betaFeedback

WorksEngineering

Density-First Design of Anti-Jamming Sparse Spreading Codes for NB-IoT

Abstract

As narrowband Internet of Things (NB-IoT) systems continue to be deployed in increasingly heterogeneous and demanding wireless environments, including those characterized by interference, high mobility, and potentially hostile conditions, enhancing the resilience of the physical-layer waveform has become an important and timely design consideration. Motivated by this need, the present paper introduces a density-first framework for the design of anti-jamming sparse spreading codes in NB-IoT systems. Building on this framework, we propose a generalized method for constructing sparse codes that decouples column density from the conventional Alamouti-paired structure, thereby enabling both disjoint-support and deterministic code designs. In addition, we develop a generative machine-learning approach based on a variational autoencoder for synthesizing jamming-robust sparse spreading codes. To evaluate the performance of sparse spreading under conditions of additive white Gaussian noise, fading, and pulse jamming, we present a bit-error-rate (BER) analysis framework. Extensive Monte Carlo simulations, conducted across a wide range of density, channel, and jamming conditions, show that the performance of sparse spreading codes is determined primarily by column density rather than by optimization of code coefficients. In particular, highly sparse codes with disjoint support offer the greatest robustness to jamming while avoiding the unfavorable intermediate-density regime. Relative to the prior-art Alamouti-paired low-density spreading code, the recommended uniform-density design reduces pulse-jamming BER by roughly 5 times; a deterministic Lempel-Costas construction attains the empirical optimum in closed form, while the variational-autoencoder generator yields a further 45% improvement over the best algebraic design.

Introduction

Narrowband Internet of Things (NB-IoT) has become a de-facto low-power wide-area technology for industrial monitoring, smart utilities, and tactical communications [1, 2], and is increasingly deployed in environments exposed to co-channel interference, deliberate jamming [3], and high mobility with Doppler-induced channel variation [4]. Direct-sequence spread spectrum (DSSS) is the workhorse mechanism for jamming-resilient communication and code-division multiple access [4, 5]: it leverages processing gain from chip-rate spreading codes, suppresses narrowband interferers, and separates concurrent users via low cross-correlation signatures. The choice of spreading code governs anti-jamming performance and multiple-access capacity. Classical sequences such as Walsh, Hadamard, Gold, and orthogonal variable spreading factor (OVSF) codes offer predictable correlation properties but are designed for benign channels, and their robustness in highly time-varying or adversarial settings is limited [4, 5].

Literature Review

A more recent line of research has focused on low-density spreading codes (LDSCs), the sparse signature matrices C considered in this paper. Originally motivated by sparse code multiple access (SCMA) and code-domain non-orthogonal multiple access (NOMA), LDSCs allow many devices to share spectrum with low cross-correlation [6, 7, 8]. Recent work has shown that LDSCs are also useful as DSSS waveform spreading sequences [9, 10], providing structural diversity that classical sequences lack.

Sparse signatures for multiple access were introduced through the SCMA framework of Nikopour and Baligh [7], in which information bits are mapped to multi-dimensional sparse codewords drawn from a per-user codebook, allowing message-passing detection at the receiver with reduced complexity relative to dense codes. Codebook design was subsequently formulated as an optimization problem by Taherzadeh et al. [6], who maximized the minimum Euclidean distance over rotated constellations, and by Cai et al. [11], who refined the design via constellation rotation and interleaving. Chen and Chen [8] addressed the design of near-optimal SCMA codebooks under shaping constraints, and Yu et al. [2] surveyed SCMA’s role in upcoming 6G wireless systems.

A complementary line of work, our prior contribution [12], formalized LDSC design under both a minimum-distance criterion and a Gaussian separability margin, and introduced an iterative TSC- (and SINR-) minimization algorithm for joint code construction. The principal distinction between SCMA and the present setting lies in what the sparse structure encodes. In SCMA, the sparse codebook is the modulation:

M. Kulhandjian is with the Department of Electrical and Computer Engineering, Rice University, Houston, TX 77005, USA (e-mail: [email protected]).

H. Kulhandjian is with the Department of Electrical and Computer Engineering, California State University, Fresno, CA 93740, USA.

D. J. Jakubisin is with the Hume Center for National Security and Technology, Virginia Tech, Blacksburg, VA 24061, USA (e-mail: [email protected]).

T. A. Tsiftsis is with the Department of Informatics and Telecommunications, University of Thessaly, Lamia 35100, Greece (e-mail: [email protected]).

A.-A. A. Boulogeorgos is with the Department of Electrical and Computer Engineering, Democritus University of Thrace, Greece (e-mail: [email protected]). information bits directly select a multi-dimensional sparse codeword, so constellation shaping and sparsity are jointly designed. In our setting, symbols are first modulated (QPSK) and then spread by a fixed sparse signature matrix C; the sparsity governs the spreading and is designed for diversity and anti-jamming rather than for codebook distance. The two frameworks share column-sparsity for tractable multi-stream detection, but differ in whether the sparse object carries the message, as in SCMA, or disperses an already-modulated symbol, as here, and hence in their design metrics and operating constraints. Liu and Yang [13] analyzed the sparse-vs-dense trade-off in code-domain NOMA and concluded that the optimal density depends on the receiver complexity budget. The present work extends this view by showing that under hostile-jamming and high-mobility threats, the optimal density additionally depends on the dominant impairment.

LDSC has also been investigated specifically for multiple-access applications. Chen et al. [9] proposed sparse-code spreading-aided multi-carrier differential chaos shift keying for multiuser transmissions, while Zheng, Wu, and Zhang [10] combined polar coding with sparse spreading for massive unsourced random access. Liu et al. [14] extended sparse code division multiple access using space-time-coded generalized spatial modulation, and Islam et al. [15] surveyed power-domain NOMA in 5G. Most closely related to the generative half of our framework, Duan et al. [16] proposed a GAN-based generator (SCMA-TPGAN) for sparse codebook design in UAV systems, targeting NOMA capacity rather than anti-jamming robustness.

The waveform-design problem becomes substantially harder when one targets explicit anti-jamming performance. Conventional design metrics, such as total squared correlation (TSC), maximum cross-correlation, processing gain, and peak-to-average ratio, characterize benign-channel behavior but do not directly capture vulnerability to specific jammer types. Performance studies on chaotic and pseudo-random spreading sequences have shown that even small structural changes can significantly affect jamming behavior [17, 18, 19].

The DSSS literature classically employs Walsh–Hadamard, Gold, Kasami, and OVSF sequences for their well-characterized auto- and cross-correlation properties [4], [5]. These fixed sequences, however, are optimized for benign channels and offer limited robustness against deliberate jamming or strong non-stationary interference [3]. To improve anti-jamming behavior, Rovatti et al. [17] and Gong and Shaoqian [18] investigated chaotic spreading sequences and demonstrated that suitably chosen chaotic codes can match or exceed the performance of pseudo-random codes against random multiple-access interference. Guo and Verdú [19] provided an asymptotic analysis of randomly spread CDMA via methods from statistical physics, deriving large-system performance bounds. Somayazulu [20] compared time-hopping versus direct-sequence spreading in UWB systems. Despite their breadth, these works do not address the systematic optimization of a sparse spreading matrix against jammer-specific impairments, nor do they characterize the adverse cross-talk regime that emerges at intermediate column densities, both of which are central in this paper.

Machine learning (ML) has made significant inroads into physical-layer wireless research [21], [22], including channel estimation, decoding, beamforming, and waveform learning. Closer to our setting, the ML-guided LDSC framework of Yang et al. [23] introduced a feature-extraction layer mapping sparse code matrices to a small set of scalar metrics (density, clustering coefficient, auto-correlation, spectral density variance), trained a neural-network regression model to predict weighted BER from these features, and used the trained predictor as the fitness function inside a genetic algorithm (GA) over the LDSC search space. The GA’s mutation mechanisms operated at the row and element level on Alamouti-stacked matrices.

The NB-IoT physical layer was originally specified for benign LPWA scenarios and is, therefore, not natively jamming-tolerant. Cheng et al. [1] introduced an NB-IoT resilience framework employing both DSSS and frequency-hopping spread spectrum (FHSS) on top of a 3GPP-compliant protocol stack, achieving substantial robustness improvements under narrowband interference. The 3GPP technical specifications for E-UTRA radio transmission [24], [25] provide the LTE channel and protocol baseline that our simulator inherits. Yang et al. [23] extended this framework with the ML-LDSC design methodology described above. Our work builds on both [1] and [23]: we adopt the same NB-IoT DSSS waveform structure and use a closely matched simulator (§ 2), but generalize the LDSC design space, replace the predictive-ML pipeline with a generative one, derive analytical BER expressions where none existed, and characterize design trade-offs across an explicit threat matrix (pulse jamming, multi-tone jamming, Doppler).

Contributions

While the existing literature establishes the potential of sparse spreading for resilient wireless communication, an important question remains unresolved: what is the fundamental design variable that determines the anti-jamming performance of a sparse spreading code? Prior LDSC studies primarily treat the spreading-code coefficients and their structural arrangements as the search space. In contrast, this paper is motivated by the hypothesis that the column density itself is a dominant determinant of performance, and that code coefficients and support arrangements should be considered within an appropriately selected density regime.

Based on this density-first perspective, we broaden the design space along the column-density axis, generalize beyond the Alamouti family to the Free family and to deterministic algebraic constructions, contrast predictive and generative machine learning for code synthesis, and establish an analytical foundation for the observed trends. These directions lead to the following contributions.

Accordingly, this paper makes four main contributions:

  1. Density-axis characterization and identification of the cross-talk regime. We systematically characterize the column-density dimension of the sparse spreading-code design space and identify an adverse cross-talk regime around the density corresponding to the number of chips

per subcarrier (M≈L/KM \approx L/K). This result places the prior-art LDSC-M30M30 construction within the identified unfavorable regime and establishes column density as a first-order design parameter for anti-jamming performance.

  1. Generalization beyond the Alamouti construction. We introduce the Free family of sparse spreading codes with disjoint per-column supports and arbitrary density and evaluate a range of structural and deterministic algebraic constructions. The results identify Lempel–Costas as a strong deterministic alternative while revealing the spectral-resonance vulnerability of the Paley quadratic-residue construction under periodic pulse jamming.

  1. Comparison of predictive and generative machine learning. We demonstrate that predictive ML models based on aggregate code features fail to reliably predict per-matrix BER in the considered finite-sample regime, whereas a linear variational autoencoder can generate sparse spreading codes whose periodic-pulse-jamming BER (0.011) is the lowest among all evaluated designs (Table III). This establishes a distinction between predictive and generative ML approaches for sparse spreading-code synthesis.

CodePulse 10k JSR=10 dBMulti-tone JSR=10 dBDoppler 200 Hz SNR=10 dBDoppler 500 Hz SNR=10 dB
Walsh0.330.290.400.51
LDSC-M30 [23]0.120.190.110.33
LDSC-Free uniform M=2M = 20.0180.210.120.31
Lempel–Costas M=2M = 20.0200.290.110.31
VAE-best M=2M = 20.0110.250.120.30
Free-overlap M=180M = 180 (dense)0.240.280.0890.31

Table III. Recommended-code BER summary (lower is better).

  1. Analytical BER characterization. We derive and validate closed-form BER expressions for AWGN, flat Rayleigh fading, and sparse pulse-jamming conditions and develop an effective processing-gain interpretation (Geff=L/(2K)G_{\rm eff}=L/(2K)) that accounts for the composite-power SNR normalization and pilot-aided zero-forcing receiver. The resulting analysis agrees closely with simulation across the considered operating range.

Together, these contributions lead to a simple and actionable design principle: the spreading-code density should be selected according to the dominant threat environment, the adverse intermediate-density regime should be avoided, and support structure should subsequently be optimized within the selected density regime. For the considered NB-IoT configuration, the recommended sparse design uses uniformly spaced, disjoint supports and achieves substantially lower pulse-jamming BER than the prior LDSC-M30M30 baseline while maintaining competitive performance under Doppler.

