Monotonic Optimization Method for Achieving the Maximum Weighted Sum-Rate in Multicell Downlink MISO Systems

Abstract

A monotonic optimization method is for achieving the maximum weighted sum-rate in multicell downlink MISO systems. It belongs to the wireless communications field. The weighted sum-rate maximization is viewed as a monotonic optimization problem over the achievable rate region. A sensible search scheme, a sequential partition method as well as a vertex relocation method are designed to reduce the complexity and accelerate the convergence.

Description

BACKGROUND OF THE PRESENT INVENTION

1. Field of Invention

The present invention describes a monotonic optimization method for achieving the maximum weighted sum-rate in multicell downlink MISO systems, which is a technique of wireless communications.

2. Description of Related Arts

Coordinated multicell downlink transmission is a promising technique for suppressing co-channel interference and improving the system performance. Among the different kinds of design criteria, weighted sum-rate maximization is an active research topic, which has attracted significant attention from both academics and industries. However, achieving the global optimum is very difficult as the capacity region remains unknown. A more pragmatic method that adopts single-user detection by simply treating interference as noise has stimulated some efficient algorithms.

In the conventional techniques, the reference L. Venturino, N. Prasad, and X. Wang, “Coordinated linear beamforming in downlink multi-cell wireless networks,” IEEE Transactions on Wireless Communications, vol. 9, no. 4, pp. 1451-1461, April 2010, presents an iterative coordinated beamforming algorithm to solve the Karush-Kuhn-Tucker conditions of the weighted sum-rate maximization problem.

The reference Q. J. Shi, M. Razaviyayn, Z. Q. Luo, and C. He, “An iteratively weighted MMSE approach to distributed sum-utility maximization for a MIMO interfering broadcast channel,” IEEE Transactions on Signal Processing, vol. 59, no. 9, pp. 4331-4340, September 2011, proposes a weighted minimum mean square error method which iteratively update the transmitter, the receiver and the weight matrix.

However, the achievable rate region under single-user detection remains non-convex. For MISO systems, the reference R. Zhang and S. G. Cui, “Cooperative interference management with MISO beamforming,” IEEE Transactions on Signal Processing, vol. 58, no. 10, pp. 5454-5462, October 2010, proves that beamforming techniques can achieve the Pareto-optimal points and describes a method to characterize the Pareto-boundary of the achievable rate region. Although there are already some techniques obtaining Pareto-optimal solutions, it is still a big challenge to search for the global optimality.

The reference L. Liu, R. Zhang, and K. Chua, “Achieving global optimality for weighted sum-rate maximization in the K-user Gaussian interference channel with multiple antennas,” IEEE Transactions on Wireless Communications, vol. 11, no. 5, pp. 1933-1945, 2012, views the weighted sum-rate maximization as a monotonic optimization problem and adopts the outer polyblock approximation algorithm combined with the rate profile technique to approaching the global optimal solution. To improve the efficiency of the monotonic optimization, another reference E. Bjornson, G. Zheng, M. Bengtsson, and B. Ottersten, “Robust monotonic optimization framework for multicell MISO systems,” IEEE Transactions on Signal Processing, vol. 60, no. 5, pp. 2508-2523, 2012, provides a branch-reduce-and-bound based method. Although it converges faster than the outer polyblock approximation algorithm, it still needs too many iterations as the number of users increase.

SUMMARY OF THE PRESENT INVENTION

Based on the conventional techniques, the present invention proposes a novel monotonic optimization method to achieve the maximum weighted sum-rate in multicell downlink MISO systems. The present invention utilizes a sensible search scheme, a sequential partition method and a vertex relocation method to accelerate the convergence.

