Ep 49. Insights from the NYU Wireless Workshop [Wireless Future Podcast]

追加: 2026-05-20

[00:00:03][music] Welcome to Wireless Future episode 49 and I'm Emil Bernson and as usual I'm here with Eric Larson. How are you?

[00:00:16]>> Hi Emily. Good. How are you?

[00:00:18]>> I'm great.

[00:00:20]So um what should we talk about today?

[00:00:23]Have you been traveling anything lately?

[00:00:26]Oh indeed I I have and um I attended u a few events recently and uh in fact one that I would like to tell you about Emil is the uh 6G workshop organized by the New York University in in Broccling New York that I attended last month.

[00:00:47]Um all right. Yeah. Have you have you ever been there? No, actually not. I've been invited a couple of times, but uh uh yeah, I'm trying to keep down my traveling a bit and unfortunately haven't fitted my schedule, but it looks every time like a very interesting event. I think during the pandemic, I actually joined online a few times.

[00:01:07]>> Yeah, but this time it was in real life and it was the first time. But I I mean totally I agree. I also try to keep down the [laughter] travel, but then it's of course difficult to resist when you get an invitation to speak at an event like that. and and uh it was a really great conference. I think in fact it was u something like a role model for how a scientific conference should be run and organized and executed >> and NYU that is like New York University in Brooklyn or >> that's right in Brooklyn. So they they've got this uh research center for wireless NYU wireless that that organized the workshop and this year the theme was on u 20 years of massive myo what's next.

[00:01:54]>> Wow 20 years so it really has been some time. Uh you can always debate of course when was the real birthday of of massive myo but what I think is less open to debate is the enormous impact that this technology has had and so that was like the leading theme of the event and there were uh like three main sessions and each session consisted of a lineup of invited speakers and then there was a short break with coffee and after the take there was a uh let's say free speech open mic session for about an hour where anybody in the audience could uh ask questions to the speaker or um ask general open questions or make statements and also the chair of the session had some questions for the speakers but more interestingly also uh for the audience. So the chair of the session might cast a an open question and then uh point to somebody in the audience. for example, Emil, could you opine on this uh this question or this issue? And then you were expected to grab the mic and and stand up or or walk up front and uh and deliver your your viewpoints um or or counterpoints, which I think is exactly the way that uh science should evolve and that this sort of um events should be run. So that was a lot of fun. Yeah, I think this is definitely a type of event that is really important for our research community particularly when these big conferences have been growing to such a large extent that uh uh yeah you can't really fit everyone in one place and uh no one really have time to listen to any paper presentations. So having this really discussions which gives you some value that you can't get from just reading people's papers that sounds like really what we are after in our community. Yeah, that's what we're after and what we are to some degree I think missing. I mean some of the big conferences might be better than others in terms of actually offering this kind of discussion for but um but by large I think we could do a lot better.

[00:04:13]>> Yeah. and and I think the setup that there actually are discussions and that is important part of it. I think that that's really good organized because I think often panel discussion a big conference is like okay you have one and a half hour and everyone gives an inter speech for 10 minutes which turns into 15 minutes and then there is like uh 30 minutes of discussions and only predefined questions so then it doesn't really turn into what you described here. Yeah.

[00:04:41]>> Ah yeah feel a lot I should have been there. So what did you really talk about or was there for example anything new related to myo there? I mean uh so um there was definitely a lot of discussion about myo and also I gave a a a talk that touched upon some aspects of of myo >> and one thing where I started off was that there is quite a bit of talking in the community and especially among operators that massive myo doesn't really deliver and then the question is is this true to start with and if it is true then what are the root causes of this observation? And uh in fact it seems that a lot of times we just got a single user in the cell and and and obviously if you only got a single user in the cell then massive myo isn't of much utility because the entire idea behind massive myo is the spatial multiplexing capability where you can serve you know in the textbooks I think we had something like 30 users at the same time with 100 antennas. maybe in practice you can't get 30 but you could at least get 10 or or you know maybe 20 or something.

[00:05:59]Um so that was one talking point >> right and uh I mean when it comes to that uh topic to my understanding uh it's a I mean that is true but it's not exactly like that my impression is that we have expanded the bandwidth of the system by five times or so compared to 4G systems and there might be many users in the cell multiple ones but it is enough to just give them different chunks of the spectrum and So on every subcarrier there it's roughly one user but not uh >> that's right and and that's such an important point and in fact this was also discussion point raised by the audience that you know maybe is it that the operators just have gotten too much spectrum >> so that we we aren't really air quotes forced or encouraged to to implement and make use of spatial multiplexing capabilities but we are still we're interlocked in is thinking of chopping the the the frequency and multiplexing wi with FDMA and so forth.

