CSI Feedback (CQI / PMI / RI) in 5G NR

Introduction

Channel State Information is the downlink's eyes. Because the propagation channel is only observable at the receiver, the gNB cannot know how to transmit efficiently unless the UE measures the channel and reports back. CSI is the standardised feedback that carries that measurement — a bundle of quantities the UE derives from downlink reference signals and returns on the uplink, defined in TS 38.214.

It runs throughout the connected lifetime of the UE. Once RRC has configured the measurement resources and the report, the UE measures continuously and reports either on a fixed cadence, on demand, or in switchable bursts. Every downlink transmission decision the scheduler makes — the MCS for link adaptation, the precoder and rank for MIMO, the serving beam in a beamformed cell — is seeded by the most recent CSI report.

CSI matters because it is the difference between transmitting blind and transmitting to the channel. Get it right and the network runs close to the channel's capacity; get it stale, mis-configured, or wrongly scoped and the scheduler either wastes spectrum by playing safe or triggers retransmissions by overreaching. Understanding CSI is understanding how the downlink is steered.

Why CSI is needed

Concretely, four downlink decisions are impossible to make well without feedback. The gNB must pick a rate (modulation and code rate) — too high and the block fails, too low and capacity is wasted. It must pick a precoder to steer energy along the channel's strong spatial directions. It must pick a rank — how many independent streams the channel can actually carry. And in a beamformed cell it must pick the right beam. Each of these depends on the instantaneous channel, which only the UE can see.

CSI supplies exactly these four answers in one report: CQI for rate, PMI for the precoder, RI for the rank, and CRI for the beam (with LI as a fifth, smaller hint about where to place phase-tracking pilots). It turns the UE's private, unobservable view of the radio link into a standardised, quantised recommendation both ends interpret identically — letting the scheduler match the transmission to the channel in rate, spatial dimension, and direction simultaneously.

What CSI is and why it exists

Downlink transmission is decided entirely at the gNB — modulation, code rate, precoder, number of layers, transmit beam — but the propagation channel is something only the receiver can observe. Without feedback the gNB would be guessing at every one of those choices, and a wrong guess costs throughput (too conservative) or a decode failure and retransmission (too aggressive). CSI is the mechanism that turns the UE's private view of the channel into something the scheduler can act on.

A crucial point of philosophy: CSI is a recommendation, not a command. The gNB is free to follow the report verbatim, adjust it with its own outer-loop algorithms, or ignore it. What the report guarantees is a grounded, standardised snapshot of the radio link computed against agreed assumptions, so that both ends interpret it identically.

The CSI quantities

Each quantity targets a different scheduling decision. A given CSI-ReportConfig selects which of them the UE reports through its reportQuantity field — for example cri-RI-PMI-CQI for full MIMO feedback, or cri-RSRP for pure beam reporting. Not every report carries every quantity.

CRI, L1-RSRP and L1-SINR belong to the beam-management side of the story. When the gNB sweeps several narrow beams (each carried on its own NZP-CSI-RS resource), the UE measures all of them and reports the best index plus its power/quality. That is how the network keeps a moving mmWave UE pointed at the right beam without dropping the link — see Beam Management.

Measuring channel and interference: NZP-CSI-RS and CSI-IM

Everything in a CSI report is derived from measurements, and CSI cleanly separates the two things worth measuring: the wanted channel and the interference the UE will actually see. Getting both right is what lets CQI reflect true operating conditions rather than an interference-free ideal.

The wanted channel is measured on NZP-CSI-RS (Non-Zero-Power CSI Reference Signal). These are known pilot symbols the gNB transmits in a configured pattern of resource elements; because the sequence is known, the UE can estimate the complete channel matrix — amplitude and phase on every transmit-receive antenna pair — from them. An NZP-CSI-RS resource can span 1, 2, 4, 8, 12, 16, 24 or 32 antenna ports, and its density and periodicity are configurable, trading measurement accuracy against overhead.

Interference is measured on CSI-IM (CSI Interference Measurement) resources. A CSI-IM is deliberately a set of resource elements on which the serving cell sends nothing — so whatever power the UE observes there is pure interference plus noise from neighbours and other transmissions. Pairing a channel measurement with an interference measurement lets the UE compute a realistic SINR, and therefore a realistic CQI. Alternatively, interference can be measured on a dedicated NZP-CSI-RS-for-interference resource that emulates an interfering stream.

