Impact of PA Nonlinearities on AFDM Sensing: A Matched Filtering Perspective

In AFDM, a one-dimensional vector $\mathbf{s}\in\mathbb{C}^{N\times 1}$ is multiplexed into the twisted time-frequency chirp domain using the inverse discrete affine Fourier transform ($\mathrm{IDAFT}$) [16],

where $(\cdot)^{H}$ is the conjugate transpose, $\mathbf{A}=\mathbf{\Lambda}_{c_{2}}\mathbf{F}\mathbf{\Lambda}_{c_{1}}\in\mathbb{C}^{N\times N}$ is the forward $N$-point discrete $\mathrm{AFT}$ ($\mathrm{DAFT}$) matrix, and $\mathbf{\Lambda}_{c_{i}}=\mathrm{diag}\big[e^{-j2\pi c_{i}0^{2}},\dots,e^{-j2\pi c_{i}(N-1)^{2}}\big]\in\mathbb{C}^{N\times N}$ is a diagonal chirp matrix with central digital frequency $c_{i}$. $\mathbf{F}\in\mathbb{C}^{N\times N}$ is the $N$-point discrete FT ($\mathrm{DFT}$) matrix, and hence, $\mathbf{F}^{H}$ is the $N$-point inverse $\mathrm{DFT}$ ($\mathrm{IDFT}$) matrix. To combat inter-symbol interference between multiple received symbol copies due to multipath effects, a chirp-periodic prefix (CPP) of length $L$ is appended [6].

Due to PA nonlinearity at the transmitter, the amplified signal is subject to distortion. Without loss of generality, we approximate the output signal using the Bussgang theorem. For sufficiently large block lengths $N$, the central limit theorem (CLT) implies that the TD signal converges in distribution to a circularly symmetric complex Gaussian random variable with variance $\sigma_{x}^{2}$. Hence, the output of the nonlinear function, i.e., nonlinear PA, can be expressed according to the Bussgang approximation as [17],

where $\mathbf{d}$ is the PA-induced distortion term uncorrelated with $\mathbf{x}$, and with variance $\mathbf{\sigma}^{2}_{d}=\mathbb{E}[|\mathbf{d}|^{2}]$. Its $\mathrm{DAFT}$-domain representation is $\mathbf{t}=\mathbf{A}\mathbf{d}$. Bussgang gain $\kappa$ is defined as,

Assuming a sparse multipath environment with $P$ reflectors, the received signal can be expressed after CPP removal as,

where $\mathbf{w}\sim\mathcal{CN}(\mathbf{0},\sigma_{w}^{2}\mathbf{I}_{N})$ is the additive white Gaussian noise, and $\mathbf{H}$ is the discrete channel matrix given by [3],

The complex path gain, the delay and Doppler indices are denoted as $h_{i}$, $l_{i}$ and $k_{i}$, respectively. $\mathbf{\Pi}_{l_{i}}\in\mathbb{C}^{N\times N}$ is a circular permutation matrix corresponding to delay $l_{i}$, and $\mathbf{\Delta}_{k_{i}}=\mathrm{diag}\!\left(e^{-j2\pi k_{i}n/N}\right)$ is the Doppler modulation matrix [3]. Substituting (2) in (4), the received signal becomes

For sensing, to maximize the SNR after (6), the receiver usually processes the received signal with a matched filter (MF) first, where delays and Doppler frequency shifts will be observed. For AFDM, matched filtering in the $\mathrm{DAFT}$ domain is applied to estimate the delay-Doppler parameters for target range and velocity detection, by exploiting the AFDM-induced cyclic shifts and 2D linear phase rotations along propagation paths [6].

For a transmitted frame of $M$ consecutive AFDM symbols, we define the transmitted symbol matrix $\mathbf{S}=[\mathbf{s}_{0},\mathbf{s}_{1},\dots,\mathbf{s}_{M-1}]\in\mathbb{C}^{N\times M}$, and the corresponding TD signals $\mathbf{X}=[\mathbf{x}_{0},\mathbf{x}_{1},\dots,\mathbf{x}_{M-1}]\in\mathbb{C}^{N\times M}$, where each column $\mathbf{s}_{m}$ contains the $N$ data symbols of the $m^{\text{th}}$ AFDM symbol, and $\mathbf{x}_{m}=\mathbf{A}^{H}\mathbf{s}_{m}$ is the corresponding TD transmitted signal as in (1).

At the receiver side, the $\mathrm{DAFT}$ domain frame can be written as,

where $\tilde{\mathbf{D}}=\mathbf{A}\mathbf{H}\mathbf{D}$, $\tilde{\mathbf{W}}=\mathbf{A}\mathbf{W},$ are the PA-induced distortion and noise representations in the DAFT domain, and $\mathbf{R},\mathbf{D},\mathbf{W}\in\mathbb{C}^{N\times M}$ are the received signal, distortion, and noise frames, respectively.