[00:07:06]>> Um it's possible. I mean it it really seems that if we can afford not to use the special multiplexing capabilities of massive MI or maybe spectrum has just gotten too cheap or there is too much of it in relation to the bandwidth I mean in relation to the data transmission requirements that we've gotten. Uh the other point I think is that a lot of the high layer protocols are just very poorly designed uh in that they don't they aren't constructed to make good use of what massive myo can offer at the physical layer because if at in fact a lot of the bits that are transmitted in the network come from services that are fundamentally streaming um at Maybe not completely at a fixed rate, but at least streaming. And there's just no reason why those should why those sources of data should be chopped up into small packets that get scheduled in the time frequency domain and so on. No, they should be just be sent. Everybody should get an wire in the air that's offered by the massive myospatial multiplexing and then as little retransmission protocols as possible as long packets as possible coding over as large time frequency bandwidth as possible which is something that seems to be just misdesigned in the standards and my understanding is that this is all because of legacy because we have this TCP IP underlying protocol stack and so forth. And we got HARK process on the application layer. We got HARK process on the access layers. And then we got another HAR process. Uh this is completely I think we're completely in need of a redesign of not only the physical layer but the entire protocol stack for high volume streaming data where massive myo really will shine. So that was like one of the points that I also >> um try to convey in uh in my talk and related to that. Yeah. No go ahead.

[00:09:16]>> Yeah. And I think people have for a long period of time talked about cross layer optimization as that it's important to optimize things across the different layers of the protocol stack. But I'm not sure how much of I mean and maybe some of those things are have been happening at the lower layers. You have optimized the resolocation and the physical transmissions. But yeah, I think people who are actually building the applications, they just try to do something that they think is good for their users in terms of that if you are scrolling through Instagram then they want the videos to to show just a few seconds automatically directly. So it downloads those things and they they do something there that they think are optimized for their application and the user experience but it might not be good for the network as a whole. Well, I mean it doesn't make use of what the physical layer and what massive myo has to offer, right? So that's the point and I mean just as you said cross layer optimization has been trendy for at least 20 years and and uh so basically we know we know how to do these things.

[00:10:14]uh it's there in it's it's in the papers in the textbooks but it isn't implemented and as a further um consequence of all this chopping of I mean you got data packets that are fundamentally large entire chunks of videos that you you know you're going to stream and you chop them into short packets that you schedule in the time frequency domain back and forth and so on and you don't make use of special multiplexing. So an additional consequence of that is that the signal to interference and noise ratio when you look at the OFDM time frequency grid becomes very random and fluctuates up and down which makes link adaptation extremely difficult if not impossible because the whole idea with link adaptation in massive myo is that you should schedule users for long enough so that you get a stable effective SINR that you can easily measure like reasonably accurately. measure and then select an MCS I mean modulation encoding scheme appropriate for that. So so that's another one you know so here we we could go on about this I guess and we are not I suppose but [laughter] >> no but I think this also somehow u yeah really reinforce the point that sometimes you people do something to make it simple but then it turns out that they have large scale consequences.

[00:11:35]So I think when people have been designing sort of the the scheduleuler for these systems they say we would like them to have as much mandate to do whatever decisions they want. So uh you shouldn't expect that just because I give you some time frequency resources right now you will get the adjacent ones uh in time or frequency in the next round. to there for you. You don't really know how your channel will evolve. And in the same way, even if you happen to know who will be scheduled and what interference that will cause in the neighboring sale right now, you have no idea what will happen in the future. And and this makes the whole scheduling much uh easier because there is no dependencies. You don't need to care about anything. But of course, you you get the problem that you described and later on when you try to guess what performance you can deliver right now because you have no idea about your interference.

[00:12:21]>> Exactly. I mean you want to you want to take out as much as possible of the randomness right so that you can predict what is the SINR I'm actually going to have over this time frequency block and and uh so you can adapt your transmission beforehand um um but it sounds what you are advocating for is a little bit how things were done maybe prior to 3G where you have this kind of circuit switch things where you you created in that case over frequency you created like here are the subcarrier mobile user. Would you like to have something that beam and those frequencies even if you don't have any data to transfer? I I I mean, let's face it, a lot of data is more suitable for packet based transmission. There's no question. But it's also true that the vast majority of the bits sent over the network actually come from sources that are fundamentally circuit switched or streaming like video transmission I think is a good example.

