Digital Post-Distortion for 5G Receivers

• DPoD is a receiver-side signal processing technique that compensates for PA nonlinearity and memory effects in 5G uplink systems.
• It leverages the base station's superior computational power to perform complex compensation algorithms, reducing UE complexity and power consumption.
• DPoD architectures—including time-domain, frequency-domain, and DFT-s-domain approaches—effectively mitigate nonlinear distortions and improve BER performance.

Digital Post-Distortion (DPoD) is a receiver-side signal-processing technique that compensates for nonlinear and memory effects of the user-equipment (UE) power amplifier (PA) after the signal has been received at the base station (BS). In the uplink of 5G-Advanced and beyond, it is motivated by the fact that the PA in the UE dominates uplink power consumption, especially at the cell edge, while the BS has abundant processing power and can therefore host more complex compensation algorithms. In a second, distinct usage within PA linearization, “post-distortion” also denotes an inverse learned from PA output data and then reused as a transmitter-side predistorter; under the condition that a post-distorter exists, the inverse is unique and satisfies $Pr = Po = F^{-1}$ (Schäufele et al., 15 Aug 2025, Jiang et al., 2013).

1. Formal problem and conceptual scope

A formal DPoD problem is defined by a true PA mapping $q:\mathbb{C}^N\to\mathbb{C}^N$, an intended transmit symbol vector $\mathbf{s}\in\mathbb{C}^N$, and a received vector $\mathbf{y} = h(\mathbf{s}) + \mathbf{n}$, where $h$ denotes the channel and $\mathbf{n}$ noise. The objective is to find a post-distorter $g:\mathbb{C}^N\to\mathbb{C}^N$ such that

$$\hat{\mathbf{s}} = g(\mathbf{y}) \approx \mathbf{s}.$$

This formulation treats DPoD as compensation of PA-induced impairments after reception rather than before transmission (Schäufele et al., 15 Aug 2025).

The contrast with Digital Pre-Distortion (DPD) is operational. DPD applies the inverse nonlinearity at the transmitter side, requiring a feedback loop in the UE to measure its own PA characteristics; this increases UE complexity and battery drain. DPoD moves the compensation to the BS, avoiding UE-side feedback and enabling more complex algorithms. In single-carrier transmission with frequency domain equalization (FDE), a two-stage approach is adopted, in which the linear communication channel is equalized at the first stage, and it is followed by a post-distortion where nonlinear distortion is reduced (Schäufele et al., 15 Aug 2025, Salman et al., 2020).

A recurrent ambiguity in the literature is that “digital post-distortion” may refer either to a receiver algorithm or to an indirect-learning step for transmitter predistortion. The latter usage is formalized by the operator statement $Po\circ F = I$ on the admissible input space; if such a post-distorter exists, then $F$ is a bijection from $q:\mathbb{C}^N\to\mathbb{C}^N$0 onto $q:\mathbb{C}^N\to\mathbb{C}^N$1, its inverse exists, and that inverse is exactly the predistorter, $q:\mathbb{C}^N\to\mathbb{C}^N$2 (Jiang et al., 2013). This suggests that the term spans two related but not identical engineering contexts.

2. Architectural forms in multicarrier uplink systems

For a 5G DFT-spread-OFDM uplink system with FFT size $q:\mathbb{C}^N\to\mathbb{C}^N$3, DFT size $q:\mathbb{C}^N\to\mathbb{C}^N$4, and guard carriers zeroed out, three principal DPoD architectures are compared: time-domain, frequency-domain, and DFT-s-domain. Let $q:\mathbb{C}^N\to\mathbb{C}^N$5 be the DFT-s domain data, $q:\mathbb{C}^N\to\mathbb{C}^N$6 its frequency-domain vector, $q:\mathbb{C}^N\to\mathbb{C}^N$7 its time-domain vector, $q:\mathbb{C}^N\to\mathbb{C}^N$8 the PA output, and $q:\mathbb{C}^N\to\mathbb{C}^N$9 the channel-convolved signal (Schäufele et al., 15 Aug 2025).

Architecture	Processing statement	Reported characterization
Time-domain DPoD	channel equalization $\mathbf{s}\in\mathbb{C}^N$0 inverse-nonlinearity $\mathbf{s}\in\mathbb{C}^N$1 $\mathbf{s}\in\mathbb{C}^N$2 DFT-s demapper $\mathbf{s}\in\mathbb{C}^N$3 decoder	good balance between low computational complexity and efficient nonlinearity compensation
Frequency-domain DPoD	FFT $\mathbf{s}\in\mathbb{C}^N$4 channel equalization $\mathbf{s}\in\mathbb{C}^N$5 inverse nonlinearity across all subcarriers jointly	the nonlinearity in time-domain “spreads” over all $\mathbf{s}\in\mathbb{C}^N$6
DFT-s-domain DPoD	subcarrier demapping and IDFT$\mathbf{s}\in\mathbb{C}^N$7, then per-symbol Volterra or kernel method	each symbol is distorted in a different pattern