A preliminary version of this work was presented in part at the IEEE Military Communications Conference [26]. The present paper substantially extends that work with complete derivations and proofs, the generalized Free family and the structural and algebraic construction study, the comparison of predictive and generative machine learning, the Doppler-robustness and implementation-complexity analyses, and expanded analytical validation.

Paper Organization

The remainder of the paper is organized as follows. Section II presents the system model; Section III the spreading-code design framework; Section IV the closed-form BER analysis; Section V the numerical results; and Section VI concludes.

Notation

Table I summarizes the notation used throughout the paper.

SymbolDescription
LLPer-symbol spreading length (chips per data symbol), set by the NB-IoT narrowband bandwidth constraint
KKNumber of NB-IoT subcarriers, equal to the number of columns of C\mathbf{C}
NsN_sNumber of data symbols per subframe
MMColumn density (non-zero elements per column ck\mathbf{c}_k); the principal design variable, 2≤M≤L2 \le M \le L
ck∈CL\mathbf{c}_k \in\mathbb{C}^{L}kk-th column of C\mathbf{C}, unit-norm (∥ck∥2=1\lVert\mathbf{c}_k\rVert_2=1)
ak,n∈Xa_{k,n} \in\mathcal{X}Modulated symbol on subcarrier kk at symbol index nn; X\mathcal{X} is QPSK constellation
γ\gammaLinear chip-level signal-to-noise ratio (SNR)
JSRLinear jamming-to-signal ratio

Table I. Summary of notation.

System Model

We adopt the NB-IoT DSSS architecture of Cheng et al. [1]. After standard NB-IoT downlink-shared-channel encoding (NDLSCH), narrow-band physical-DL-SCH modulation (NPDSCH), and resource-element mapping, the resource grid is multiplied chip-by-chip by a spreading matrix C∈CL×K\mathbf{C}\in\mathbb{C}^{L\times K} in which each column ck\mathbf{c}_k carries the spreading sequence assigned to subcarrier kk. The chip-spread waveform is transmitted over the wireless channel, optionally subjected to deliberate jamming, despread at the receiver, and demodulated. We consider a downlink scenario.

Transmit Signal Model

For symbol index n∈{1,…,Ns}n \in\{1,\ldots,N_s\} and subcarrier k∈{1,…,K}k \in\{1,\ldots,K\}, the transmitter modulates a QPSK symbol ak,na_{k,n}1 and spreads it to a length-LL chip vector ckak,n\mathbf{c}_k a_{k,n}. The composite chip stream for symbol nn is the per-subcarrier sum and can be expressed as

xn=∑k=1Kckak,n∈CL.(1)\mathbf{x}_n = \sum_{k=1}^{K} \mathbf{c}_k a_{k,n} \in\mathbb{C}^{L}. \tag*{(1)}

The full subframe waveform is the concatenation x=[x1T ⋯ xNsT]T∈CLNs\mathbf{x}=[\mathbf{x}_1^{\mathsf T}\ \cdots\ \mathbf{x}_{N_s}^{\mathsf T}]^{\mathsf T}\in\mathbb{C}^{LN_s}. With unit-norm columns and unit-energy QPSK symbols, the average per-chip transmit power is Px=K/LP_x=K/L, reflecting KK unit-energy streams sharing LL chip positions per symbol; this normalization is re-derived in the BER analysis of Section IV.

Channel Models

We consider three channel scenarios:

  1. AWGN: The received chip stream is

y=x+n,n∼CN(0,σ2I),(2)\mathbf{y}=\mathbf{x}+\mathbf{n},\qquad\mathbf{n}\sim\mathcal{CN}(\mathbf{0},\sigma^2\mathbf{I}), \tag*{(2)}

where σ2=Px/γ\sigma^2=P_x/\gamma is the per-chip noise variance calibrated to the transmit signal power and γ=10SNR/10\gamma=10^{\mathrm{SNR}/10} is the linear SNR. This is used to validate baseline processing-gain behavior.

  1. Flat Rayleigh Fading with AWGN: For analytical comparison, we additionally consider one independent unit-variance complex Gaussian fade per subframe. In this case, the received chip stream can be expressed as

y=hx+n,(3)\mathbf{y}=h\mathbf{x}+\mathbf{n}, \tag*{(3)}

with h∼CN(0,1)h\sim\mathcal{CN}(0,1) and n∼CN(0,σ2I)\mathbf{n}\sim\mathcal{CN}(0,\sigma^{2}\mathbf{I}). The noise variance σ2\sigma^{2} is held fixed (independent of hh), so that fading attenuates the received SNR.

  1. ETU Multipath with Doppler: For realistic propagation we use the 3GPP Extended Typical Urban (ETU) tapped-delay-line model specified in [24, 25]. The ETU profile has nine paths with delay set D\mathcal{D} and relative-power profile P\mathcal{P}. Doppler is modeled per the Jakes spectrum at configurable maximum Doppler frequency fd∈{5,70,200,500}f_d\in\{5,70,200,500\} Hz [24].

Jammer Models

Two deliberate-jamming models are superimposed on the received signal:

  1. Periodic Pulse Jammer: A pulse jammer of frequency fpf_p and duty cycle ρ∈(0,1]\rho\in(0,1] adds a chip-level interference signal jp\mathbf{j}_p that is non-zero during the ON-portion of each period. The pulse can be written as

jp(t)=1ON(t)PJρ w(t),(4)\mathbf{j}_p(t)=\mathbf{1}_{\mathrm{ON}}(t)\sqrt{\frac{P_J}{\rho}}\,w(t), \tag*{(4)}

where 1ON(t)\mathbf{1}_{\mathrm{ON}}(t) is the ON-state indicator with duty cycle ρ\rho and period 1/fp1/f_p, w(t)∼CN(0,1)w(t)\sim\mathcal{CN}(0,1) is independent and identically distributed (i.i.d.) complex Gaussian, and the average jammer power satisfies E[∣jp∣2]=PJ=JSR Px\mathbb{E}[|j_p|^2]=P_J=\mathrm{JSR}\,P_x. The factor 1/ρ1/\rho ensures that the time-averaged jammer power equals PJP_J regardless of duty cycle.

  1. Multi-Tone Jammer: A multi-tone jammer with KtK_t tones at frequencies {ωt}t=1Kt\{\omega_t\}_{t=1}^{K_t} adds

jmt(τ)=PJKt∑t=1Ktexp⁡(i(ωtτ+ϕt)),(5)j_{\mathrm{mt}}(\tau)=\sqrt{\frac{P_J}{K_t}}\sum_{t=1}^{K_t}\exp\left(\mathrm{i}(\omega_t\tau+\phi_t)\right), \tag*{(5)}

where each tone has equal power PJ/KtP_J/K_t, the frequencies ωt\omega_t are drawn uniformly in the signal band, and ϕt∼U[0,2π)\phi_t\sim\mathcal{U}[0,2\pi) are i.i.d. uniform phases.

Receiver Model

The receiver uses a coherent matched-filter despreader followed by pilot-aided zero-forcing (ZF) channel-equalization. For symbol index nn and subcarrier kk, the despread sample can be obtained as

y^k,n=ckHyn=∥ck∥2⟨h⟩nak,n+∑k′≠k⟨ck,ck′⟩⟨h⟩nak′,n+n~k,n+j~k,n,(6)\begin{aligned} \hat{y}_{k,n} &=\mathbf{c}_k^{H}\mathbf{y}_n\\ &=\|\mathbf{c}_k\|^2\langle h\rangle_n a_{k,n} +\sum_{k'\ne k}\langle\mathbf{c}_k,\mathbf{c}_{k'}\rangle\langle h\rangle_n a_{k',n}\\ &\quad+\tilde{n}_{k,n}+\tilde{j}_{k,n}, \tag*{(6)} \end{aligned}

where ⟨h⟩n\langle h\rangle_n denotes the effective channel coefficient for symbol nn after spreading-bandwidth filtering, ⟨ck,ck′⟩=ckHck′\langle\mathbf{c}_k,\mathbf{c}_{k'}\rangle=\mathbf{c}_k^{H}\mathbf{c}_{k'} is the cross-correlation between columns kk and k′k', and n~k,n\tilde{n}_{k,n} and j~k,n\tilde{j}_{k,n} are the despread-domain noise and jamming components, respectively. For unit-norm columns and disjoint per-column supports ⟨ck,ck′⟩=0\langle\mathbf{c}_k,\mathbf{c}_{k'}\rangle=0 for k≠k′k\ne k', eliminating the cross-talk term; for overlapping supports the cross-talk term is non-zero and is one of the mechanisms underlying the M≈L/KM\approx L/K bad-zone.

A pilot symbol p0p_0 is placed at ZZ locations defined by the set of indices (k,n)∈Sp(k,n)\in\mathcal{S}_p where Z=∣Sp∣Z=|\mathcal{S}_p| to support channel estimation. The pilot-derived channel estimate is

h^=1Z∑(k,n)∈Spy^k,np0,(7)\hat{h}=\frac{1}{Z}\sum_{(k,n)\in\mathcal{S}_p}\frac{\hat{y}_{k,n}}{p_0}, \tag*{(7)}

and ZF equalization divides every despread cell by h^\hat{h}, as

y~k,n=y^k,nh^with k=1,…,K, n=1,…,Ns.(8)\tilde{y}_{k,n}=\frac{\hat{y}_{k,n}}{\hat{h}}\quad\text{with }k=1,\ldots,K,\ n=1,\ldots,N_s. \tag*{(8)}

The pilots are excluded from BER computation. The equalized samples y~k,n\tilde{y}_{k,n} are passed through a Gray-coded QPSK demodulator to recover the bit estimates b^\hat{b}.

Performance Metrics

We use three operational performance metrics.

  1. BER: The fraction of incorrectly demodulated bits, computed at a specific operating point (SNR, JSR, fdf_d, jammer type). Reported with 95% Wilson-score confidence intervals where applicable.

  1. Anti-Jamming Margin (AJM): For a given jammer scenario, the AJM is defined as the JSR at which the BER first crosses a threshold BERth\mathrm{BER}_{\mathrm{th}}:

AJM(BERth)=arg⁡min⁡JSRdB{BER(JSRdB)≥BERth}.(9)\mathrm{AJM}(\mathrm{BER}_{\mathrm{th}})=\underset{\mathrm{JSR}_{\mathrm{dB}}}{\arg\min}\left\{\mathrm{BER}(\mathrm{JSR}_{\mathrm{dB}})\ge\mathrm{BER}_{\mathrm{th}}\right\}. \tag*{(9)}

Higher AJM indicates a more jammer-resilient code.

  1. Weighted BER: We compute the weighted BER over a set of SNR points S={SNRi}\mathcal{S}=\{\mathrm{SNR}_i\} as

wBER=∑i∈Swi BER(SNRi)with wi=10SNRi/10.(10)\mathrm{wBER}=\sum_{i\in\mathcal{S}}w_i\,\mathrm{BER}(\mathrm{SNR}_i)\quad\text{with }w_i=10^{\mathrm{SNR}_i/10}. \tag*{(10)}

This weighting is proportional to the linear SNR and emphasizes high-SNR performance.

Spreading Code Design Framework

We organize the spreading-code design space along two orthogonal axes: the per-column density MM, which stands for the number of non-zero entries per column ck\mathbf{c}_k, and the structural placement of the non-zero positions within the LL-chip support (the secondary axis). The section first reviews the prior-art Alamouti-paired construction of [23], provides our generalized disjoint-support Free family, specific structural variants, as well as deterministic algebraic constructions, and reports both predictive and generative machine-learning approaches. The section closes with a unified set of code-characterization metrics that we use throughout the rest of the paper.

