MIMO, Spatial Multiplexing & Diversity in LTE 4G

Introduction

MIMO — Multiple-Input Multiple-Output — is the use of several antennas at the transmitter and receiver to do more with the same spectrum than a single antenna ever could. In LTE it is the single biggest lever on downlink performance: it is how a cell reaches its headline peak rates, how it keeps a cell-edge UE connected, and how it stretches coverage without more power. Every LTE cell with two or more antennas is doing MIMO of some kind on almost every subframe.

MIMO comes into play the moment a UE is in RRC_CONNECTED and being scheduled on the shared data channel. The scheduler chooses, per UE and per subframe, which MIMO technique to use based on the channel feedback the UE reports — and the network commits that choice to the UE as a Transmission Mode, configured by RRC. So MIMO is not one fixed behaviour; it is a family of techniques the network switches between as radio conditions change.

It matters because the three things multiple antennas can buy — reliability, rate, coverage — are in tension: you cannot have all three at once from the same array in the same subframe. Understanding MIMO is really understanding that trade-off, the feedback (RI/PMI/CQI) that drives it, and the reference-signal design that makes each technique decodable at the UE.

Why MIMO is needed

Concretely, a single antenna forces one compromise for every situation, and the radio channel refuses to cooperate. A deep fade can wipe out the only path you have; a strong, richly scattering channel offers parallel routes a single antenna cannot exploit; a distant UE needs more received signal power than one omnidirectional antenna delivers. Multiple antennas turn each of these into an opportunity: redundant copies survive fades, independent streams multiply throughput, and coherent combining concentrates energy. Because you cannot maximise reliability, rate, and coverage simultaneously, LTE needs a way to select per UE and per subframe — and that selection, bound together with the reference signal and control format it implies, is exactly what a Transmission Mode encodes.

What Extra Antennas Actually Buy You

The single most important idea in LTE MIMO is that multiple antennas do not do just one thing. Depending on how you drive them, the same 2- or 4-antenna array delivers a completely different benefit. Getting these three apart is the whole game — and the choice is made afresh per UE, per subframe, by the scheduler.

Transmit Diversity — Reliability First

Transmit diversity sends the same data over multiple antennas, coded so the UE can combine the copies even if one path is deeply faded. There is no rate gain here — you still carry one codeword through one stream — but the probability that the transport block decodes correctly rises sharply. This is TM2, the robust fallback every LTE UE supports, and it operates entirely open-loop: the UE feeds back nothing that steers it.

LTE builds transmit diversity from two techniques. With two antenna ports it uses SFBC (Space-Frequency Block Coding), the frequency-domain cousin of Alamouti coding: a pair of complex symbols is spread across two adjacent subcarriers and two antenna ports in an orthogonal pattern, so the receiver can recover both symbols cleanly even after independent fading on each path. Alamouti's original scheme was space-time; LTE moves it to space-frequency so the two symbols share the same OFDM symbol and the channel is essentially flat across the adjacent subcarrier pair.

With four antenna ports pure SFBC does not extend cleanly to four, so LTE combines SFBC with FSTD (Frequency-Switched Transmit Diversity). The four ports are worked in pairs: SFBC is applied across ports 0 and 2 on one pair of subcarriers, then across ports 1 and 3 on the next pair, cycling so that every port contributes and only two ports are ever active on any given subcarrier. This SFBC+FSTD arrangement keeps the orthogonality SFBC needs while spreading diversity across all four physical paths. Because it demodulates against the always-on cell-specific reference signals, four-port transmit diversity is fully backward compatible with a legacy UE.

Spatial Multiplexing — Rate First

Spatial multiplexing is where MIMO earns its throughput reputation. Instead of one stream, the transmitter sends several independent data streams — called layers — on the same time/frequency resources at the same time. A rich, scattering channel makes the antenna paths look different enough that the receiver can untangle the layers with linear algebra (an MMSE or ML receiver). Two layers over a 2x2 channel roughly double the peak rate; four layers roughly quadruple it.

The number of layers is the rank. In Rel-8 the rank can be up to 4 (a 4x4 configuration); Rel-10 (LTE-Advanced) pushes it to 8 layers on the downlink and 4 on the uplink. Rank is not fixed — it depends on how well the channel can separate streams, so a UE in a rich urban channel may support rank 4 while the same UE in a line-of-sight rural cell may only support rank 1. The UE recommends the rank it can currently sustain through its RI report.

Between the data and the antennas sit two mapping steps defined in TS 36.211. First, codeword-to-layer mapping. LTE carries up to two codewords — independently coded transport blocks, each with its own MCS and its own HARQ process, so each can be acknowledged and adapted separately — and spreads them across up to four layers. Rank 1 uses one codeword on one layer; ranks 2–4 use two codewords, splitting the higher-rank cases so a codeword feeds two layers. The exact map is fixed by the standard:

Second, precoding: the layers are multiplied by a precoding matrix W that maps them onto the physical antenna ports, either to match the channel (closed loop) or by a fixed cycling pattern (open loop). The chain below traces a two-codeword, two-layer transmission from transport blocks all the way to the antenna ports.