[00:13:24]Uh anyways and another point with massive myimo that's sometimes raised is that it doesn't work well at the cellage because the reciprocity based beam forming in TDD which as we know is the the the way to implement massive myo then uh the term the mobiles don't have enough uh transmit power >> um such that the the received pilots are too weak so reciprocity based beam forming doesn't work and I I think that's a legitimate concern. I mean, it's more than a concern. It's just the fact of physics that if you have a a a limited power, you spread it over a very wide bandwidth and the SNR per dimension goes down and at some point the reciprocity the base channel estimation will stop working. But [snorts] to me, the root root cause here really seems to be that you know maybe we are spreading over too large bandwidth. Yeah, we're coming back to that the operators might have too much spectrum because I think what typically happens is that the the mobile phones it has a particular transmit power. It is allowed to use say 100 200 matt and there are safety reasons and battery reasons why it shouldn't go beyond that. While the base stations their power typically can grow with a much amount of spectrum that they're having. So they have been growing while you haven't grown in the same way on the mobile phone and it creates these issues. That's right. The uplink down link imbalance probably has increased even and and and um you know it's a fact of physics that if you spread your fixed power over more and more bandwidth SNR per degree of freedom decreases so the only way is to use less bandwidth and more aggressive special multiplexing instead right and then of course one could then say that what they have been been doing now is to say if the amount of traffic that people are having is not high enough. So you can give them different chunks of subcarriers instead and just divide up like that. Well then they focus their power on those subcarriers spread it out over time and in a way that is a reasonable response to that problem even if it means to that we are not really utilizing the spatial multiplexing. That's right. Because then effectively, I mean, you're you're singling out parts of the spectrum or you spend your power and then you regain your SNR per degree of freedom. So that reciprocity based um general estimation works again, >> right? But I think that also then the fact that the I mean when you deploy a new network and if it has say five times the amount of spectrum and then hopefully some four or eight times more capabilities in terms of spatial multiplexing than in the past then it's not like you immediately will have more traffic that was just waiting to happen.

[00:16:16]So it it sort of uh makes good sense that the traffic will grow up over time and then maybe eventually we will need these capabilities more and more and then we will really see when the network gets stress tests that if uh they actually work well in terms of spatial multiplexing and I think what you heard was this kind of feeling that in the cells where the traffic has grown so much already uh yeah the implementations doesn't really um yeah work as well as theories predict. I think that that's what it seems like. Yeah. Right. Um Yeah. So then I mean related to this there was actually both in in my own uh talk but also in some of the other presentations and in the discussion there was a lot of talking about channel state information >> which uh kind of warmed my heart because uh [laughter] as we know I think it was actually Tom Marcelo who coined this phrase like 10 15 years ago now that channel state information isn't everything. it's the only thing and this has always been through and it remains true more so than ever and uh I mean it's really about getting the CSI for beam forming of course on uplink that's not an issue u because in principle you can decode blindly but on down link uh you need to know where and how to beam your power >> and uh we should keep in mind that there are a lot of different sources of errors in the channel state information or in the channel estimates, right? I mean, there is noise obviously on the pilots.

[00:17:48]There is contamination on the pilots from the just from the sheer fact that pilot sequences are used in other cells and so on. Uh there is uh channel aging because we estimate a channel and then um the mobiles uh I mean the terminals move around and then the estimates get outdated. So we have to reestimate and also in uh in cellf free or or distributed myo there might be phase errors uh between the different antenna panels depending on how they are phase aligned. I mean do they have like a common reference clock that they distill their phase reference from or do they have free running oscillators that have to be periodically aligned and so on. Um and uh the accuracy of this J state information is really a limiting factor which is a point that uh I took the opportunity to one more time uh drive home. But there were also other talks.

[00:18:44]There was one for example by uh uh Gilinia from China mobile who also reinforced this exact uh um conclusion and um you know especially in in cell andd my mode this is the number one technical problem we've got to solve in order to make the systems uh work and uh then I also had some interesting offline discussions about channel state information because you know I mean even with pointto-point myo of course channel state information [clears throat] is crucial right I mean you got an antenna array and uh you got the receiver and uh you want to uh achieve capacities and then we know the optimal thing to do is the singular value de composition and to do that you need channel estimates and uh I also had some discussions with some folks about OAM which is a particular instance or form of um pointto-point myo And I think an important takeaway here is that chance state information is as important for pointto-point myo for OAM and so on. And in some cases channel state information just means that you know you know how to point your antenna, right? I mean it's important even for the dish antenna.

[00:20:02]>> So it's important for a fully digital mray. It's important for an OAM. It's important for a dish antenna. And for the dish antenna on the OM it might just mean that you know you need to know how to precisely point your your your transmitter which is you know you got to know the asimote the elevation angles and so on that [snorts] is what channel state information mean in in in that yeah and um this can't be said enough many times I think because [laughter] yeah and that also connect to some of the topics we talked about in in recent episodes when we talked about sensing systems for example those that are system that are used by the militaries to really do sensing they can actually physically move around that is the horizontally and and point beams uh in or point their directivity in that way and uh we talked I think in the last episodes about movable antennas as one topic and I think one particular instance of that that people are researching is rotatable antennas you would have some base station that can actually rotate in different ways and can optimize and then of course you need an even deeper kind of channel state information uh for that to happen.

[00:21:12]>> Yeah, that's right. I mean >> and I think engineers they in 3DP so they are doing their best to do something out of this then I don't know exactly how well things are actually working in in practical systems if we are limited by fundamental problems or if it's just implementation challenges but I I read up on all the the things that they have been trying to do in terms of uh channel state information feedback because the alternative way of this reciprocity method that you are describing thing is that you transmit in a down link where you have more power.