Algorithm 1  Free LDSC Generator

Input: $L$, $K$, density vector $\{M_k\}_{k=1}^{K}$ with $\sum_k M_k \le L$

Output: $\mathbf{C} \in \mathbb{C}^{L \times K}$ with disjoint supports, unit-norm columns

1: Draw a random permutation $\pi : \{1,\ldots,L\} \to \{1,\ldots,L\}$.
2: $\mathbf{C} \leftarrow \mathbf{0}_{L \times K}$; $j \leftarrow 1$.
3: for $k=1$ to $K$ do
4: Assign support $\mathcal{I}_k \leftarrow \{\pi(j),\pi(j+1),\ldots,\pi(j+M_k-1)\}$.
5: Draw values $\{c_{k,i}\}_{i\in\mathcal{I}_k}$ i.i.d. from $\mathcal{CN}(0,1)$.
6: Normalize: $c_{k,i} \leftarrow c_{k,i}/\sqrt{\sum_{i'\in\mathcal{I}_k}|c_{k,i'}|^2}$.
7: $j \leftarrow j+M_k$.
8: end for
9: return $\mathbf{C}$.

Alamouti-Stacked LDSC

The construction of [23] starts from a unit-energy Alamouti seed [27], i.e.,

A(c1,c2)=(c1c2c2∗−c1∗)with∣c1∣2+∣c2∣2=1,(11)\mathbf{A}(c_1,c_2)=\begin{pmatrix}c_1 & c_2 \\ c_2^* & -c_1^*\end{pmatrix} \quad\text{with}\quad|c_1|^2+|c_2|^2=1, \tag*{(11)}

where c1,c2c_1,c_2 are i.i.d. complex Gaussian and renormalized. A K×KK \times K block-diagonal matrix C′∈CK×K\mathbf{C}' \in\mathbb{C}^{K \times K} is built from K/2K/2 Alamouti blocks. To construct the full LDSC of dimension L×KL \times K at column density MM (with MM even), the seed matrix is stacked M/2M/2 times vertically, zero-padded to length LL, and row-permuted with a random index permutation π:{1,…,L}→{1,…,L}\pi:\{1,\ldots,L\}\to\{1,\ldots,L\}:

C=π([C′T⋯C′T0T]T).(12)\mathbf{C}=\pi\left(\begin{bmatrix}\mathbf{C}'^{\mathsf{T}}&\cdots&\mathbf{C}'^{\mathsf{T}}&\mathbf{0}^{\mathsf{T}}\end{bmatrix}^{\mathsf{T}}\right). \tag*{(12)}

Each column of the resulting matrix has exactly MM non-zero entries; columns from the same 2×22\times2 Alamouti block (i.e., consecutive column pairs {2j−1,2j}\{2j-1,2j\}) share the same non-zero row positions. This pairing is structurally enforced and constrains the search space to a sub-manifold of the larger sparse-code family. Following [23], we label this baseline by its stacking count as LDSC-M30 (30 stacked Alamouti blocks, hence 60 non-zeros per column, i.e., M=60M=60 in the per-column-density convention used throughout).

Free Family: Disjoint Supports

To remove the column-pairing constraint, we introduce the Free family of LDSC constructions in Algorithm 1 in which each column ck\mathbf{c}_k is supported on its own (column-specific) subset Ik⊆{1,…,L}\mathcal{I}_k \subseteq\{1,\ldots,L\} of ∣Ik∣=Mk|\mathcal{I}_k|=M_k row positions, and the supports are mutually disjoint: Ik∩Ik′=∅\mathcal{I}_k\cap\mathcal{I}_{k'}=\varnothing for k≠k′k\ne k'. Disjoint supports are feasible whenever ∑k=1KMk≤L\sum_{k=1}^{K}M_k\le L, with the maximum uniform density Mmax⁡=⌊L/K⌋M_{\max}=\lfloor L/K\rfloor.

The Free family eliminates inter-column cross-talk by construction: ⟨ck,ck′⟩=0\langle\mathbf{c}_k,\mathbf{c}_{k'}\rangle=0 for any k≠k′k\ne k' when supports are disjoint, regardless of the values placed on the support. We instantiate two concrete sub-families:

  • Free-Uniform-M: Mk=MM_k=M for all kk, fixed.

  • Free-Variable-M: Mk∼U{Mmin⁡,…,Mmax⁡}M_k\sim\mathcal{U}\{M_{\min},\ldots,M_{\max}\} drawn independently per column.

For the dense regime M>L/KM>L/K, where disjoint supports are infeasible, we adopt the Free-Overlap variant in which each column independently selects MM random positions from {1,…,L}\{1,\ldots,L\} without disjoint-support enforcement; the resulting columns are not mutually orthogonal, and ⟨ck,ck′⟩\langle\mathbf{c}_k,\mathbf{c}_{k'}\rangle scales as O(M/L)\mathcal{O}(M/L).

Structural Placement Variants

Within a chosen density MM, the placement of non-zero positions can be random or follow a deterministic structural rule. We evaluate four structures, each replacing line 1 of Algorithm 1 with a structure-specific position generator:

  1. Uniform spacing: The KMKM positions are placed at evenly-spaced locations {i⌊L/(KM)⌋}i=0KM−1\{i\lfloor L/(KM)\rfloor\}_{i=0}^{KM-1} (rounded to nearest integer). After deduplication and shuffling, the positions are partitioned into KK disjoint MM-subsets, one per column.

  1. Golden-ratio spacing: The positions are generated via pi=⌊iφL/(KM)⌋ mod Lp_i=\lfloor i\varphi L/(KM)\rfloor\bmod L, where φ=(1+5)/2\varphi=(1+\sqrt{5})/2 is the golden ratio. This yields a quasi-uniform distribution that avoids exact periodic structure.

  1. Chirp spacing: Positions follow a quadratic law pi=αi2 mod Lp_i=\alpha i^2\bmod L with α=⌊L/(KM)⌋\alpha=\lfloor L/(KM)\rfloor. This is the LFM-radar-inspired structure with a wider band ambiguity behavior.

  1. mm-sequence thinning: A maximal-length pseudorandom binary sequence of length 2n−1≥L2^n-1\ge L is generated via an nn-stage linear-feedback shift register (LFSR), truncated to length LL, randomly phase-shifted, and the positions of the non-zero entries are randomly subsampled to obtain the KMKM targeted positions.

In every variant, after position assignment the values follow line 5 of Algorithm 1: i.i.d. complex-Gaussian, unit-normalized per column. Thus structure modifies where non-zeros sit, not what value they take.

Algebraic Constructions

We evaluate two deterministic constructions inspired by classical combinatorial-design theory.

  1. Paley Quadratic Residue (Paley-QR): For prime p≤Lp\le L with p≡3(mod4)p\equiv3\pmod4, the set of quadratic residues Qp={i2 mod p:1≤i≤(p−1)/2}\mathcal{Q}_p=\{i^2\bmod p:1\le i\le(p-1)/2\} is a (p,(p−1)/2,(p−3)/4)(p,(p-1)/2,(p-3)/4) cyclic difference set [28]. We use p=359p=359 (the largest prime ≤L=360\le L=360), which gives ∣Qp∣=179|\mathcal{Q}_p|=179 candidate positions, ample for KM≤60KM\le60 disjoint allocations. The KMKM positions are drawn disjointly from Qp\mathcal{Q}_p and partitioned across columns.

  1. Lempel–Costas: For prime pp and a primitive root α\alpha of pp, the Lempel construction defines positions pi=(αi+σ−1) mod pp_i=(\alpha^{i+\sigma}-1)\bmod p for i=1,…,KMi=1,\ldots,KM, where σ\sigma is a random starting offset. Removing the zero entry and uniqueness-deduplicating yields KMKM distinct positions that are partitioned across columns. The Lempel construction is closely related to Welch–Costas [29] (which uses pi=αi mod pp_i=\alpha^i\bmod p) but exhibits different per-symbol autocorrelation.

Machine Learning Approaches

  1. Predictive ML: We first replicate the predictive-ML pipeline of [23]: a feature extraction layer mapping each

candidate code CC to a feature vector ϕ(C)\phi(C) of length either 4 (the paper’s selected subset: density, clustering coefficient, autocorrelation, power spectral density variance) or 15 (the full set including total squared correlation, maximum cross-correlation, processing gain, peak-to-average ratio, etc.); a feed-forward neural network with hidden-layer widths {64,32,16,8,4,1}\{64, 32, 16, 8, 4, 1\} and ReLU activations; trained on N=210N = 210 codes generated from Algorithm 1 of [23] at stacking multipliers μ∈{1,…,30}\mu\in\{1, \ldots, 30\} (i.e., 2μ2\mu non-zeros per column) with 7 random row-permutations each, with weighted BER as the target. The training-test split is 70/30.

We additionally evaluate a random forest [30] regressor with 200 trees as an alternative non-parametric predictor. We document that both predictors achieve R2<0R^2 < 0 on the held-out test set, worse than predicting the mean. The cause is that per-matrix wBER variance from finite-sample Monte Carlo simulation dominates the aggregate-feature signal at the available sample size; the predictor cannot extract a generalizing signal at N=210N = 210 regardless of feature engineering.

  1. Generative ML: Linear VAE: In the same sample regime, we replace the predictive pipeline with a generative one. We train a linear variational autoencoder (VAE) [31] (a principal-component-analysis-based VAE with a Gaussian latent prior) on the same 210-matrix dataset. In contrast to generative adversarial networks [32] and sparse-coding VAE variants such as SC-VAE [33], this generator targets matrix-valued spreading signatures with hard column-support sparsity under a communication-theoretic objective. The encoder maps each flattened real-imaginary code vector R2LK\mathbb{R}^{2LK} to a Gaussian latent N(μz,Σz)\mathcal{N}(\mu_z, \Sigma_z) in Rdz\mathbb{R}^{d_z} with dz=32d_z = 32; the decoder is the corresponding linear inverse. After training, we sample NgN_g latent vectors from the learned Gaussian prior, decode them back to the code-vector space, and sparsify each candidate to the target density MM by retaining the top-MM entries (in magnitude) per column and zeroing the rest. Each sparsified candidate is re-normalized to unit column norm.

The sampled candidates are pre-screened by a composite quality score (combining merit factor, spectral flatness, and clustering coefficient) and the top candidates are evaluated via direct BER simulation. The best generated code is denoted CVAE_{\mathrm{VAE}}.

This pipeline succeeds where the predictive pipeline failed because the generative model learns the structural manifold of valid sparse codes (a low-dimensional subspace of R2LK\mathbb{R}^{2LK} characterizing typical sparse-orthogonal patterns) rather than attempting to predict a noise-dominated scalar target.

Code Characterization Metrics

We use the following metrics to characterize sparse codes; we partition them into design metrics (used as optimization or selection criteria) and diagnostic metrics (used to report code properties but not to drive design).

Density and Sparsity: Per-column density MkM_k and average M=Ek[Mk]M = \mathbb{E}_k[M_k]; normalized density DENS =M/L= M/L; Hoyer sparseness measure [34] S1,2(c)=(L−∥c∥1/∥c∥2)/(L−1)S_{1,2}(\mathbf{c}) = (\sqrt{L}-\lVert\mathbf{c}\rVert_1/\lVert\mathbf{c}\rVert_2)/(\sqrt{L}-1), ranging from 0 (uniform) to 1 (single-spike).

Cross-Correlation: Total squared cross-correlation TSC =∑i≠j∣ciHcj∣2= \sum_{i\ne j} \lvert\mathbf{c}_i^{H}\mathbf{c}_j\rvert^2 [12]; RMS cross-correlation Irms=TSC/(K(K−1))I_{\mathrm{rms}} = \sqrt{\mathrm{TSC}/(K(K-1))}; maximum cross-correlation Imax⁡=max⁡i<j∣ciHcj∣I_{\max} = \max_{i<j}\lvert\mathbf{c}_i^{H}\mathbf{c}_j\rvert (also called coherence); Welch-bound [35] ratio Irms/WI_{\mathrm{rms}}/W with W=(K−L)/(K−1)LW = \sqrt{(K-L)/(K-1)L} for L<KL<K, else trivially zero; Gaussian separability margin Δavg\Delta_{\mathrm{avg}} from [12].

