OTFS versus OFDM in mobile wireless: where the high-mobility gap actually appears

The baseline problem: OFDM's orthogonality breaks under Doppler

OFDM has been the default multicarrier waveform for LTE, Wi-Fi and 5G NR because it is simple, spectrally efficient and robust against multipath fading, but it transmits symbols in the time-frequency domain and is therefore sensitive to rapidly time-varying channels [1]. High Doppler shifts cause inter-carrier interference (ICI), which degrades OFDM performance in high-mobility scenarios [1]. This is not a niche concern: satellite links experience LEO Doppler shifts up to roughly ±50 kHz, which destroys OFDM subcarrier orthogonality and produces ICI, channel-estimation errors and key-agreement breakdown [6]. Free-space optical multicarrier links show the same mechanism in a different medium, where normalized Doppler produces constellation rotation plus ICI-driven spreading, and residual ICI leaves a BER floor even after common phase error correction [3].

The important nuance from the optical study is that not all Doppler damage is equal: a simple per-symbol phase correction largely removes common phase error at moderate normalized Doppler, but the residual ICI remains and requires ICI-aware equalization based on a time-varying channel model [3]. That distinction — correctable phase rotation versus irreducible subcarrier mixing — is the physical reason a delay-Doppler representation is attractive in the first place.

Why delay-Doppler mapping changes the channel model

OTFS maps information symbols onto a two-dimensional delay-Doppler grid instead of the time-frequency plane, converting a time-varying multipath channel into an approximately invariant two-dimensional response and allowing symbols to benefit from the full time-frequency diversity of the channel [1]. In implementation terms, OTFS can be layered as an additional block around an OFDM multicarrier signal, which is what makes it a plausible upgrade path rather than a clean-sheet waveform [1]. The transform chain is explicit: an inverse symplectic finite Fourier transform maps delay-Doppler symbols to time-frequency, the Heisenberg transform produces the time-domain signal, and the receiver reverses this with the Wigner transform and SFFT [1]. A lower-complexity route uses the inverse discrete Zak transform to map delay-Doppler symbols directly to discrete time [1].

This representation has been picked up well beyond terrestrial cellular. Satellite physical-layer key generation now uses OTFS delay-Doppler grids with ephemeris-driven time-hopping to build high-entropy, high-reciprocity randomness sources for LEO/MEO links [6]. Visible-light and terahertz work has also adopted delay-Doppler mapping: an ORIS-aided NOMA-MIMO VLC system using DFT-s-OTFS reports better BER than RIS-assisted OFDM, especially as RIS element count grows [7], and a MIMO CV-MDI-QKD terahertz system uses OTFS delay-Doppler mapping to resist multipath and Doppler while MIMO beamforming extends secure distance [11]. The concept is therefore established; what has been less settled is a systematic, quantitative boundary for when it actually pays off.

What the new comparison actually measured — and where the gap appears

The anchor study evaluates OTFS and OFDM under perfect channel state information so that channel estimation is removed as a confounder and the comparison isolates modulation behavior [1]. It sweeps four dimensions in one framework: modulation order (4-QAM, 16-QAM, 64-QAM, 128-QAM), mobility (static 0 km/h, mid-mobility 127 km/h, high-mobility 380 km/h), number of multipath components (2, 5 and 8), and equalizer type for OTFS (LMMSE versus message passing) [1]. The headline result is conditional rather than universal: OTFS outperforms OFDM consistently in high-mobility and multipath environments, mainly at low and mid modulation orders (4 and 16), where it keeps BER low even in extreme vehicular and high-speed-train-like conditions [1]. At high modulation orders (64 and 128), OTFS only outperforms OFDM across high-SNR scenarios [1]. Depending on modulation order, the gap between the two can be very close or many orders of magnitude [1].

The equalizer result is the second half of the story. At low and mid SNR, the non-linear message-passing equalizer performs better than the traditional linear LMMSE equalizer for OTFS, with the gain concentrated between roughly 0 and 21 dB [1]. Mechanistically this fits the delay-Doppler channel structure: LMMSE regularizes inversion to limit noise amplification, while message passing iteratively updates symbol beliefs over the sparse delay-Doppler channel matrix and exploits that sparsity, at the cost of higher computational complexity and convergence that is not always guaranteed and may depend on channel sparsity and iteration count [1].

