Beamforming in 5G: Analog, Digital, Hybrid, and Keeping the Beam Aligned

An omnidirectional antenna radiates energy in every direction, which means almost all of it goes somewhere no one is listening. Beamforming concentrates that energy toward a specific receiver.

The mechanism is constructive interference. Feed the same signal to multiple antenna elements with carefully chosen phase offsets, and the wavefronts add up in one direction and cancel in others.

In 4G, beamforming was an enhancement. In 5G — particularly above 24 GHz — it's a precondition. An FR2 cell without beamforming has essentially no usable coverage.

---

Where the gain comes from

An N-element array provides roughly 10·log₁₀(N) dB of array gain. Eight elements gives ~9 dB, 64 gives ~18 dB, 256 gives ~24 dB.

That's not free power — it's the same total energy redistributed. What you gain in the beam direction, you lose everywhere else.

Path loss scales with frequency squared, so moving from 2 GHz to 28 GHz costs roughly 23 dB of additional free-space loss. Array gain is what buys it back. At 28 GHz a 64-element array is doing the work that made the link possible at all — which is exactly why beamforming stops being optional in FR2.

The second benefit is often underrated: energy that isn't radiated toward a receiver isn't interfering with anyone else. In dense deployments, beamforming is as much an interference management tool as a coverage tool.

---

Three architectures

The distinguishing question: how many RF chains does the array have?

An RF chain is the expensive part — data converters, mixer, amplifier, filtering. Cost, power consumption, and capability all scale with the count.

Analog beamforming

One RF chain feeding all elements, with phase shifters in the analog domain.

Cheap and power-efficient. The constraint is decisive: one RF chain means one beam at a time. Every UE served by that array shares one beam direction, so you're time-multiplexing users across directions. No spatial multiplexing, no simultaneous multi-user service.

Digital beamforming

One RF chain per element, with beamforming applied as digital precoding in baseband.

Maximum flexibility. Multiple simultaneous beams, full spatial multiplexing, per-subcarrier beamforming — different beams on different parts of the band within one OFDM symbol, which analog fundamentally cannot do.

The cost is brutal at scale. A 256-element array needs 256 complete RF chains, 256 ADC/DAC pairs, and the baseband throughput to process them all. At mmWave bandwidths the data converter power alone becomes prohibitive.

Hybrid beamforming

Analog phase shifting within subarrays, digital precoding across them.

A 64-element array might use 8 RF chains, each driving an 8-element subarray. You get 8 independent digital streams — enough for meaningful multi-user MIMO — at one eighth the RF chain cost.

The pattern in practice: FR1 uses digital (64 chains at 3.5 GHz is manageable), FR2 uses hybrid (256+ elements with full digital is not).

---

Beam sweeping and initial access

A directional system has a bootstrapping problem. The gNB can't beamform toward a UE it doesn't know about, and the UE can't report a beam it hasn't measured.

NR solves it with beam sweeping on synchronisation signals.

The gNB transmits SS/PBCH Blocks (SSBs) in different beam directions in sequence within an SS burst set, repeating every 20 ms by default. Up to 4 SSBs below 3 GHz, 8 up to 6 GHz, and 64 in FR2 — the count scales with how narrow the beams are.

Each SSB carries an implicit index, so measuring an SSB tells the UE which beam it came from. The UE measures all of them, picks the strongest, and uses the associated RACH occasion — which itself signals the chosen beam back to the gNB. Beam selection is complete before any dedicated signalling exists.

The trade-off: 64 SSBs consume noticeably more overhead than 4, and the full sweep takes longer. That's the price of narrow beams.

---

Beam management

Selecting a beam once isn't enough. UEs move, rotate, and get blocked. NR defines a continuous process, conventionally described in three procedures:

P1 — initial beam acquisition through SSB sweeping. Coarse.

P2 — gNB-side refinement. The gNB sweeps narrower CSI-RS beams around the selected direction while the UE keeps its receive beam fixed.

P3 — UE-side refinement. The gNB holds its beam steady while the UE sweeps its own receive beams. Important because the UE also has an array and also has to point it.

Refinement matters because SSB beams are deliberately wide — they have to cover the sector with a manageable number of transmissions. CSI-RS beams are narrow, and the additional gain from refinement is often several dB.

Beam failure recovery

At mmWave, a hand or a passing vehicle can drop the link in milliseconds.

NR specifies an explicit procedure. The UE monitors beam failure detection reference signals; consecutive failure instances above a threshold declare beam failure. The UE then searches candidate beams from a configured list, and on finding a viable one sends a beam failure recovery request via a dedicated RACH resource. The gNB responds and the link continues.

Crucially, this happens without radio link failure or RRC re-establishment. Beam failure is treated as a routine, recoverable event — which it is, at mmWave, potentially several times a minute.

---

Beamforming vs precoding vs spatial multiplexing

These three get used loosely and mean different things.

Beamforming shapes the radiation pattern to direct energy.

Precoding applies weights to multiple data streams before transmission. It's the more general operation — beamforming is the special case where those weights produce a directional pattern for a single stream.

Spatial multiplexing sends different data on different spatial paths, multiplying capacity rather than improving link quality.

The relationship: beamforming improves link quality, spatial multiplexing improves capacity, and modern systems do both — beamform toward each of several UEs simultaneously, each on its own spatial stream. That's MU-MIMO, and it's what makes massive MIMO worth deploying.