[00:21:47]You transmit a number of different signals. You measure the channel and you defeat it back and then the the challenge instead becomes how do you describe your channel in a confined way so you can feed it back over this low capacity uplink uh uh link that you're having before you you you know the channel. They've been doing all kinds of things like determining how frequently in time should you sound your channel and estimate things how frequently in frequency domain depending on the frequency variations and uh what kind of cor can you divide up things in in spatial domain. So you say I know this link will only be described by five main directions representative and physical objects. There are all of those type of things there in the standard that you can implement. But I don't think it has been implemented because I think just the basic features there are >> feeding back a vector occasional to describe the the directions that consumes a lot of uplink feedback as well. So you can do it but it's always a research question. You can also send a lot of uplink pilots in order to build up SNR but it would eat up such a big fraction of your resources that it's not worth it.

[00:22:58]>> That's right. But I mean that's an important point. Of course rather than a reciprocity based beam forming you could send this feedback and then I guess machine learning would come to the rescue right I mean fundamentally measuring the CSI now at the mobiles and then compressing it is a data compression problem which we know AI is excellent at. So using some kind of neural network to do this it seems to be the way to go. My my point is rather that nature and physics gave us reciprocity for free. So why not exploit it? I mean could even be combined with with feedback, right? So yeah. Anyway, maybe we should not digress. [laughter] >> So you mean now that the reciprocity isn't everything is the only thing?

[00:23:43][laughter] >> It might not be the only thing. you know, it's it's something that's out there that nature gave us for free and we know it's a very juicy and good thing. It's a very useful thing. So, we should we should exploit it. I think that's Yeah. Yeah. And for those who who also wonder where does this kind of quote come from, there's actually an American football coach who once said, "Winning isn't everything, it's the only thing." So, some >> Yeah. And I think there are many versions of this ones if you Google around that people have been adapting this for a different topic. Mhm.

[00:24:16]>> Yeah. Wow. So, >> so what else did people talk about?

[00:24:21]Anything sensing related?

[00:24:23]>> Yeah, I mean uh of course sensing because it seems like it'll become an important aspect of of 6G with integrated sensing and communications.

[00:24:32]And one point there I think that also I made is that wireless sensing integrates very naturally together with dynamic duplex >> in in um in in in in myo. Um >> so what is dynamic duplex now again?

[00:24:50]>> So dynamic duplex is when you shift the switching point so that normally everybody switches from up link to down link at the very same instant in time.

[00:24:58]But with dynamic duplex then you shift so that not everybody switches from up link to down link at the same time. So you create like some part of the slot where access points or base stations actually can can send to each other and you you can serve up link and down link users simultaneously. And you know there are some capacity advantages or some advantages in terms of capacity of doing this if you have a lot of imbalance between uplink and down link traffic for example.

[00:25:29]>> Would this be another source of this kind of interference that we talked about before that is unbalanced >> totally there will be interference. I mean in dynamic duplex there is interference between users and between access points and so on. But but actually it is this interference between access points that could be exploited for sensing right because you would then create a part of the slot where one access point transmits another one receives and that's exactly what you want in a bi static sensing setup. So yeah, no I've been thinking about a simple version of these type of things myself that the natural thing in such distributed M setup would be that uh when you're transmitting in the down link from your base stations a few of them are turning themselves into uplink mode and just listen instead and of course you can do this even more dynamically by switching in different ways uh as well.

[00:26:19]>> Totally. I mean that's that's you know I think uh again dynamic duplexing comes very naturally together goes hand inhand with with wireless sensing. So that was one thing I think looking at the future it's very likely that that a lot of data in future networks will uh come from um AI related applications.

[00:26:46]Mhm.

[00:26:47]>> Uh the argument that um I've been making is that AI has gotten extremely impressive. I mean it's basically read the entire internet and everything that any human has has you know written or spoken or every every blog out there every show on YouTube every podcast um every scientific paper probably. So the only way for AI fundamentally to get any better will be that we feed it with more data.

[00:27:23]>> Mhm.

[00:27:23]>> And there is only so much data that humans can generate.

[00:27:29]So all this data is more likely to come from sensors of various kinds. I mean could be cameras obviously, could be radio, could be RF, could be anything.

[00:27:37]But bottom line very likely this data it's got to be transmitted somehow. And a lot of that transmission is going to happen over wireless. So I think if if this AI revolution is to continue then it will necessiate um [gasps] highcapacity wireless networks that can deal with all the information. And then we should keep in mind that a lot of this data I mean raw sensor data for one thing but also you know when you train machine learning models and you send updates and use gradient descent and you send these gradient models and so on.

[00:28:10]this kind of data is fundamentally different from sending [laughter] like the kind of bits that we are used to transmitting. M >> um so that in turn might very well entail also a redesign of the the entire protocol stack from the physical layer and up and spec one one specific technique that I found a bit find a bit fascinating is this overthe-air data aggregation where you basically rely on superposition principle of the wave equation to you know everybody sends like data and it kind of adds up in the air so you're interested not in the individual streams but only in the sum Um and interestingly I mean if you want to make that work with coherent superposition then we are back at these channel state information and then phase drift and phase noise issues um uh that we that we mentioned earlier.

