>> Nikhil Devanur Rangarajan: Welcome everyone. It's my great pleasure to welcome back Shuchi Chawla from the University of Wisconsin-Madison. Shuchi is no stranger to people here. She spent a year as a visiting researcher between Europe and Microsoft and we missed her very much since she left. I hope she missed us too. >> Shuchi Chawla: I came back. >> Nikhil Devanur Rangarajan: Yeah. So we're glad to see her back, and she's going to tell us something about auctions with interdependent values. And this is something she did during her visit here. So over to Shuchi. >> Shuchi Chawla: Thanks, Nikhil. Definitely a pleasure to be back. So this is joint work with Anna Karlin, who is right there in the audience and Hu Fu who is just finishing a post doc at MSR in New England and is going to be in the neighborhood soon at UBC. And also I wanted to very much acknowledge Microsoft's support. This was work done when all three of us were visitors here at Redmond. So it's been really great. So let me start by talking about the classical auction design problem. So here's the kind of setup that I want you to keep in mind. There's a seller who wants to sell some items or some services and there are a number of buyers that arrive interested in buying these goods or services and each buyer has some value for this item which is the maximum amount of money that this buyer is willing to pay for this item. And so what we like to do here is to come up with a mechanism or a protocol that looks at these values that the buyers have and figures out who to give the item to or how to allocate the service to and also figures out how much to charge each person. And one of the things that this protocol should satisfy is that here, the buyer's values for the item are like an input to this auction and the buyers can game the system by misreporting their values. So we're going to ask for a truthful auction, namely one that incentivizes people to report their values truthfully. And you can go beyond this but for this talk, we'll just think about truthful auctions. And subject to these constraints, the seller might have some objective that he wants to optimize. For example, the seller might want to maximize amount of revenue he gets or the economic efficiency of allocation meaning that people that desire the item or service the most are the ones to get it. This is also called social welfare and there may be other different kinds of objectives. These are the two objectives that we will look at in this talk. So this is the classical setting of auction design. And we make an important assumption which is that the seller might not know the values that the buyers assign to the item or service but does know the distribution from which these values are drawn. And there's a lot of work on the setting. It's a very well understood problem. Now, what often happens in the real world is that buyers may not in fact know how much value they want to assign to the item. Okay. So they look at the item. They try to get as much information as they can about it and this informs their value. And so they might have some sort of guess or estimate as to what their true value is but not know what the value is. And in fact, their value might depend on how other people value this item so if they don't have full information about the item, looking at other buyers' behavior might change their estimate of their own value over the length of the protocol or the auction. And so the question that we want to ask is how do we design an auction in such a setting where buyers don't know their values precisely. Let me give two examples of where this kind of setting arises. The first is a classic example by Milgrom and Weber. And this is a setting where multiple companies are competing for the rights to drill at a particular site for oil. And the value that any particular company would have for this depends on the amount of oil that they would find at this particular site but until they get the rights to drill and they actually do the drilling, they cannot know the exact amount of oil at that particular site. So typically, what happens is different companies have different ways of estimating what the total amount of oil is, and different people might get different estimates but their values are all the same and it's the total value of oil at that particular place. So here's a formalization of this sort of setting so let's say that V is the common value that is the true value of the resource and every bidder or every company gets some noisy signal of what these values are. Now, these signals are what they get to see and they don't get to see this value and so over the course of a protocol or an auction, if they get to see some information about other people's signals, then their best guess of what they value might look something like an average of these signals. Okay. So this is a setting in which every everyone's best guess as to their own value is a function of all of the information that not just they have on their own but also what everyone else gets. Okay. Here's another more contemporary example of where such a situation may arise. So imagine that you are a user who is searching for something on a search website. And one of the things that search engines do when you search for a term is to figure out what adds to display on this page and we do this by running an auction over different advertisers. Okay. And so when they run this auction, they use cookies to figure out what kind of user is performing the search. And they share this information with the advertisers. And having this information better informs the advertisers how much value they see from this particular customer viewing their ad. Okay. And this is a setting where the values of these advertisers are correlated in this manner by having been informed by the same piece of information. Okay. So here, the advertisers look at this information and derive their value. We will think of this as a