Spectral and Spatial Distribution: Spectral flatness measure (SFM) =exp⁡(E[log⁡P(ω)])/E[P(ω)]= \exp(\mathbb{E}[\log P(\omega)])/\mathbb{E}[P(\omega)]; spectral entropy H=−∑kpklog⁡pkH = -\sum_k p_k \log p_k on normalized PSD; merit factor F=L2/(2∑k≠0∣ak∣2)F = L^2/(2\sum_{k\ne0}\lvert a_k\rvert^2) based on aperiodic autocorrelation aka_k [36]; peak-to-sidelobe ratio (PSL) in dB.

Sparse-Pattern Statistics: Clustering coefficient (CLUST), gap variance (GapVar == variance of zero-run lengths between non-zeros), maximum gap (MaxGap), nearest-neighbor distance (NND), minimum pairwise support distance min⁡i<j∣Ii△Ij∣\min_{i<j}\lvert\mathcal{I}_i \mathbin{\triangle} \mathcal{I}_j\rvert (symmetric set difference).

Operational Metric: Anti-jamming margin AJM(BERth_{\mathrm{th}}) defined in (9); a single deployment-relevant number per (code, jammer) pair.

In § 8 we identify which subset of these metrics actually drive performance differences in our regime, a key practical contribution, since the prior art [23] relied on density and clustering coefficient alone.

BER Analysis

This section derives closed-form BER expressions for the three operating regimes. A key contribution is the corrective effective processing gain

Geff≜L2K,(13)G_{\mathrm{eff}} \triangleq\frac{L}{2K}, \tag*{(13)}

which replaces the classical single-stream matched-filter gain LL when both (i) the receiver operates in a KK-stream composite signal regime with measured-power-calibrated additive noise, and (ii) the channel-equalization stage is pilot-aided zero-forcing. The factor decomposes as L/K×1/2L/K \times1/2: the first factor converts the composite-power SNR reference into the per-stream reference used at the detector (Section IV-A), and the second factor captures the 3 dB noise penalty incurred by ZF division by a noisy pilot estimate. We discuss each factor in turn before stating the closed-form BER expressions.

SNR Reference and Composite-Power Normalization

Because the effective gain in (13) depends entirely on how transmit power and SNR are referenced, we fix these conventions before deriving it. Two distinct SNR references arise in this system:

  • Composite per-chip SNR γ≜Px/σ2=10SNR/10\gamma\triangleq P_x/\sigma^2 = 10^{\mathrm{SNR}/10}, referenced to the aggregate per-chip power Px=K/LP_x = K/L of the superimposed KK-stream waveform (14). This is the quantity fixed by the simulator’s measured-power noise calibration (15), and it is therefore the SNR reported on the horizontal axis of every BER figure.

  • Per-stream SNR γstream≜Px,k/σ2≜γ/K\gamma_{\mathrm{stream}} \triangleq P_{x,k}/\sigma^2 \triangleq\gamma/K, referenced to the power Px,k=∥ck∥2/L=1/LP_{x,k} = \lVert\mathbf{c}_k\rVert^2/L = 1/L carried by a single

column, the power that actually conveys the information symbol at the detector.

The two references differ by exactly the stream count, γstream=γ/K\gamma_{\mathrm{stream}}=\gamma/K, because the composite reference charges each stream with the power of all KK streams while detection is performed one stream at a time. We emphasize that this factor is a power-normalization convention, not a physical multi-stream penalty: the disjoint-support construction makes the inter-stream interference identically zero (the cross-correlation terms vanish in (16)), so no signal energy is lost to aggregation. Accordingly, all BER expressions below are stated per information symbol, per stream, and plotted against the composite reference γ\gamma.

The per-chip transmit power of the composite stream xn=∑k=1Kckak,n\mathbf{x}_n=\sum_{k=1}^{K}\mathbf{c}_k a_{k,n} (1) follows from unit-norm columns (∥ck∥2=1\lVert\mathbf{c}_k\rVert^2=1) and unit-energy QPSK symbols (E∣ak,n∣2=1\mathbb{E}|a_{k,n}|^2=1):

Px=1LE[∑i=1L∣xi,n∣2]=1LE[∑i=1L∣∑k=1Kck,iak,n∣2]=1L∑k=1K∑i=1L∣ck,i∣2E∣ak,n∣2=KL,(14)\begin{aligned} P_x=\frac{1}{L}\mathbb{E}\left[\sum_{i=1}^{L}|x_{i,n}|^2\right]=\frac{1}{L}\mathbb{E}\left[\sum_{i=1}^{L}\left|\sum_{k=1}^{K}c_{k,i}a_{k,n}\right|^2\right] \\ &=\frac{1}{L}\sum_{k=1}^{K}\sum_{i=1}^{L}|c_{k,i}|^2\mathbb{E}|a_{k,n}|^2=\frac{K}{L}, \tag*{(14)} \end{aligned}

where step (∗)(*) uses the disjoint-support assumption (cross terms E[ak,nak′,n∗]=0\mathbb{E}[a_{k,n}a_{k',n}^{*}]=0 for k≠k′k\ne k' since data symbols are independent and zero-mean). The factor K/LK/L reflects KK unit-energy streams sharing LL chip positions per symbol; it is the composite reference power against which the noise is calibrated in (15), in contrast to the classical single-stream reference 1/L1/L.

Under the measured-power AWGN model adopted by the simulator (mirroring the practice of [1, 23]), the noise variance is set such that the per-chip SNR matches the prescribed value:

σ2≜Pxγ=KL10−SNR/10.(15)\sigma^2 \triangleq\frac{P_x}{\gamma}=\frac{K}{L}10^{-\mathrm{SNR}/10}. \tag*{(15)}

For column ck\mathbf{c}_k with disjoint support relative to all other columns, the matched-filter despread output of (6) simplifies to

y^k,n=∥ck∥2ak,n+ckHn=ak,n+n~k,n,(16)\hat{y}_{k,n}=\lVert\mathbf{c}_k\rVert^2a_{k,n}+\mathbf{c}_k^{H}\mathbf{n}=a_{k,n}+\tilde{n}_{k,n}, \tag*{(16)}

since the cross-correlation terms ⟨ck,ck′⟩=0\langle\mathbf{c}_k,\mathbf{c}_{k'}\rangle=0 vanish on disjoint supports. The despread noise is n~k,n∼CN(0,σdesp2)\tilde{n}_{k,n}\sim\mathcal{CN}(0,\sigma_{\mathrm{desp}}^2), with σdesp2=∥ck∥2σ2=σ2\sigma_{\mathrm{desp}}^2=\lVert\mathbf{c}_k\rVert^2\sigma^2=\sigma^2 (preserved by unit-norm matched-filter combining). The instantaneous despread SNR is therefore

γdesp=∣ak,n∣2σdesp2=1σ2=LK⋅10SNR/10.(17)\gamma_{\mathrm{desp}}=\frac{|a_{k,n}|^2}{\sigma_{\mathrm{desp}}^2}=\frac{1}{\sigma^2}=\frac{L}{K}\cdot10^{\mathrm{SNR}/10}. \tag*{(17)}

Thus the matched-filter despread SNR per stream is γdesp=1/σ2=(L/K)γ=Lγstream\gamma_{\mathrm{desp}}=1/\sigma^2=(L/K)\gamma=L\gamma_{\mathrm{stream}}: the classical processing gain LL [5] is recovered exactly (the special case K=1K=1, where the two references coincide), while the factor L/KL/K in (17) arises solely from re-expressing this gain against the composite reference γ=Kγstream\gamma=K\gamma_{\mathrm{stream}}. For the K=12K=12 NB-IoT setting this shifts the curve by 10log⁡10K≈1110\log_{10}K\approx11 dB relative to a single-stream plot; it is a change of SNR reference, not a loss of signal energy, since the disjoint supports render the streams orthogonal and no inter-stream interference is present.

Pilot-Aided ZF Noise Penalty

The pilot estimate h^\hat{h} from (7) is computed from the despread output at the pilot positions (k,n)∈Sp(k,n)\in\mathcal{S}_p. In the AWGN regime where the true channel coefficient is h˙=1\dot{h}=1, applying (16) at the pilot position yields

h^=1Z∑(k,n)∈Spp0+n~k,np0=1+n~pZp0.(18)\hat{h}=\frac{1}{Z}\sum_{(k,n)\in\mathcal{S}_p}\frac{p_0+\tilde{n}_{k,n}}{p_0}=1+\frac{\tilde{n}_p}{Zp_0}. \tag*{(18)}

where n~k,n∼CN(0,σ2)\tilde{n}_{k,n}\sim\mathcal{CN}(0,\sigma^2) is the despread-domain noise on the pilot and n~p=1/Z∑(k,n)∈Spn~k,n\tilde{n}_p=1/Z\sum_{(k,n)\in\mathcal{S}_p}\tilde{n}_{k,n} is the mean of the despread noise terms at each pilot location. With p0=1p_0=1, the pilot estimate is h^∼CN(1,σ2/Z)\hat{h}\sim\mathcal{CN}(1,\sigma^2/Z).

The ZF step in (8) divides every other despread cell by h^\hat{h}. Writing the ratio as a perturbation series in n~p/p0\tilde{n}_p/p_0 (valid for small noise relative to unit pilot magnitude), we obtain

1h^=11+n~p/p0≈1−n~pp0+O(∣n~p∣2/∣p0∣2).(19)\frac{1}{\hat{h}}=\frac{1}{1+\tilde{n}_p/p_0}\approx1-\frac{\tilde{n}_p}{p_0}+O(|\tilde{n}_p|^2/|p_0|^2). \tag*{(19)}

The post-ZF data sample at (k,n)∉Sp(k,n)\notin\mathcal{S}_p is therefore

y~k,n=ak,n+n~k,nh^≈(ak,n+n~k,n)(1−n~p/p0)=ak,n−ak,nn~pp0+n~k,n+O(∣n~∣2).(20)\begin{aligned} \tilde{y}_{k,n}=\frac{a_{k,n}+\tilde{n}_{k,n}}{\hat{h}} \\ &\approx(a_{k,n}+\tilde{n}_{k,n})(1-\tilde{n}_p/p_0) \\ &=a_{k,n}-a_{k,n}\frac{\tilde{n}_p}{p_0}+\tilde{n}_{k,n}+O(|\tilde{n}|^2). \tag*{(20)} \end{aligned}

The error y~k,n−ak,n\tilde{y}_{k,n}-a_{k,n} contains two zero-mean independent noise contributions: the original despread noise n~k,n\tilde{n}_{k,n}, and the pilot-induced multiplicative term −ak,nn~p/p0-a_{k,n}\tilde{n}_p/p_0. Their variances add (independence), giving

E[∣y~k,n−ak,n∣2]≈σ2+E∣ak,n∣2σ2Z=(1+1Z)σ2.(21)\mathbb{E}\left[|\tilde{y}_{k,n}-a_{k,n}|^2\right]\approx\sigma^2+\mathbb{E}|a_{k,n}|^2\frac{\sigma^2}{Z}=\left(1+\frac{1}{Z}\right)\sigma^2. \tag*{(21)}

where we used ∣p0∣2=1|p_0|^2=1 and unit-energy data symbols. When a single pilot symbol is used for channel estimation (Z=1Z=1), the post-ZF noise variance is exactly twice the despread noise variance; equivalently, the SNR is halved. The remainder of the paper assumes a single spread pilot (Z=1Z=1).

Combining this 3 dB penalty with the composite-power normalization factor of (17) gives the principal analytical result of this paper:

γeff=12σ2=L2K10SNR/10=Geffγ.(22)\gamma_{\mathrm{eff}}=\frac{1}{2\sigma^2}=\frac{L}{2K}10^{\mathrm{SNR}/10}=G_{\mathrm{eff}}\gamma. \tag*{(22)}

Theorem 1 (Effective processing gain). In the KK-stream disjoint-support sparse-code DSSS system with unit-norm columns, AWGN measured-power calibration, and a single-pilot zero-forcing receiver, the effective processing gain at the data-decision stage is

Geff=L2K.G_{\mathrm{eff}}=\frac{L}{2K}.