The present invention is implemented as the following steps of:

Step 1: setting system parameters comprising: a cell/base station number M, a user number in an m-th cell K_m, an antenna number of an m-th base station T_m, a maximum transmission power of the m-th base station P_m, wherein m=1, a 1 by T_n channel vector from a n-th base station to an m_k-th user (i.e. the k-th user in the m-th cell) h_mk,n, wherein m,n=1, . . . , M, k=1, . . . , K_m, a variance of a zero-mean complex Gaussian additive noise at the m_k-th user σ_mk², a weight of the m_k-th user α_mk, wherein m=1, . . . , M, k=1, . . . , K_m;

Step 2: defining R_mk as a rate of the m_k-th user,

$R_{m_k}={\log }_2\left(1+\frac{{h_{m_k,m}w_{m_k}}^2}{\sum _{\left(n,j\right)\ne \left(m,k\right)}{h_{m_k,n}w_{n_j}}^2+{\sigma }_{m_k}^2}\right)$

wherein w_mk is the T_m by 1 beamformer for the m_k-th user, wherein m=1, . . . , M, k=1, . . . , K_m;

formulating a achievable rate vector

$r=\left(R_{1_1},\dots ,R_{1_{K_1}},\dots ,R_{M_1},\dots ,R_{M_{K_M}}\right)\in {ℝ}_{+}^K$

as well as a weighted sum-rate function

$f\left(r\right)=\sum _{m,k}{\alpha }_{m_k}R_{m_k},$

wherein ε₊^K indicates that r is a positive real vector of a length K with K=Σ_m=1^MK_m;

Step 3: defining a box [a,b]={xε₊^K|a≦x≦b}; initializing the set of boxes as ={[0,b₀]}; assuming K₀=0, a (Σ_i=0^m−1K_i+k)-th element (associated with the m_k-th user) of b₀ is log₂(1+P_mνh_mk,m^H∥²/σ_mk²), wherein m=1, . . . , k=1, . . . , K_m; setting a termination accuracy η and a line search accuracy δ;

Step 4: initializing an upper bound f_max and a lower bound f_min of the weighted sum-rate, wherein

f _max =f(b₀), f_min=max(α·b₀)

wherein α is a weight vector comprising weights for all the users, i.e.

$\alpha =\left({\alpha }_{1_1},\dots ,{\alpha }_{1_{K_1}},\dots ,{\alpha }_{M_1},\dots ,{\alpha }_{M_{K_M}}\right)\in {ℝ}_{+}^K;$

Step 5: choosing a box [a,b] from that satisfies f(b)=f_max, and then checking feasibility of a: whether a locates in a achievable rate region or not; wherein the feasibility is determined via a problem φ(a);

the specified problem φ(a) is:

• • maximize 0
  • subject to √{square root over (β_mk)}∥A_mkx+n_mk∥≦√{square root over (1+β_mk)}(h_mk,mS_mkx),∀m,k
  • p^Tx=0
  • ∥G_mx∥≦√{square root over (P_m)},∀mwith the following notations

$x_m={\left[w_{m_1}^H,\dots ,w_{m_{K_m}}^H\right]}^H,m=1,\dots ,M,x={\left[x_1^H,\dots ,x_M^H,0\right]}^H,n_{m_k}={\left[0,\dots 0,{\sigma }_{m_k}\right]}^T,S_{m_k}=\left[00\dots I_{T_m}\dots 0\right],A_{m_k}=\mathrm{diag}\left(\underset{\underset{K_1}{}}{h_{m_k,1},\dots ,h_{m_k,1}},\dots ,\underset{\underset{K_M}{}}{h_{m_k,M},\dots ,h_{m_k,M}},0\right),G_m={\left[S_{m_1}^H\dots S_{m_{K_m}}^H\right]}^H,{\beta }_{m_k}=2^{a_{m_k}}-1,p={\left[0,\dots 0,1\right]}^T$

wherein σ_mk is a standard deviation of the zero-mean complex Gaussian additive noise at the m_k-th user, I_Tm denotes a identity matrix with dimension T_m;

Step 6: if the problem φ(a) is feasible, conducting a sensible search scheme for the box [a,b] to obtain a partition point r;

wherein if the problem φ(a) is infeasible, updating the box set as =\[a,b] and calculating the upper bound f_max=max_[a,b]εf(b), then going back to the Step 5;

the sensible search scheme is:

denoting l_ab as the line connecting a and b; finding the intersection point c on the hyperplane {r|f(r)=f_min} with the line l_ab, i.e.

