Sounding, and why a phone switches antennas

A base station is about to transmit to a phone. To do it well — to point energy at that phone rather than spraying it, to pick a modulation the link can actually carry — it must know what the radio channel is doing right now. But the channel it needs to know about is the downlink, and the only place the downlink can be measured is inside the phone.

This note is about the trick 5G uses to get around that, and about the piece of hardware the trick requires. The trick is called sounding. The hardware is a switch.

In one line. The network learns the downlink channel by listening to the uplink. That only works for antennas the phone actually transmits from — and a phone has more antennas than transmitters. So it transmits from each of them in turn, through a switch. Everything else here is detail.

The whole answer is in a title. The procedure lives in TS 38.214 clause 6.2.1.2. Secondary sources cite that number constantly and almost never say what the clause is called. It is called UE sounding procedure for DL CSI acquisition — uplink sounding, for downlink channel state. Once you have read the title, antenna switching is not a curiosity any more. It is the only way the sentence can be true.

How to read this note. Like its predecessor, it is built as a ladder and gets harder on purpose, in one direction only. Nothing later changes anything earlier, so you can stop at the end of any part. Every fact carries a clause number, so you can check it.

The question the network cannot answer on its own

What the transmitter needs to know

A radio channel is not one number. Between each transmit antenna and each receive antenna there is a complex gain — how much the signal shrinks, and how much its phase turns — and that gain is different at every frequency and changes as things move.

Collect those numbers into a matrix and you have what 3GPP calls channel state information, CSI. With it, a base station can do three things it otherwise cannot:

All three need the channel from the base station's antennas to the phone's antennas — the downlink. And that is measurable only at the phone.

Two ways to find out: ask, or measure

There are exactly two strategies, and 5G uses both.

Ask the phone. The network transmits a known signal — a CSI reference signal — the phone measures it, and reports back what it saw. This works in any band. But the report has to be compressed into a few bits from a fixed codebook, sent up on a control channel, and it is out of date by the time it arrives. The overhead grows with the number of antennas, which is a problem precisely when antennas are numerous and beamforming is most valuable.

Measure the phone. The phone transmits a known signal — a sounding reference signal, SRS — and the base station measures it directly with its own receivers. Nothing is quantised, nothing is reported, and the base station ends up holding the raw channel rather than a phone's opinion of it.

The second is better whenever it is available. The catch is in the word available, and it is the subject of the next two sections.

Reciprocity — why an uplink measurement answers a downlink question

Sounding measures the uplink: phone to base station. The thing we wanted was the downlink. So why is this allowed at all?

Because of reciprocity. The propagation between two antennas — the reflections, the delays, the phase turns of the air itself — is the same in both directions when it is the same frequency at the same moment. Radio physics does not care which end is talking.

So the chain of reasoning is now complete, and it runs backwards from what the network wants:

Read the last two boxes again. Everything up to "the phone must transmit from that antenna" is physics and arithmetic. The last two boxes are about what is inside a handset — and that is where this note stops being about a standard and starts being about a product.

The asymmetry inside every phone

Here is the fact the whole procedure is built around, and it is a commercial fact before it is a technical one.

A phone has more receive antennas than transmit chains.

So the ordinary handset is something like one or two transmitters and four receivers. The network, meanwhile, would like downlink CSI for all four receive antennas, because all four are what it will eventually transmit into.

That is the gap. The phone can measure four antennas but can only speak from one or two. And the network can only learn about antennas the phone speaks from.

The resolution is time. If the phone cannot sound four antennas at once, it can sound them one after another: transmit from antenna 0, throw a switch, transmit from antenna 1, and so on, until the base station has seen all four. The channel must not change much while this happens, which is why the whole cycle is measured in tens of microseconds rather than milliseconds (§8).

That is antenna switching. The rest of this note is what 3GPP had to write down to make it work, and what it costs.

What a sounding reference signal actually is

Before the procedure, the object. An SRS is defined in TS 38.211 clause 6.4.1.4, and it is smaller and simpler than its reputation.