A CSI-RS resource can also be time-configured as periodic, semi-persistent, or aperiodic — independently of the report timing. Periodic CSI-RS is always present; aperiodic CSI-RS is transmitted only when a report is triggered, saving overhead. The report and its resources do not have to share the same timing mode, but valid combinations are restricted (for instance, an aperiodic report can use periodic, semi-persistent or aperiodic resources).

The configuration framework: ReportConfig and ResourceConfig

Two RRC structures wire the whole thing together, and keeping them straight is the key to understanding CSI. One says what to measure; the other says what to report and how.

CSI-ResourceConfig defines the measurement side. It groups one or more resource sets — nzp-CSI-RS-ResourceSet(s) for channel measurement, and optionally csi-IM-ResourceSet(s) for interference. Each set in turn lists the individual resources. This is the "what the UE measures" container.

CSI-ReportConfig defines the reporting side. It references the relevant CSI-ResourceConfig(s) — one for channel measurement (resourcesForChannelMeasurement), and optionally ones for interference (csi-IM-ResourcesForInterference or nzp-CSI-RS-ResourcesForInterference) — then selects the reportQuantity, the codebookConfig (Type I or Type II and its parameters), whether feedback is wideband or subband (reportFreqConfiguration), and the timing mode (reportConfigType). It is the "what to report and how" container.

A UE can be configured with many CSI-ReportConfigs at once — for example one periodic wideband report as a cheap baseline, plus an aperiodic subband Type II report the gNB triggers only when it wants to set up high-order MU-MIMO. Each report is independent and points at its own resources.

Report timing: periodic, semi-persistent, aperiodic

CSI reporting comes in three timing modes, each carried on a different uplink channel and triggered differently. The choice trades feedback freshness and detail against uplink overhead.

Periodic CSI is the simplest: RRC configures a period and offset, and the UE reports on PUCCH at that cadence forever until reconfigured. It is cheap and predictable, but because PUCCH has limited payload it usually carries only compact wideband feedback.

Semi-persistent CSI behaves like periodic reporting that can be switched on and off. On PUCCH it is activated and deactivated by a MAC CE; on PUSCH it is activated by a DCI and then recurs on scheduled PUSCH occasions. This lets the network pay for detailed periodic feedback only during the intervals it matters — a bursty large download, say — and stop it otherwise.

Aperiodic CSI is a single on-demand report carried on PUSCH. A DCI (typically an uplink grant, DCI format 0_1/0_2) contains a CSI request field; each nonzero codepoint maps to one or more configured CSI-ReportConfigs (a "trigger state" in CSI-AperiodicTriggerStateList). When the UE sees the trigger, it measures and appends the report to the granted PUSCH. Because PUSCH can carry a large payload, aperiodic reporting is where the heavy, subband, Type II reports live.

CSI Part 1 and Part 2, and how it maps to the channel

A CSI report can be large and, worse, its size can depend on its own content — the number of subband CQIs and the PMI payload depend on the reported RI, which the gNB does not know until it decodes the report. To make this decodable, larger reports are split into two parts:

• CSI Part 1 has a fixed, known size. It carries the RI (and CRI), a wideband CQI, and — for Type II — an indication of the number of non-zero coefficients. Because its size is fixed, the gNB can always decode Part 1.
• CSI Part 2 carries the rest: the PMI, subband CQI, LI, and Type II amplitude/phase coefficients. Its size is derived from Part 1, so the gNB knows how many bits to expect only after decoding Part 1.

On PUSCH, Part 1 and Part 2 are encoded separately, and if the grant is too small Part 2 can be partially omitted according to a defined priority order (wideband survives, the least important subband information is dropped first). On PUCCH, small reports fit in one part; larger periodic reports use the two-part structure across the PUCCH resource. This two-part design is the reason a variable-size report never becomes undecodable.

Codebooks and frequency granularity: Type I vs Type II, wideband vs subband

The PMI is an index into a codebook — a standardised set of candidate precoding matrices both ends agree on. NR defines two families, trading feedback overhead against precoding accuracy.

Type I codebooks report a single preferred beam direction per layer, chosen from a DFT-based grid of beams, with a co-phasing term across polarisations. The feedback is compact and the resulting precoder is essentially "point the array in the best direction." Type I is the workhorse for single-user MIMO (SU-MIMO), where you mainly need to steer energy toward one UE.

