Every time a new block is added on top of block chain, the miners have to restart their work because the next block has to have a proper reference to previous block.
Let's suppose that there is some nonce for each block such that there is also sufficiently small hash (smaller than target). In general, there are 2256 possible hashes. Let the target be t. The target can be also understand as a number of acceptable hashes. So there is t / 2256 probability to find a proper hash in each try i.e. to find a block.
The number of unsuccessful tries before a block follows a geometrical distribution with parameter p=t/2256. The expected value of a variable following such a distribution is EX = 1/p = 2256/t. So, each mining pool has to spend 2256/t tries in average to find a block.
How is it that concurrent mining can be efficient, providing that each time some pool publishes a new block all other pools have to restart their work and thus throw out their tries on blocks that now can't be used any more?
Note: Please be a bit detailed. I've already read explanations like: Every try has equal chance to success. But I can't get it from such short hints.