Proof. Combining the composite-power normalization γdesp=1/σ2\gamma_{\mathrm{desp}}=1/\sigma^2 of (17) with the pilot-aided ZF noise penalty (21) (which doubles the effective noise variance) gives γeff=1/(2σ2)=Geffγ\gamma_{\mathrm{eff}}=1/(2\sigma^2)=G_{\mathrm{eff}}\gamma with Geff=L/(2K)G_{\mathrm{eff}}=L/(2K), as in (22). □\square

This is a ∼14\sim14 dB reduction relative to the classical LL processing-gain prediction for the parameters of this paper (L=360L=360, K=12K=12, hence Geff=15G_{\mathrm{eff}}=15). (22) is conservative: as γ→∞\gamma\to\infty, the O(∣n~∣2)O(|\tilde{n}|^2) terms in (19) vanish faster than the leading n~/p0\tilde{n}/p_0 term, and γeff→(L/K)γ\gamma_{\mathrm{eff}} \to(L/K)\gamma asymptotically. The SNR-dependent transition between L/(2K)L/(2K) and L/KL/K accounts for the residual analytical-vs-simulation discrepancy at high SNR.

Closed-Form BER Expressions

AWGN

Because the ZF stage divides the despread data by the random pilot estimate h^∼CN(1,σ2)\hat{h} \sim\mathcal{CN}(1,\sigma^2) of (18), the post-correction decision statistic z=(ak,n+n~k,n)/h^z=(a_{k,n}+\tilde{n}_{k,n})/\hat{h} is only conditionally Gaussian given h^\hat{h}. The AWGN bit-error probability is therefore properly written as an expectation over the estimate,

PbAWGN(γ)=Eh^[Pb(γ∣h^)],(23)P_b^{\mathrm{AWGN}}(\gamma)=\mathbb{E}_{\hat{h}}\left[P_b(\gamma\mid\hat{h})\right], \tag*{(23)}

rather than by substituting a single effective SNR into the QPSK QQ-function. For Gray-coded QPSK the quadrant decision on zz is invariant to the estimate magnitude ∣h^∣|\hat{h}| and depends only on the pilot phase error ϕ≜arg⁡h^\phi\triangleq\arg\hat{h}; conditioning on ϕ\phi and using 1/σ2=(L/K)γ1/\sigma^2=(L/K)\gamma gives the conditional bit-error probability

Pb(γ∣ϕ)=12Q((cos⁡ϕ+sin⁡ϕ)LKγ)+12Q((cos⁡ϕ−sin⁡ϕ)LKγ),(24)P_b(\gamma\mid\phi)=\frac{1}{2}Q\left((\cos\phi+\sin\phi)\sqrt{\frac{L}{K}\gamma}\right)+\frac{1}{2}Q\left((\cos\phi-\sin\phi)\sqrt{\frac{L}{K}\gamma}\right), \tag*{(24)}

which is averaged over ϕ=arg⁡(1+n~p/p0)\phi=\arg(1+\tilde{n}_p/p_0) to evaluate (23). This makes the role of the noisy channel estimate explicit: in AWGN (h=1h=1) the pilot penalty is a phase-error effect, and it vanishes as σ2→0\sigma^2\to0 (i.e. ϕ→0\phi\to0), where (24) recovers the perfect-CSI value Q((L/K)γ)Q(\sqrt{(L/K)\gamma}).

In the high-pilot-SNR regime (σ2≪1\sigma^2\ll1), a first-order expansion of the noisy-ZF division (equivalently, absorbing the phase-error averaging into the effective 3 dB noise penalty of Section IV-B) collapses (23) to the compact approximation

PbAWGN(γ)≈Q(Geffγ)=Q(L2Kγ),(25)P_b^{\mathrm{AWGN}}(\gamma)\approx Q\left(\sqrt{G_{\mathrm{eff}}\gamma}\right)=Q\left(\sqrt{\frac{L}{2K}\gamma}\right), \tag*{(25)}

which is the baseline plotted in Figure 1a. We stress that (25) is not a universal closed form for sparse spreading: it is the small-estimation-error limit of (23), valid only under the measured-power SNR convention of Section IV-A and the single-pilot ZF receiver of Section IV-B. Under these conditions it matches simulation within 1 dB across SNR ∈[−20,−4]\in[-20,-4] dB; the ≈30%\approx30\% lower simulated BER at SNR ≥−2\ge-2 dB reflects the higher-order (phase-averaging) terms dropped in passing from (23) to (25), consistent with the L/(2K)→L/KL/(2K)\to L/K transition noted after (22).

BER versus SNR for AWGNBER versus SNR for flat RayleighBER versus JSR for periodic pulse

Figure 1. Closed-form analytical BER vs. simulation for the uniform M=2M = 2 Free design (95% Wilson-score CI): (a) AWGN; (b) flat Rayleigh in fixed-noise mode; (c) periodic pulse (fp=10f_p = 10 kHz, ρ=2%\rho= 2\%). Simulation tracks analysis within 1–2 dB, validating the Geff=L/(2K)G_{\mathrm{eff}} = L/(2K) correction.

Flat Rayleigh Fading with AWGN

Under flat fading (3) the pilot no longer estimates a known unit channel: from (7) the estimate is h^=h+n~p/p0\hat{h}=h+\tilde{n}_p/p_0 with n~p∼CN(0,σ2)\tilde{n}_p\sim\mathcal{CN}(0,\sigma^2), so the true channel hh and its estimate h^\hat{h} are jointly complex Gaussian with correlation

ρ≜E[hh^∗]E∣h∣2E∣h^∣2=11+σ2.(26)\rho\triangleq\frac{\mathbb{E}[h\hat{h}^{*}]}{\sqrt{\mathbb{E}|h|^2\mathbb{E}|\hat{h}|^2}}=\frac{1}{\sqrt{1+\sigma^2}}. \tag*{(26)}

As in the AWGN case, the ZF detector divides by the random estimate h^\hat{h}, so the BER must be averaged over the joint law of (h,h^)(h,\hat{h}), not over ∣h∣2|h|^2 alone:

PbRay(γ)=Eh,h^[Pb(γ∣h,h^)].(27)P_b^{\mathrm{Ray}}(\gamma)=\mathbb{E}_{h,\hat{h}}\left[P_b(\gamma\mid h,\hat{h})\right]. \tag*{(27)}

Conditioned on the realized channel hh and estimate h^\hat{h}, the quadrant decision on z=(hak,n+n~k,n)/h^z=(ha_{k,n}+\tilde{n}_{k,n})/\hat{h} depends on the true fading magnitude ∣h∣|h| and the estimation phase error ψ≜arg⁡h−arg⁡h^\psi\triangleq\arg h-\arg\hat{h}; using 1/σ2=(L/K)γ1/\sigma^2=(L/K)\gamma,

Pb(γ∣h,h^)=12Q(∣h∣(cos⁡ψ+sin⁡ψ)LKγ)+12Q(∣h∣(cos⁡ψ−sin⁡ψ)LKγ),(28)P_b(\gamma\mid h,\hat{h})=\frac{1}{2}Q\left(|h|(\cos\psi+\sin\psi)\sqrt{\frac{L}{K}\gamma}\right)+\frac{1}{2}Q\left(|h|(\cos\psi-\sin\psi)\sqrt{\frac{L}{K}\gamma}\right), \tag*{(28)}

the direct fading analogue of the AWGN conditional law (24), which it recovers for ∣h∣=1|h|=1.

The closed-form baseline plotted in Figure 1a follows, exactly as in the AWGN case, by replacing the conditional law (28) with a single effective per-realization SNR γinst=∣h∣2Geffγ\gamma_{\mathrm{inst}}=|h|^2G_{\mathrm{eff}}\gamma (i.e., folding the estimation phase penalty into the 3 dB factor of GeffG_{\mathrm{eff}}) and averaging Q(∣h∣2Geffγ)Q(\sqrt{|h|^2G_{\mathrm{eff}}\gamma}) over ∣h∣2∼Exp⁡(1)|h|^2\sim\operatorname{Exp}(1) [4 Eq. (3.7)]:

PbRay(γ)≈12(1−μ1+μ),μ=Geffγ2.(29)P_b^{\mathrm{Ray}}(\gamma)\approx\frac{1}{2}\left(1-\sqrt{\frac{\mu}{1+\mu}}\right),\qquad\mu=\frac{G_{\mathrm{eff}}\gamma}{2}. \tag*{(29)}

Like (25), this is an approximation of the exact expectation (27), not a general closed form: it collapses the estimation phase error of (28) into the fixed penalty GeffG_{\mathrm{eff}}. The error-free estimate limit (ρ→1,ψ→0\rho\to1,\psi\to0) of (28) instead removes the 3 dB penalty and gives μ=Lγ/(2K)\mu=L\gamma/(2K); retaining Geff=L/(2K)G_{\mathrm{eff}}=L/(2K) (hence μ=Lγ/(4K)\mu=L\gamma/(4K)) is the effective-penalty form the noisy single-pilot receiver is observed to follow across the operating range. At high SNR (μ≫1\mu\gg1), (29) yields PbRay≈1/(4μ)=K/(Lγ)P_b^{\mathrm{Ray}}\approx1/(4\mu)=K/(L\gamma), the classical first-order Rayleigh diversity slope.

Deep-fade regime. The approximation (29) is weakest when the fade is comparable to the noise, ∣h∣≲σ|h|\lesssim\sigma: the pilot estimate is then noise-dominated, the ZF ratio h/h^h/\hat{h} becomes heavy-tailed (a quotient of complex Gaussians, with no finite second moment) rather than a fixed 3 dB noise scaling, and the phase error drives Pb(γ∣h,h^)→12P_b(\gamma\mid h,\hat{h})\to\frac{1}{2} for the affected symbols, an error-floor contribution rather than an SNR shift. The ∣h∣2|h|^2-only form (29) omits these effects and under-predicts BER in the deep-fade tail, where it should be replaced by numerical evaluation of the exact expectation (27).

(29) further requires the noise variance to be fixed in absolute terms (independent of hh), as adopted in our Rayleigh simulation per Section II-B2. Under the alternative measured-power model, the effective per-frame SNR is held constant by automatic gain control (AGC)-like re-calibration, eliminating the fading penalty.

Periodic Pulse Jamming on AWGN

Sparse (M+1)(M+1)-State Model

For sparse codes with MM non-zero entries per column, the classical two-state Bernoulli pulse-jammer model [3] , which assumes the receiver-integrated jamming power averages over many chips per symbol, substantially underestimates BER because only the MM non-zero chip positions per column carry that stream's signal energy: a pulse hit on a zero-position has no effect, while a pulse hit on a non-zero position contributes disproportionately to the despread output. We refine it into an (M+1)(M+1)-state model that enumerates the number of non-zero chip positions hit by the pulse during a given symbol.

