NR SRS: The Uplink Signal That Shapes the Downlink

Most reference signals exist so a receiver can demodulate what it is already being sent. SRS is different: the UE transmits it so the network can learn a channel the UE is not currently using.

That sounds like a detail of uplink link adaptation, and in FDD it largely is. In TDD it is the mechanism that makes massive MIMO practical, and understanding why is the single most useful thing to take from this channel.

---

Reciprocity: one measurement, two directions

A TDD cell uses the same frequency in both directions, separated in time. The physical propagation — the paths, the delays, the phase relationships — is the same either way over any interval short compared with how fast the channel changes.

So a gNB that measures the uplink channel from a UE has, to a good approximation, also measured the downlink channel to it.

This is worth stating plainly because of what it avoids. A 64-antenna array needs a precoding weight per antenna per subcarrier group to point energy at a UE. Getting that from the UE by feedback would mean the UE measuring 64 channels and reporting them — an amount of uplink signalling that scales with the array and quickly becomes absurd.

Reciprocity replaces all of it with one uplink transmission the UE was going to be capable of anyway. The array grows, the feedback does not. That is why large arrays are deployed in TDD bands and why the same array in FDD is a much harder proposition.

Where reciprocity stops being true

Two things break it, and both matter operationally.

The first is that it only holds for the radio channel, not the radio. The UE's transmit chain and its receive chain are different hardware with different gain and phase; the same is true at the gNB. The propagation is reciprocal, the end-to-end response is not, and closing that gap needs calibration. Uncalibrated arrays produce beams that point slightly wrong in a way that degrades gracefully enough to be easy to miss.

The second is time. The estimate ages. If the UE is moving, the channel it sounded is not the channel the downlink will use a few milliseconds later, and the faster it moves the shorter the useful life of the measurement. This is why SRS periodicity and mobility are linked, and why high-speed scenarios lean more on feedback-based schemes than reciprocity.

In FDD, reciprocity does not apply at all — the uplink and downlink are in different bands, so the fading is uncorrelated. There, SRS still tells the gNB useful things about the uplink, but the downlink precoder has to come from CSI reports.

One signal, four jobs

SRS resources are grouped into sets, and each set is configured with a usage that tells the gNB what the measurement is for. The physical signal does not change; the interpretation does.

codebook — the network will pick an uplink precoder from a standard set, and SRS tells it which one fits.

nonCodebook — the UE proposes candidate precoders by transmitting SRS through them, and the gNB indicates which it liked. This is the reciprocity-based uplink path: the UE works out its own beamforming from the downlink it has received, and SRS is how it shows its work.

antennaSwitching — the interesting one. A UE with more receive antennas than transmit chains cannot sound all of its antennas simultaneously, so it sounds them in turn across separate SRS resources. Without this the gNB only ever learns the channel to the antennas the UE happens to transmit on, and reciprocity gives it a partial picture of a downlink that will use all of them. A "1T4R" UE that is not configured for antenna switching is a common and quiet cause of downlink MIMO underperforming.

beamManagement — SRS transmitted on different UE beams so the gNB can compare them, used in FR2 where the UE beamforms too.

Periodic, semi-persistent, aperiodic

The three time-domain behaviours map onto three different jobs.

Periodic SRS transmits on a configured schedule indefinitely. Simple, predictable, and it consumes uplink resource whether the measurement is needed or not.

Semi-persistent is configured by RRC and activated and deactivated by MAC control elements — periodic while it is on, off the rest of the time.

Aperiodic is triggered by a field in DCI: one transmission, on demand. This is what a scheduler uses when it is about to make a decision that needs a fresh estimate and does not want to pay for a standing configuration.

The trade is the familiar one. Frequent sounding gives the scheduler accurate, current channel knowledge and costs uplink capacity that could have carried data. Infrequent sounding is cheap and gives the beamformer a stale channel — which, at the moment it matters most, is a channel for where the UE used to be.

Why SRS is expensive in a loaded cell

SRS occupies the last symbols of a slot, and those symbols are uplink capacity. In a cell with many UEs each sounding regularly, the aggregate cost is not small.

Worse, SRS from different UEs has to be kept separable — by comb offset, by cyclic shift, by sequence, or by being scheduled at different times. The available combinations are finite. A cell can run out of ways to distinguish simultaneous SRS transmissions before it runs out of uplink resource to carry them, and the symptom is beamforming quality that degrades as the cell fills, which is easy to misread as interference.

This is also where the tension with carrier aggregation shows up: a UE that must sound several carriers with one transmit chain sounds each of them less often.

What goes wrong in practice

Antenna switching not configured. The single most common way to leave downlink MIMO gain on the table in TDD. The UE reports the capability; the network has to act on it.

Periodicity set for a stationary UE and applied to everyone. Reciprocity-based precoding on a moving UE with slow sounding produces beams aimed behind it. The failure looks like mobility performance and is a sounding configuration.

Array calibration assumed rather than verified. Reciprocity is a statement about propagation, not about hardware. Calibration drift shows up as a gradual loss of beamforming gain with no obvious event to blame.

---

The mental model

SRS lets the gNB measure a channel the UE is not using. That is the whole idea.

TDD reciprocity turns one uplink transmission into a downlink precoder, which is what makes 64-antenna arrays deployable without absurd feedback.

Reciprocity covers the channel, not the transceivers. Calibration is not optional, and the estimate ages.

antennaSwitching exists because most UEs receive on more antennas than they transmit from. Skip it and reciprocity sees half the picture.

Sounding costs uplink capacity and separability. Cells run out of the second one first.

Further reading

• 3GPP TS 38.211 §6.4.1.4 — SRS sequence generation and resource mapping
• 3GPP TS 38.214 §6.2.1 — SRS procedures and usage configuration
• 3GPP TS 38.331 — SRS-Config, resource sets and the usage field
• ShareTechnote — 5G SRS, for resource mapping diagrams and configuration examples