How the precoder is chosen splits spatial multiplexing into two flavours. Closed-loop spatial multiplexing (TM4, and its single-layer relative TM6) uses a PMI reported by the UE, so the eNB precodes to match the channel — best when the feedback is fresh and reliable. Open-loop spatial multiplexing (TM3) ignores the PMI and instead cycles deterministically through a set of precoders combined with large-delay CDD (Cyclic Delay Diversity). Large-delay CDD applies a layer-dependent phase ramp across subcarriers before precoding, which artificially decorrelates the layers and averages the channel so no single precoder choice is critical. The UE still reports RI and CQI, but no PMI — open loop wins when the UE moves fast and any reported precoder would be stale before use.

The Precoding Codebooks

For closed-loop operation the eNB and UE must agree on which precoding matrices are even allowed, otherwise the UE could not report a precoder in a handful of feedback bits. LTE solves this with standardised codebooks in TS 36.211: a finite, numbered set of precoding matrices for each antenna count and each rank. The UE reports only the index of the matrix it prefers — that index is the PMI.

The two-antenna codebook is small. For rank 1 it holds four column vectors — essentially (1/√2)[1 1]ᵀ, [1 -1]ᵀ, [1 j]ᵀ, and [1 -j]ᵀ — so a 2-bit PMI selects among them. For rank 2 only a couple of matrices are usable (the identity-like pairs), so the report shrinks to roughly one bit. This codebook is what a classic 2-port cell uses with cell-specific reference signals on ports 0 and 1.

The four-antenna codebook is far richer: 16 precoding matrices per rank, indices 0–15, so the PMI is 4 bits. These matrices are generated by a Householder construction from 16 base vectors, which keeps them constant-modulus (every antenna transmits at equal power) and nested across ranks. The same 16 indices describe rank 1 through rank 4; the reported rank (via RI) tells the eNB how many columns of the selected matrix to use.

The eight-antenna case (LTE-Advanced) does not scale the flat codebook to hundreds of entries. Instead it uses a dual-stage codebook: the precoder factors as W = W1·W2, where W1 captures the wideband, slowly-varying spatial correlation (a beam group) and W2 selects and co-phases a beam within that group on a narrower, faster basis. The UE reports two indices — a wideband PMI1 and a subband PMI2 — which keeps feedback overhead manageable while addressing an 8-port array. This structure is the direct ancestor of the Type I / Type II codebooks in 5G NR.

Closing the Loop — RI, PMI, CQI

For spatial multiplexing to work well, the transmitter often wants to know the channel. The UE measures the downlink reference signals and feeds back three quantities that together steer the link. They are reported on PUCCH (periodic, low-rate) or PUSCH (aperiodic, triggered by the eNB when it wants a fuller report).

The RI (Rank Indicator) tells the eNB how many layers the channel can currently support — effectively the recommended rank. The PMI (Precoding Matrix Indicator) points to an entry in the standardised codebook: the eNB and UE share the same finite set of precoding matrices, so the UE just reports the index of the one that best matches the channel. The CQI (Channel Quality Indicator) is a 4-bit value (0–15) reporting the highest modulation and coding rate the channel can carry at a target 10% block error rate, which the eNB maps to an MCS. With two codewords the UE can report a separate CQI per codeword, because the two layers may see different post-equalisation SINR.

The three are computed in order and are conditional: the UE first decides the rank it can sustain, then picks the best precoder for that rank, then computes the CQI it could decode assuming that precoder is applied. This is why they are always reported as a consistent triple. In open loop the PMI is dropped because a fast-moving UE's reported precoder would already be stale; the eNB cycles precoders itself and the UE reports a CQI that averages over that cycling.

Single-User, Multi-User, Beamforming & Antenna Ports

So far the layers all go to one UE — that is single-user MIMO (SU-MIMO), where one UE receives every layer and gets the full rank benefit. But the eNB can instead point different layers at different UEs sharing the same resources: multi-user MIMO (MU-MIMO), which in Rel-8 is TM5. Here the spatial dimension multiplies cell capacity rather than one UE's rate — two UEs with well-separated spatial channels each get a stream on the same PRBs, and the eNB relies on their spatial separation (aided by their reported PMIs) to keep the inter-user interference low.

