On Worst-case Regret of Linear Thompson Sampling