setting where the values are private. They're known to the advertiser but they're correlated. Okay. So this is already a little bit different from the classical setting where everyone independently obtains their value from something. Okay. In contrast, consider the following sort of setup where again the search engine obtains some information about the user that is searching on this web page, but now sends different pieces of information to the different advertisers. And once again, based on these signals that the advertisers get, they form some estimate of their value for this particular consumer. However, now, their true value is a function of this entire information put together but at the beginning of the auction, they only have a limited amount of information. So again, formalizing this, we can think of the advertisers as obtaining some signals from a joint distribution, and then their true values being a particular function of all these signals and not just their own. Okay. And this is the kind of setting that we call an interdependent value setting, values of all of the buyers in the system depend on each other's information. >> [Indiscernible]. >> Shuchi Chawla: That's right. Yeah. They get to see the signal. Then they place the bid. But they get to place a bid without knowing what their true, you know, without knowing their exact value because this depends on information that they haven't received so far. Yeah. Okay? >> [Indiscernible]? >> Shuchi Chawla: So it is true that the search engine shares some information and not necessarily all, but I'm not sure that in practice what happens is that they share different information. So just a hypothetical example of how it might be. >> [Indiscernible]. It seems like the value is some function of the signals and something that happens eventually. >> Shuchi Chawla: Right. So the value eventually might also be a random variable, but here we're thinking of the value as being the buyer's expected value given all of the information they could possibly obtain through the auction. Yeah. Yeah. So at any point of time, the buyer can think about their expected value given all they know and this could evolve as they gather more information. >> So this is kind of most accurate estimate of the value [indiscernible] auction. >> Shuchi Chawla: >> That's right. Before you actually discover -- >> Shuchi Chawla: Before you discover more, that's right. Yeah. Yeah. So the question you want to ask is how should we design auctions or run auctions within this setting? Is this setting similar to the classical setting in how we approach auction design or not? Okay. So I'm going to start with a quick primer on classical auction theory just to make sure we're all on the same page and then we'll talk about the two objectives of welfare maximization and revenue maximization. As I mentioned before. Okay. So let's talk about the classical setting where buyers know their values. And these values are independent of each other's values. First we're going to talk about the welfare maximization objective. Recall that I very vaguely defined this before. So here, we are trying to maximize the total value that buyers achieve by the allocation that the auction generates. So in other words, we want to give the goods or services to people that value them the most. Okay. So how can we maximize this objective in the context of auction design? Well, we could, to begin with, we could figure out what is this allocation that maximizes the value that the buyers get in total. Okay. And the question is: Is it possible to come up with payments that the buyers make in such a way that this encourages buyers to report their values truthfully. And it turns out that the VCG mechanism is a classic mechanism that gives a payment scheme for supporting precisely these welfare maximizing allocations. Okay. And it's based off of this informal statement like any truthful auction, any auction that aims to incentivize buyers to report their values truthfully must be of the following form. It offers the item or service to each agent at a price that is independent of the agent's own value. So it comes up with a price that depends on other people's reported values and then just offers the item or service to the agent at that value. Okay. And so in the context of the VCG mechanism, it turns out that every buyer pays the minimum value that he needs to bid in order to win the item. And if he can afford to pay this, then he wins the item. This is also called the critical value. And let me illustrate through a quick example. So let's say we have three buyers competing for a single item. And these are the values. Then we can figure out for each particular buyer, we want to give the item to the buyer that desires it the most, that has the largest value for this item. So for each buyer, we can ask the question of how high he needs to bid for the item in order to win the item. So this guy clearly needs to bid $100 to beat this other guy. So we offer the item to him at $100. And this guy can't afford to pay that much. This guy needs to bid $100 to win the item, can't afford it. This guy only needs to bid $15 to beat the other two. So we offer the item to the third guy at $15 and he accepts and in fact, the way that we set up these offers, the buyers can make this accept or reject design without having to -- these offers are independent of their own values and so the buyers will behave in a manner that the auction wants them to behave in. So in that sense, this is truthful. And achieves what we want to achieve, mainly giving the item to the person that values it the most. Okay. So in fact, this is an example of what's called a Vickrey auction. This is precisely the Vickrey auction. The highest bidder wins the item and pays the second highest value that was reported. Okay. So this is the classic auction here. Very beautiful result. You can do welfare maximization in this nice manner. What about revenue