Scaling Bitcoin workshop : Stanford 2017

How to charge lightning

How to charge Lightning

Aviv Zohar

Our next talk is getting into layer 2. Leaving hwat needs to done at th...

I am going to start with maybe a simple beginning talking about a single channel between Alice and Bob. We heard about channels yesterday so I am not going to go into details. Alice is not going to cheat Bob in my scenario. Everything is going to run as if everyone honest and doing their best effort, for my example. They open a channel and insert 10 coins. Every time Alice moves her coins to Bob or the other way around, the liquidity within the channel shifts. So you can think of it as if we are in a state space between 0 and 10. If we are just transacting in single units of bitcoin we are just moving around there. I am going to assume that just for the sake of modeling that Alice cannot predict whether she will be paying Bob next or the other way or whatever.

Some basic fact about transaction channels... are that you can start to do the math and for random walks this is well-known stuff. If Alice starts to send money to Bob at a certain rate and Bob sends money back at a certain rate, then we can look at the expected lifetime of a channel. There's a formula and we can look at it in a graph format. And it's slightly better. Over here is biased transfers on a channel if we're slowly moving money then there's an optimal point as which we can fund the channel. The x axis is balance that Alice has, we would like to begin with most of the money on Alice's side. If we're doing, around here, if the transfers are biased, so that, I think this is actually Bob's side, so she is paying Alice slowly and money will flow as we slip down this curve, right? We can even talk about balanced channels where transfers go back and forth at equal probability. Benefits of lightning show up here- if we put double the amount of money in the channel, then the optimal way to start the channel is with equal fundings from both participants, and the channel lifetime scales up with the square of double. If you double the amount of money then you can quadruple the channel lifetime. This makes lightning interesting, transfers get to cancel each other out and the channel lives longer.

Once you start to look at this you must ask yourself-- we don't assume that Alice and Bob are transacting a single point between them, they are probably sending more between themselves. Sometimes we make very small transfers, sometimes large ones. I often buy coffee or pay for a meal, but only rarely do I buy a car. So there's a probability that I will be doing a large transaction. And if you look at data for sizes of transactions you see powe rlaws. The probability of a large payment scales maybe inversely and quadratically with the size of the payment. Right? So this is usually stuff we see in data. The exponent will be different from 2, it could be larger, it could be smaller. This gives rise to distributions of money transfers that are large.

And then you start to think about channels and you wonder, when do you restart a channel? How much money would you put into it to get the maximum utility from the channel? Beneficial to reset the channel even if you didn't reach the border of the channel. Sometimes there's a large hop that gets you in this yellow over here. But you might want to restart the channel at this point. The next transaction that shows up might be a large one, and it won't be usable in the channel because you don't have the equilibrium in the channel funds.

If you simulate this stuff, you find that there's a-- what you're seeing is the number of blockchain hits versus the radius at which I reset the channel. There's an optimal point, other than waiting for the channel to completely run out, you reset before it gets to the border.

We looked at basic channel behavior and we asked how much do channels cost? There are two main costs to a channel. If you increase the amount of funding in a chanenl, you increase its lifetime. Can I put in as much money as I want, can I just pour a lt of money in there? The limitation is that I can always get more money but I have to borrow it and pay interest rates on it. So if I want to fund a lot of channels and lock up funds in there, they are going to be moving back and forth over the channel but never out of the channel. They are losing potential interest if they had put the money elsewhere. There's another cost, which is channel setup and settlement which has to touch the blockchain and it has fees and fees might be high. So you are trying to mitigate between these two costs. These are the things we will consider.

The next thing to ask is how are fees collected? What would a lightning transaction pay and what would a bitcoin transaction pay? For regular on-chain transactions you have fee rate volatility. There's a somewhat fixed fee for bitcoin blockchain use that's how it's built. On lightning, it's somewhat unknown. If you think about it, you will notice that if you do a large transaction in a ightning channel, you shorten the lifespan more than you do the small transaction. So you will incur a cost if you reset the channel and estbalish it again. It makes a lot of sense, at least in my view, to charge lightning transactions not a flat fee, but a proportional fee to the amount of money being used. In some instances, a large transaction should cost superlinearly in some sense, since it hurts the channel much more than just its size.