Derivation. Assume, as an approximation to the periodic pulse jammer (4), that each non-zero chip position is hit independently with probability ρ\rho (its duty cycle). For column ck\mathbf{c}_k with MM non-zero positions, the random number of hits i∈{0,…,M}i \in\{0,\ldots,M\} then follows a binomial distribution, which can be described by

Pi=(Mi)ρi(1−ρ)M−i.(30)P_i = \binom{M}{i}\rho^i(1-\rho)^{M-i}. \tag*{(30)}

This is consistent with (4) in the marginal: the ON-state indicator of (4) has duty cycle ρ\rho, so each individual non-zero chip is struck with marginal probability ρ\rho, matching the per-chip hit rate of the jammer. The independence across the MM positions is the approximating step. Equation (4) is deterministic and periodic with a contiguous ON-interval of duration ρ/fp\rho/f_p; hit events are therefore in general correlated, through the shared pulse phase and, for closely spaced positions, through a single ON-burst spanning several consecutive chips. Independence is recovered in the limit where the pulse phase is uniform relative to the code grid and the non-zero positions are spaced wider than the ON-interval, so that no burst covers two of them; the correlated regime and its effect on (32) are quantified under Model assumptions and scope below. Conditional on ii hits, the despread output gains an additional jamming-power contribution σJ2=(i/M)Jon\sigma_J^2=(i/M)J_{\mathrm{on}} where Jon=JSR σ2/ρJ_{\mathrm{on}}=\mathrm{JSR}\,\sigma^2/\rho is the on-state jammer power per chip. The 1/ρ1/\rho factor compensates for the ρ\rho-fraction of time the jammer is on, giving average power JSR σ2\mathrm{JSR}\,\sigma^2. Combining with the AWGN noise variance σ2\sigma^2 and applying the effective processing gain GeffG_{\mathrm{eff}}, the state-conditioned post-ZF SNR is

γi=Geffγ1+(i/M) JSR/ρ,i∈{0,1,…,M}.(31)\gamma_i=\frac{G_{\mathrm{eff}}\gamma}{1+(i/M)\,\mathrm{JSR}/\rho},\qquad i\in\{0,1,\ldots,M\}. \tag*{(31)}

For QPSK with these state-conditional SNRs, the total bit-error probability is the binomial-weighted average, which can be obtained as

Pbpulse(γ,JSR)=∑i=0M(Mi)ρi(1−ρ)M−iQ(γi).(32)P_b^{\mathrm{pulse}}(\gamma,\mathrm{JSR})=\sum_{i=0}^{M}\binom{M}{i}\rho^i(1-\rho)^{M-i}Q(\sqrt{\gamma_i}). \tag*{(32)}

Special case M=2M=2 (the recommended sparse design): For three states i∈{0,1,2}i\in\{0,1,2\}, (32) expands to

Pbpulse∣M=2=(1−ρ)2Q(Geffγ)+2ρ(1−ρ)Q(Geffγ1+JSR/(2ρ))+ρ2Q(Geffγ1+JSR/ρ).(33)\begin{aligned} P_b^{\mathrm{pulse}}\big|_{M=2} &=(1-\rho)^2Q\left(\sqrt{G_{\mathrm{eff}}\gamma}\right)\\ &\quad+2\rho(1-\rho)Q\left(\sqrt{\frac{G_{\mathrm{eff}}\gamma}{1+\mathrm{JSR}/(2\rho)}}\right)\\ &\quad+\rho^2Q\left(\sqrt{\frac{G_{\mathrm{eff}}\gamma}{1+\mathrm{JSR}/\rho}}\right). \tag*{(33)} \end{aligned}

At low duty cycle (ρ≪1\rho\ll1), Pbpulse∣M=2≈PbAWGNP_b^{\mathrm{pulse}}\big|_{M=2}\approx P_b^{\mathrm{AWGN}}, recovering the unjammed limit. As ρ→1\rho\to1, Pbpulse∣M=2→Q(Geffγ/(1+JSR))P_b^{\mathrm{pulse}}\big|_{M=2}\to Q\left(\sqrt{G_{\mathrm{eff}}\gamma/(1+\mathrm{JSR})}\right), recovering the always-jammed AWGN+JSR limit. Between these extremes, the binomial mixture captures the sparse-code-specific behavior that the classical two-state model misses.

Model assumptions and scope. The binomial law (30) treats the MM hits as independent with duty-cycle probability ρ\rho, which holds when the pulse on-time is short relative to the spacing between a column’s non-zero chips, or equivalently when the pulse phase is uniform from symbol to symbol. It is therefore an ensemble model for the JSR-averaged BER of a spectrally flat code and does not resolve two regimes: (i) long (correlated) pulses, where one pulse strikes several consecutive non-zero chips, making the hits bursty and sub-binomial so that the tails of (32) broaden (a Gilbert–Elliott block model would be needed, and the independent-hit form is its decorrelated limit); and (ii) deterministic spectral alignment, namely the Paley–QR failure (Section V-F), a frequency-domain resonance in which the quadratic-residue spectral peaks coincide with the 10 kHz pulse harmonics, a mechanism invisible to the flat-spectrum model (32) and characterized separately via Figure 4.

Chip-rate FFT magnitude of representative $M=2$ columns

Figure 4. Chip-rate FFT magnitude of representative M=2M=2 columns. All two-tap columns share the same peak-to-average ratio (≈4\approx4 dB); the constructions differ in peak spacing. The Paley-QR peaks (red) recur finely and coincide with the 10 kHz pulse harmonics (dashed), whereas the Lempel–Costas (blue) and Free-uniform (green) peaks are coarser and fall between them.

Comparison to the classical Bernoulli model: The two-state classical Bernoulli model is the M→∞M\to\infty limit of (32), where the law of large numbers averages the binomial mixture into a single weighted-noise floor. In this case,

Pbclassical=ρQ(Geffγ1+JSR/ρ)+(1−ρ)Q(Geffγ).(34)P_b^{\mathrm{classical}}=\rho Q\left(\sqrt{\frac{G_{\mathrm{eff}}\gamma}{1+\mathrm{JSR}/\rho}}\right)+(1-\rho)Q\left(\sqrt{G_{\mathrm{eff}}\gamma}\right). \tag*{(34)}

For M=2M=2, the difference between (33) and (34) is O(ρ2)O(\rho^2) per (33)’s third term (the quadratic-in-ρ\rho “both hit” state). The discrepancy is one of our analytical contributions: the sparse pulse model corrects the classical Bernoulli expression in the M→1M\to1 limit, where binomial finite-MM effects matter.

Cross-Talk in the Overlap Regime

With KK columns each holding MM non-zeros, the total occupancy KMKM cannot exceed LL if columns are to remain disjoint. The bound KM≤LKM\le L is tight at M=L/KM=L/K (equal to 30 for our NB-IoT parameters L=360L=360, K=12K=12), beyond which disjoint allocation is combinatorially infeasible. For random allocation beyond this threshold, a standard collision-probability approximation gives the per-pair overlap probability

Povr(M)≈1−(1−ML)M,(35)P_{\mathrm{ovr}}(M)\approx1-\left(1-\frac{M}{L}\right)^M, \tag*{(35)}

which transitions from ≪1\ll1 to ∼1\sim1 as MM crosses L≈19\sqrt{L}\approx19 for our parameters (well before L/K=30L/K=30) and saturates at 1 for M≥L/KM\ge L/K. The implication is that the M≈L/KM\approx L/K transition is not a coincidence of the simulation parameters but a structural threshold dictated by the (L,K,M)(L,K,M) combinatorics.

For the overlap regime where disjoint supports are infeasible, the despread output (6) carries an inter-column cross-talk term whose magnitude is governed by the off-diagonal entries of the Gram matrix CHC\mathbf{C}^{H}\mathbf{C}. For columns drawn with i.i.d. unit-norm complex-Gaussian values on independently chosen MM-subsets of {1,…,L}\{1,\ldots,L\}, the expected total squared correlation obeys the Welch-type estimate [35]

E[∑i≠j∣ciHcj∣2]≈K(K−1)M2L2.(36)\mathbb{E}\left[\sum_{i\ne j}\left|\mathbf{c}_i^{H}\mathbf{c}_j\right|^2\right]\approx\frac{K(K-1)M^2}{L^2}. \tag*{(36)}

Eq. (36) is a leading-order estimate for moderate overlap and is monotonically increasing in MM; it captures the onset and initial growth of inter-column cross-talk once disjoint supports become infeasible (M>L/KM > L/K), but it does not by itself predict a peak or a subsequent decay. The non-monotone BER measured in § 8-C (a sparse-regime minimum (M=2M = 2–5), a bad-zone peak at M≈60M \approx60–120, and a partial recovery for M≥180M \ge180) arises instead from the interplay between this rising cross-talk and the per-column averaging that suppresses it. Just past the disjoint threshold the overlaps are few and concentrated, so the cross-talk adds coherently and inflates BER; at high density each column integrates many non-zero chips, averaging the same cross-talk toward a smaller noise-like residual while the per-column signal budget grows. Thus (36) and the overlap probability (35) fix the location of the bad-zone onset at M=L/KM = L/K, whereas the recovery at large MM is an averaging effect beyond this leading-order estimate; a quantitative peak-and-recovery model is left to future analytical refinement.

Numerical Results

This section presents simulations that (i) validate the analytical framework, and (ii) characterize the design space. Subsection V-A describes the simulation configuration. Subsection V-B reports the analytical-vs-simulation overlay. Subsections V-C–V-H present the empirical findings on density, structure, algebraic constructions, ML approaches, and Doppler. Subsection V-I summarizes the recommended designs across all four threat regimes.

Simulation Configuration

The simulator is implemented in MATLAB using the LTE Toolbox, with NB-IoT parameters following [1, 24, 25]. The simulation parameters are summarized in Table II. Each BER point is averaged over a deterministic seed-set of frames and reported with 95% Wilson-score confidence intervals where applicable; pilot-aided ZF equalization uses a single pilot symbol at (k,n)=(1,1)(k,n)=(1,1), with h^\hat{h} computed from (7) and applied via (8), and AJM is computed by linear interpolation of log⁡10(BER)\log_{10}(\mathrm{BER}).

ParameterValue
Spreading length LL360 chips per data symbol
Subcarriers KK12
Symbols per subframe NsN_s14
ModulationQPSK
Frames per BER point20–1000
Confidence intervals95% Wilson score
Fading channel3GPP ETU, fd∈{5,70,200,500}f_d \in\{5, 70, 200, 500\} Hz
Pilot / equalizationsingle pilot p0=1p_0=1 at (1,1)(1,1); ZF
Pulse jammer (fp,ρ)(f_p,\rho)(10 kHz, 2%), (20 kHz, 2%), (50 kHz, 1%)
Multi-tone jammer KtK_t{1,2,3,4}\{1, 2, 3, 4\} tones
AJM threshold BERth\mathrm{BER}_{\mathrm{th}}10−210^{-2}

Table II. Simulation parameters.

Equal-power assumption and scope. All K=12K = 12 streams are modeled with equal received power (unit-energy symbols and unit-norm columns, giving the composite power Px=K/LP_x=K/L of Section IV-A), so the disjoint-support designs carry zero inter-stream interference and the overlap-regime cross-talk is treated as an aggregate noise contribution. This is appropriate for the synchronous, power-controlled NB-IoT downlink considered here, but it does not capture the near–far regime of dense NOMA/SCMA deployments: with unequal user powers {Pj}\{P_j\} the interference seen at column kk is the power-weighted ∑j≠kPj∣cjHck∣2\sum_{j\ne k}P_j\lvert\mathbf{c}_j^H\mathbf{c}_k\rvert^2 rather than a uniform aggregate, so the cross-correlation impact (and any near–far disparity) becomes strongly dependent on the user power profile. Extending the density-first analysis to unequal-power and asynchronous NOMA/SCMA profiles is left to future work.

Analytical Validation

We first validate the closed-form analytical expressions of Section IV on the recommended uniform-M=2M = 2 Free code (the design that the rest of the section identifies as the optimal sparse choice). The 1000-frame simulation comprises 334,000 payload bits per BER point.

Figure 1 overlays the closed-form curves on simulation for the AWGN (a), flat-Rayleigh (b), and periodic-pulse (c) cases. The analytical curves (solid black) are computed from (25), (29), and (32); simulation points (blue dashed line, filled circles) are shown with 95% Wilson-score CI error bars. Across the meaningful operating range (BER∈[10−4,10−1]\mathrm{BER}\in[10^{-4},10^{-1}]), simulation tracks analytical within 1–2 dB, validating the Geff=L/(2K)G_{\mathrm{eff}}=L/(2K) correction. The minor discrepancies reflect (i) the SNR-dependent ZF penalty, (ii) deep-fade undersampling at high SNR, and (iii) periodic-pulse position correlation, all consistent with the analytical model’s stated approximations.

For the periodic-pulse case, the difference between the sparse (M+1)(M+1)-state model (32) and the classical two-state Bernoulli expression (34) is O(ρ2)O(\rho^2): a small effect at ρ=2%\rho=2\% but a large one on the analytical-vs-simulation agreement at high JSR. The Rayleigh validation, in turn, requires a fixed-noise simulation mode: a simulator that applies awgn('measured') on the post-fading waveform holds the per-frame SNR constant and produces results indistinguishable from AWGN, masking the true Rayleigh impact.

