Type I Codebooks in MIMO and Coding Theory

• Type I codebooks are structured families using DFT-based steering vectors for planar arrays, enabling efficient beamforming in 3GPP NR systems.
• They leverage a Kronecker-product construction and prune evanescent beams, reducing codebook size by up to 21.5% without throughput loss.
• The framework extends to algebraic coding, achieving asymptotically optimal cross-correlation properties and connecting to binary self-dual codes.

A Type I codebook refers to a structured family of codebooks or frames achieving either algebraic, combinatorial, or spatial optimality under specific system-theoretic or correlation constraints in various fields such as wireless communication (notably beamforming for MIMO/NR systems), finite-field constructions, or coding theory. In the context of 3GPP New Radio (NR) through Release 18, “Type I codebook” specifically denotes a Kronecker-product-based family of DFT steering vectors for planar (often dual-polarized) arrays, serving as the core low-overhead, wideband beamforming design for FDD and TDD protocols. In algebraic coding theory, “Type I” also designates singly-even self-dual binary linear codes.

1. Construction of Type I Codebooks in 3GPP MIMO Systems

Type I codebooks in 3GPP Release 16/17 are defined for dual-polarized uniform planar arrays (UPA) of size $N_1 \times N_2$. Each codeword is indexed by $(l, m)$, representing horizontal and vertical spatial dimensions, with associated oversampling factors $(O_1, O_2)$. The basis vectors are one-dimensional DFT (discrete Fourier transform) steering vectors for each axis: $$\mathbf f^{(N_1,O_1)}_l = \frac{1}{\sqrt{N_1}}\left[1, e^{j\frac{2\pi l}{N_1O_1}}, \ldots, e^{j\frac{2\pi l}{N_1O_1}(N_1-1)}\right]^T$$

$$\mathbf f^{(N_2,O_2)}_m = \frac{1}{\sqrt{N_2}}\left[1, e^{j\frac{2\pi m}{N_2O_2}}, \ldots, e^{j\frac{2\pi m}{N_2O_2}(N_2-1)}\right]^T$$

The full set of spatial codewords is given by the Kronecker product: $$\mathcal{CB}_{N_1,N_2}^{O_1,O_2} = \left\{ \mathbf v_{l,m} = \mathbf f^{(N_1,O_1)}_l \otimes \mathbf f^{(N_2,O_2)}_m ~|~ l=0,\ldots,N_1O_1-1;\, m=0,\ldots,N_2O_2-1 \right\}$$ Such codebooks enable simple, low-overhead feedback via the Precoding Matrix Indicator (PMI), since each codeword is referenced by integer indices $(l, m)$ (Yang et al., 2024, Ning et al., 8 Jan 2026).

2. Evanescent Components and Their Redundancy

Within the DFT-Kronecker construction, not every codeword $\mathbf v_{l,m}$ represents a physically valid beam into the array's half-space. The correspondence to physical angles $(\theta, \varphi)$ is governed by the equations: $$\sin \theta \cos \varphi = \frac{1}{\alpha_1} \frac{l'}{N_1O_1} \quad \sin \theta \sin \varphi = \frac{1}{\alpha_2} \frac{m'}{N_2O_2}$$ with $\alpha_i = d_i/\lambda$ and integer-wrapped indices $l', m'$.

A codeword is termed evanescent (nonradiating) if

$$\left(\frac{l'}{\alpha_1 N_1O_1}\right)^2 + \left(\frac{m'}{\alpha_2 N_2O_2}\right)^2 > 1$$

Such codewords impose spatial frequencies exceeding the free-space cutoff; they do not correspond to far-field propagating waves and exhibit negligible matched power when projecting the physical channel (Yang et al., 2024). Simulations show evanescent beams have much lower broadside gain (≈10 dB down) and essentially zero selection count in beam search heuristics.

3. Codebook Pruning and Protocol Implications

Eliminating indices $(l, m)$ that violate the ellipse criterion above can shrink the Type I codebook size by up to 21.5% for half-wavelength antenna spacings, yielding direct reductions in feedback overhead and computational search time for PMI selection. Wideband NR systems should perform this pruning per sub-band since the visibility region depends on array geometry and carrier frequency.

