Ferkans — Interactive Telecom Tutor

Why FDD Still Matters

TDD massive MIMO elegantly sidesteps the CSI acquisition problem through uplink–downlink reciprocity: the base station estimates the channel from uplink pilots and uses the same channel for downlink precoding. But a large fraction of deployed cellular spectrum is FDD — paired bands where the uplink and downlink frequencies differ by tens or hundreds of MHz. In these bands, the channel at the downlink frequency cannot be inferred from the uplink, and the base station must obtain downlink CSI through a fundamentally different mechanism: transmit downlink pilots, let each UE estimate its own channel, and receive quantized feedback over the uplink control channel. The cost of this mechanism scales with $N_t$ , and understanding precisely how — and how to mitigate it — is the subject of this chapter.

Definition:
FDD Massive MIMO System Model

Consider a single-cell downlink with a base station (BS) equipped with $N_t$ antennas serving $K$ single-antenna users. The system operates in FDD mode with downlink carrier frequency $f_{\text{DL}}$ and uplink carrier frequency $f_{\text{UL}}$ , where $|f_{\text{DL}} - f_{\text{UL}}| \gg B_c$ (the frequency gap far exceeds the coherence bandwidth). The downlink received signal at user $k$ is

$y_k = \mathbf{H}_{k}^{H} \sum_{j=1}^{K} \mathbf{v}_{j} s_j + w_k,$

where $\mathbf{H}_{k} \in \mathbb{C}^{N_t}$ is the downlink channel vector, $\mathbf{v}_{j} \in \mathbb{C}^{N_t}$ is the precoding vector for user $j$ , $s_j$ is the data symbol with $\mathbb{E}[|s_j|^2] = 1$ , and $w_k \sim \mathcal{CN}(0, \sigma^2)$ is AWGN.

The BS requires knowledge of $\{\mathbf{H}_{k}\}_{k=1}^{K}$ to design the precoders. In FDD, this knowledge is obtained through a three-step protocol:

Downlink training: The BS transmits $\tau_d$ pilot symbols.
Channel estimation: Each UE estimates $\mathbf{H}_{k}$ from the received pilots.
Uplink feedback: Each UE quantizes its estimate and feeds back $B_{\text{fb}}$ bits.

The frequency gap $|f_{\text{DL}} - f_{\text{UL}}|$ is typically 45 MHz in LTE Band 1 and 190 MHz in Band 7. Since the coherence bandwidth in urban environments is 1–10 MHz, the UL and DL channels are statistically independent — reciprocity does not hold.

Frequency Division Duplex (FDD)

A duplexing mode in which uplink and downlink transmissions occupy different frequency bands simultaneously. Because the channel realizations at the two frequencies are independent when the duplex gap exceeds the coherence bandwidth, the BS cannot exploit uplink–downlink reciprocity and must rely on downlink training and uplink feedback for CSI acquisition.

Theorem: Downlink Pilot Overhead Scaling

For a BS with $N_t$ antennas serving $K$ users, the minimum downlink pilot overhead satisfies $\tau_d \geq N_t$ . Combined with the uplink pilot overhead $\tau_u \geq K$ , the fraction of the coherence interval $T_c$ available for data transmission is

$\eta_{\text{data}} = 1 - \frac{\tau_d + \tau_u}{T_c} \leq 1 - \frac{N_t + K}{T_c}.$

For $N_t \gg 1$ , the overhead $\tau_d / T_c$ can dominate the coherence block, leaving negligible room for data.

Each UE must estimate an $N_t$ -dimensional channel vector. To identify all $N_t$ components, the BS must transmit at least $N_t$ linearly independent pilot vectors — one per antenna dimension. In TDD, the UEs transmit the pilots (only $K$ needed), and the BS estimates the $N_t$ -dimensional channel directly. The asymmetry is stark: TDD overhead scales with $K$ , FDD overhead scales with $N_t$ .

Proof

Minimum pilot dimension

Let $\mathbf{P} \in \mathbb{C}^{N_t \times \tau_d}$ be the pilot matrix transmitted by the BS. User $k$ observes $\mathbf{y}_k^{\text{pilot}} = \mathbf{P}^H \mathbf{H}_{k} + \mathbf{w}_{k}$ . For $\mathbf{H}_{k}$ to be identifiable from $\mathbf{y}_k^{\text{pilot}}$ , the matrix $\mathbf{P}^H$ must have rank $N_t$ , which requires $\tau_d \geq N_t$ .