Density-Axis Sweep and the Bad-Zone Regime

We sweep the per-column density MM from 2 (extreme sparse) to 360 (fully dense), with M∈{2,5,10,20,30,60,120,180,240,300,360}M\in\{2,5,10,20,30,60,120,180,240,300,360\}. For each M≤L/K=30M\le L/K=30 we use the disjoint Free-Uniform construction; for M>30M>30 we use the Free-Overlap construction that allows column-support overlap.

Figure 2 presents the central empirical result of this paper. Three operating curves are overlaid: clean (no jammer at SNR=5 dB), pulse jamming (10 kHz, 2%, JSR=10 dB), and multi-tone jamming (2 tones, JSR=10 dB). The vertical dashed line at M=L/K=30M=L/K=30 marks the disjoint/overlap boundary.

BER versus column density M under clean, pulse, and multi-tone jamming

Figure 2. BER vs. column density MM under clean, pulse, and multi-tone jamming. The dashed line marks the disjoint/overlap boundary M=L/K=30M = L/K = 30; the bad zone appears at M≈60M \approx60–120120.

The figure reveals three distinct regimes:

Sparse regime (M=2M=2–5): pulse BER reaches 0.024 at M=5M=5, the empirical minimum. Multi-tone BER is also at its minimum at M=2M=2 (0.20). Disjoint column supports eliminate inter-column cross-talk, leaving only intra-column noise to defeat the receiver.

Bad zone (M≈60M \approx60–120120): immediately past the disjoint/overlap boundary, the pulse BER jumps to 0.30 (≈10×\approx10\times the sparse-regime optimum). Cross-column overlap introduces inter-column cross-talk per (36), but the column averaging required to suppress it is not yet available because density is still moderate.

Dense regime (M≥180M \ge180): pulse BER recovers to 0.15–0.18 as cross-talk gets averaged across the many non-zero chips per column. Doppler performance is dominantly determined by per-column averaging budget and is best at full density.

The LDSC paper [23] operating point of M=30M = 30 via Alamouti pairing (yielding 60 non-zeros per column) sits exactly in the upper edge of the bad zone. This finding explains the relatively poor pulse-jamming performance reported there relative to a properly sparse design.

Free Family vs. Alamouti-Paired LDSC

We compare the Free family against the density-matched Alamouti-paired LDSC of [23], both at 30 non-zeros per column. Averaged over seven jamming scenarios (three pulse and four multi-tone configurations) at JSR = 10 dB, the Free-Uniform construction achieves a mean BER of 0.18 versus 0.23 for the Alamouti-paired construction, a ∼20%\sim20\% reduction. The improvement is largest at pulse jamming, where the Free construction’s disjoint column supports eliminate the inter-column cross-talk that plagues the paired construction.

Structural Placement Comparison

Within the sparse Free regime (M=2,3,5M = 2, 3, 5), four structural placement variants are compared: random, uniform, golden-ratio, and chirp. Across 1000-frame BER measurements at JSR = 10 dB, the structural variants differ by 1–2 dB in pulse jamming AJM. The uniform-spacing variant emerges as a marginally best deterministic default, although the differences are within MC variance for several of the comparisons. We adopt uniform M=2M = 2 as the recommended Free-family configuration for the remainder of the paper.

Algebraic Constructions: Lempel–Costas versus Paley-QR

Algebraic constructions provide deterministic, reproducible alternatives to the random free family. We evaluate Paley-QR and Lempel-Costas at M=2M = 2 and M=5M = 5, alongside Walsh and the original LDSC-M30M30 as baselines. Figure 3 shows the BER vs. JSR for these eight codes across pulse jamming (a), multi-tone jamming (b), and ETU + 200 Hz Doppler (c).

BER under pulse, multi-tone, and ETU plus Doppler jamming

Figure 3. Algebraic constructions vs. baselines under (a) pulse, (b) multi-tone, and (c) ETU 200 Hz Doppler jamming.

The headline finding is that Paley-QR at M=2M = 2 degrades to the random-decision floor at all JSR levels (BER ≈0.48\approx0.48), while Lempel-Costas at M=2M = 2 matches the empirical optimum (BER = 0.020 at JSR = 10 dB, vs. uniform M=2M = 2’s 0.018). The Paley-QR failure is attributable to the spectral structure of the QR set. Because a sparse M=2M = 2 column is a two-tap sequence, its chip-rate spectrum is a raised cosine whose peak-to-average ratio is fixed (near 4 dB) for every M=2M = 2 code; the distinguishing feature is therefore not the peak height but the peak spacing, set by the gap Δ\Delta between the two non-zero positions, with peaks recurring every fs/Δf_s/\Delta in frequency (fsf_s the chip rate). The quadratic-residue positions mod 359 produce closely spaced spectral peaks that overlap the harmonic comb of the 10 kHz periodic pulse, coupling the jammer energy into the despread output, a resonance. Lempel–Costas, by its log-domain construction, produces a coarser, offset spectrum whose peaks largely fall between the jammer harmonics, avoiding the resonance. Figure 4 shows this directly: the Paley-QR peaks recur several times more finely than the Lempel–Costas and Free-uniform peaks and coincide with the 10 kHz harmonics (marked), explaining the catastrophic BER inflation in Figure 3(a).

The Lempel-Costas M=2M=2 result (Figure 3) demonstrates that a closed-form, deterministic algebraic construction can achieve the same jamming-robustness performance as the empirically-optimal Free design, with the practical advantage of full reproducibility.

Predictive and Generative Machine Learning

We evaluate both predictive and generative ML approaches per the methodology of Subsection III-E.

  1. Predictive ML: Figure 5 shows the predictor performance scatter: the trained neural-network regressor (and an alternative random-forest regressor) achieve test R2<0R^2 < 0 on held-out data, indicating that predictions are worse than predicting the test mean. The cause is that the per-matrix wBER variance from finite-sample Monte Carlo simulation overwhelms the signal at N=210N=210 training samples.

Predicted versus actual wBER and permutation feature importance

Figure 5. Predictive-ML failure: predicted vs. actual wBER (left) and permutation feature importance (right); test R2<0R^2 < 0 on the 210-matrix set.

  1. Generative ML: In contrast, the linear VAE generator produces sparse codes that match or exceed the best deterministic algebraic construction across all three threat regimes. Figure 5 compares VAE-best vs. Lempel-Costas M=2M=2 vs. uniform M=2M=2 in the multi-threat operating regime at 1000 frames per point.

The interpretation is that generative ML, by learning the structural manifold of valid sparse codes, is able to navigate around the noise that defeats the predictive pipeline. The tradeoff is that the generative result is non-deterministic (dependent on the random latent sample); for deployment use cases requiring reproducibility, Lempel-Costas M=2M = 2 remains the recommended deterministic choice.

BER comparison under pulse, multi-tone, and ETU 200 Hz Doppler jamming

Figure 6. Linear-VAE generator vs. Lempel-Costas M=2M = 2 and uniform M=2M = 2 under (a) pulse, (b) multi-tone, and (c) ETU 200 Hz Doppler jamming.

Doppler Robustness

The complement to jamming-axis performance is mobility-axis performance. Under ETU with 200 Hz Doppler, the dense regime (M≥180M \ge180) achieves the lowest BER (0.089 at SNR = 10 dB), the sparse regime (M≤5M \le5) is roughly 20% higher (0.108–0.137), and the bad zone (M=60M = 60) is intermediate (0.13). The mechanism is per-symbol channel-estimate averaging across the MM non-zero positions: under i.i.d. Jakes-spectrum statistics the post-spreading estimate variance scales as σh^2(M)∝1/M\sigma_{\hat{h}}^{2}(M) \propto1/M, predicting the sparse-to-dense improvement. At 500 Hz, all densities hit the single-pilot-ZF receiver-saturation floor (≈0.30\approx0.30), so Doppler beyond approximately 300 Hz is receiver-limited rather than code-limited.

Final Comparison and Recommendation

Table III summarizes the BER performance of the leading candidate codes across the four-threat operating matrix.

All reported BER differences in Table III exceed the 95% Wilson confidence intervals at the corresponding operating points, confirming statistical significance. Although the density-axis sweep (Figure 2) shows a random-allocation pulse minimum at M=5M = 5 (BER 0.024), the structured uniform-M=2M = 2 design achieves a strictly lower pulse-jamming BER (0.018) with smaller code storage and lower despread complexity (column support K⋅2K \cdot2 non-zeros vs K⋅5K \cdot5 at M=5M = 5), motivating M=2M = 2 as the recommendation over M=5M = 5. The recommended design is uniform-spacing M=2M = 2 on disjoint supports (or, for deterministic deployment, Lempel-Costas M=2M = 2): both achieve a ∼5×\sim5\times reduction in pulse-jamming BER compared to the LDSC-M=30M = 30 baseline of [23] while keeping multi-tone and Doppler BER within ∼20%\sim20\% of the dense optimum. The generative-ML candidate (VAE-best) is even better for pulse jamming but lacks reproducibility; we recommend it as

a complementary option when training data is available and reproducibility is not required.

Implementation complexity. Beyond BER, deployment also depends on code-generation, storage, and receiver-side complexity. Table IV summarizes these for the leading candidate designs. The sparse-M=2M = 2 designs (Free uniform and Lempel–Costas) are uniformly cheaper than the prior-art M=30M = 30 Alamouti construction across all three axes: code generation drops from O(KM)=O(L)O(KM) = O(L) (one Alamouti seed per block, K/2K/2 blocks of MM vertical stacks) to O(K)O(K) (one disjoint MM-position assignment per column), per-code storage drops from O(KM)O(KM) to O(K)O(K), and the matched-filter despreader shrinks from O(KM)O(KM) to O(K)O(K) multiply-accumulates per symbol. The Lempel–Costas variant adds O(log⁡p)O(\log p) for the primitive-root construction but avoids the random row-permutation needed by the Alamouti family. These gains are independent of the BER advantages and reinforce the practical case for sparse density selection.

Conclusion and Future Work

We have presented a systematic study of sparse spreading-code design for anti-jamming NB-IoT direct-sequence spread spectrum communications. Departing from the prior-art ML-guided LDSC framework of [23], we generalize the design space along a column-density axis, characterize an empirically-discovered adverse density regime at M≈L/KM \approx L/K, and propose a Free family of codes (disjoint per-column supports with arbitrary density) that outperforms the density-matched Alamouti-paired construction by approximately 20% in mean BER across the pulse and multi-tone scenarios. We further demonstrate that deterministic algebraic constructions (Lempel-Costas) match the empirical optimum, while the closely related Paley-QR family degrades to the random-decision floor under periodic pulse jamming. On the ML axis, predictive wBER regression fails on a 210-sample dataset, but generative modeling (linear VAE) successfully synthesizes jamming-robust codes that beat the best algebraic construction by 45%. We derive and validate three closed-form BER expressions incorporating a corrective effective processing gain Geff=L/(2K)G_{\mathrm{eff}} = L/(2K) that accounts for the composite-power SNR

TABLE III: Recommended-code BER summary (lower is better). All values at the indicated operating points with 1000-frame averaging. Best per scenario column in bold. Pulse 10k: periodic pulse jammer at 10 kHz, 2% duty. Multi-tone: 2-tone jammer with random tone frequencies. Doppler 200/500 Hz: ETU multipath fading channel with maximum Doppler fd=200f_d = 200 and 500 Hz, respectively.

TABLE IV: Implementation complexity of candidate spreading designs (per symbol). L=360L = 360, K=12K = 12; MM is the column density. Generation is per-code, executed once at provisioning.

DesignGenerationStorageDespread
Walsh / OVSFO(Llog⁡L)O(L \log L)O(KL)O(KL)O(KL)O(KL)
LDSC-M30 [23]O(KM)=O(L)O(KM)=O(L)O(KM)O(KM)O(KM)O(KM)
LDSC-Free uniform M=2M = 2O(K)O(K)O(K)O(K)O(K)O(K)
Lempel–Costas M=2M = 2O(K+log⁡p)O(K+\log p)O(K)O(K)O(K)O(K)
VAE-best M=2M = 2training-boundO(K)O(K)O(K)O(K)
Free-overlap M=180M = 180 (dense)O(KM)O(KM)O(KM)O(KM)O(KM)O(KM)

Table IV. Implementation complexity of candidate spreading designs (per symbol).

normalization of the KK-stream waveform and pilot-aided ZF noise.