After pruning, all regular (non-evanescent) beams are indexed contiguously, reducing CBSR signaling and expediting beam-training (SRS, CSI-RS), with simulation evidence of zero degradation in aggregate throughput (Δ < 0.1%) (Yang et al., 2024).

4. Mathematical Formulation and Feedback Quantization

In 5G NR, the gNB first configures one or more CSI-Reference-Signal (CSI-RS) resources, with each CSI-RS mapped to logical antenna ports. The UE computes the maximally received SNR for each codeword $\mathbf v_{l,m}$ and selects the pair maximizing this value (Mode 1, wideband). For two streams, the selection is complemented by a quantized cross-polarization phase shift $\phi_n = e^{j\pi n/2}$ and possible small integer offsets in beam indices.

The total Mode 1 PMI feedback overhead is

$$\text{bits} = \lceil \log_2(N_1O_1) \rceil + \lceil \log_2(N_2O_2) \rceil + 3$$

Mode 2 adds three more bits per subband for adaptive selection (Ning et al., 8 Jan 2026).

5. Performance, Complexity, and Applicability

Type I codebooks achieve single-beam per stream quantization, with linear search complexity over $(N_1O_1)(N_2O_2)$ candidates. Benchmarking over different array sizes shows that for small arrays ($N_1N_2=4$), Type I and high-resolution Type II codebooks are comparable, while for larger arrays, Type II can outperform Type I by up to 2 dB in beamforming gain. Type I codebooks result in ≈10× lower feedback overhead than the most complex Type II codebooks, at the expense of ~0.5–1 dB performance loss for large arrays.

Type I codebooks are especially suitable for FDD duplexing (where feedback overhead is critical) or low-complexity UE scenarios, and serve as a robust fallback under high mobility. In TDD systems or when array size grows large, alternative codebooks (Type II/port-selection) may be preferred for beam granularity (Ning et al., 8 Jan 2026, Li et al., 2020).

6. Algebraic Type I Codebooks and Connections to Finite Geometry

In algebraic coding and signal design, “Type I codebook” can also denote a construction asymptotically achieving the Welch bound on maximum cross-correlation. Two such constructions (via trace functions and finite field character sums) provide new families of codebooks with parameters $(N, K)$:

• Construction I: Codewords indexed by multiplicative characters over GF$(q)$, restricted to a subset $D$ determined by trace and quadratic character constraints, generalizing cyclotomic frames (Wu et al., 2019).
• Construction II: Codewords indexed by additive characters over GF$(q^2)$, with selection driven by a bent-function trace constraint.

In both cases, the parameter $\lambda$ for maximum cross-correlation satisfies $\lambda/W(N,K) \to 1$ as $q \to \infty$ (where $W(N,K)$ is the Welch bound), confirming the asymptotic optimality of these Type I codebooks.

7. Connections to Binary Type I Self-Dual Codes

A distinct but historically related usage is in the theory of binary self-dual codes:

• A Type I self-dual binary code is singly-even: it is self-dual, all codewords have even weight, but at least one has weight equivalent to $2 \pmod 4$ (Hannusch et al., 2021).
• These codes are classified by length, congruence, and minimum distance and play a role in the taxonomy of code families relevant to codebook constructions.

A Type I code in this sense is defined algebraically and is not directly used for spatial beamforming but informs the combinatorial underpinnings of optimal correlation codebook design.

---

References:

• “Revealing the evanescent components in Kronecker-product based codebooks: insights and implications” (Yang et al., 2024)
• “Precoding Matrix Indicator in the 5G NR Protocol: A Tutorial on 3GPP Beamforming Codebooks” (Ning et al., 8 Jan 2026)
• “Pushing The Limit of Type I Codebook For FDD Massive MIMO Beamforming: A Channel Covariance Reconstruction Approach” (Li et al., 2020)
• “Two constructions of asymptotically optimal codebooks via the trace functions” (Wu et al., 2019)
• “The search of Type I codes” (Hannusch et al., 2021)