This is also going into our model. And of course what we want t od ois we want to give transactions a choice. I am assuming that people are going to have this really nice wallet, and UX/UI wallet problems will go away. And you are asking to send money from here to there, and it's using the channel if it's cheaper, or use the blockchain if it's cheaper, or you might see the fees and you say it's too expensive and you say you're going to use something else like VISA or Western Union or myabe you don't make a transaction at all. So this is our model.

Now I am ready to show you the approach going from this micromanagement of channels to what happens to the economy. What we're going to do, in general, is to think about a pattern of moving money. It's going to be really simple to begin with. And then I am going to talk about channel management and topology and so on and then the market of delivery for these fees. And basically, right, the intuition that I want you to have in your head before I show you some of the results is that we're going to have these large transfers that really prefer to go on the blockchain because on the blockchain they pay a flat fee, and on a lightning channel network they pay proportional to their size. Small transactions will prefer the lightning network. Together they share the cost of establishing the channel and so on. Both of these sources of transactions are going to touch the blockchain for different reasons, like establishing channels and direct transactions. Do the prices interact with eahc other? Are large transactions common enough to squeeze out the lightning transactions? This is the exact question that interests me.

Here are some made-up parameters that I picked. I am assuming every person from now that I talk about has 10 tx/day. That's all they do in expectation. We draw the amounts they use from a parallel distribution like I defined before. Each of them are willing to pay up to 1% of what they are transacting, in fees. There is a 4% yearly interest rate. I know it's lower now, it could be higher, I don't know. It's made up. And there are 288,000 on chain transactions that we can do every day. This is baed on blocksize before segwit because we don't have exact numbers on what segwit is going to do on transactions until it gets fully adopted I don't have a good estimate I guess.

In terms of topology what am I going to assume about the world? What are people trying? I am going to assume a nieve model which is just that people are paired up together. And they are just Alice and Bob are sending money back and forth. This is actually designed to be nice to lightning. We know exactly which lightning channel to open, and no multi-hop. And their partner there, they can send money back in forth. In reality, it's much more complex. Another thing that you might consider is some uniform topology where each person could be paying another with equal probability and we don't know which 2 people are going to be transacting. In this world, what is the topology of the lightning network? One obvious solution is to model it with a central hub, which simplifies the math and makes it easier for lightning to work efficiently. This is the model that I am going to be talking about. These things look different in how much money they end. But really they are not too far apart. Because the model with the pairs has one transaction channel for every 2 people. And the model hub has one channel per person. So it's a factor of 2 on maintaining transaction channels if we had exactly the same behavior and same funding in both. I am going to talk about the pairs model and just as a walk-through model, and we have done more in the paper that will come out. Let me walk you through it.

Here's the behavior of individuals. Here's the funding of a channel. We're assuming we have some blockchain fee, could be in BTC or something. This is just an example. Given this fee, I can decide on how much channel capacity to choose. I can make the channel large with a bunch of BTC or small. I am funding it optimally where both sides have equal funding and we are assuming transfers back and forth with equal probability. As I grow the channel capacity, the red curve is cost, it's one as the channel capacity grows, I have to touch the blockchain less often (green line goes down), I do less resets onthe channel. The blue line goes up the more money I put into the channel, I pay more in interest opportunity cost. There's an optimal point where I want to be around to fund the channel and therefore minimize my cost.

We can have a curve like this for every curve. So what happens to the channel capacity when we change the fees? if I increase the fees slowly, the optimal channel capacity that I am going to choose goes up like the square root of the fee times some constant. This is the exact benefit of lightning- I double the fees, I only need the channel to have a acpacity square root of 2 times larger to manage it and have good behavior. And from this we can derive demand for blockchain records. If I put some of my funding into the channel, I can now figure out which transactions find it expensive to go through the channel, and from my different payments that come out of different sizes, and I have this demand curve, and if you look at it, this is the exact curve on log-log scale which is a straight line, it's a power law. So I have the demand scaling down as one over the fee, and htis is for a world that has no lightning in it at all. If I can't use lightning because of the way I chose my distribution I would get this demand curve. In a world with lightning, the curve looks a little bit like the same, but log-log, instead of 1/fee it's more like 1/sqrt(fee) and that's a minor difference. What would it do to us? We'll see.