Ports, symbols, and where it sits in the slot

The unit of configuration is an SRS resource. One resource is (clause 6.4.1.4.1):

That backwards count is not a quirk of notation. An SRS block is anchored to the end of its slot, not to the beginning. Sounding is the last thing an uplink slot does — the data and the control have already had their turn — and anchoring to the tail means the sounding block keeps its relationship to the slot boundary whatever else the slot contains.

The sequence itself is chosen for the amplifier's sake. Clause 6.4.1.4.2 builds the SRS out of $r_{u,v}^{(\alpha,\delta)}(n)$, which is defined in clause 5.2.2 — Low-PAPR sequence generation, type 1. That is the same PAPR problem the frame-structure note ends on (§9.2 there): a reference signal is transmitted at the edge of coverage by a battery-powered device, so it is built from sequences with a small peak-to-average ratio. The hardware is visible in the standard again, one clause down.

The comb, and how many phones fit on one symbol

Sounding is pure overhead — it carries no user data at all. So the design question is: how many phones can sound the same symbol without interfering?

The first answer is the transmission comb. An SRS does not occupy every subcarrier. It occupies every $K_{\text{TC}}$-th one, where $K_{\text{TC}} \in {2, 4, 8}$ (clause 6.4.1.4.2), with an offset $k_{\text{TC}}$ that says which of the interleaved sets this resource gets.

The second answer is the cyclic shift. Two phones on the same comb offset can still be separated, because the base sequence can be rotated, and rotations of it stay orthogonal. The number of rotations available depends on the comb (Table 6.4.1.4.2-1):

Forty-eight phones can sound the same OFDM symbol. That is the number that makes sounding affordable at all, and it is why the standard spends a clause on combs rather than simply giving each phone its own symbol.

There is a price, and it is the last column. A comb-8 resource samples the channel at only one point every eight subcarriers. The channel must be smooth enough across that gap for the measurement to mean anything — which is another way of saying the echoes must be short enough. The comb is therefore a frequency-resolution setting disguised as a multiplexing setting.

When it is sent, and what it is for

Periodic, semi-persistent, aperiodic

Resources are gathered into an SRS resource set, and the set carries a resourceType (TS 38.211 clause 6.4.1.4.4, TS 38.214 clause 6.2.1):

The three exist for the same reason bandwidth parts exist: battery. A phone that is not being scheduled should not be sounding, and a phone that is about to be scheduled should sound immediately rather than waiting for the next tick.

usage — one signal, four jobs

Every SRS resource set carries a parameter called usage, and it decides what the sounding is for. This is the parameter that matters most for this note, because only one of its values is about antenna switching — and the existence of the other three is why SRS is in the standard even where reciprocity does not apply.

Read the last column downwards. The first three are questions about the uplink. Only the last one is a question about the downlink — and that is exactly the difference that forces a switch into the phone.

Antenna switching, read from the title of its clause

Here is the clause, named in full:

Take the title apart. UE sounding procedure — the phone transmits. for DL CSI acquisition — so that the network learns the downlink. A procedure in which the device transmits, in order that the network may learn what the device would have received.

That sentence contains its own hardware requirement. If the network is to learn the downlink at UE antenna 3, then UE antenna 3 must at some point transmit. And if the phone has only one transmitter, the only way antenna 3 can transmit is if the transmitter is temporarily connected to it — which is a switch.

Look at what the picture says about cost. Three of those four antennas exist for receiving — that is what pays for them. Antenna switching borrows them, one at a time, for a single OFDM symbol each, purely so the network can find out what they are hearing.

The clause never says "switch". It says the SRS port of each resource "is associated with a different UE antenna port". That is the standard being careful: 3GPP specifies the observable behaviour — different resources must come out of different antennas — and leaves the phone to arrange it. A switch is how it is done, not what is required. But the guard period in §8 is sized for one, which is how you know.

The tXrY grammar

A phone declares what it can do with a capability called supportedSRS-TxPortSwitch, whose values look like t1r4 and t2r4. The grammar is simple once seen:

t transmitters r receivers. t1r4 = one transmit chain, four receive antennas. t2r4 = two and four.