maximization? What if the seller wanted to earn as much as possible? Well, in this case, the seller might ask for as much as a hundred dollars in revenue but it turns out that this is not something that the seller can hope to obtain in a truthful manner. So for revenue maximization, it turns out that we need to make the assumption that the seller knows the distribution from which the values are drawn and then designs the auction based on that distribution. And in this case, it turns out that one can slightly generalize the VCG mechanism that we described here in order to maximize the seller's revenue and expectation over the value distribution. And here's what it looks like. So you ask the buyers to report their values. You look at their value distributions and use those to apply a certain transform to the values giving you virtual values. And then we run the VCG mechanism over these virtual values. And this turns out to be optimal for revenue maximization. Don't worry if you don't quite get this fully. The one point that I want to make about this mechanism is that these virtual value functions can be strange functions and so the mechanism can look strange and complex for that reason. It can also give the appearance of being unfair in the sense that sometimes a person with a small value wins the item whereas a person with a much larger value doesn't. So it can look like a strange mechanism. It turns out that in fact there's a tighter connection between revenue maximization and welfare maximization of the following sort. So if the buyer can set result prices, meaning that if everybody bids below a certain price, then nobody wins the item, then welfare maximization with these result prices is near optimal. Doesn't give you the optimal mechanism but it brings you very close to optimal. So for example, in the case where the seller has a single item to sell, we saw that welfare maximization can be done with this Vickrey auction. Here, if you set an appropriate result price in the Vickrey auction, this gives you a two approximation to revenue, to the expect revenue. So this is nice. It nicely aligns revenue maximization with welfare maximization, gives a simple mechanism that has nice fairness properties. So we also understand revenue maximization in this respect very well. Now, let's move on to a more general setting where values of different people are correlated. So we didn't use anything about the distributions of values when we talked about the VCG mechanism or the Vickrey auction. So indeed, it doesn't matter what distribution values are picked from. These mechanisms continue to be optimal for welfare maximization. But revenue maximization relied on looking at the value distributions of the agents and in fact, the mechanisms that I talked about earlier don't work anymore when we go to a setting with correlated values. So what happens here, so intuitively, the objective of revenue maximization is difficult for the seller because of this lack of information that the seller has about the agent's value. And but when the values are correlated, that means that the values of other agents give you some information about the value of any particular agent. And this extra information should only help the seller to obtain more revenue from the buyer. Okay. And so indeed this happens. So under some very mild assumptions, it's known that you can in fact extract the entire value of the buyers as revenue, meaning that this is the most possible, the largest possible price that a buyer might be willing to pay for the item and so you can do the best possible thing by exploiting this correlation. The caveat is that sometimes the buyers pay more than they value the item for and sometimes they pay less and so at the end of the auction, a buyer might be unhappy if they paid more than how much they value the item. And it also tends to be a very complicated mechanism. Okay. So you might ask, well, what can we do if we impose the restriction that the buyer never pays any more than how much they value the item and then this problem becomes NP-hard. Okay. So in terms of trying to optimize for revenue in this correlated setting, we understand a little bit less than our understanding for the independent case. But fortunately, the connection between approximate revenue maximization and welfare maximization that I mentioned earlier continues to work even in this correlated private value setting. So in the single item special case, once again, one can show that the Vickrey auction with appropriately set result prices is 2-approximation to expected revenue. I'm going to talk about why this -- I'm going to show proof of this result a little bit later on. Okay. So great. We understand the independent value setting. We understand the correlated value setting. >> [Indiscernible]. >> Shuchi Chawla: Right. So this is NP-hard relative to the optimal -- the revenue optimal mechanism. So meaning that the optimization problem, what is the mechanism that maximizes expect revenue over the class of all truthful mechanisms, that problem is NP-hard. So there is going to be a gap between the total expected value in the system and the revenue that the seller can extract but that's not what we are comparing against. We are just comparing against the revenue optimal mechanism. What's the most revenue you can obtain truthfully. So I'm going to go on to interdependent settings now. And first describe what challenges we face when we talk about welfare maximization in these settings. So here is the formal model that we're going to assume. So for the next portion of the talk, I just want to think about a setting where a seller has a single item to sell to the buyers and the buyers are competing for this one item. So we're going assume that the buyers receive some signals about their value. And these signals are drawn from some drawing distributions, some arbitrary distribution. And then as I said before, the value of every buyer is going to be a