It's not really all that difficult. When you bisect a network, such as a data center, the bisection bandwidth is the bandwidth usable to get from one side of the bisection to the other, possibly asymmetric values, depending on the topology. There may be links between the two sides which are unavailable for use due to things like STP blocking, and they don't count.
One definition of full bandwidth would be a 1:1 ratio between the aggregate bandwidth which the hosts in one section have to the available bandwidth between the two sections.
Another definition of full bandwidth is that there is no unavailable bandwidth between the two sections, regardless of the subscription ratio.
Over-subscription is where the aggregate bandwidth of the hosts in one section is greater than the available bandwidth to the other section.
Under-subscription is where the aggregate bandwidth of the hosts in one section is less than the available bandwidth to the other section.
Ideally, you would have full bandwidth between any two sections, no matter how you bisect the network. Practically, this is often not feasible, due to things like cost. Studies have been done on what level of over-subscription is acceptable at various points in a network.