$c=a+\left(b-a\right)\times \frac{f_{\min }-f\left(a\right)}{f\left(b-a\right)}$

checking feasibility of c via the problem φ(a) in the Step 5 with a=c;

wherein if φ(c) is feasible, a bisection line search is conducted along the line l_cb to obtain an intersection point on a Pareto-boundary; given a line search accuracy δ, two points r_min and r_max are acquired; the partition point is set as r=r_max while the lower bound is updated as f_min=f(r_min);

if φ(c) is infeasible, just set the partition point as r=c;

Step 7: based on the partition point r, dividing the box [a,b] into K new boxes [a⁽ⁱ⁾,b⁽ⁱ⁾], i=1, . . . , K using a sequential partition method; then updating the boxes set as

$=\mathrm{\backslash [}a,b]\bigcup \left\{\bigcup _{i=1,\dots ,K}\left[a^{\left(i\right)},b^{\left(i\right)}\right]\right\}$

the sequential partition method is:

i) firstly generating a set of K new vertices {b⁽¹⁾, . . . , b^(K)} based on the partition point r, wherein

b ⁽ⁱ⁾ =b−(b_i−r_i)e_i, i=1, . . . ,K

wherein the subscript i indicates the i-th element of the vector and e_i is a vector with the i-th element being 1 and the others being 0;

ii) sorting the K vertices as {b⁽ⁱ¹⁾, . . . , b^(iK)} in ascending order of the achievable weighted sum-rate, wherein i_s denotes an original index of the s-th vertex;

ill) sequentially determining the corresponding vertices {a⁽ⁱ¹⁾, . . . , a^(iK)} as

$a^{\left(i_s\right)}=\{\begin{array}{cc}a,& s=1\\ a^{\left(i_{s-1}\right)}+\left(r_{i_{s-1}}-a_{i_{s-1}}\right)e_{i_{s-1}},& s>1\end{array};$

Step 8; calculating f(b) for each box [a,b];

wherein if f(b)>f_min, the associated vertex a is relocated as

$a_i=b_i-\min \left(\frac{f\left(b\right)-f_{\min }}{{\alpha }_i\left(b_i-a_i\right)},1\right)\times \left(b_i-a_i\right),i=1,\dots ,K$

wherein, ã_i is an i-th element of the relocated vertex;

if f(b)≦f_min, the box [a,b] is removed from the box set, i.e. =\[a,b];

Step 9: resetting the upper bound f_max as f_max=max_[a,b]εf(b); and

Step 10: checking a relative error of the upper and lower bound;

• • wherein if (f_max−f_min)/f_min>η, going back to the Step 5, otherwise, returning f_max and r_min

The monotonic optimization method proposed in the present invention adopts a sensible search scheme, a sequential partition method and a vertex relocation method to reduce the number of checking feasibility, therefore the computational complexity is decreased and the convergence is accelerated.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 is an illustration of an iteration for K=3;

FIG. 2 is a diagram of comparison of convergence performance of the embodiment of the present invention, the conventional outer polyblock approximation method and branch-reduce-and-bound method in the scenario wherein M=3, K_m=1, P_m=5, T_m=4;

FIG. 3 is a diagram of comparison of convergence performance of the embodiment of the present invention, the conventional outer polyblock approximation method and branch-reduce-and-bound method in the scenario wherein M=2, K_m=2, P=10, T_m=4

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENT

The details of the embodiment of the present invention are specified below. The embodiment is based on the technical content of the present invention. The specific implementing process is presented. The scope of the present invention is not limited to the following embodiment.

