Abstract: In order to solve the problem of limited snapshots in the estimation of source number in the case of large antenna array, a subspace averaging algorithm based on Toeplitz optimization is proposed in this paper. To address the inherent limitations of conventional subspace averaging algorithms that solely rely on forward smoothing for sub-covariance matrix acquisition, this paper proposes a Toeplitz-constrained structural refinement framework for forward-backward smoothed sub-covariance matrices. Firstly, the forward /backward smoothing algorithm is introduced to obtain the forward/backward smoothing sub-covariance matrix. The bidirectional smoothing sub-covariance matrix is summed after Toeplitz optimization, and is used as the sub-covariance matrix of multiple subspaces. Secondary, the discrete probability distribution is established through eigenvalue decomposition, where the subspace ensemble is constructed via a non-replacement stochastic sampling mechanism. Experimental results demonstrate that the enhanced algorithm achieves superior performance compared to conventional approaches under limited snapshots and low signal-to-noise ratio (SNR) conditions. Notably, the performance gain becomes particularly pronounced in large-scale array configurations, while maintaining robust estimation capabilities in moderate-scale array scenarios.