function of all of these signals. And we're going to assume that the value is an increasing function of the buyer's own signal. Now, here's the first problem that arises in this setting. Before we talked about the notion of truthfulness which is also called incentive compatibility for auctions meaning that a buyer, once he knows what format the auction is going to have, is going to be happy to reveal his true value for the item for this auction. But here, if the buyer doesn't even know his true value, how do we define this notion of truthfulness? In fact, indeed, the auction may have a format where the buyer learns more and more than information through the auction, so what he thinks about his value at the beginning of the auction may be different from what he thinks about his value at the end of the auction. Okay. So we side step this issue a little bit by looking at the notion of ex-post incentive compatibility, which is the following: So we require the mechanism to have the property that at the end of the auction, once everything has been revealed, a buyer is satisfied with what he bid, what strategy he followed through the auction, even after everything has been revealed about his value. So for every possible setting of other people's private information, when other people reveal their private information and the buyer knows what his value is, he is still satisfied with his strategy at the end of the auction. This is called ex-post incentive compatibility. And with this notion of incentive compatibility, the following generalized version, one can define the following generalized version of the VCG mechanism. So recall what the VCG mechanism does for private values. It asks people to reveal their values. Here we're going to ask people to reveal their signals and then compute their values from their signals. And then the VCG mechanism serves the welfare maximizing set. Here again, we're going to serve the welfare maximizing set, meaning give the item to the person with the highest value. And now the question is what prices should we charge? Now, before, in the private values setting, we said that the VCG mechanism should charge any person the minimum amount they needed to have bid in order to win the item. Here, we're going to look at the minimum signal that the buyer needs to report entered to win the item. Okay. And then so let's call this the critical signal. This buyer, we're going to compute the value of the buyer at this critical signal given the signals that other people reported, and this is the price that we will charge. Okay. So and in the single item case, this is the generalized Vickrey auction. Again, we asked people to report their signals for the buyer with the maximum value given these signals. We ask what is the smallest signal this person should have reported entered to win, and then we recompute his value at that critical signal and that's the price that we offer the item at. So pictorially, let's imagine you have two buyers in the system. Their signals are S1 and S2. Now, let's suppose we want to try selling the item to buyer one. We're going to fix the second person's signal, so S2 is fixed. And we're going to look at how the values of the two buyers evolve when S1 changes. Okay. So this is the signal that the first buyer reports. Now, note here that this is the point at which the first buyer's value starts to become bigger than the second buyer's value. So this is going to be his critical signal. He better report a signal larger than this point here in order to have a higher value. Okay. And so we're going to charge this person the price, the value corresponding to this critical signal. That's the price that we offer the item at. Okay. And this is the generalized victory auction. Does anyone see a problem with this auction or with this description here? >> [Indiscernible]. >> Shuchi Chawla: Yes. So the buyer was revealing -- so let's again think about the private value setting where everyone knows their value. They reveal their values, but we want to charge prices, so to person one, I want to allocate him the item at a price that's independent of his value. But only if his value is the highest. So how am I going to do that? I'm going to take the highest of everyone else's value and set that as the price for this person. And if his value exceeds that price, he's going to buy and but the price is independent of his own value. Okay. Likewise, here, now, I'm going to ping in terms of signals. So everyone reports their signals to the auction. And I look at -- I compete every person's value. And I want to allocate the item to the person who has the highest value given these reported signals. Now, let's say this is buyer one. And I want to ask, well, what is the smallest signal that this person should have reported in order to have the highest value? Okay. And if they wanted to lie to me, and they knew what everyone else is going to report, they could have reported that signal. So I better not charge them anything more than how much I compute their value to be at that particular signal, critical signal. And that's what this mechanism is going charge. Okay. So it's going to pretend that they reported this critical signal. It's going to compute their value at this critical signal and that's the price that we charge. Yeah? >> [Indiscernible]? >> Shuchi Chawla: I'm sorry? Yeah. So the value functions are known. That's right. Yeah. Yeah. So we don't know the signals, but we know how the values depend on the signals for instance in the oil field setting, every person's value is an average of everyone's signals. >> This is not exposed. [Indiscernible] the other signals [indiscernible]. >> Shuchi Chawla: So this is ex-post in the sense that given what other people reported, now, assuming that to be the truth, that's my best information about my value, I'm going to be happy to pay this amount. Yeah. >> And here showing as [indiscernible]. >> Shuchi Chawla: That's exactly