> The current default Bitcoin implementation makes no effort to solve or even really approximately this problem
I think this was somewhat out of date even when it was written. :(
When I've sampled it before the solver in Bitcoin reliably produces results which are very close to the best results that an external MILP solver produces... and since a year and a half ago it even considers dependencies in that analysis.
The knapsack part of the problem is not that impacting because a block is composed of thousands of transactions, which are mostly very small compared to the block size... and for the lower fee transaction the slope of the fees per unit for the available options is not very high. Basically, so long as you get the high fee transactions it usually doesn't much matter if you fail to eek the last few bytes out of a block.
I'm sure you know this, but NP-hardness results only matter for worst-case instances. Even if Bitcoin's algorithm is roughly close to optimal on real-world instances, it doesn't mean somebody won't be able to generate a series of transactions that cause the algorithm to exhibit worst-case behaviour.
Also, the NP-hardness results would apply even if the Bitcoin algorithm is updated.
Oh sure, I wasn't intending to comment on the NP-hardness but only on the accuracy of the approximation in practice.
There is a lovely paper disproving the all-public-information efficient market hypothesis, assuming computationally bound actors by showing how to embed market inefficiency into a set of trades that can only be removed by solving a NP-hard problem. :)
Many miners require a minimum fee-per-byte for including a transaction in their memory pool of transactions awaiting inclusion in a block. So they would for instance ignore a transaction of B bytes if its fee is less than B * 100 satoshi (at 10^-8 BTC, the smallest unit of account).
Then the miner will often find that its entire memory pool fits inside the 1MB block limit, and the knapsack problem disappears (still leaving the NP-hard conflict resolution problem).
Just cutting off low fee TX is a worse result than even the most approximate knapsack solver you could use. :) (Take transactions in order of fee per unit-limit, skip ones that won't fit, until you're full or you've skipped too many times sequentially-- which is what we do... with a too many threshold of a few thousand IIRC)
The reason for the floor is primarily to avoid wasting resources trying to verify transactions which are never going to confirm.
The floor rate used in the network today is 1e-8 BTC per byte, though it automatically goes up if the backlog of transactions gets large (over about 150MB of transactions). That limiting works by nodes keeping 150MB of transactions and if they gain more they drop the lowest feerate transaction and set their minimum to that value. The minimum then decays back to 1e-8 BTC/byte at a speed which depends on how much below the limit the node is...
[ these parameters don't need to be completely consistent in the network, nodes will tell their peers what fee levels they're willing to process]
> NP-hard conflict resolution problem
That is ignored in Bitcoin implementations today because double spends are rare, interesting ones are extra rare with complex dependency graphs, and there is a social good to behaving in simple an explicable ways.
Considering how centralized Bitcoin has become minors are incentivized to raise the threshold for inclusion in a block even if they make less per block in the short term artificial scarcity can become extremely valuable in the mid term.
The conflict resolution problem was traditionally solved just by taking the first seen valid transaction and disregarding any subsequent conflicting transactions. That of course is not necessarily profit maximizing, so current implementations generally allow replacing the transaction as long as it pays an incrementally higher fee than the one it's replacing (the increment usually being the same as the minimum fee you mentioned).
Non-mining nodes have to implement a similar policy, as it's also used for preventing DoS on the P2P network.
I think this was somewhat out of date even when it was written. :(
When I've sampled it before the solver in Bitcoin reliably produces results which are very close to the best results that an external MILP solver produces... and since a year and a half ago it even considers dependencies in that analysis.
The knapsack part of the problem is not that impacting because a block is composed of thousands of transactions, which are mostly very small compared to the block size... and for the lower fee transaction the slope of the fees per unit for the available options is not very high. Basically, so long as you get the high fee transactions it usually doesn't much matter if you fail to eek the last few bytes out of a block.