Let X be a centred random vector in R^n. The L_p norms that X endows on R^n are defined by \|v\|_{L_p}= (E|<X,v>|^p)^{1/p}. The goal is to approximate those L_p norms, and the given data consists of N independent sample points X_1,...,X_N distributed as X. More accurately, one would like to construct data-dependent functionals \phi_{p,\epsilon} which satisfy with (very) high probability, that for every v in R^n, (1-\epsilon) \phi_{p,\epsilon} \leq E|<X,v>|^p \leq (1+\epsilon) \phi_{p,\epsilon}.
I will show that the functionals \frac{1}{N}\sum_{j \in J} |<X_j,v>|^p are a good choice, where the set of indices J is obtained from \{1,...,N\} by removing the c\eps^2 N largest values of |<X_j,v>|. Under mild assumptions on X, only N=(c^p)\epsilon^{-2} n measurements are required, and the probability that the functional performs well is at least 1-2\exp(-c\epsilon^2 N).