Beamforming is the third use of the array. Instead of picking from a fixed codebook, the eNB shapes a beam toward the UE using its own antenna weights — which it can compute from an uplink sounding measurement rather than any codebook — and transmits a UE-specific reference signal precoded with the same weights. The UE then estimates the effective (precoded) channel directly from that reference signal and demodulates without ever knowing the beam weights. LTE calls these UE-specific reference signals demodulation reference signals (DM-RS). This underpins TM7 (single-layer beamforming, DM-RS port 5), TM8 (dual-layer beamforming, DM-RS ports 7–8), and TM9/TM10 (up to 8 layers, DM-RS ports 7–14, the LTE-Advanced modes).

This exposes a key split: CRS-based versus DMRS-based demodulation. In TM1–TM6 the UE demodulates the PDSCH using the always-on cell-specific reference signals (CRS) and must be told the precoder explicitly (via PMI signalling), because the CRS are not precoded — they measure the raw per-port channel. In the beamforming modes TM7–TM10 the UE demodulates using precoded UE-specific DM-RS, so the precoder can be any weight vector the eNB likes and never has to be signalled or standardised. This is exactly why beamforming escapes the codebook: the reference signal already carries the precoding.

All of this rests on the antenna port abstraction. An antenna port in LTE is not a physical antenna — it is defined by its reference signal: two signals are "the same port" if the channel one experiences can be inferred from the other. CRS define ports 0–3, DM-RS define ports 5 and 7–14, and CSI-RS define ports 15 and up. The mapping from ports to actual antenna elements is left to the implementation, which is what lets a single physical array present a different port set in different transmission modes.

The transmission mode is the knob that binds all of this together — each TM fixes the diversity/multiplexing/beamforming behaviour, the demodulation reference signal, and the DCI formats the UE must decode. The full TM1–TM10 catalogue is covered on the dedicated Transmission Modes page; here it is enough to know that the CRS-vs-DMRS split cleaves the modes into a legacy half and a beamforming half.

LTE-Advanced Enhancements and TDD Reciprocity

Rel-10 (LTE-Advanced) stretches every dimension of MIMO. On the downlink it raises spatial multiplexing to 8 layers (8x8) through TM9, doubling the Rel-8 peak. On the uplink, which in Rel-8 was single-antenna transmission only, it introduces closed-loop spatial multiplexing on PUSCH with up to 4 layers, so a capable UE can now send two codewords upward much as the downlink does — complete with its own uplink precoding codebook and PMI signalled in the uplink grant.

Making 8-port downlink measurement work required a new reference signal. The always-on CRS could not scale to eight ports without crippling overhead, so Rel-10 splits the two jobs the CRS used to do. CSI-RS (Channel State Information Reference Signals) are sparse, low-duty-cycle signals dedicated to measurement: the UE uses them to compute RI/PMI/CQI for up to 8 ports without needing a dense pilot in every subframe. DM-RS handle demodulation, precoded with the actual transmission weights. This measurement/demodulation separation — CSI-RS to measure, DM-RS to demodulate — is the architectural template that 5G NR adopts wholesale.

TM10 extends TM9 for Coordinated Multi-Point (CoMP), letting multiple transmission points cooperate; it adds the notion of multiple CSI processes so a UE can report channel state for several candidate points at once, and it carries a QCL indication so the UE knows which reference signals share channel properties.

Finally, TDD reciprocity gives TDD systems a shortcut the codebook feedback loop cannot match. Because a TDD cell transmits and receives on the same frequency, the uplink and downlink channels are (after calibration) the same channel. The eNB can therefore estimate the downlink channel directly from an uplink Sounding Reference Signal (SRS) the UE transmits, and compute beamforming weights from it — no PMI report, no codebook quantisation, and full (not quantised) channel knowledge. This is why TDD deployments lean heavily on the beamforming modes (TM7–TM9): reciprocity makes non-codebook beamforming both cheap and accurate. FDD, with its separated uplink and downlink bands, has no such reciprocity and must rely on codebook feedback.

Summary

MIMO in LTE is the disciplined use of multiple antennas to buy one of three things — reliability (transmit diversity via SFBC, and SFBC+FSTD for four ports), rate (spatial multiplexing of up to 4, later 8, independent layers), or coverage/SINR (beamforming) — plus a fourth, cell capacity, through multi-user MIMO. You cannot have all of them at once, so the scheduler chooses per UE and per subframe and commits the choice as a Transmission Mode. The layer chain (two codewords → layer mapping → precoder W → ports) and the feedback triple (RI how many layers, PMI which precoder, CQI which MCS) are the two mechanisms that make it all run.

The deepest structural line to remember is the CRS-vs-DMRS split: with unprecoded cell-specific pilots the precoder must be a signalled codebook entry, which caps how finely the eNB can steer; with precoded UE-specific DM-RS the pilot itself carries the beam, so beamforming escapes the codebook entirely. That split, together with the CSI-RS/DM-RS measurement-versus-demodulation separation introduced in LTE-Advanced, is precisely the architecture 5G NR adopted wholesale — so understanding LTE MIMO is most of the way to understanding NR MIMO.

Quick Q&A