The requirement is channel state information. Without knowing the channel, the array can't compute weights. In TDD, reciprocity lets the gNB infer the downlink channel from uplink SRS — which is a major reason 5G mid-band deployments are overwhelmingly TDD. In FDD, the UE must measure and report CSI explicitly, costing overhead and adding delay.

---

Beam failure and recovery

A beam is a narrow thing, and narrow things break. A hand over the antenna, a bus passing between device and cell, a turn of the wrist — any of these can drop a mmWave link in tens of milliseconds. The specification treats this as an expected event rather than a fault, and defines a dedicated recovery procedure.

Detection runs continuously in the physical layer. The UE monitors a configured set of reference signals and compares the hypothetical error rate against a threshold. Each time it crosses, the PHY sends a beam failure instance indication to MAC. MAC counts these, and when the count reaches beamFailureInstanceMaxCount within the running beamFailureDetectionTimer, beam failure is declared.

Recovery then proceeds in three steps. The UE searches its configured candidate list for a beam whose RSRP exceeds the threshold. It sends a beam failure recovery request, usually on a contention-free PRACH resource reserved for exactly this purpose. The network responds on the recovery search space, and the link resumes on the new beam.

The whole sequence typically completes in a few tens of milliseconds. If no candidate beam is found, or recovery times out, the UE declares radio link failure and falls back to RRC re-establishment — which is far more expensive, and is what beam failure recovery exists to avoid.

Why the candidate set matters

The candidate beam list is configured by the network, and its composition determines whether recovery succeeds or escalates. Too few candidates and the UE may find nothing usable. Too many and the search takes longer than the failure timer allows.

This is one of the more consequential tuning parameters in a mmWave deployment, and it interacts with cell geometry: a cell whose beams all point in similar directions offers poor recovery diversity, because whatever blocked one beam likely blocks its neighbours too.

---

Analog beamforming's hidden constraint

The three architectures differ in more than cost, and the operational difference is easy to underestimate.

An analog array has one RF chain and one set of phase shifters, which means it can form exactly one beam at a time. Not one beam per user — one beam, full stop. Serving four users in different directions means serving them in four different time slots.

That constraint propagates upward into the scheduler. With digital beamforming the scheduler can pair spatially separated users in the same slot, because separate RF chains produce separate beams simultaneously. With analog beamforming the scheduler is choosing directions in time, and multi-user MIMO is simply unavailable within a panel.

Hybrid architectures sit in between in a specific way: the number of simultaneous beams equals the number of RF chains, not the number of antenna elements. A 64-element panel with four RF chains forms four beams, each shaped by sixteen elements. The elements buy narrowness; the chains buy simultaneity. They are different resources, and confusing them is the most common error in reading a hybrid design.

---

What it costs

Beamforming isn't free, and the costs are worth naming:

Reference signal overhead. More beams means more SSBs and CSI-RS, consuming resources that could carry data.

Measurement and reporting. CSI feedback is continuous overhead — and CSI compression is one of the more active AI-in-RAN research areas precisely because of it.

Mobility complexity. Handover now involves beam selection at the target cell, not just cell selection.

Calibration. Phase coherence across elements must be maintained. Drift degrades the beam pattern in ways that are hard to diagnose from network counters.

Power and cost. RF chains, phase shifters, and calibration circuitry are real bill-of-materials.

Blockage sensitivity. A narrow beam is precisely a beam that's easy to interrupt.

---

Uplink beamforming is the harder problem

Almost everything written about beamforming describes the downlink, where a base station with generous power and space forms beams toward devices. The uplink is the direction that actually constrains coverage, and it is meaningfully harder.

The device has a handful of elements, not sixty-four. It has a fraction of the transmit power. It is held in a hand that blocks part of the sphere, and its orientation changes continuously with no relationship to the network's geometry.

NR handles this with codebook-based and non-codebook-based uplink transmission. In the codebook case the gNB measures the UE's SRS and tells it which precoder to use, chosen from a standard set — the network decides, the device follows. In the non-codebook case the UE chooses its own precoder by measuring a downlink reference signal and exploiting channel reciprocity, which works in TDD where both directions share a frequency.

Panel selection sits underneath both. A phone carries several antenna panels facing different directions precisely because any one of them may be against a palm. The device selects among them, and the network's SRS configuration must give it enough resources to sound each.

The practical consequence is that uplink beam management is often the limiting factor in mmWave coverage. The downlink beam may be excellent while the uplink cannot close, and the cell edge in a mmWave deployment is usually an uplink edge.

---

The mental model

Array gain ≈ 10·log₁₀(N) dB. It's energy redirected, not created.

RF chain count determines the architecture. One = analog, one per element = digital, one per subarray = hybrid.

FR1 goes digital, FR2 goes hybrid — because 256 full RF chains at mmWave is impractical.

SSB sweeping bootstraps initial access. Up to 64 beams in FR2.

P1/P2/P3 handle acquisition, gNB refinement, and UE refinement. Beam management is continuous.

Beam failure recovery is routine, not an error path.

Beamforming improves link quality; spatial multiplexing multiplies capacity. Both need channel knowledge.

Beamforming is what makes mmWave viable and what makes dense mid-band deployments manageable. It also introduces an entire layer of procedures — sweeping, refinement, failure recovery — that simply didn't exist in a network of omnidirectional cells.

---

Further reading

• 3GPP TS 38.214 — Physical layer procedures for data, CSI and beam management
• 3GPP TS 38.213 — Physical layer procedures for control, beam failure recovery
• 3GPP TS 38.211 — Physical channels, SSB structure and mapping
• 3GPP TR 38.802 — Study on new radio access technology, physical layer aspects