Type II codebooks report a linear combination of several beams per layer, each with its own quantised amplitude and phase. This produces a far more accurate reconstruction of the channel's spatial structure — precise enough for the gNB to null interference between users — which is exactly what multi-user MIMO (MU-MIMO) needs. The price is a much larger report, which is why Type II typically rides aperiodic PUSCH. An enhanced variant (eType II) adds frequency-domain compression to cut that overhead.

Independently of codebook type, the report has a frequency granularity. Wideband feedback gives one value across the whole bandwidth part — one CQI, one PMI. Subband feedback gives per-subband values (a subband is a small group of resource blocks — typically 4, 8, 16 or 32 RBs, its size set by BWP width), letting a frequency-selective scheduler assign each UE the resource blocks where its channel is best. Subband feedback costs proportionally more uplink bits, so it is reserved for reports where the gain justifies it. The reportFreqConfiguration in CSI-ReportConfig selects wideband vs subband for CQI and PMI separately.

CQI in depth: the quality-to-MCS bridge

The CQI is the quantity most directly tied to throughput. The UE reports the index of the highest entry in a CQI table such that a hypothetical PDSCH using that entry's modulation and code rate would decode with block error rate not exceeding the target — 10% (0.1) for normal traffic, or 0.001 for the low-spectral-efficiency (URLLC) table. The three CQI tables mirror the MCS tables: one topping out at 64QAM, one reaching 256QAM, and a low-SE table for high-reliability services.

Each index encodes a modulation order and a target code rate, from which a spectral efficiency (bits/RE) follows. The excerpt below is from the 256QAM CQI table (TS 38.214, Table 5.2.2.1-3); index 0 always means "out of range — no reliable transmission possible."

The CQI can be wideband (one index for the whole BWP) or subband (a wideband reference plus small per-subband differentials, each a 2-bit offset), controlled by cqi-FormatIndicator. Note that the UE computes CQI conditioned on the reported RI and PMI and on the assumed number of layers — it is the quality you would get if the gNB used the recommended rank and precoder, not a raw SNR.

How CSI drives link adaptation and MIMO precoding

Once a report lands, the scheduler splits it across two decisions. CQI sets the starting point for link adaptation — it maps to an MCS that the outer loop then trims using HARQ feedback. RI and PMI drive MIMO: the rank fixes how many transport layers to send, and the precoder shapes the transmit antenna array to concentrate energy along the channel's strong directions (and, with Type II, to null energy where other users sit). CRI picks the serving beam in a beamformed cell, and LI tells the transmitter which layer to hang phase-tracking reference signals on. The outcome is a transmission matched to the channel simultaneously in rate, spatial dimension, and direction.

Because the loop has latency — measure, compute, wait for an uplink occasion, report, decode, schedule — the channel may have moved by the time the report is used, especially for fast-moving UEs. This is another reason the gNB treats CSI as advisory and leans on HARQ-driven outer-loop correction: the report tells it where the channel was, and the ACK/NACK stream tells it how well its guess about now is holding up.

Summary

CSI is how the downlink is steered. Only the UE can observe the channel, so it measures known reference signals — NZP-CSI-RS for the wanted channel, CSI-IM for interference — and returns a bundle of quantities that answer the scheduler's four questions: CQI (how good → MCS), PMI (how to point → precoder), RI (how many layers → rank), and CRI (which beam), with LI flagging the strongest layer for PTRS. The quantities are computed conditionally in the order CRI → RI → PMI → CQI → LI, so every CQI means "quality given this rank and precoder," never a raw SNR.

Two RRC structures wire it up — CSI-ResourceConfig (what to measure) and CSI-ReportConfig (what to report and how) — and reporting comes in periodic, semi-persistent and aperiodic modes on PUCCH or PUSCH, with a fixed-size Part 1 / derived-size Part 2 split that keeps a content-dependent payload decodable. Codebooks trade overhead for accuracy (Type I for SU-MIMO steering, Type II for MU-MIMO nulling), and frequency granularity trades uplink bits for frequency-selective scheduling (wideband vs subband). Finally, because the loop has latency, the gNB treats every report as advisory and corrects it with HARQ-driven outer-loop link adaptation — the report says where the channel was, the ACK/NACK stream says how the guess about now is holding up.

Quick Q&A