The next question is what happesn when I scale up. This demand function is for a single Alice-Bob hcannel and at differnet fees how many times you will want to touch the blockchain. So now I am adding lots of Alices and Bobs, maybe 100 million, each who want a channel with their partner. Here are the fees, without lightning, when we scale up the number of users, we go to 100s of millions, the fees go up. The fee for settlement and blockchain space, goes up. Don't get hung up on the numbers, because remember we made up the distributions. But look at it qualitatively. The next thing of course is what happens with lightning. This gets interesting. The demand curve had a different shape with it... You can tell probably that the we're having very different fees. One thing that is obvious to begin with is that we reah really high fees at the end. Lightning tends to collect more money from people and they are willing to pay it because it amortizes cost over the lifetime of the channel, they are willing to pay more to estbalish the channel. Between 0 and 20 million people, we're actually making less money with lightning because we added a tech that allows us to move some stuff off the chain and there's less competition for the chain space. So there's lower fees. And then it gets higher after 20 million people.

For miner's reveneu this is exactly proportional because there's a fixed amount of spots on the blockchain. So if the fees are lower, miner revenue is lower. If the fees are higher, then security goes up. So we're seeing the effect of lightning here. I know this other thing is controversial, but let's talk about the block size- here is the world with lightning, what happens if we scale it up with segwit? I know this is highly volatile and it's a hard topic to talk about, but with an extra increase in block size, the fees drop. Instead of 25 here, there's 10. It drops by more than a factor of 2. So there's more space because of the increase. The total revenue of miners has gone down by the block size increase. But we will getting more throughput, but the security in this case, for this model, for these assumptions, has gone down.

Here's the conclusion. I hope I hae interested you enough to come and discuss this with me later. Lightning definitely helps. But now you have to decide for yourself whether it's a lot or a little. I am not sure what to think. I was expecting lightning to do a lot more in some sense. I wanted a factor of 100x on transaction throughputs but we got something less than half of what was already there. Maybe we can do micropayments and suddenly it's different. A 2x block size increase helps but not by very much. It's not that amazing. It's not surprising though because I showed graphs that took user count from 10 million to 100 million, and I only increased block size by a factor of 2, so obviously that's not going to help a lot. There's a lot that I didn't show you, which will be in our paper. We are looking at heterogenous population, but really, some people are moving more money than others, some will push others out of the blockchain... the results will be different. There are more complex patterns of flow. The pairs model is super simplistic. There are other distributions of flows, like multi-hop networks, circular where people don't just send money back and forth. And usually people get money from their job. How would lightning channels look like in that scenario? And in general I think I would like to voice my concerns.. I don't know how fragile my models are, how much these assumptions are bad? Would a change in interest rates do a huge shift in the results? Would we see an economic effect? The fees are very important there. Everything in the model was, you know, the average or expected behavior-- what about fluctuations? We have those often in the economy, right. So I'll leave you with more questions than I hope we started. Thank you all.

Q&A

Q: I am wondering how viable do you think it would be that when we have to move to the chain because your channel becomes too big to be feasible on lightning. Why not just purchase multiple services from the counterparty? And what about resets? I send a little more money, and I have my counterparty reset my channel. Wouldn't it help me a lot to increase the apacity to use the on-chain for the reset and the payment?

A: Yes we have assumed something like that. If you do two things with your ttransaction it's going t be larger, but then youre paying for it more because you're paying per bytes and not per transaction. There are some constants that need fine-tuning. Paying someone is a different transaction size from setting up a channel. That would effect the model at the edges. We used constants, because we didn't know what values to choose.

Q: I'e been thinking about lightning and UX and transactions coming out of our exchange. One thing I was wondering about... what are your thoughts on what you think transaction olume would be effected by people wh want to send on-chain and .. top-up their channel or do some sort of channel operation like when you add a UTXO to a channel you're sending a coin to the multisig on the blockchain... and adding an UTXO and saying x amount is the other person's channel. Sending on the chain, but also adding to the channel. Do you hae insights into what that would appear like on the blockchain, and whether you can put that into your models at all?

A: We thought about this a lot. There are a lot of interesting questions exactly there. It depends on how cooperativ vv ve the othe rparty is... if you're putting more money in the channle, that's not going to happen often, you're locking more money in there. People might send their counterparty directly, and then maybe someone send sit on the channel, to reset, we're doing it in both layers. We thought about this. Our reset transactions are 1 tx, because we thought people would be trying to save money.