, we evaluate the expected objective value
E ( x ^ Q x ^ ) = E ( Q ⋅ x ^ x ^ T ) = Q ⋅ E ( x ^ x ^ T ) = Q ⋅ 2 π arcsin ⁡ [ X ∗ ] = 2 π ( Q ⋅ arcsin ⁡ [ X ∗ ] ) \mathbb{E}(\hat{\mathbf{x}}\mathbf{Q}\hat{\mathbf{x}})=\mathbb{E}(\mathbf{Q}\cdot\hat{\mathbf{x}}\hat{\mathbf{x}}^T)=\mathbf{Q}\cdot\mathbb{E}(\hat{\mathbf{x}}\hat{\mathbf{x}}^T)=\mathbf{Q}\cdot\frac{2}{\pi}\arcsin[\mathbf{X}^*]=\frac{2}{\pi}(\mathbf{Q}\cdot\arcsin[\mathbf{X}^*]) E(x^Qx^)=E(Qx^x^T)=QE(x^x^T)=Qπ2arcsin[X]=π2(Qarcsin[X])
where arcsin ⁡ [ X ∗ ] ∈ S n \arcsin[\mathbf{X}^*]\in\mathcal{S}^n arcsin[X]Sn, and further
arcsin ⁡ [ X ∗ ] − X ∗ ⪰ 0 \arcsin[\mathbf{X}^*]-\mathbf{X}^*\succeq \bf{0} arcsin[X]X0
from Q ⪰ 0 \mathbf{Q}\succeq \bf{0} Q0, we have
Q ⋅ arcsin ⁡ [ X ∗ ] ≥ Q ⋅ X ∗ = z S D P ≥ z ∗ \mathbf{Q}\cdot\arcsin[\mathbf{X}^*]\geq \mathbf{Q}\cdot \mathbf{X}^*=z^{SDP}\geq z^* Qarcsin[X]QX=zSDPz

相關文章