[00:29:06]So also there it seems that channel state information is going to become one of the limiting factors. So this u punchline on on on channel state information that we talked earlier [laughter] you know remains also pertinent there. So that was one point.

[00:29:26]Yeah. Yeah. And if I understand this correctly, it's like uh you have some data that you have collected and you want to transfer to other place and if the what you will do as soon as you get that data at in as input to an AI training or so is that you will just combine the data in a particular way. Uh then it unnecessary to send them as parallel non interfering streams over there and then combine them at the receiver. You want to do it over over the air there.

[00:29:55]>> That's right. If you're only interested in the let's say arithmetic sum of the data, right, which is the case in some of these machine learning applications where you train models using the centralized data and so on, >> right? But but then you also come back to this situation where you uh in order to do that uh you want them the different streams to be uh yeah having similar sync to noise ratios so that they are adding up uh not only in phase but also with the similar strength and I think that would be a challenge in practice.

[00:30:25]>> Yeah it is a ch it's a major challenge.

[00:30:28]I mean you're right. I mean not only you want the phase to align to to align uh but you also need a scale the amplitudes essentially you got to invert a channel right for this to happen >> in its most basic form at least. So uh so that that's like one general topic that was also discussed a bit and that I that I talked in in in in my uh own presentation.

[00:30:51]But do you think there is anything one can do uh when it comes to overtheair computation which doesn't require this phase synchronization? I mean at the end of the day you you want to transmit something you want at the other side to compute an estimate of maybe the sum or something. That's right. and and you can perhaps extract something anyway without face synchronization or what do you think >> one can yeah I mean there are non-coherent versions of it where you basically send a a um you know everybody sends a sequence weighed by an amplitude and the receiver estimates the power which you [clears throat] get very noisy measurements but it still works to some degree in in in some applications especially when you feed these estimates to to like this gradient updates to a machine learning algorithm, they tend to work fairly well because those algorithms are so resilient to noise.

[00:31:43]It's is is pretty much enough that you know whatever you receive on average has is has the right sign. So the the expected value is positive and then [laughter] okay it's it's horribly noisy but so what it still works.

[00:31:58]>> Um so yeah maybe we shouldn't go there today. I mean but it's a a whole uh you know very technical topic which is quite interesting in itself. So >> yeah.

[00:32:08]>> All right. So did anything else you picked up from this sessions?

[00:32:13]>> I mean there were a lot of other interesting now we talked a lot about points that that I brought um and there were a lot of other interesting talks as well. For example, there was one by my my colleague Sarah Hamar who spoke about um fading models and especially about the iid fading that I think you and me have debated in the past [laughter] quite a bit and and and the punch line here was that it's a very rare species.

[00:32:37]I mean you have to look very hard to find it. But he actually did find it in some measurements where they were in like an indoor environment with really a lot of scattering, a lot of metal that really creates this kind of isotropic environment where you get almost iid fading which is kind of neat. I mean because a lot of times as we know you have more like correlated fading and you have line of sight path that can be quite strong and so on. And also in the same talk there was the point being made that well now if we don't have iid really fading maybe we can create it somehow and how would we do that I mean one way is this notion this concept of swarm repeaters to deploy network controlled repeaters uh calibrate and configure them so that they are reciprocal and deploy them in in large numbers so that they basically each one of them acts as a scattering point in the channel but with significant amplification and didn't we talk about that Emil the other month on the podcast? I think we we did right. Uh >> maybe at least we have talked about yes swarm repeaters maybe it wasn't in last year but uh but when you describe this from from from the perspective of wanted to create ID fading is the goal then to have like a pointto-point link with multiple antennas on both sides and give you a a good channel condition there so you can send multiple data stream between two places. I mean I I think it's more generally to of course you can never create IID fading but you can at least enhance the richness of the channel right so basically you can of course improve coverage if you got like coverage holes or or your blocking you know um poor SNR because of blocking but as importantly I mean you can uh you could create additional propagation pads that increase the rank of the channel which would be mostly important as you said for pointto-point um or when you have very closely separated receivers so that you start with a a channel that has low rank and now you add additional paths to it uh to increase its rank.

[00:34:53]>> Yeah. Yeah, I was asking that question because I think uh when people have been working with theory, I think there there's a good understanding and even mathematical proof of that if you want to have a high capacity for a given SNR or pointto-point link, you want to have something ID like so. So you get the many different paths but then when you do multi-user transmissions uh then each individual link doesn't necessarily need to have ID fading if the people are physically separated involve different set of scattering objects. might even be better because they become more isolated from each other.

[00:35:27]>> Yeah, you're I mean that's a good point.