right. So that is the problem with this mechanism is that if let's say that the value functions as a function of S1, looked something like this, well, now, the critical signal is not well-defined. So what should the critical signal be? This one or this one? It's challenging. These are both places at which V1 crosses V2 becomes the higher value. In fact, it turns out that if value functions look like this, they cross each other many times, then we don't know -- in fact, there's no mechanism that is ex-post incentive compatible and is always going to allocate the item to the person with the highest value. Okay? And so we're going to make this assumption that for any particular signal, for any vector of signals of other people, if I vary agent I's signal, then his value function crosses any other at only one point. This is called the single crossing and this is a pervasive assumption in all of the work on interdependent value settings and given that we don't even -- we cannot maximize welfare without this assumption exposed, it seems to be challenging to work without this assumption. So we're going to assume this for the rest of the talk. Okay? Okay. Once we have signal crossing then, the generalized Vickrey mechanism as I define, it becomes well defined, it's ex-post incentive compatible and it also always maximizes welfare. Okay. So now, let's go on to revenue maximization. So for revenue maximization, we might start out doing -- taking the same sort of approach that Meyerson's optimal mechanism took for the independent private value setting. Okay. So that is certainly possible to do here. You can define certain virtual value functions such that maximizing welfare over virtual values would be equivalent to maximizing revenue over the actual values. Okay. However, this does not lead to a way of finding the optimal mechanism and that is because these virtual value functions can be very strange looking functions and so in Meyerson's setting, once you define these functions, there is a natural optimal mechanism that comes out of that approach. Here, you need a lot of very strong assumptions in order for this approach to result in an optimal exposed incentive compatible mechanism. So assumptions such as seeing that different buyers look symmetric, their values are drawn from -their signals are drawn from a symmetric distribution. Their value functions are symmetric. There is a certain assumption on the distribution that's called affiliation. I won't define this. There are other assumptions on the signal distributions that some of you are familiar with but don't worry, if you don't know what these are, regularity hazard rate assumption but in fact, much stronger version of these assumptions that apply to conditional values and conditional signals as well. And so characterizing the optimal mechanism requires very, very strong assumptions on the setting and we don't really understand this very well. Okay. What about near optimal mechanisms? You might ask, well, what about, you know, we talked earlier about VCG with reserves and this turns out to be near optimal in the private value setting. So does the generalized Vickrey auction with appropriately set result prices, is that approximately optimal. And it turns out that it is. But again, you require a pretty strong assumption this general mono tone hazard rate assumption, and you can get this sort of a result also under some very strong symmetry assumptions on this setting. Okay. And in general, so, again, I want to point out that in private value settings this sort of Vickrey with results being approximately optimal happens in the absence of any assumptions. But here, it turns out that in general, this mechanism can be really, really bad, really far from optimal if you don't have these assumptions and I'll show you an example where this happens. Okay. And so to summarize what's going on, we don't really have a good understanding so far of revenue maximization in these interdependent settings in the absence of some very strong assumptions. Question? >> [Indiscernible]? >> Shuchi Chawla: So the generalized assumption is that conditioning on any vector of signals of other buyers, if you look at the distribution of value of any particular buyer, then this satisfies monotone hazard rate assumption. So every conditional distribution of values satisfies this property. And likewise, the generalized regularity assumption is similar. Okay. So to summarize what we have so far in the private value setting, we understand welfare maximization, revenue maximization, and nice property that we achieve are that revenue and welfare are roughly aligned. In the interdependent value setting, we don't understand the revenue objective very well. And furthermore, so far, we don't know how to maximize revenue by appealing to welfare maximization which we do understand. Okay. So what do we do in this work? It's taken me a long time to get there. We've come up with a new tool that we can exploit here. And we call this admission control and I will describe the mechanism very shortly. And we show that this slight modification to the Vickrey auction or the VCG mechanism leads to approximate optimality and this is a way of reducing our aligning revenue maximization with welfare maximization. So this alignment happens, just not in the obvious way. Needs a little tweak to welfare maximization. And furthermore, we require no assumptions on the value distribution. We do need the signal crossing assumption and we need one more technical assumption that I will come to. Okay. Okay. So before we go there, I'm going to show you why revenue and welfare align in private value settings. So why is it that the Vickrey auction reserves gives you approximately optimal revenue and why it fails in interdependent settings and that will naturally lead us to this modification that gives us an approximation. Okay. So let's again recall this private correlated setting. We have a single item to