Data efficiency fraction

In a coherence block of $T_c$ symbols, $\tau_d$ are used for DL pilots and $\tau_u$ for UL pilots. The remaining $T_c - \tau_d - \tau_u$ symbols carry data. Hence $\eta_{\text{data}} = \frac{T_c - \tau_d - \tau_u}{T_c} = 1 - \frac{\tau_d + \tau_u}{T_c}.$ Substituting the minimum values $\tau_d = N_t$ and $\tau_u = K$ yields the stated bound.

Scaling implication

For massive MIMO with $N_t = 256$ and a typical urban coherence block $T_c \approx 200$ symbols (5 MHz coherence BW, 1 ms coherence time), the DL pilot overhead alone exceeds $T_c$ — no symbols remain for data. This is the fundamental FDD bottleneck. $\blacksquare$

,

Definition:
Uplink Feedback Overhead

After estimating the downlink channel $\mathbf{H}_{k} \in \mathbb{C}^{N_t}$ , user $k$ must communicate this estimate to the BS over the uplink control channel. The feedback overhead is measured by the number of bits $B_{\text{fb}}$ per user per coherence block.

Unstructured (naive) feedback: Quantizing each complex entry of $\mathbf{H}_{k}$ to $b$ bits per real dimension requires $B_{\text{fb}} = 2 b N_t$ bits per user. For $N_t = 64$ and $b = 5$ , this is 640 bits per coherence block — a substantial burden on the uplink control channel.

Total feedback load: With $K$ users, the aggregate feedback is $K B_{\text{fb}}$ bits per coherence block. This competes with uplink data for control channel resources.

In LTE, the PUCCH (Physical Uplink Control Channel) supports at most a few hundred bits per TTI. With $N_t = 64$ and $K = 16$ , the naive feedback scheme requires over 10,000 bits per TTI — far exceeding the PUCCH capacity.

CSI Feedback

The process by which a UE communicates its estimated downlink channel state information to the BS via the uplink control channel. In FDD systems, CSI feedback is the primary mechanism for the BS to acquire downlink channel knowledge. The feedback may take the form of explicit channel coefficients, codebook indices (PMI), or compressed representations.

FDD vs TDD Overhead Comparison

Compare the data efficiency $\eta_{\text{data}}$ for FDD and TDD as the number of BS antennas $N_t$ grows. In TDD, pilot overhead scales with $K$ ; in FDD, it scales with $N_t$ . The plot reveals the point where FDD overhead consumes the entire coherence block.

Parameters

N_t

64

Number of BS antennas

K

16

Number of users

T_c

200

Coherence interval (symbols)

b

5

Bits per real dimension for naive feedback

Example: FDD Overhead in 5G NR Frequency Bands

A 5G NR base station with $N_t = 64$ transmit antennas operates in FDD Band n1 (DL: 2110–2170 MHz, UL: 1920–1980 MHz, duplex gap = 190 MHz). The subcarrier spacing is $\Delta f = 30$ kHz with $T_{\text{slot}} = 0.5$ ms. The coherence bandwidth is $B_c = 5$ MHz (urban macro) and the coherence time is $T_c^{\text{time}} = 2$ ms. Compute: (a) the coherence block size $T_c$ , (b) the minimum DL pilot overhead, (c) the data efficiency $\eta_{\text{data}}$ for $K = 16$ users.

Solution

Coherence block size

The number of coherent subcarriers is $N_c = \lfloor B_c / \Delta f \rfloor = \lfloor 5000/30 \rfloor = 166$ . The number of coherent OFDM symbols is $N_t^{\text{time}} = \lfloor T_c^{\text{time}} / T_{\text{slot}} \rfloor \times 14 = 4 \times 14 = 56$ (14 symbols per slot in NR). The coherence block size is $T_c = N_c \times N_t^{\text{time}} = 166 \times 56 \approx 9296$ resource elements.

DL pilot overhead

Minimum DL training requires $\tau_d \geq N_t = 64$ pilot resource elements. With $K = 16$ , the UL pilot overhead is $\tau_u = 16$ .