A weighted sum-rate maximization problem is

$\mathrm{maximize}\sum _{m,k}{\alpha }_{m_k}R_{m_k}=\sum _{m,k}{\alpha }_{m_k}{\log }_2\left(1+\frac{{h_{m_k,m}w_{m_k}}^2}{\sum _{\left(n,j\right)\ne \left(m,k\right)}{h_{m_k,n}w_{n_j}}^2+{\sigma }_{m_k}^2}\right)$ $\mathrm{subject}\mathrm{to}\sum _{k=1}^{K_m}{w_{m_k}}^2\le P_m,\forall m$

wherein K_m is a number of users in an m-th cell, T_m is an antenna number of an m-th base station, P_m denotes a maximum transmission power of the m-th base station, h_mk,n, is channel vector from a n-th base station to an m_k-th user, σ_mk² represents a variance of a zero-mean complex Gaussian additive noise at the m_k-th user and α_mk is a weight of the m_k-th user, wherein m,n=1, . . . , M, k=1, . . . , K_m;

The monotonic optimization problem is

$\mathrm{maximize}f\left(r\right):=\sum _{m,k}{\alpha }_{m_k}R_{m_k}$ $\mathrm{subject}\mathrm{to}r\in \Re$

wherein

$r=\left(R_{1_1},\dots ,R_{1_{K_1}},\dots ,R_{M_1},\dots ,R_{M_{K_M}}\right)\in {ℝ}_{+}^K$

is the achievable rate vector and denotes the achievable rate region, i.e.

$\Re =\bigcup _{\sum _{k=1}^{K_m}{w_{m_k}}^2\le P_m,\forall m}\left\{\left(r_{1_1},\dots ,r_{M_{K_M}}\right)|0\le r_{m_k}\le R_{m_k},\forall m,k\right\}$

The present invention includes the following steps of:

Step 1: setting system parameters comprising: a cell/base station number M, the user number in the m-th cell K_m, the antenna number of the m-th base station T_m, the maximum transmission power of the m-th base station P_m, wherein m=1, . . . , M, a 1 by T_n channel vector from the n-th base station to the m_k-th user (i.e. the k-th user in the m-th cell) h_mk,n, wherein m,n=1, . . . , M, k=1, . . . , K_m, the variance of the zero-mean complex Gaussian additive noise at the m_k-th user σ_mk², the weight of the m_k-th user α_mk, wherein m=1, . . . , M, k=1, . . . , K_m;

In this embodiment, two scenarios are considered: M=3, K_m=1, P_m=5, T_m=4 and M=2, K_m=2, P_m=10, T_m=4, wherein m=1, . . . , M.

In this embodiment, every entry of h_mk,n is a zero-mean unit-variance complex Gaussian random variable, σ_mk²=1, α_mk=1, wherein m,n=1, . . . , M, k=1, . . . , K_m.

Step 2: defining R_mk as a rate of the m_k-th user,

$R_{m_k}={\log }_2\left(1+\frac{{h_{m_k,m}w_{m_k}}^2}{\sum _{\left(n,j\right)\ne \left(m,k\right)}{h_{m_k,n}w_{n_j}}^2+{\sigma }_{m_k}^2}\right)$

wherein W_mk is the T_m by 1 beamformer for the m_k-th user, wherein m=1, . . . , M, k=1, . . . , K_m;

formulating a achievable rate vector

$r=\left(R_{1_1},\dots ,R_{1_{K_1}},\dots ,R_{M_1},\dots ,R_{M_{K_M}}\right)\in {ℝ}_{+}^K$

as well as a weighted sum-rate function

$f\left(r\right)=\sum _{m,k}{\alpha }_{m_k}R_{m_k},$

wherein ε₊^K indicates that r is a positive real vector of a length K with K=Σ_m=1^MK_m;

Step 3: defining a box [a,b]={xε₊^K|a≦x≦b}; initializing the set of boxes as ={[0,b₀]}; assuming K₀=0, a (Σ_i=0^m−1K_i+k)-th element (associated with the m_k-th user) of b₀ is log₂(1+P_m∥h_mk,m^H∥²/σ_mk²), wherein m=1, . . . , k=1, . . . , K_m; setting a termination accuracy η and a line search accuracy δ;

In this embodiment, the accuracy parameters are set as η=0.01, δ=0.01.