The zero-Doppler cut can also be regarded as the autocorrelation function of the TD signal. In this section, we study the autocorrelation properties of the AFDM in the presence of PA-induced nonlinearities. The zero-Doppler cut can be expressed from (12) as

where $\chi_{x}(l,0)$ denotes the autocorrelation of the input signal, $\chi_{d}(l,0)$ that of the distortion, and $\kappa\chi_{x,d}(l,0)+\kappa^{\ast}\chi_{d,x}(0,l)$ is the cross-correlation term.

Under the Bussgang assumptions, the distortion component $\mathbf{d}$ is uncorrelated with the input signal $\mathbf{x}$, i.e., $\mathbb{E}[\mathbf{x}\mathbf{d}^{H}]=\mathbf{0}$. Therefore, the expectation of cross-terms $\mathbb{E}[\chi_{x,d}(l,0)]=\mathbb{E}[\mathbf{x}^{H}\mathbf{J}_{l}\mathbf{d}]=0$ and $\mathbb{E}[\chi_{d,x}(l,0)]=\mathbb{E}[\mathbf{d}^{H}\mathbf{J}_{l}\mathbf{x}]=0$. We obtain,

For AFDM, the $\mathrm{IDAFT}$ modulation matrix does not diagonalize the time-shift matrix $\mathbf{J}_{l}$ for non-zero $c_{1}$ [5]. Consequently, the intrinsic delay sidelobes $\chi_{x}(l,0)$ remain dominant, unlike in OFDM, where the $\mathrm{DFT}$ diagonalizes $\mathbf{J}_{l}$ and suppresses delay sidelobes. Although the distortion term $\chi_{d}(l,0)$ may introduce additional peaks, their impact is negligible compared to the elevated sidelobe floor already caused by the non-diagonal structure of $\mathbf{B}_{(l,0)}$. This is consistent with our findings in [14], where QAM-based OFDM was shown to be relatively insensitive to PA nonlinearities. Since AFDM exhibits even higher intrinsic sidelobes than OFDM [5], the effect of PA-induced distortion is expected to be even less significant.

We now characterize the impact of nonlinear PA distortion on the zero-delay cut using the Bussgang decomposition. By fixing $l=0$ in (12), the zero-delay cut of the amplified signal is expressed as

where $\Delta_{\text{Doppler}}=\kappa\chi_{x,d}(0,k)+\kappa^{\ast}\chi_{d,x}(0,k)+\chi_{d}(0,k)$ is the nonlinear contribution. Under the Bussgang assumptions $\mathbb{E}[\chi_{x,d}(l,0)]=0$, and $\mathbb{E}[\chi_{d,x}(l,0)]=0$.

Moreover, since $\mathbf{t}$ is a unitary transform of the i.i.d. TD distortion $d(n)$, its components $t(n)$ are also independent and identically distributed (i.e., $\mathbb{E}[\mathbf{t}\mathbf{t}^{H}]=\sigma_{t}^{2}\mathbf{I}$). Therefore, we obtain,

From (11), it is easy to prove that the diagonal terms of $\mathbf{B}_{(0,k)}$ are null for nonzero integer Doppler shifts [8], which implies that in (18), the contribution of the nonlinear term to the existing sidelobes is zero in expectation. Whereas, the distortion term $\mathbb{E}[\Delta_{\text{Doppler}}]$ contributes exactly $\sigma_{t}^{2}\operatorname{tr}(\mathbf{B}_{(0,0)})=\sigma_{t}^{2}N$, to the mainlobe. As a result, the Doppler sidelobes decreases after nonlinear amplification, relatively to the mainlobe, which in result decreases the $\mathrm{PSLR}$.

Figure 1 represents the zero-Doppler cuts before and after nonlinear amplification for both AFDM and OFDM waveforms using PSK and QAM constellations. A clear increase of the sidelobes is observed for OFDM under strong nonlinear distortion (low IBO). In particular, the sidelobes increase noticeably in the OFDM case, especially with PSK modulation, indicating a higher sensitivity to PA nonlinearities in terms of ranging performance. In contrast, AFDM shows a remarkable insensitivity where the sidelobes remain nearly unchanged before and after amplification. The sidelobe level at the input is initially high, which makes the contribution of the distortion negligible. Furthermore, for AFDM, both PSK and QAM constellations yield almost identical sidelobe levels. This indicates that, under AFDM signaling, the ranging sidelobes are primarily governed by the waveform modulation scheme rather than by the specific constellation geometry.