In time-domain DPoD, after perfect channel inversion, one obtains $\mathbf{s}\in\mathbb{C}^N$8 with $\mathbf{s}\in\mathbb{C}^N$9, and a Volterra-type inverse model of order $\mathbf{y} = h(\mathbf{s}) + \mathbf{n}$0 and memory $\mathbf{y} = h(\mathbf{s}) + \mathbf{n}$1 is

$\mathbf{y} = h(\mathbf{s}) + \mathbf{n}$2

In frequency-domain DPoD, the compensation must act jointly across subcarriers because the nonlinearity “spreads” over all $\mathbf{y} = h(\mathbf{s}) + \mathbf{n}$3: $\mathbf{y} = h(\mathbf{s}) + \mathbf{n}$4 In DFT-s-domain DPoD, after subcarrier demapping and IDFT$\mathbf{y} = h(\mathbf{s}) + \mathbf{n}$5, one defines $\mathbf{y} = h(\mathbf{s}) + \mathbf{n}$6 and applies a per-symbol Volterra or kernel method,

$\mathbf{y} = h(\mathbf{s}) + \mathbf{n}$7

The reported finding is that implementing DPoD in the time-domain, complemented by frequency-domain channel equalization, strikes a good balance between low computational complexity and efficient nonlinearity compensation (Schäufele et al., 15 Aug 2025).

3. Memory effects and nonlinear inter-symbol interference

A central technical point is that memory has to be taken into account regardless of the memory of the PA. Even a memoryless PA followed by ideal subcarrier mapping and demapping and low-pass filtering induces effective memory in the composite nonlinearity, because the low-pass filter is an FIR. Hence, any inverse model must include memory taps to correctly invert the PA$\mathbf{y} = h(\mathbf{s}) + \mathbf{n}$8filter chain (Schäufele et al., 15 Aug 2025).

An analogous conclusion appears in single-carrier systems. Even though $\mathbf{y} = h(\mathbf{s}) + \mathbf{n}$9 is memoryless, pulse-shaping and matched filtering create inter-symbol interference. A discrete-time equivalent shows

$h$0

so neighboring received symbols must be exploited in order to suppress nonlinear interference (Salman et al., 2020).

This point addresses a common misconception: a memoryless PA does not imply a memoryless receiver compensation problem. In DPoD, the received or intermediate samples are therefore arranged into memory-augmented vectors. One formulation uses

$h$1

while a symbol-rate single-carrier receiver uses

$h$2

and then learns a post-distorter $h$3 such that $h$4 (Schäufele et al., 15 Aug 2025, Salman et al., 2020).

4. Volterra representations, real Hilbert spaces, and kernel methods

One line of DPoD research develops the inverse in a real Hilbert space. The isomorphism is defined by

$h$5

so that $h$6 with inner product $h$7 is isometric to $h$8 with the usual Euclidean inner product. With

$h$9

any PA-inverse function $\mathbf{n}$0 satisfying $\mathbf{n}$1 can be written as $\mathbf{n}$2 and $\mathbf{n}$3 for some real-valued odd function $\mathbf{n}$4 (Schäufele et al., 15 Aug 2025).

Within that formulation, a real-domain Volterra series of odd orders $\mathbf{n}$5 is

$\mathbf{n}$6

A complexified Volterra series up to order $\mathbf{n}$7 and memory $\mathbf{n}$8 may also be written as

$\mathbf{n}$9

For coefficient estimation, one stacks $g:\mathbb{C}^N\to\mathbb{C}^N$0 training pairs, forms a regressor matrix $g:\mathbb{C}^N\to\mathbb{C}^N$1 with $g:\mathbb{C}^N\to\mathbb{C}^N$2, and solves

$g:\mathbb{C}^N\to\mathbb{C}^N$3

The equivalent kernel method replaces explicit Volterra coefficient enumeration by an RKHS model. With polynomial kernel

$g:\mathbb{C}^N\to\mathbb{C}^N$4

one solves

$g:\mathbb{C}^N\to\mathbb{C}^N$5

By the representer theorem,

$g:\mathbb{C}^N\to\mathbb{C}^N$6

The reported per-sample complexity is $g:\mathbb{C}^N\to\mathbb{C}^N$7 real multiplies for Volterra and $g:\mathbb{C}^N\to\mathbb{C}^N$8 real multiplies for the kernel method, while training complexity is $g:\mathbb{C}^N\to\mathbb{C}^N$9 for Volterra LS and $$\hat{\mathbf{s}} = g(\mathbf{y}) \approx \mathbf{s}.$$0 for the kernel solve (Schäufele et al., 15 Aug 2025).

5. Symbol-rate post-distortion in single-carrier receivers

A concrete receiver realization is a two-stage architecture for single-carrier transmission with FDE under PA nonlinearities. After matched filtering and sampling, $$\hat{\mathbf{s}} = g(\mathbf{y}) \approx \mathbf{s}.$$1 sampling branches are formed as

$$\hat{\mathbf{s}} = g(\mathbf{y}) \approx \mathbf{s}.$$2

Using a known training block, one estimates the channel in each branch by least squares,

$$\hat{\mathbf{s}} = g(\mathbf{y}) \approx \mathbf{s}.$$3

then applies MMSE-style FDE in the frequency domain,

$$\hat{\mathbf{s}} = g(\mathbf{y}) \approx \mathbf{s}.$$4

The post-distorter then operates at symbol rate on memory-augmented vectors $$\hat{\mathbf{s}} = g(\mathbf{y}) \approx \mathbf{s}.$$5 (Salman et al., 2020).

