Semidefinite Relaxation codes for the discrete Integer Least Squares problem

 
 

About

Consider the following linear model:

system model

where

  • H is a real or complex mixing matrix of dimensions m by n
  • s is a state vector of dimensions n by 1
  • y is an observation vector of dimensions m by 1
  • v is a real or complex random vector with Gaussian, zero-mean, unit variance entries
  • rho is quality of observation y (signal-to-noise ratio)

Integer Least Squares problem (Maximum-Likelihood detection) is given by:

maximum-likelihood detection problem

where C^n is a given discrete set (constellation). Integer Least Squares problem is, in general, (strongly) NP-hard. Exhaustive search over all possible vectors s from C^n can be used to solve the problem with exponential complexity. In many practical scenarios a suboptimal strategy with reduced complexity is desired. This website is dedicated to efficient implementations of suboptimal strategies based on semi-definite relaxation of the original ML problem. The software presented on the website delivers near-ML probability of error with polynomial running time.

SDR Detector is based on a convex semifinite relaxation of the Integer Least Squares problem for binary and lattice constellations (BPSK, QPSK, (2^M)-PAM, (2^2M)-QAM, M > 0) and provides the following guarantees:

  • theoretically proven worst-case polynomial complexity: O(n^3.5)
  • exact solution with high probability
  • constant factor approximation algorithm for the Integer Least Squares problem
  • exact solution for a given channel realization with sufficiently high signal-to-noise ratio

PSK Detector is based on a low-rank (non-convex) semidefinite relaxation of the Integer Least Squares problem for binary, circular and lattice constellations (BPSK, QPSK, M-PSK, (2^2M)-QAM) and provides the following guarantees:

  • theoretically proven worst case quadratic complexity: O(n^2)
  • exact solution with high probability
  • every local minimizer of the relaxed problem is 0.5-minimizer for M-PSK, M > 2 constellations
  • every local minimizer is the global minimizer for BPSK (2-PSK) constellation