[00:35:29]We got to be careful to distinguish multi-user myo from pointtooint myo here. And one thing we know is that multi-user myo if you get single antenna terminals works pretty much as well in line of sight as it works in iid fading.

[00:35:44]So then there wouldn't be a lot of point in enhancing the the rank of the channel. But when you got multi- antenna terminals then effectively you create a pointto-oint myo channel to each one and then it still pays off to to to to increase the rank.

[00:36:02]>> Um so I think by large I mean scattering is a good thing and especially when you can introduce scattering that has significant amplification of the pads which is precisely the point of this horn repeater concept.

[00:36:15]>> Yeah. And I mean if you're serving two uses that are relatively well separated from the beginning, you put this a repeater in between them that amplifies the mix of the signals. You can actually degrade the performance unless you do cross layer optimization as we talked about before. So it's sort of it's not necessary that you just throw out hardware there and everything becomes better. Yeah. Channel state information and the good design of them is important to really get what you want out of it.

[00:36:42]>> Yeah. No, I agree. I mean, it's reassuring to also hear you saying this, Emil, because it means that there are still there's still a lot of good topics for uh research and development in wireless cons.

[00:36:57]>> Uh in fact, yeah.

[00:36:59]>> So, what about frequency bands? Uh are we stuck below 7 GHz?

[00:37:03]>> Yeah, I mean, so [laughter] there's this endless debate, right? How high should we go up in frequency? And um of course millimeter wave didn't really become such a great success for 5G as we know because I think some countries even installed equipment and then decommissioned it because there was not enough uh like need for it. But what we don't know obviously is in the future what new applications that will come along and that will really require this enormous data rates that we can offer when we have this at these high frequencies.

[00:37:39]Um so in the workshop there was also a lot of talks and discussion about higher frequencies and especially terahertz which is something where I mean I was really in the learning mode here because I don't have like much working experience on it and I found it pretty interesting and uh specifically there were there were presentations that I found quite interesting about moving as much as possible of the signal processing saying as close to the antennas as as you can. I mean using low precision 80 converters obvious which is by the way not a new concept as we know but still it can become very useful in this context and uh there was also a very interesting talk by Daniela Cabridge on um using true time delay beam formers at terahertz rather than just building a phase array and shifting the phase you actually implement a small configurable time delay per antenna element which brings huge benefits when you have wide bandwidths and these large arrays, >> right? Yeah. And I think that is also a concept that has been around for 20 plus years. But the question is whether one can now actually build such unit in a power efficient way.

[00:38:56]>> Yeah. But but I mean to be fair I think a lot of the ideas that we are talking about in coms engineering and and myo and so so on. and they have roots that go back very far in time. And maybe it is that for the terahertz, you know, particularly definitely there will be signal processing challenges um that are novel. But above all, it seems to me that these are implementation challenges of actually building the micro electronics and getting that to work. And then maybe in the end it will be like you know here is what is actually feasible to do in terms of of of of constructing the hardware and that hardware will have certain limitations and then we've got signal processing problems that comms engineers can [laughter] come in and and and attack.

[00:39:55]For example, how to given that we have true time delays that have certain capabilities that you know you probably will have like a maximum delay, you will have quization steps and so on. Uh what is the best way of using such an architecture and I think there are still interesting uh partly open problems to tackle. So pretty interesting topic indeed.

[00:40:17]>> Yeah. No and I think if we talked earlier about that maybe operators have too much bandwidth and don't need spatial multiplexing. I think if we would build uh links at terahertz frequencies we will have immense amount of bandwidth there. So it's it's mainly about closing the link pointing the beams towards each other and and overcoming those challenges not to >> but are there any other phenomena that are showing up there that are interesting beyond sort of just pointing beams towards each other and make use of all the spectrum. I mean there is some talking about uh the fact that we in a lot of these systems will be in the in the geometric near field of the antenna array. So the beams don't exactly look like in a lot of the text books. I mean you got this like very cartoonish uh but that's not what the beams look like and depending on how you configure your phase uh shifts or your your time delays you can get beams that take on different shapes. And um you know particularly you can get like beams that look a little bit like curved like this and so on which is quite nice. But what we should keep in mind here is and there was also a point made by by somebody in the audience. I was about to state this point but then somebody else made it before me that you know look the oldfashioned time reversal beam forming conjugate I mean time reversal in the time domain uh which is the conjugate beam forming or maximum rate transmission in frequency domain is always optimal you can never do better so given that you have accurate enough chance state information which means in this context that you know you basically know if you're in in like free space line of sight. It kind of just means to know the the angle and the distance. If you got blocking objects, you got to know where are these objects and so on. Um so that you can compute the channel impulse response from your target point to each one of the elements in your array. You can never beat maximum radio transmission.

[00:42:17]Uh also known as carne beam forming or time reversal beam forming. Uh I [snorts] think that's important to keep in mind here. uh which again is something that's been known for decades.