Two post-distortion techniques are described. In Gaussian process regression (GPR), the in-phase and quadrature components are modeled separately,

$$\hat{\mathbf{s}} = g(\mathbf{y}) \approx \mathbf{s}.$$6

with Gaussian kernel

$$\hat{\mathbf{s}} = g(\mathbf{y}) \approx \mathbf{s}.$$7

The predictive posterior mean is

$$\hat{\mathbf{s}} = g(\mathbf{y}) \approx \mathbf{s}.$$8

and segmentation reduces complexity by training small GPRs and fusing the estimates by Best Linear Unbiased Estimator (BLUE). In the neural network approach, the ARVTDNN uses the same $$\hat{\mathbf{s}} = g(\mathbf{y}) \approx \mathbf{s}.$$9, one hidden layer with $Po\circ F = I$0 neurons and activation $Po\circ F = I$1, and outputs

$Po\circ F = I$2

with analogous $Po\circ F = I$3. Training minimizes the MSE

$Po\circ F = I$4

Detection is then performed with a fractional-delay FDE bank and a Gaussian-approximate decision metric. Stacking the $Po\circ F = I$5 post-distorted soft outputs as $Po\circ F = I$6, one models

$Po\circ F = I$7

learns $Po\circ F = I$8 and $Po\circ F = I$9, and decides

$F$0

This is equivalent to whitening and maximum-ratio combining across the $F$1 branches, giving more weight to the less-distorted timing offsets (Salman et al., 2020).

6. Reported gains, trade-offs, and terminological boundaries

The reported evaluation metrics for receiver-side DPoD include uncoded BER, BLER with 5G-LDPC, EVM, and ACLR in multicarrier studies, and bit error rate together with achievable information rate metrics in single-carrier studies (Schäufele et al., 15 Aug 2025, Salman et al., 2020). In an AWGN + memoryless-PA scenario, time-domain DPoD with symmetric memory recovers near-ideal BER with a $F$2–$F$3 dB backoff penalty. Under TDL-D fading + GMP-PA, DPoD with the Volterra or kernel approach achieves $F$4 dB SNR improvement at $F$5 BLER over the best previously published receiver-side MP method and $F$6 dB over a DFT-s-domain NN; the residual SNR penalty versus a linear PA is only $F$7 dB (Schäufele et al., 15 Aug 2025).

For the single-carrier FDE-bank receiver, the simulation setup includes 1024-QAM in AWGN with a Saleh PA, 256-QAM for a dispersive COST-207/Rayleigh channel with realistic GaN PA models, oversampling $F$8, RRC pulse-shaping with roll-off $F$9, slow-time training $q:\mathbb{C}^N\to\mathbb{C}^N$00, fast-time training $q:\mathbb{C}^N\to\mathbb{C}^N$01, GPR with memory depth $q:\mathbb{C}^N\to\mathbb{C}^N$02 and $q:\mathbb{C}^N\to\mathbb{C}^N$03 segments, and ARVTDNN with a single hidden layer of $q:\mathbb{C}^N\to\mathbb{C}^N$04. In AWGN + Saleh PA, NN and GPR post-distortion achieve near-ideal BER floors down to $q:\mathbb{C}^N\to\mathbb{C}^N$05 at $q:\mathbb{C}^N\to\mathbb{C}^N$06 dB, versus $q:\mathbb{C}^N\to\mathbb{C}^N$07 dB for memoryless and $q:\mathbb{C}^N\to\mathbb{C}^N$08 dB for the conventional receiver. In a dispersive channel with GaN PA, the bank-of-FDE plus DA-SSD with NN or GPR reaches BER close to the linear-PA baseline, and the post-distortion plus FDE-bank attains within $q:\mathbb{C}^N\to\mathbb{C}^N$09 b/s/Hz of the linear-PA capacity over a wide OBO/SNR range (Salman et al., 2020).

The principal trade-off reported for DPoD is between nonlinear compensation quality and receiver complexity. Existing transmitter-side solutions such as DPD require complex feedback mechanisms for optimal performance, leading to increased UE complexity and power consumption, whereas DPoD leverages the superior computational resources of the BS (Schäufele et al., 15 Aug 2025). At the same time, the older PA-linearization literature establishes a precise boundary condition for the alternative usage of “post-distorter”: if a post-distorter exists, then it is also a predistorter, and therefore the predistorter and postdistorter are equivalent (Jiang et al., 2013). This distinction is terminological rather than contradictory: receiver-side DPoD denotes BS-based nonlinear compensation after reception, while post-distorter-based indirect learning denotes identification of an inverse that may later be deployed at the transmitter.