Figure 2 shows the corresponding zero-delay cuts for PSK signaling. Interestingly, both waveforms show comparable Doppler-domain characteristics. Moreover, the amplified signals even display reduced sidelobe levels, consistent with our earlier observations; this reduction can be attributed to compression effects, which smooth out large instantaneous peaks [14]. It should be noted that, despite the significant reduction of the far sidelobes, the sidelobes in the vicinity of the main lobe are slightly reduced.

While qualitative inspection of the ambiguity plots provides useful intuition, a systematic comparison requires quantitative metrics. For this purpose, we evaluate the PSLR, summarized in Table I, for both the delay and Doppler domains. First, in the delay domain, OFDM achieves lower PSLR than AFDM. However, OFDM experiences a noticeable degradation after nonlinear amplification (approximately 2 dB at IBO = 1 dB), whereas AFDM shows only marginal variation (less than 0.5 dB). This confirms that AFDM delay sidelobes are governed by the $\mathrm{IDAFT}$ spreading and are therefore largely unaffected by PA distortions. Second, in the Doppler domain, both waveforms exhibit nearly identical PSLR values, which are reduced after nonlinear amplification.

It is important to note that averaging the AF over many realizations may smooth out rare nonlinearity occurrences and therefore underestimate instantaneous unfavorable ambiguity instances. Motivated by this observation, we present in Fig. 3 the complementary cumulative distribution function (CCDF) of the PSLR for the zero-Doppler and zero-delay cuts, i.e., $\Pr\left(\mathrm{PSLR}>\mathrm{PSLR}_{0}\right)$, which represents the outage probability that the ambiguity sidelobes exceed a certain threshold $\mathrm{PSLR}_{0}$. Since larger $\mathrm{PSLR}$ values correspond to higher sidelobes, a left-shifted CCDF indicates improved ambiguity characteristics. Accordingly, for a fixed sidelobe threshold, a lower CCDF value implies a lower outage probability and thus a more robust matched filter response under nonlinearities.

Fig. 3(a) shows the outage probability associated with the zero-Doppler cut. A clear performance gap between OFDM and AFDM can be observed. Moreover, the AFDM curves undergo only a slight shift after amplification, revealing a significantly smaller increase in outage probability and therefore a lesser sensitivity of the ranging sidelobes to nonlinearities. The outage probability interpretation is particularly meaningful when a target sidelobe level is imposed by the ISAC system design, as it is directly reflected in the range-Doppler response of the matched filter. For instance, near $\mathrm{PSLR}_{0}=-14$ dB, the OFDM curves exhibit lower outage probabilities than the AFDM curves, meaning that the OFDM sidelobes are less likely to surpass this threshold; however, these probability values are highly sensitive to nonlinear amplification. A different behavior is observed in Fig. 3(b), corresponding to the zero-delay cut. The output curves are left-shifted, confirming our earlier observations and indicating a smaller outage probability after nonlinear amplification, which is always favorable.

For the practical detection, we consider the monostatic case where we rely on the knowledge of the transmit signal at the receiver. The carrier frequency $f_{c}$ is fixed to 24 GHz. For comparison, the OFDM waveform is processed using a matched filter in the time-frequency domain, followed by a 2D RD map construction by the 2D-$\mathrm{FFT}$ [6].

Fig. 4 shows the map SNR, which is the ratio between the peak caused by the target and the average noise level in the RD map, before and after nonlinear amplification, versus normalized Doppler shift (with respect to $\Delta f$ the chirp subcarrier spacing). AFDM exhibits a nearly constant SNR across Doppler shifts, while OFDM shows a pronounced drop near one subcarrier spacing due to inter-carrier interference. Under nonlinear amplification, the radar map SNR degrades with AFDM by approximately 1 dB, compared to about 2 dB for OFDM, and identical curves are observed for PSK and QAM, consistent with the AF cuts’ observations.

Finally, to evaluate the average MF response under PA nonlinearities, we show in Fig. 5 the probability of detection $P_{d}$ versus SNR for $P=3$ resolvable targets, using 16-PSK symbols. It is computed as the average per-target successfully detected over 1k Monte Carlo realizations. Detection is performed using a CFAR detector [6] over the RD map with a fixed false alarm probability of $10^{-4}$, and SNR is defined as the inverse of the noise variance. At high velocities, OFDM fails to detect the targets. The solid lines correspond to moderate velocities ($f=0.1\Delta f$), while the dashed lines correspond to high velocities ($f=0.98\Delta f$). The substantial drop in SNR (as seen in Fig. 4) due to Doppler-induced inter-carrier interference causes the CFAR detector to trigger on noise peaks rather than the true target peaks. In contrast, AFDM manages to maintain a similar detection probability. As observed in Fig. 4, the map SNR is constant and remains sufficiently high for the CFAR threshold to detect the targets reliably. Here, the PA introduces an additive distortion term that slightly raises the noise floor, but the peak positions in the RD map remain unbiased, so only minor adjustments to the CFAR threshold are needed.