[00:42:27]I mean, so you know, [laughter] >> yeah, and and I think sometimes people are just describing this, oh, as a collection of different kinds of other beam shapes that are better in other ways. But yeah, I think you're totally right that at the point where you want to focus it, what you describe is the optimal way. It doesn't matter what be if you want to create some other beam shapes, you should have other motivation for doing that. You basically say, here is my focal point. I can trade away so that I could get a little bit less power there. And what I gain is that I can shape the beam and that could potentially be useful if you don't really know where your true focal point will be and you want to widen this area somehow. Yeah, absolutely Emil and and that's a really interesting relevant topic I think also to work on that you know we we aren't only interested in maximizing the S&R at the at the target or focal spot but we want to make it like as wide as possible but also there we should keep in mind that there is a huge body of literature on you know robust beam forming techniques and all sort of things let's not go there today but uh but I'm sure more could be done and especially at these high bandwidths and you know high data rates and also high Dopplers. We need computationally efficient algorithms to do the job. So I think there's ample of opportunity here for comms engineers and signal processing engineers to invent new algorithms.

[00:43:55]>> Right. So if we then leave the the wireless common sensing were anything about computations or anything?

[00:44:04]>> Yeah, definitely. I mean of course there's a lot of talking about uh again AI decentralized and federated learning as possible drivers of um new networks design for if not for 6G maybe for who knows what I mean I suppose there will be a 7G and 8G and development is not going to stop right and but it was also a thought-provoking talk that basically asked the question whether Um the demand for wireless coms is becoming saturated.

[00:44:40]So that innovation will rather be driven by desire for cost reduction more so than getting more bits, getting like better coverage and and all of that. And I have less intuition for this. I mean, I can somehow sympathize with the sentiment that now we got this technology which after all works very well and it's quite cheap for consumers.

[00:45:10]Um and if we are to you know roll it out uh in more places in the world and we are going to improve coverage in in like weak spots then maybe we don't want to pay a lot more and that means we need to make it more cost efficient. So that was one thing and there was some talking about um spot markets for spectrum among operators. Could we trade spectrum rights on a millisecond level basis >> which is not a new idea you know we had a European project on this that was now if I remember well 15 years ago Sapphire was the name where the idea was to trade spectrum on on a slot basis between operators and these ideas seem to be coming back periodically and uh it's difficult to tell the realism in it because it would entail a lot of regulatory change presumably it's quite interesting to think about and how would you integrate it with modern myo how would you integrate this with massive myo I mean that I think is largely an open question so that was a bit interesting yeah and I think this also comes down to who uh is owning the customer that uh are then paying something for getting this kind of access because we we have with regulation try to created markets by uh saying that well we should have some three, four different national telecom networks in in each country to create some kind of competition. And if you compare this with like a uh the power grid, yeah, we have multiple companies in Sweden that build power plants and they are building power lines and so on, but we don't they don't cover the same regions. Don't build four different power grids that are on top of each other. So uh and what people are then doing is that they they get a contract from a company that is then buying power from these different places and put together a good portfolio for you and hopefully give you a good price. So potentially the business models for for for access should be be better or different in the future where you you buy connectivity that can come from any network but of course that would be very radical change for these uh companies that you are today having your contract with and owns these physical networks.

[00:47:29]Yeah, probably. I mean, and there was actually also a question uh cast by someone to the audience that, you know, look, we got all these services, right?

[00:47:38]And as a consumer, say Emil, that you were to pay another $5 a month for your, you know, cell phone subscription, what service would motivate you to actually um would make you feel that this extra cost is worthwhile given what we have and how how well it it works. and somewhat startingly you know nobody in the audience could really deliver a a uh a compelling answer to that. So I think here is something to think about. Anyways, [snorts] so it was altogether a good discussion on on some uh let's say more softer maybe topics uh that uh are are definitely relevant for for any 6G discussion any any beyond beyond current standards discussion.

[00:48:31]>> Yeah. And it's it's definitely important also from this viewpoint that we can design technology we can standardize it but if it's never deployed or very uh limited areas where the deployments are happening then uh yeah is it really going to be an important and have an impact in the future and I think technologies that can bring 60 connectivity out in the countryside and provide good services there of course the capacity doesn't need to be the same because the demand is not the same But if you can go around anywhere and feel that my service is the same from my viewpoint, then yeah, I think that would have a much bigger impact than maybe squeezing out a few extra bit in the city center where you anyway have can just deploy things more densely.

[00:49:17]>> Yeah. [snorts] Um, you know, totally. So, >> so, so finally then, do you have any other thoughtprovoking conversations that you would like to share with the audience? I I mean I I had a lot of good conversations on different topics uh you know ranging from like more general aspects of next generation wireless to more specifics and what one conversation that I remember especially um vividly is uh at the lunch table together with Hansen he's a an expert on electromagnetics and Ruth GMI is a a grad student of Thomas Marcetta. We were sitting together and uh we were talking about because I both Ruth and I raised the point coincidentally we were sitting on the same question because a lot of papers make statements about uh super directivity in uh antenna arrays and about mutual coupling between antennas and so on. And a lot of people claim that super directivity phenomenon is is a consequence of mutual coupling which I could never get my head around. I mean I felt like why would it be the case and I don't understand you know what they're saying like this. So I brought this point and u it is very clarifying because super directivity has nothing to do with mutual coupling and it's certainly important thing to understand that it's something that seems a lot of people has gotten wrong at least it's wrong in a lot of papers. Yeah, I could probably have said something like that myself without having studied the topic.