Since the iterations of the embodiment and the conventional methods are different, it is not meaningful to compare the number of iterations. As the main complexity derives from checking the feasibility of a point, the convergence performance is evaluated as a function of the feasibility checking times. The convergence performance is presented via the relative errors of the upper and lower bound with the optimal value, i.e.

(f_max−f_opt)/f_opt,(f_min−f_opt)/f_opt

wherein f_opt is the obtained optimal value of the weighted sum-rate;

Step 4: initializing an upper bound f_max and a lower bound f_min of the weighted sum-rate, wherein

f _max =f(b₀),f_min=max(α·b₀)

wherein α is a weight vector comprising weights for all the users, i.e.

$\alpha =\left({\alpha }_{1_1},\dots ,{\alpha }_{1_{K_1}},\dots ,{\alpha }_{M_1},\dots ,{\alpha }_{M_{K_M}}\right)\in {ℝ}_{+}^K;$

Step 5: choosing a box [a,b] from that satisfies f(b)=f_max, and then checking feasibility of a: whether a locates in a achievable rate region or not; wherein the feasibility is determined via a problem φ(a);

the specified problem φ(a) is:

• • maximize 0
  • subject to √{square root over (β_mk)}∥A_mkx+n_mk∥≦√{square root over (1+β_mk)}(h_mk,mS_mkx),∀m,k
  • p^Tx=0
  • ∥G_mx∥≦√{square root over (P_m)},∀mwith the following notations

$x_m={\left[w_{m_1}^H,\dots ,w_{m_{K_m}}^H\right]}^H,m=1,\dots ,M,x={\left[x_1^H,\dots ,x_M^H,0\right]}^H,n_{m_k}={\left[0,\dots 0,{\sigma }_{m_k}\right]}^T,S_{m_k}=\left[00\dots I_{T_m}\mathrm{\dots 0}\right],A_{m_k}=\mathrm{diag}\left(\underset{\underset{K_1}{}}{h_{m_k,1},\dots ,h_{m_k,1}},\dots ,\underset{\underset{K_M}{}}{h_{m_k,M},\dots ,h_{m_k,M}},0\right),G_{m_1}={\left[S_{m_1}^H\dots S_{m_{K_m}}^H\right]}^H,{\beta }_{m_k}=2^{a_{m_k}}-1,p={\left[0,\dots 0,1\right]}^T$

wherein σ_mk is a standard deviation of the zero-mean complex Gaussian additive noise at the m_k-th user, I_Tm denotes a identity matrix with dimension T_m;

Step 6: if the problem φ(a) is feasible, conducting a sensible search scheme for the box [a,b] to obtain a partition point r;

wherein if the problem φ(a) is infeasible, updating the box set as =\[a,b] and calculating the upper bound f_max=max_[a,b]εf(b), then going back to the Step 5;

the sensible search scheme is:

denoting l_ab as the line connecting a and b; finding the intersection point c on the hyperplane {r|f(r)=f_min} with the line l_ab, i.e.

$c=a+\left(b-a\right)\times \frac{f_{\min }-f\left(a\right)}{f\left(b-a\right)}$

checking feasibility of c via the problem φ(a) in the Step 5 with a=c;

wherein if φ(c) is feasible, a bisection line search is conducted along the line l_cb to obtain an intersection point on a Pareto-boundary; given a line search accuracy δ, two points r_min and r_max are acquired; the partition point is set as r=r_max while the lower bound is updated as f_min=f(r_min);

if φ(c) is infeasible, just set the partition point as r=c;

Step 7: based on the partition point r, dividing the box [a,b] into K new boxes [a⁽ⁱ⁾,b⁽ⁱ⁾], i=1, . . . , K using a sequential partition method; then updating the boxes set as