Data efficiency

$\eta_{\text{data}} = 1 - \frac{64 + 16}{9296} = 1 - \frac{80}{9296} \approx 0.991.$ $The overhead is modest — about 0.9%. This is because the coherence block is large (wideband, low mobility). The FDD bottleneck becomes severe when either (a)$ N_t $grows much larger, (b) the channel is highly time-varying (small$ T_c^{\text{time}} $), or (c) the bandwidth is narrow (small$ N_c$).

Common Mistake: Partial Reciprocity Is Not Reciprocity

Mistake:

Assuming that the angles of arrival/departure are the same at UL and DL frequencies, and therefore TDD-style beamforming can be applied in FDD. While the scattering geometry is indeed shared, the exact channel coefficients (phases, small-scale fading) differ because the wavelength changes by $\Delta \lambda / \lambda \approx \Delta f / f_0$ .

Correction:

The spatial covariance matrix $\mathbf{R}_k = \mathbb{E}[\mathbf{H}_{k} \mathbf{H}_{k}^{H}]$ is approximately frequency-independent (it depends on angles and array geometry, not on wavelength to first order). This "partial reciprocity" of the second-order statistics is precisely what JSDM exploits in Section 5. But the instantaneous channel realization $\mathbf{H}_{k}$ at the DL frequency is statistically independent of the UL realization — the BS cannot use it for coherent precoding without explicit DL-based feedback.

Definition:
Coherence Block and Spectral Efficiency with Overhead

A coherence block is a time–frequency region over which the channel can be approximated as constant. If the coherence time is $T_c^{\text{time}}$ seconds and the coherence bandwidth is $B_c$ Hz, the coherence block contains approximately

$T_c \approx T_c^{\text{time}} \cdot B_c$

independent uses of the channel (in degrees of freedom). Within each coherence block, the system must allocate resources for:

$\tau_d$ downlink pilot symbols (FDD only),
$\tau_u$ uplink pilot symbols (both TDD and FDD),
$T_c - \tau_d - \tau_u$ data symbols.

The net spectral efficiency per user $k$ is

$\bar{R}_k = \left(1 - \frac{\tau_d + \tau_u}{T_c}\right) R_k,$

where $R_k$ is the per-symbol rate (bits/s/Hz) achieved with the available CSI.

Key Takeaway

The FDD bottleneck is twofold. (1) The DL pilot overhead $\tau_d \geq N_t$ consumes coherence block resources that scale with the number of BS antennas — not the number of users. (2) The feedback overhead $B_{\text{fb}} \propto N_t$ strains the uplink control channel. Both problems are absent in TDD, where the pilot overhead scales with $K \ll N_t$ and no feedback is needed. The rest of this chapter develops techniques to reduce these overheads without sacrificing too much CSI quality.

Quick Check

In an FDD massive MIMO system with $N_t = 128$ antennas and a coherence block of $T_c = 200$ symbols, what fraction of the coherence block is consumed by DL pilots alone (assuming optimal $\tau_d = N_t$ )?

12.8%

64%

100%

25%

Correction:

64%

$\tau_d / T_c = 128/200 = 0.64 = 64\%$ . Nearly two-thirds of the coherence block is consumed by pilots — leaving little room for data.

Historical Note: The TDD vs FDD Debate in Massive MIMO

2010–2018

When Thomas Marzetta introduced the massive MIMO concept in 2010, he explicitly assumed TDD operation, arguing that the FDD overhead made large arrays impractical in paired bands. This sparked an intense debate in the wireless community: was FDD massive MIMO fundamentally impossible, or merely harder? The 3GPP standardization of NR (Release 15, 2018) included both TDD and FDD MIMO configurations, with codebook-based feedback for FDD. The debate continues to motivate research on reducing the FDD overhead gap — from compressed sensing to deep learning to JSDM — which we survey in the remainder of this chapter.

