We define a numerical method that provides a non-parametric estimation of the kernelshape in symmetric multivariate Hawkes processes. This method relies on second orderstatistical properties of Hawkes processes that relate the covariance matrix of theprocess to the kernel matrix. The square root of the correlation function is computedusing a minimal phase recovering method. We illustrate our method on some examples andprovide an empirical study of the estimation errors. Within this framework, we analyzehigh frequency financial price data modeled as 1D or 2D Hawkes processes. We find slowlydecaying (power-law) kernel shapes suggesting a long memory nature of self-excitationphenomena at the microstructure level of price dynamics.