The recommended design (uniform-spacing M=2M = 2 on disjoint supports for empirical-optimum performance, or Lempel-Costas M=2M = 2 for a deterministic closed-form alternative) achieves a 5× reduction in pulse-jamming BER relative to the prior-art LDSC-M30M30 baseline. More broadly, our results suggest that future anti-jamming sparse spreading systems should prioritize density selection before pursuing machine-learning-based coefficient optimization, since operating near the overlap transition can dominate performance regardless of optimization strategy.

Future work directions, in order of practical impact:

  1. Multi-pilot or data-aided channel tracking to unlock the dense-code mobility regime beyond the current ∼300 Hz Doppler ceiling. A ZZ-pilot generalization of the effective processing gain follows the form Geff(Z)=L/[K(1+1/Z)]G_{\mathrm{eff}}(Z) = L/[K(1 + 1/Z)], which reduces to our single-pilot L/(2K)L/(2K) at Z=1Z = 1 and recovers the classical L/KL/K in the limit Z→∞Z \to\infty. The underlying ZZ-dependent pilot-aided ZF noise penalty (1+1/Z)σ2(1 + 1/Z)\sigma^2 is already derived for general ZZ in Section IV-B; the remaining throughput/coverage trade-off analysis (optimal ZZ and pilot placement) is in preparation.

  1. Joint position-and-value optimization via gradient-based or semi-definite-programming methods, particularly for the algebraic constructions.

  1. Specialized algebraic constructions (Singer difference sets, perfect difference sets, Reed–Muller thinning) tailored to specific jammer-type vulnerabilities.

  1. Adversarial / smart-jammer scenarios where the jammer has knowledge of the code (sweep, reactive, follower, DRFM-based jammers); a game-theoretic minimax design framework against this class. The closed-form GeffG_{\mathrm{eff}} and sparse pulse-jamming expressions provide the analytical baseline against which adaptive-jammer strategies can be benchmarked.

  1. Generative ML expansion: nonlinear VAE/GAN, conditional models trained on specific jammer-class datasets, reinforcement-learning-driven code synthesis.

  1. Unequal-power / near–far NOMA/SCMA: relaxing the equal-power assumption of Section V-A to power-profile-dependent operation, where the interference at column kk becomes the power-weighted ∑j≠kPj∣cjHck∣2\sum_{j\ne k} P_j |c_j^{H}c_k|^2 and near–far disparities dominate the cross-correlation impact; joint power-control and sparse-code co-design for dense SCMA deployments.

  1. Real-channel measurements for V2X, UAV, and satellite links instead of statistical LTE models.

  1. Field testing / over-the-air validation of the recommended design on hardware.

0pt plus .5pt 0pt plus 0.2pt

References

  1. [1]X. Cheng, H. Yang, D. J. Jakubisin, N. Tripathi, G. Anderson, A. Wang, Y. Yang, and J. H. Reed, “5G Physical Layer Resiliency Enhancements with NB-IoT Use Case Study,” in Proc. IEEE Military Commun. Conf. (MILCOM), 2022, pp. 379–384.DOI
  2. [2]L. Yu, Z. Liu, M. Wen, D. Cai, S. Dang, Y. Wang, and P. Xiao, “Sparse Code Multiple Access for 6G Wireless Communication Networks: Recent Advances and Future Directions,” IEEE Commun. Standards Mag., vol. 5, no. 2, pp. 92–99, 2021.arxiv.org/abs/2104.01402
  3. [3]R. A. Poisel, Modern Communications Jamming Principles and Techniques, 2nd ed. Artech House, 2011.
  4. [4]D. Tse and P. Viswanath, Fundamentals of Wireless Communication. Cambridge University Press, 2005.DOI
  5. [5]D. Torrieri, Principles of Spread-Spectrum Communication Systems, 1st ed. Springer, 2005.
  6. [6]M. Tazebayeh, H. Nikopour, A. Bayesteh, and H. Baligh, “SCMA Codebook Design,” in Proc. IEEE 80th Veh. Technol. Conf. (VTC2014-Fall), 2014, pp. 1–5.
  7. [7]H. Nikopour and H. Baligh, “Sparse Code Multiple Access,” in Proc. IEEE 24th Annual Int. Symp. Personal, Indoor, Mobile Radio Commun. (PIMRC), 2013, pp. 332–336.DOI
  8. [8]Y.-M. Chen and J.-W. Chen, “On the Design of Near-Optimal Sparse Code Multiple Access Codebooks,” IEEE Trans. Commun., vol. 68, no. 5, pp. 2950–2962, 2020.DOI
  9. [9]Z. Chen, L. Zhang, Z. Wu, L. Wang, and W. Xu, “Reliable and Efficient Sparse Code Spreading Aided MC-DCSK Transceiver Design for Multiuser Transmissions,” IEEE Trans. Commun., vol. 68, no. 12, pp. 7402–7416, 2020.DOI
  10. [10]M. Zheng, Y. Wu, and W. Zhang, “Polar Coding and Sparse Spreading for Massive Unsourced Random Access,” in Proc. IEEE 92nd Veh. Technol. Conf. (VTC2020-Fall), 2020, pp. 1–5.arxiv.org/abs/2007.05739
  11. [11]D. Cai, P. Fan, X. Lei, Y. Liu, and D. Chen, “Multi-Dimensional SCMA Codebook Design Based on Constellation Rotation and Interleaving,” in Proc. IEEE 83rd Veh. Technol. Conf. (VTC Spring), 2016, pp. 1–5.
  12. [12]M. Kulhandjian, H. Kulhandjian, C. D’Amours, H. Khamyromeglu, and L. Hanzo, “Low-Density Spreading Codes for NOMA Systems and a Gaussian Separability-Based Design,” IEEE Access, vol. 9, pp. 33 963–33 993, 2020.DOI
  13. [13]Z. Liu and L.-L. Yang, “Sparse or Dense: A Comparative Study of Code-Domain NOMA Systems,” IEEE Trans. Wireless Commun., vol. 20, no. 8, pp. 4768–4780, 2021.arxiv.org/abs/2009.04148
  14. [14]Y. Liu, L. Xiang, L.-L. Yang, and L. Hanzo, “Space-Time Coded Generalized Spatial Modulation for Sparse Code Division Multiple Access,” IEEE Trans. Wireless Commun., vol. 20, no. 8, pp. 5359–5372, 2021.DOI
  15. [15]S. M. R. Islam, N. Avazov, O. A. Dobre, and K.-S. Kwak, “Power-Domain Non-Orthogonal Multiple Access (NOMA) in 5G Systems: Potentials and Challenges,” IEEE Commun. Surveys Tuts., vol. 19, no. 2, pp. 721–742, 2017.DOI
  16. [16]C. Duan, S. Zhang, P. Yin, X. Li, and J. Luo, “SCMA-TPGAN: A New Perspective on Sparse Codebook Multiple Access for UAV System,” Computer Communications, 2023, generative model for SCMA codebook design (NOMA setting).DOI
  17. [17]R. Rovatti, G. Setti, and G. Mazzini, “Chaotic Complex Spreading Sequences for Asynchronous DS-CDMA. Part II: Some Theoretical Performance Bounds,” IEEE Trans. Circuits Syst. I, Fundam. Theory Appl., vol. 45, no. 4, pp. 496–506, 1998.
  18. [18]L. Gong and L. Shaoqian, “Chaotic Spreading Sequences with Multiple Access Performance Better than Random Sequences,” IEEE Trans. Circuits Syst. I, Fundam. Theory Appl., vol. 47, no. 3, pp. 394–397, 2000.DOI
  19. [19]D. Guo and S. Verdú, “Randomly Spread CDMA: Asymptotics via Statistical Physics,” IEEE Trans. Inf. Theory, vol. 51, no. 6, pp. 1983–2010, 2005.arxiv.org/abs/cs/0503063
  20. [20]V. S. Somayazulu, “Multiple Access Performance in UWB Systems Using Time Hopping vs. Direct Sequence Spreading,” in Proc. IEEE Wireless Commun. Networking Conf. (WCNC), vol. 2, 2002, pp. 522–525.DOI
  21. [21]O. Simeone, “A Very Brief Introduction to Machine Learning with Applications to Communication Systems,” IEEE Trans. Cogn. Commun. Netw., vol. 4, no. 4, pp. 648–664, 2018.arxiv.org/abs/1808.02342
  22. [22]G. Zhu, D. Liu, Y. Du, C. You, J. Zhang, and K. Huang, “Toward an Intelligent Edge: Wireless Communication Meets Machine Learning,” IEEE Commun. Mag., vol. 58, no. 1, pp. 19–25, 2020.DOI
  23. [23]H. Yang, M. Kulhandjian, H. Kulhandjian, D. J. Jakubisin, X. Cheng, K. S. Bondada, and Y. Yang, “ML-Based Low-Density Spreading Code Design,” in Proc. 58th Asilomar Conf. Signals, Systems, and Computers. IEEE, 2024, pp. 1–5.DOI
  24. [24]3GPP, “ETSI TS 136 101: Evolved Universal Terrestrial Radio Access (E-UTRA); User Equipment (UE) Radio Transmission and Reception,” ETSI Technical Specification, 2023.
  25. [25]——, “ETSI TS 136 104: Evolved Universal Terrestrial Radio Access (E-UTRA); Base Station (BS) Radio Transmission and Reception,” ETSI Technical Specification, 2023.
  26. [26]M. Kulhandjian, H. K. Kulhandjian, and D. J. Jakubisin, “Density-First Sparse Spreading Code Design for Anti-Jamming NB-IoT with Closed-Form BER Analysis,” in Proc. IEEE Military Commun. Conf. (MILCOM), Washington, DC, USA, Oct. 2026, to appear.
  27. [27]S. M. Alamouti, “A Simple Transmit Diversity Technique for Wireless Communications,” IEEE J. Sel. Areas Commun., vol. 16, no. 8, pp. 1451–1458, 1998.DOI
  28. [28]R. E. A. C. Paley, “On Orthogonal Matrices,” Journal of Mathematics and Physics, vol. 12, no. 1, pp. 311–320, 1933.DOI
  29. [29]J. P. Costas, “A Study of a Class of Detection Waveforms Having Nearly Ideal Range-Doppler Ambiguity Properties,” Proc. IEEE, vol. 72, no. 8, pp. 996–1009, 1984.
  30. [30]L. Breiman, “Random Forests,” Machine Learning, vol. 45, no. 1, pp. 5–32, 2001.
  31. [31]D. P. Kingma and M. Welling, “Auto-Encoding Variational Bayes,” in Proc. Int. Conf. Learning Representations (ICLR), 2014.arxiv.org/abs/1312.6114
  32. [32]I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y. Bengio, “Generative Adversarial Nets,” in Adv. Neural Inform. Process. Syst. (NeurIPS), 2014, pp. 2672–2680.
  33. [33]P. Xiao, P. Qiu, S. Ha, A. Bani, S. Zhou, and A. Sotiras, “SC-VAE: Sparse Coding-based Variational Autoencoder with Learned ISTA,” Pattern Recognition, 2024, also available as arXiv:2303.16666.arxiv.org/abs/2303.16666
  34. [34]P. O. Hoyer, “Non-negative Matrix Factorization with Sparseness Constraints,” Journal of Machine Learning Research, vol. 5, pp. 1457–1469, 2004.DOI
  35. [35]L. Welch, “Lower Bounds on the Maximum Cross-Correlation of Signals,” IEEE Trans. Inf. Theory, vol. 20, no. 3, pp. 397–399, 1974.DOI
  36. [36]M. J. E. Golay, “The Merit Factor of Long Low Autocorrelation Binary Sequences,” IEEE Trans. Inf. Theory, vol. 28, no. 3, pp. 543–549, 1982.DOI

Paper details

vlnc:W_6wzvfgrw
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Engineering

Not recorded for this paper

View statement
Looking for related papers…
Contents