3GPP writes the same thing in the clause as 1T4R and 2T4R. When the two numbers are equal it writes 1T=1R, 2T=2R, 4T=4R — and the equals sign is doing real work, because those are the cases where no switching is needed at all: every receive antenna already has its own transmitter.

The invariant is worth memorising, because it is the whole idea in one line:

So a 1T4R phone needs four turns, a 2T4R phone two, and a 4T4R phone one — which is another way of saying that buying transmitters buys back time.

What the standard actually makes you transmit

Clause 6.2.1.2 is long, and most of its length is combinations. Stripped down, it obliges four things.

And then a fifth requirement, which is the one this note is really about:

5 · A gap between the resources. "The UE is configured with a guard period of Y symbols, in which the UE does not transmit any other signal, in the case the SRS resources of a set are transmitted in the same slot. The guard period is in-between the SRS resources of the set."

Not a gap in which the phone sends something else. A gap in which it sends nothing at all. Airtime, deliberately wasted, in a standard that fights for every symbol elsewhere.

Why would a specification insist on silence? §8 answers it, and the answer is not in this document.

The guard period is a switch

This section joins two tables in two different specifications and gets a piece of silicon out of them. It is the reason this note exists.

The table, and the oddity in it

Clause 6.2.1.2 ends by giving the length of that mandatory silence:

Two things about that table should stop you.

First, $Y$ is not constant. Every other quantity in NR's timing is a fixed count of symbols — fourteen symbols to a slot, at every numerology, always (§4 of the frame-structure note). A count that changes with $\mu$ is unusual, and it means the thing being counted is not really symbols.

Second, $\mu = 4$ is missing. There is no row for 240 kHz. That spacing exists in the numerology table and is used for synchronisation signals, but it is not a spacing you sound at, so the row would have no meaning.

A count of symbols that grows as symbols get shorter is the fingerprint of something measured in seconds and merely expressed in symbols. So: how long is $Y$ in real time?

Where the real number comes from

It is not in TS 38.214 at all. It is in the RF specification, and it is stated plainly.

Ten microseconds is the answer. That is how long the standard allows for the phone's transmit path to change state — for the switch to move, the amplifier to settle, the power to come up on the new antenna and be clean enough to measure. It is not a protocol number. It is a settling time.

Why a physical switch needs microseconds. The switching element itself is fast. What is slow is everything hanging off it: the amplifier's bias must re-establish, matching networks ring down, control lines have to be driven, and the automatic gain and power-control loops must reach a steady value — because the ON power is specified as a mean over the symbol, and a symbol that is still settling has the wrong mean. The number is a system settling time, not a device switching time, which is why it is far larger than a datasheet figure for the switch alone.

The guard is constant in microseconds

Now put the two documents together. Take $Y$ from TS 38.214 Table 6.2.1.2-1, multiply by the symbol duration at that numerology — which is TS 38.211's business, derived in §10 of the frame-structure note — and see what falls out.

Look down the fifth column. 71 µs, 36 µs, 18 µs, 18 µs, 16 µs, 16 µs. It falls while symbols are long and there is slack, and then it stops falling and sits between 15 and 18 µs no matter how much faster the numerology gets.

And the confirmation is printed in the RF spec itself: TS 38.101-1 Figure 6.3.3.6-4 — the time mask for consecutive SRS at 60 kHz — is annotated 17.84 µs. That is not a number 38.101-1 derives; it is the 60 kHz symbol duration, computed from TS 38.211's constants and printed in a document about amplifiers. The two halves of the standard are looking at the same clock.

Check it yourself. The whole claim is six lines of arithmetic — the symbol length from TS 38.211 clause 5.3.1, and $Y$ from TS 38.214 Table 6.2.1.2-1.

from fractions import Fraction as F
Tc = F(1, 480000 * 4096); k = 64            # TS 38.211 clause 4.1
Y  = {0: 1, 1: 1, 2: 1, 3: 2, 5: 7, 6: 14}  # TS 38.214 Table 6.2.1.2-1
for mu, y in Y.items():
    sym = float((F(2048 * k, 2**mu) + F(144 * k, 2**mu)) * Tc) * 1e6
    print(mu, f"{sym:7.2f} us x {y:2d} = {sym * y:6.2f} us")

It prints 71.35, 35.68, 17.84, 17.84, 15.61, 15.61.