$=\mathrm{\backslash [}a,b]\bigcup \left\{\bigcup _{i=1,\dots ,K}\left[a^{\left(i\right)},b^{\left(i\right)}\right]\right\}$

the sequential partition method is:

i) firstly generating a set of K new vertices {b⁽¹⁾, . . . ,b^(K)} based on the partition point r, wherein

b ⁽ⁱ⁾ =b−(b_i−r_i)e_i, i=1, . . . ,K

wherein the subscript i indicates the i-th element of the vector and e_i is a vector with the i-th element being 1 and the others being 0;

ii) sorting the K vertices as {b⁽ⁱ¹⁾, . . . , b^(iK)} in ascending order of the achievable weighted sum-rate, wherein i_s denotes an original index of the s-th vertex;

iii) sequentially determining the corresponding vertices {a⁽ⁱ¹⁾, . . . , a^(iK)} as

$a^{\left(i_s\right)}=\{\begin{array}{cc}a,& s=1\\ a^{\left(i_{s-1}\right)}+\left(r_{i_{s-1}}-a_{i_{s-1}}\right)e_{i_{s-1}},& s>1\end{array};$

Step 8; calculating f(b) for each box [a,b];

wherein if f(b)>f_min, the associated vertex a is relocated as

${\tilde{a}}_i=b_i-\min \left(\frac{f\left(b\right)-f_{\min }}{{\alpha }_i\left(b_i-a_i\right)},1\right)\times \left(b_i-a_i\right),i=1,\dots ,K$

wherein ã_i is an i-th element of the relocated vertex;

if f(b)≦f_min the box [a,b] is removed from the box set, i.e. =\[a,b];

Step 9: resetting the upper bound f_max as f_max=max_[a,b]εf(b); and Step 10: checking a relative error of the upper and lower bound;

wherein if (f_max−f_min)/f_min>η, going back to the Step 5, otherwise, returning f_min, f_max and r_min.

FIG. 2 compares the convergence performance of the embodiment of the present invention, the conventional outer polyblock approximation method (refer to Polyblock in figure) and branch-reduce-and-bound method (refer to BRB in figure) in the scenario wherein M=3, K_m=1, P_m=5, T_m=4;

FIG. 3 compares the convergence performance of the embodiment of the present invention, the conventional outer polyblock approximation method (refer to Polyblock in figure) and branch-reduce-and-bound method (refer to BRB in figure) in the scenario wherein M=2, K_m=2, P_m=10, T_m=4.

From FIGS. 2 and 3, it is observed that the embodiment of the present invention converges and achieves the global optimal solution faster compared with the conventional methods.

While only selected embodiment has been chosen to illustrate the present invention, it will be apparent to those skilled in the art from this disclosure that various changes and modifications can be made herein without departing from the scope of the present invention.

Claims

1. A monotonic optimization method for achieving a maximum weighted sum-rate in multicell downlink MISO systems, wherein the method utilizes a sensible search scheme, a sequential partition method and a vertex relocation method, specifically comprising steps of: Step 1: setting system parameters comprising: a cell/base station number M, a user number in an m-th cell Km, an antenna number of an m-th base station Tm, a maximum transmission power of the m-th base station Pm, wherein m=1, . . . , M, a 1 by Tn channel vector from a n-th base station to an mk-th user hmk,n, wherein m,n=1, . . . , M, k=1, . . . , Km, a variance of a zero-mean complex Gaussian additive noise at the mk-th user σmk2 a weight of the mk-th user αmk, wherein m=1, . . . , M, k=1, . . . , Km;Step 2: defining Rmk as a rate of the mk-th user,

2. The monotonic optimization, as recited in claim 1, wherein the problem φ(a) in the Step 5 is: maximize 0subject to √{square root over (βmk)}∥Amkx+nmk∥≦√{square root over (1+βmk)}(hmk,mSmkx),∀m,kpTx=0∥Gmx∥≦√{square root over (Pm)},∀m

3. The monotonic optimization, as recited in claim 1, wherein the sensible search scheme in the Step 6 is: denoting lab as the line connecting a and b; finding the intersection point c on the hyperplane {r|f(r)=fmin} with the line lab,

4. The monotonic optimization, as recited in claim 1, wherein the sequential partition method in the Step 7 comprising steps of i) firstly generating a set of K new vertices {b(1), . . . , b(K)} based on the partition point r, wherein b(i)=b−(bi−ri)ei, i=1, . . . ,K