Let’s start with this this (stolen) image of a multitree:
The nodes (letters from A to U) are the transaction records. Let’s say ours is G, and only A, B, and C exist at the moment.
So, I want to add a record of my transaction (G) to the ledger. I attach the original details, but I also add a reference to B and C. I sign it and I throw it in the pool. With this:
- I proved that this transaction was added to the ledger after both B and C existed
- I proved that both B and C existed before G, which I’m aware is the same thing kinda, but it is semantically different
In the meantime came a couple other transactions: D, E, … H – they also signed some or all of the existing three. and this further increases our assurance about their place in the timeline.
With enough references between the nodes, it becomes very hard to pretend something was not where everybody else is saying it was.
When M and L arrives, my G is getting pinned down from the future as well, not just the past. Then comes P and Q and R, and S and T and U, and now good luck if somebody’s trying to say G didn’t come before M: nobody can disregard this many independent (as we want to believe they are) proofs.
Anyway, it’s a graph, in general, and a multi-tree, specifically: every node can have any number of parents and any number of children, and references can only go in one direction (you can’t reference a parent node.)
Word 
I hope you don’t mind my more rambling illustration of the concept.