Detector evidence converges: linear front-ends leave performance on the table

Independent MIMO-OTFS work reinforces the equalizer finding with a different method. A residual-learning MMSE neural detector, tested under Rayleigh and Rician fading with up to 20% channel estimation error, reaches a BER of 10⁻³ at roughly 9–10 dB SNR under Rayleigh with 20% channel error, versus about 18–20 dB for zero-forcing and 16–17 dB for MMSE; under Rician fading it reaches the same BER at about 6–7 dB versus 16–18 dB for ZFE and 14–15 dB for MMSE [2]. The reported gains are 10–13 dB over zero-forcing and MMSE at BER 10⁻³, with 3–6 dB lower SNR than ML, QRM-MLD and deep-learning baselines, plus over 60% lower inference latency than LSTM/Bi-LSTM detectors and 5–8 dB lower out-of-band emissions [2]. Those numbers come from a different configuration — large-scale MIMO-OTFS with imperfect CSI — so they are not directly transferable to the anchor study's SISO perfect-CSI setting, but the direction agrees: linear detection is the weak link, and non-linear or learned correction recovers substantial SNR.

The same pattern appears in RIS-assisted OTFS, where an extreme learning machine channel estimator improves NMSE and BER over plain OTFS, RIS-OTFS and OTFS-with-ELM baselines, while remaining slightly inferior to heavier deep-learning estimators such as ReEsNet and Channelformer at high SNR but with far fewer complex multiplications and much shorter running time [8]. That study also reports robustness of symbol detection against variations in modulation order, maximum velocity and number of sub-surfaces [8] — a useful counterpoint to the anchor paper's finding that modulation order sharply modulates the OTFS advantage.

Where the conclusion stops: simulation scope, hardware, and competing waveforms

The anchor paper's own framing limits its reach: results are simulation-based under perfect CSI, and the authors explicitly isolate modulation effects by idealizing channel estimation [1]. That means the reported OTFS-versus-OFDM gaps should not be read as deployed cellular or standardized-system performance. The same caution is standard in waveform benchmarking: a unified OFDM/FBMC/F-OFDM/GFDM/UFMC assessment with perfect channel knowledge and ideal synchronization explicitly describes its BER results as controlled relative trends rather than complete claims about deployed performance, and notes that carrier-frequency offset, phase noise, timing errors and imperfect channel estimates may affect candidates differently, particularly those built around long filters or non-orthogonal pulses [4]. That study also found no single waveform winning every early-6G service, favoring an air interface that can switch waveforms with the operating requirement [4] — a direct argument against treating OTFS as a universal replacement.

Hardware and nonlinearity are a second boundary. OTFS channel estimation under high-power-amplifier distortion requires dedicated compensation; an IBO-driven dynamic gated Bi-GRU estimator reports logarithmic BER improvement over conventional methods, 3–4 dB over deep-learning baselines at equal SNR, and over 7 dB PAPR reduction, with roughly 26% fewer MACs and FLOPs than a standard Bi-GRU [5]. Constant-envelope OTFS work likewise treats PAPR and phase recovery as first-order design problems, and its constant-envelope claim applies only to ideal discrete complex-baseband samples, excluding pulse shaping and RF hardware [9]. A third boundary is the equalizer itself: message passing offers gains but its convergence is not always guaranteed and depends on channel sparsity and iteration count [1], and learned detectors trained offline may not adapt in rapidly time-varying environments [2].

Field evidence is sparse but not absent. An underwater acoustic OTFS receiver using a turbo decision-feedback equalizer and decoder reports better BER against long-multipath fading and severe Doppler than existing delay-Doppler-domain equalizers, and reduces the accuracy requirements of Doppler compensation relative to single-carrier coherent modulation and OFDM in lake experiments [10]. That is a different medium and a different receiver architecture, so it validates the general direction rather than the anchor paper's specific numbers. Meanwhile, OTFS is being embedded in sensing and security applications — a constant-envelope OTFS-CPM-LFM waveform for high-mobility ISAC [9], a power-domain OTFS plus FMCW superposition for UAV ISAC with an explicit rate-distortion trade-off [12], and OTFS-based satellite key generation [6] — which means the waveform's future will be judged on PAPR, complexity, sensing sidelobes and security as much as on BER. The open questions the anchor paper leaves are concrete: how the gap behaves with imperfect CSI, with OFDM variants and other equalizers, and under PAPR and computational-cost accounting [1].

About These Sources

This research page is built on 12 peer-reviewed studies — published from 2025 to 2026, 12 from 2024 or later — selected as the most relevant from 13 studies that passed quality screening, drawn from 93 papers retrieved from a database of over 500 million.

Sources used in this answer