What antenna switching costs

Nothing in a radio is free, and the folder's rule is that a trade-off without a number is not an engineering statement. Here are the three costs, in the order they can be quantified.

Airtime

A 1T4R phone must send four resources with three guards between them. At one symbol per resource that is $4 + 3Y$ symbols in which no data moves at all — the four SRS symbols carry no user data, and the guards carry nothing whatsoever.

Half a slot, at the numerology most 5G actually runs on. That is the price of a full 1T4R sounding cycle in the slot it occupies — and it is why sounding is periodic at a slow rate, or triggered only when it is about to be useful, rather than run continuously.

The last two rows explain a piece of clause 6.2.1.2 that looks arbitrary until you do this arithmetic. At 480 and 960 kHz the cycle cannot fit in one slot, so the clause provides configurations in which the four resources are spread over "two different slots", or four sets in "four different slots". That is not flexibility for its own sake. It is the only way the arithmetic closes.

Power

Sounding power is governed by TS 38.213 clause 7.3, and it contains one sentence with real consequences:

"a UE splits a linear value … of the transmit power … equally across the SRS ports of each SRS resource of an SRS resource set in a symbol for SRS transmission."

Equally. So the number of ports in a resource decides how much power each antenna gets:

That is a genuine trade, not a ranking. The four-transmitter phone is faster and its channel estimate is fresher; the one-transmitter phone gets 6 dB more signal per antenna, which at the edge of a cell may be the difference between a usable measurement and noise. The expensive front end wins on latency, not on sounding quality.

Insertion loss, and why there is no table for it here

The third cost is the one a front-end engineer would name first. Every path through that switch has a loss, the loss is different for each path, and it is paid on both the sounding and — for the antenna that also carries data — on every uplink transmission for the life of the product.

This note does not put a number on it, and that is deliberate.

Sources, and where to read more

The clauses this note stands on

They are text PDFs, so the claims above can be checked from a terminal:

cd ~/Documents/md-engine/"Courses (RAW)/5G Topics"
pdftotext -layout ts_138214v190400p.pdf - | grep -n "guard period of Y symbols"
pdftotext -layout ts_13810101v190400p.pdf - | grep -n -A3 "SRS time mask"

Where to read around it

How to check a fact in this note

The same two-minute habit as the previous note:

1. Find the claim and read the clause number beside it.
2. Open the document and go to that clause. The clause numbers here are all from Release 19 v19.4.0.
3. If the note and the standard disagree, the standard is right.
4. If a number is calculated rather than quoted, the calculation is shown — §8.3 is a worked example you can run as code.

The section-number warning applies here too. This page numbers its own sections, and so does 3GPP. "§8.3" is a section of this page; a 3GPP clause is always written with its document, as "TS 38.214 clause 6.2.1.2". Never a bare clause number.

Where this goes next

Two threads leave this note, and they go to opposite halves of the folder.

Downward, into the hardware. §9.3 stopped at the point where the standard stops. The switch in Figure 3 has a real insertion loss on every path, and that loss has to be paid out of a power budget that TS 38.101-1 has already spent. That is fe-insertion-loss.md, and then fe-antenna-switching.md — the note this folder was built for, where a clause number and a switch matrix finally meet on the same page.

Sideways, into the rest of the procedure. This note took the network's need for downlink CSI as given. How that CSI is actually used — the precoders, the rank, the reports that come back — is nr-mimo-csi.md.

What to take away, if only one thing. A specification will tell you what to transmit. It will almost never tell you why. But when a requirement is expressed in units that do not fit it — a silence measured in symbols that is really measured in microseconds — something physical is hiding inside the arithmetic, and converting the units will find it. That method is worth more than the particular answer it produced here.