A new neural network for solvinga class of constrained least square problems is pre-sented. The network is shown to be completely stableand globally convergent to the exact solutions to theconstrained least square problem. In contrast to theneural network proposed in the Ref.[1], our new neu-ral network has the following advantages in two ma-jor aspects. First, the convergent region of this newnetwork is the whole space Rn. Second, in hardwareimplementations this new network does not need theexpensive analogue multiplier for variables.