[00:50:54]Right. Now, >> here's the thing that mutual coupling is a phenomena that you you'll be subjected to in an antenna array where you got antennas that sit well distance doesn't matter but it's only significant if they're like close to each other, right?

[00:51:07]So, so let's say you got two antennas close to each other and you induce a current in one of you drive one of them with a current. Okay? uh then that antenna will radiate an electromagnetic field and that in turn will induce a current in the other antennas. They will like affect each other. They couple with each other and to model this properly is not at all trivial. I mean but the the physical phenomenon is called mutual coupling. Okay.

[00:51:37]>> And then almost unrelated to that there is the phenomenon of super directivity.

[00:51:44]Okay. which is that with an array of M antennas, let's say in in free space line of sight. Okay. Uh conventional textbooks in my typically teach that we can achieve an array gain of M.

[00:51:58]But actually the array gain that we can achieve is M squared H that's a super directivity phenomenon and this can be achieved even if the antennas are point sources and point sources can't affect each other. I mean there you know you drive a point source with a with an ideal current source it is what it is really is what it does right um >> right but if you just superimpose the signal from multiple sources uh and there are m sources then it seems like it should scale with m so so where does this m square come from >> yeah so now this getting pretty technical Emil I think it would deserve an episode in itself right but it's actually explained in uh one place where it is well explained is in the book chapter by Tom Marcetta and Hansen and and and and and and me that came out in uh let's see was the title of it's an edited book that might not have gotten a lot of visibility. I'm actually not sure but I can send you the chapter came out the other year and uh so here's the thing that in order to compute the array gain properly uh in this case what you got to do is let's say you do it for scalar field so it could be like a pressure source okay it could be like a liquid or or or air so the the field here is the air pressure so that means that a point source means that you inject a tiny amount of air or you suck out a tiny amount of air at each one of the point sources And then what you got to do to compute the array gain is that you can definitely sum up because of superposition principle the green function which is like 1 / r ether j k into r uh any point in space you can sum up from the array. Okay. But then to compute the array gain what you got to do is that you're going to compute then this amplitude. Okay. square it and then normalize it by the total radiated power. And the total radiated power is what you get if you you take this amplitude squared, okay, and you integrate it over a sphere that encloses the the source array. And that integral is a little bit uh you need a few lines of calculation to get it. But then you'll see in the end you get something which when the separation is large between the sources the gain is m but when the sources get closer and closer and closer the gain approaches m² and that is the super directivity effect which has nothing to do with mutual coupling because there is no mutual coupling between point sources. Now that said, it's also important to understand that in any practical system, if you were to build antenna, an antenna array, then you would have dipoles or patches or you know whatever. Okay. And [snorts] if you put these closer because super directivity only can happen if antennas are closer than half a wavelength. Yeah.

[00:54:46]Now if you put real antennas like dipoles closer than half a wavelength in order to achieve super directivity then mutual coupling will become very significant. So they kind of go hand in hand but one isn't the consequence of the other. Super directivity can be achieved with point sources and when you have an when the sources are close enough together and in that case mutual coupling becomes a very significant effect. Uh so uh that was a very good discussion which confirmed something that I thought you know this is what it's got to be like but a lot of papers claim something else and we felt it's very strange why they say like that.

[00:55:31]though but let's see if I I grasp this correctly. So is it so that uh you get this m square not necessarily become because you get an unnaturally strong single particular direction but just that you radiate much less than you would normally expect and therefor yeah I mean in in like information theory my books then the radiated power is the sum of the squared amplitudes of the signal you feel to the array right but the physically radiated power is can be quite different >> and when you have half wavelength spacing they coincide in in free space.

[00:56:05]Okay. But when you pack the antennas more closely, then the radiated power becomes something else than the sum of the the input amplitude squared. Yeah.

[00:56:15]I'm going to send you the chapter, Emil.

[00:56:17]It's it's quite uh comprehensive. It's only a couple of pages actually to to very lucid derivation that.

[00:56:24]>> All right. Right. That's that's really nice and very enlightening. And I I hope that the audience who stayed all the way to the end of this episode also found this enlighting as well. So thank you very much for sharing your experience from the NYU Summit there on 6G. And >> thank you Emil. So >> and uh I hope uh our audience will come back then for the next episode which will be number 50.

[00:56:50]>> Wow. Number 50.

[00:56:53]>> Yes. So thank you very much for listening and see you in the next episode. All right. Thank you, Emil.

[00:56:58]Bye-bye. [music]