How does a DFT-based codebook work?

A DFT (Discrete Fourier Transform) codebook exploits the structure of uniform linear or planar antenna arrays, where the optimal beamforming vector for a plane wave arriving from angle theta is a DFT vector with phase progression exp(j*2*pi*d*sin(theta)/lambda) across elements. For an N-element array, the basic DFT codebook contains N beams pointing in directions theta_i = arcsin(i*lambda/(N*d)) for i = 0, 1, ..., N-1. Oversampled DFT codebooks multiply the number of beams by an oversampling factor O (typically 2 to 4), creating O*N beams with finer angular spacing. In 5G NR Type I single-panel codebook for a 2D array with N1 horizontal and N2 vertical elements, the codebook has O1*N1 horizontal beams and O2*N2 vertical beams, with the PMI selecting one horizontal index (i1,1), one vertical index (i1,2), and a co-phasing index between dual-polarized panels (i2). The total number of codebook entries is O1*N1 * O2*N2 * 4 (4 co-phasing values for dual-pol). For 16 horizontal elements with O1=4 and 8 vertical elements with O2=4: 64*32*4 = 8,192 entries.

What is the difference between Type I and Type II codebooks?

Type I codebooks select a single beam direction from the DFT grid (single-panel) or one beam per panel (multi-panel), providing coarse spatial information with low feedback overhead (typically 10 to 20 bits for the PMI across wideband and subband). They work well for single-user MIMO and scenarios where the channel has one dominant path. Type II codebooks report a linear combination of L = 2 or 4 strongest DFT beams with per-beam wideband amplitude coefficients and per-subband phase coefficients, approximating the actual channel eigenvector much more accurately. This requires 50 to 200 bits of feedback (3 to 5x more than Type I) but captures multi-path channel structure, enabling the gNB to form more precise beams for multi-user MIMO where accurate spatial separation between users is critical. Type II provides 5 to 15% cell-average throughput gain and 10 to 20% cell-edge gain over Type I in multi-user MIMO scenarios. Type II is mandatory for all 5G NR UE categories but is typically used only when the gNB detects sufficient multi-user multiplexing opportunity to justify the feedback overhead.

How does codebook design affect massive MIMO performance?

In massive MIMO with 64 to 256 antenna elements, the angular resolution of the codebook directly determines how precisely the gNB can direct beams to individual users and suppress inter-user interference in MU-MIMO. With 64 elements (8x8 dual-pol) and oversampling factor 4, the Type I codebook has 32x32 = 1,024 beam directions, each covering approximately 3.5 degrees in azimuth and 3.5 degrees in elevation. This is sufficient to separate users that are more than 3.5 degrees apart but cannot distinguish closer users. Type II codebooks overcome this by combining beams: 4 DFT beams with optimized coefficients can synthesize a beam pointed in any direction with sub-degree accuracy, not just the DFT grid points. For frequency-selective channels, Type II reports different phase coefficients per subband (20 to 50 subbands across the bandwidth), capturing the fact that different frequency components arrive from slightly different effective directions due to multipath. The enhanced codebook in 3GPP Release 18 (Type II with port selection) further reduces overhead by reporting only the most significant antenna ports rather than all ports.

mmWave & 5G

Codebook

/kohd-book/

A codebook is a predefined set of beamforming weight vectors shared between transmitter and receiver for MIMO precoding. UE reports the best-matching index (PMI) instead of full channel matrix. 5G NR Type I: single DFT beam per panel, 10 to 20 bits feedback, good for SU-MIMO. Type II: linear combination of L=2 to 4 strongest DFT beams with amplitude/phase coefficients, 50 to 200 bits, 5 to 15% MU-MIMO throughput gain.

Category: mmWave & 5G

Type I: 10 to 20 bits PMI

Type II: 50 to 200 bits PMI

Understanding Codebooks

In a MIMO system with N transmit antennas, the optimal beamforming vector is an N-dimensional complex vector matched to the channel. Feeding back this vector directly would require quantizing N complex numbers (2N real values) with sufficient precision, consuming prohibitive uplink bandwidth in FDD systems where the transmitter cannot estimate the downlink channel directly. Codebooks solve this: both the gNB and UE store the same set of candidate beamforming vectors, and the UE reports only the index of the best match, reducing feedback from hundreds of bits to a compact integer.

The challenge is designing the codebook to cover the space of possible channels efficiently. Too few entries and the best match may be far from optimal, losing beamforming gain. Too many and the feedback overhead increases while the UE search complexity grows. DFT-based codebooks exploit the structure of uniform arrays where the optimal beamforming vector for a single plane wave is always a DFT vector, regardless of the arrival angle. By sampling the DFT at regular angular intervals with oversampling, the codebook covers all possible beam directions with predictable worst-case mismatch.

Codebook Formulas

DFT Beam Vector:
w(i) = [1, e^j2πi/N, e^j2π2i/N, ..., e^j2π(N-1)i/N]^T / √N

Type I Codebook Size:
|C| = O₁N₁ × O₂N₂ × 4 (dual-pol co-phasing)

Type II Feedback Bits:
B ≈ log₂(O₁N₁ × O₂N₂) + L×(3 + S×3) bits

Where O₁,O₂ = oversampling (2-4), N₁,N₂ = elements per dimension, L = beams (2-4), S = subbands (20-50). 8x8 dual-pol, O=4: Type I = 8,192 entries; Type II with L=4, S=26: ~400 bits total.

5G NR Codebook Types

Codebook	Resolution	Feedback	Complexity	Best For
Type I Single-Panel	1 beam direction	10 to 20 bits	Low	SU-MIMO, single path
Type I Multi-Panel	1 beam per panel	15 to 25 bits	Low	Distributed arrays
Type II	L=2-4 beam combination	50 to 200 bits	High	MU-MIMO, multipath
Type II Port Selection	Subset of ports	30 to 100 bits	Medium	Rel-18 overhead reduction
Enhanced Type II	Doppler+spatial	100 to 400 bits	Very high	High mobility MU-MIMO

Common Questions

Frequently Asked Questions

How does a DFT codebook work?

Exploits uniform array structure: optimal beam for angle θ is DFT vector with phase progression e^{j2πd sinθ/λ}. Oversampled by O=2 to 4x for finer angular spacing. For 16×8 dual-pol, O=4: 64×32×4 = 8,192 entries. PMI selects horizontal, vertical, and co-phasing indices.

Type I vs Type II?

Type I: single beam, low feedback (10 to 20 bits), good for SU-MIMO. Type II: L=2 to 4 beam linear combination with per-subband phase, 50 to 200 bits. Type II: 5 to 15% cell-average and 10 to 20% cell-edge throughput gain in MU-MIMO. Mandatory for all UE categories but used selectively by gNB.

How does codebook affect massive MIMO?

64 elements (8×8 dual-pol), O=4: 1,024 beam directions, ~3.5° resolution. Sufficient to separate users >3.5° apart. Type II: sub-degree accuracy by combining beams with optimized coefficients. Rel-18 port selection reduces overhead by reporting only significant antenna ports.

Related Terms

← Code Rate Detail Codebook Based Beamforming →