sell. We have a number of different buyers that are interested in this item. Their values are drawn from some drawing distribution. Okay. And we're interested in some variant of the Vickrey auction for this setting, and so in particular, let's say these are the values that the different buyers have. We're going focus on trying to sell the item to the person with the highest value. That's what the Vickrey auction tries to do. Now the question is how much should we charge this person for the item? Okay. And so we're going to use all of the information we have at our disposal in order to come up with a price for this highest value agent. What do we know about his value? Well, we can come up with a price that is based on everyone else's reported values because we're not selling to them. We're only selling to this one guy. And so we're going to use all of these other values. So this is all values other than I star and we're going to use the fact that this guy that we're selling to has the highest value. Okay. And putting all of this information together gives us some conditional distribution over this highest agent's value. Highest bidder's value. And we're going to look at this conditional distribution and we're going to pick an optimal reserve price from this distribution. We're going to pick an optimal price to sell the site to this agent from this distribution. Okay. So that's the mechanism that we're going look at. This is Vickrey with reserves. It's also called lookahead auction and it was first proposed by Ronen in 2001. Okay. So let me now convince you that this is a good auction. So let's use R to denote the revenue of this lookahead auction and let's use opt to denote the revenue of the overall optimal mechanism. And what I'm going to do is to look at how much revenue this overall optimal mechanism gets from all of the losers essentially. Everyone but the highest value guy. And then I'm going to look at the revenue that the optimal auction gets from this guy and L plus H gives me opt. Now, what did our mechanism do? Well, it focused on selling to this guy. And it used all of the information it had at its disposal in order to come up with the best price to charge to this guy. Okay. And so that tells us that this I have can be no more than R. Even optimal mechanism can get no more than how much we get because we do the best possible thing knowing that we're going to sell to this guy. Okay. On the other hand, how much money could the optimal mechanism extract from all of these losers? Well, their value is no more than the value of the second highest guy among everyone. So the highest among all of these people. Okay. That's the maximum possible value that they have. The mechanism can't possibly charge any more than that. But what did we do? We came up with a price to charge to this guy conditioned on his value being larger than this max. So one of the things we could have done is to basically sell the item at this value at this price to this guy and he would have always accepted at this price. And so our revenue is at least as large as the second highest value. Okay. And so that tells us that two times R is at least as large as opt, gives us 2-approximation to optimal revenue. And this is a simple argument but it turns out to be tight. There are settings where this mechanism will be a factor of two off from the optimal revenue. Okay. So what happens if we try to apply this mechanism to interdependent settings? Let me show you an example where it might go wrong. So now let's imagine that you have two buyers and you're selling one item. And recall that in an interdependent setting, people just get signals about this item and their value depends on their joined signals somehow. So here, I'm going to assume that only the first buyer gets any information about the item. So S1 is drawn from some distribution. And his value is equal to his signal. And the second buyer gets no information but his value is equal to the first guy's signal minus some tiny epsilon. Okay. So the first guy always has the higher value. In fact, their values are very close to each other, but let's imagine that we wanted to use some where end of the Vickrey auction, meaning that we wanted to sell to the guy with the higher value. Okay. What do we do then? Well, knowing the second guy's signal doesn't help us at all. The second guy is not getting any information. And so we try to price the item in the best possible way for the first guy. And let's say that the first guy has a signal distribution which does not get very good revenue for the seller. So this is called equal revenue distribution regardless of the price that you charge this guy. Let's say you pick a price of X. The probability that this buyer is going to buy the item is one over X. Your expected revenue is going to be one. So any auction that only sells to this highest value guy is going to get a revenue of one. But in fact, you could get much higher revenue by ignoring the first guy, selling to the second guy at a price equal to his value. Okay. The point here is that once I decide to sell to the second guy, I know his value precisely. It's what the first guy told me minus epsilon. Okay. So there's some informational asymmetry. I can get much better information about the second guy's value from the first guy's signal than I can about the first guy's value from the second guy's signal. And somehow my mechanism, my auction should take that into account. Okay. And so that's the reason that it's not always best to sell to the highest value agent. Pictorially, one of the things that's going on in this example is that we're selling to the first guy at his value at his critical signal. Okay. Whereas at his true signal, which could be here, even the second guy's value might be much larger than the price at which we are selling item to the first guy. So we would have been better off trying to exploit the second highest value. Okay. So this suggests the simple