I am no expert, but if I had to put the pieces together, I'd say... Bitcoin is not just a consensus algorithm, but a public key cryptography system.
Bitcoin clients obtain consensus on what transactions have occurred by building a blockchain where basically, the longest chain that meets the requirements set by the difficulty and other hardcoded client demands is the "winning" blockchain.
So, start with 51% of mining power, which is the first requirement.
Next, pick the transaction to erase and count up the number of blocks that are "on top" of that. Each block represents some realized investment from the continued expense of hash-power. Those will have to go as well; if we wanted to erase a transaction, it's best to have made up our minds as soon as possible to do it, since the longer that passes, the harder it will be to convince everyone who's minted a block since then to part with it. But we started with already having convinced more than 50% of everyone, so let's ignore that difficulty. Still, in reality it's just that much harder every day.
If the "hot wallet" private keys are not in our hands, we can't just "give" the seized bitcoins to someone else. They are currently protected with a cryptographic signature that says "these are mine, spend them to the FBI." With DPR's hot-wallet keys, either stolen from the FBI or obtained from a backup owned by DPR, having erased the transaction we can now spend them somewhere else.
But now we've also erased all of the _other_ transactions that have happened since then. We will want to read those out of the now defunct chain and put them back into new blocks as soon as possible, too.
(Won't we? Fortunately we don't all have to agree about the validity of those transactions, as long as 51% haven't also conspired to block some of them, we still have the signatures and they can be re-broadcasted to the network and again eventually make it into the winning chain. But for a window of time, there's now a chance that those coins are spent another way too.)
Only, that newly continuing chain shows that those timestamped blocks from before the seizure were the last to be generated. Now we're far behind target, of generating 2016 blocks every two weeks. In fact more than 10 days have passed and we have the appearance that no valid blocks have been generated in that time.
We've got some serious catching up to do.
Now we've got a situation where the next difficulty retarget is going to push the difficulty _down_ for the first time in a LONG time, because the two week target has been missed by a longshot. We've either got to find substantial _extra_ hashpower that was not there before (that's probably well over 100% of the network) or otherwise basically convince pretty much everyone to set their clocks back at the same time.
Or we can just let the difficulty plummet. That's another hard sell either way.