https://github.com/fchollet/keras/issues/4149

And refer to Traditional LSTM section of the wiki page, where all are vectors and matrices in the equatioins.
I guess this is why an LSTM layer is considered as a "cell", a fully connected bunch of nodes.

https://en.wikipedia.org/wiki/Long_short-term_memory#History


Therefore, the input vector size can be different from the hidden layer size and the output vector size is equal to the hidden layer size.