modification to the Vickrey reserve price mechanism. Here's what we're going to do. We are going to ask everyone to report their signals. And then we're going to flip a coin for every buyer and we're going reject them out right if the copy comes up tails. Okay. So everyone is rejected outright, probably one-half. Independently. And then so we retain about half of the guys in the auction and now we're going to run the Vickrey auction just as Ronen's mechanism did before. But of course, we have a bunch of rejected buyers. They did report some information to us. They reported their signals. So we are going to use those signals to compute the conditional value distributions of the remaining buyers. Okay. So I'm going to now give you a proof of the fact that this modified auction that we call the generalized Vickrey auction with reserves and admission control, so this first step is like admission control, I'm going to convince you that this gives an approximation to the optimal revenue. Okay. So again, we're going to focus on the case of two buyers which essentially has all the complexity that we need to worry about. And so let's say that at the reported signals, this is V1's value as a function of S1 once we fix S2 and this is V2's value. Okay. So once again, we are going to assume that the optimal mechanism gets some amount H from this highest value guy and some amount L from the other buyer. Okay. And we're going to think of the value of the second highest value guy as some of these two components. One is his value at the first guy's critical signal. Okay. And so that's L2. And L1 is this difference, this increase in the second guy's value from the signal S1. Okay. So what happens in our auction. Well, at the first step, with probability one-half, we reject buyer one outright. With probability one-half, we reject buyer two outright. And so with some probability, both of them are going to survive the first step. Okay. In which case we will run this Vickrey auction with reserve and our auction is going to get at least H condition on both of them surviving. Now, that takes care of this component that opt gets. In fact, it also takes care this have quantity L2 because the price that we charge to the first guy is conditioned on this guy having the higher value out of the two. And that price is going to be at least as large as this critical price here. So that takes care of this L2, our auction, this L2 can be no more than H. Okay. Now, all we need do is to make up for this quantity L1. And here, we're going to argue that our auction is going to get at least L1 in the case where two survives our admission control and one doesn't survive. Okay. And so if this happens, then the optimal auction is getting at most 2H plus L1 and we're getting at least H over 4 plus L1 over 4. Okay. And this gives an eight approximation and in fact, you can add just the probabilities in this first step I think the optimal thing to do is to reject everyone with probability one-third independently and then that gives a four and a half approximation. So you can equalize these terms to get a better approximation. We have this claim that we left and we need to prove this. So the claim once again is that there's this amount L1 which is the difference between V2's actual value and the two signals and the critical value, the value at the signal where the two values are equal. There's this difference L1 that we want to make up. And we argue that if two survives the first step and one doesn't, then we can get this value in revenue. So why is that? This requires another assumption. And the assumption is the following. So what's happening over here? We fixed the second guy's signal and this quantity L1 is the increase in the second agent's value from an increase in the first agent's signal. Okay. Now, let's say that the two signals had independent contributions to V2, for instance hypothetically let's say V2 was S1 plus S2. Okay. So here, we're just looking at the contribution to V2 from the first guy's signal, S1. The first guy's signal went up, S2 didn't change, and there's the improvement in V2's value. Now, what we are going to do for the second guy, here's one thing that we could do for the second guy, we could charge him V2 of S1, his own signal being held zero. Okay. So this value of the second guy, when his own signal is zero and the other guy's signal is S1, is something that's independent. This quantity is independent of his own signal. So we can imagine charging, you know, trying to allocate the item to the person at this price. And he's going to accept this price. Okay. And so our mechanism could get at a minimum this revenue from the second guy. Okay. And if we assume that the responsiveness of V2 to S1 is at least as large when S2 is zero as when S2 is something else, then this quantity is at least L1. Okay. So what is this assumption again? Let me restate it. So this is saying that I get some increase in my value from an increase in the other person's signal and this increase depends on what my own signal is. But I see more of an increase when my signal is low than when my signal is high. Okay. So the responsiveness of my value to someone else's signal decreases or does not increase as my own signal increases. Okay. Other people's signals become less important to me as my own signal goes up. That's this assumption and with this assumption, then the mechanism gets at least this amount of revenue. And so that put together with the proof I described earlier gives this eight approximation. Okay. So to summarize what I said so far, let me just compare the result that I just described against two other recent results on the interdependent settings. So there is a paper by Rough Garden and Talgum Cohen that characterized optimal mechanisms using a Meyerson type theorem for these settings. And there's another paper by Lee which gives an e-approximation via a Vickrey with reserve prices mechanism and we get this 4.5 