⚠️Engineering Note

FDD Spectrum in Commercial Deployments

As of 2024, the majority of sub-6 GHz cellular spectrum worldwide is allocated as FDD paired bands (e.g., LTE Bands 1, 3, 7, 20; NR Bands n1, n3, n7, n28). TDD bands (e.g., Band 41, n77, n78) are growing, particularly for 5G mid-band deployments, but the installed base of FDD infrastructure represents trillions of dollars of investment. Operators need massive MIMO to work in their existing FDD spectrum, not just in new TDD allocations. This economic reality is the primary driver for the research presented in this chapter.

Practical Constraints

•
LTE FDD bands: duplex gaps of 45–400 MHz, far exceeding coherence bandwidth
•
NR FDD supports up to 256 antenna ports with Type II CSI feedback
•
PUCCH/PUSCH feedback capacity limits the number of feedback bits per slot

📋 Ref: 3GPP TS 38.101-1

Definition:
Rate Loss from Imperfect CSI

With imperfect CSI $\hat{\mathbf{H}}_k$ at the BS (obtained from quantized feedback), the achievable rate for user $k$ under ZF precoding is

$R_k^{\text{ZF}} = \log_2\!\left(1 + \frac{P_t}{K} \cdot \frac{|\hat{\mathbf{H}}_k^H \mathbf{v}_{k}|^2}{\sigma^2 + \frac{P_t}{K} \sum_{j \neq k} |\hat{\mathbf{H}}_k^H \mathbf{v}_{j}|^2 + \frac{P_t}{K} \|\tilde{\mathbf{H}}_k\|^2 / N_t}\right),$

where $\tilde{\mathbf{H}}_k = \mathbf{H}_{k} - \hat{\mathbf{H}}_k$ is the CSI error. The last term in the denominator represents the quantization noise floor — irreducible interference caused by CSI inaccuracy. Reducing $B_{\text{fb}}$ increases $\|\tilde{\mathbf{H}}_k\|^2$ , degrading the rate.

The interplay between pilot overhead (which reduces $\eta_{\text{data}}$ ) and CSI quality (which affects $R_k$ ) creates an overhead–accuracy tradeoff: spending more resources on training and feedback improves $R_k$ but reduces $\eta_{\text{data}}$ . The optimal operating point depends on $N_t$ , $K$ , $T_c$ , and $\text{SNR}$ .

CSI Quantization Error

The difference $\tilde{\mathbf{H}}_k = \mathbf{H}_{k} - \hat{\mathbf{H}}_k$ between the true downlink channel and the BS's reconstructed estimate from quantized feedback. This error causes residual multi-user interference that does not vanish with increased transmit power, creating an interference floor analogous to pilot contamination in TDD.

The Asymmetry in Numbers

Consider a concrete scenario: $N_t = 128$ , $K = 16$ , $T_c = 200$ .

TDD: Pilot overhead $= K/T_c = 16/200 = 8\%$ . No feedback needed. Data efficiency: $92\%$ .
FDD: DL pilot overhead $= N_t/T_c = 128/200 = 64\%$ . Add $\tau_u = 16$ for UL pilots. Data efficiency: $28\%$ . Plus the UE must feed back $\sim 2 \times 5 \times 128 = 1280$ bits per coherence block.

The FDD system spends more than twice as many resources on overhead as on data — and still obtains worse CSI than TDD (quantization error). This order-of-magnitude gap motivates every technique in Sections 2–5.

Why This Matters: From FDD Overhead to Reciprocity-Based Solutions

The FDD overhead barrier established here explains why TDD is the preferred duplex mode for massive MIMO deployments (5G NR FR2, LTE-TDD). It also motivates the study of partial-reciprocity techniques that exploit the frequency-invariance of spatial statistics (angles, covariance), even when instantaneous channel coefficients differ. JSDM (Section 5) is the leading example: it uses UL-estimated covariance for DL pre-beamforming, reducing the FDD feedback problem from $N_t$ to $r_g$ dimensions.

See full treatment in JSDM as a Structured FDD Solution

FDD vs TDD: Training and Feedback Overhead

Side-by-side comparison of TDD and FDD frame structures as

N_t

grows. In TDD, uplink pilot overhead stays fixed at

K

symbols regardless of

N_t

; in FDD, downlink training scales as

N_t

and feedback bits grow proportionally. The animation makes the overhead catastrophe visually immediate: at

N_t=128

with

K=16

, FDD devotes over half the coherence block to training.

The FDD Challenge