approximation by adding admission control to that mechanism. So what are the assumptions that these results require? We all require some assumptions on the value functions themselves. Single crossing as I described earlier, all of the results need that. We need this assumption on how my value changes as a function of someone else's signal and how that change is affected as my own signal increases. This is a concavity type assumption that we need. The e-approximation does not need any such assumption but the rough garden and L paper needs several assumptions of this sort as you can see they are quite similar but in fact more restrictive than the kind of assumption that we have. >> [Indiscernible]. >> Shuchi Chawla: Yes. Yes. So this is exactly opposite, except when both are equal to zero, then both are satisfied. So yeah. >> [Indiscernible] very commonly? >> Shuchi Chawla: like convexity. >> Right. This is The first one is [indiscernible]. Yes. [Indiscernible]. >> Shuchi Chawla: >> So concavity to me seems somewhat natural. [Indiscernible] two conditions. >> Shuchi Chawla: >> Yes. So decreasing differences is concavity, right? The second condition, yeah, that is concavity. >> Shuchi Chawla: Yeah. Yeah. So this is measured along a different direction. Yeah. But it's -- the second derivative along a certain direction is less than equal to zero. Yeah. And it's a long and different, a certain direction. So yes. So these are all conditions on different -second derivatives and different directions and partials, second derivatives. So it's a little hard to compare, yeah. Does that make sense? >> Yeah. [Indiscernible]. >> Shuchi Chawla: Okay. Okay. So yeah, one thing I want to mention is that here's a sort of value function that this condition captures. So if the value of an agent is a sum over some functions of different people's signals or even a concave function of a sum of different people's signals but as long as each individual signal affects a different term, that satisfies this condition for instance. So those are the assumptions on the value functions. As I mentioned earlier, these works also require assumptions on the value distributions. So the rough garden and Al work for instance requires generalized [indiscernible] hazard rate affiliation symmetry. The e-approximation also requires generalized monotone hazard rate and our work does not need any assumptions on the value distribution. >> [Indiscernible] regularity? >> Shuchi Chawla: Not even regularity. Ronen's mechanism also doesn't need regularity, so that is really where this is coming from. Okay. Let me quickly mention that this approach extends to broader settings. So one way to extend this is to again think of every buyer desiring 1 item or 1 service. But then let's say that the seller has a feasibility constrained on which subsets of people can be allocated their items simultaneously. So in particular, there is a kind of feasibility constraint that is natural. This is called matroid constraint and some of you know what matroids are, and if you don't, that's fine, but let me give you two examples. One example of a matroid is just K-unit auctions. Where you have K units of the same item 1 is one side of a matching. So let's say you have items. Everyone has some subset of the items that they'd be satisfied by, and the people that you can serve feasibly are those that can simultaneously be matched to some items. Okay. So that's an example, another example of a matroid. So we show that VCG with reserves and admission control works also in matroid settings and here is what the mechanism looks like. You again take the buyers, ask them to report their signals, and then independently reject each one with probability one-half. And then over the remaining set that survives, we run this generalized VCG mechanism where we compute prices optimally, conditioning on all of the reported signals. And we show that this gives a constant factor approximation to the optimal revenue. So this is the result for matroids. You could ask, well, what happens if we have a more general feasibility constraint when we go beyond matroids? So it turns out that generalized VCG is no longer exposed incentive compatible so in that case, and the reason is somewhat similar to what happens if you don't have the single crossing assumption. So it's the same sort of reasoning that makes generalized VCG be no longer truthful. And so again, going beyond maitroids, we don't even know how to maximize welfare. Okay. And so to summarize, we give these natural mechanisms for interdependent settings that get constant factor approximations to optimal revenue without assumptions on value distributions and there are some very natural open questions going forward. One is to improve approximation factors, approximation 18 doesn't sound so great so can you do better? Remove the concavity type assumption that I showed that's a technical and it's possible that we don't need that assumption in order to be able do well. And then of course can we go beyond matroids, go beyond single crossing. How would we maximize welfare and expectation, let's say, and would that lead to revenue maximization and expectation? And finally, here's a very interesting open question. We kept talking about ex-post incentive compatibility and you could also define a notion of Bayesian incentive compatibility here and independent private value settings it turns out that there's no difference between the power of Bayesian IC and exposed IC mechanisms for single parameter settings and so you could ask, well, are Bayesian IC mechanisms more powerful in the interdependent settings? And if so, how much more powerful? And that appears to be a challenging open question. Okay. I'll take questions. [Applause] >> Nikhil Devanur Rangarajan: >> Shuchi Chawla: Thank you. >> Nikhil Devanur Rangarajan: >> Shuchi Chawla: [Applause] Questions? Yeah. [Indiscernible] all of next week? I'll be happy to drop more [indiscernible].