LiQSS: Post-Transformer Linear Quantum-Inspired State-Space Tensor Networks for Real-Time 6G

TL;DR

This paper targets the challenge of real-time KPI forecasting in 6G O-RAN's Near-RT RIC, where Transformer-based models are hindered by $O(L^2)$ complexity. It introduces LiQSS, a post-Transformer architecture that combines quantum-inspired linear state-space dynamics with tensor-network factorization (TT/MPS) for input embedding and readout, and lightweight cross-KPI interaction operators. By employing HiPPO-LegS mixture kernels for stable, multi-scale temporal modeling and TT-based global maps, LiQSS achieves competitive forecasting accuracy with dramatically reduced parameter counts and inference latency, demonstrating linear-time scaling in practice. Experimental results on a bespoke KPI dataset (59,441 windows across 13 KPIs) show LiQSS can be up to $155\times$ smaller in parameters and up to $2.74\times$ faster than Transformer baselines while maintaining high forecast fidelity. These findings indicate that post-Transformer, state-space, and tensor-network approaches are viable for deployment-faithful, real-time intelligence in future 6G Near-RT RIC environments.

Abstract

Proactive and agentic control in Sixth-Generation (6G) Open Radio Access Networks (O-RAN) requires control-grade prediction under stringent Near-Real-Time (Near-RT) latency and computational constraints. While Transformer-based models are effective for sequence modeling, their quadratic complexity limits scalability in Near-RT RAN Intelligent Controller (RIC) analytics. This paper investigates a post-Transformer design paradigm for efficient radio telemetry forecasting. We propose a quantum-inspired many-body state-space tensor network that replaces self-attention with stable structured state-space dynamics kernels, enabling linear-time sequence modeling. Tensor-network factorizations in the form of Tensor Train (TT) / Matrix Product State (MPS) representations are employed to reduce parameterization and data movement in both input projections and prediction heads, while lightweight channel gating and mixing layers capture non-stationary cross-Key Performance Indicator (KPI) dependencies. The proposed model is instantiated as an agentic perceive-predict xApp and evaluated on a bespoke O-RAN KPI time-series dataset comprising 59,441 sliding windows across 13 KPIs, using Reference Signal Received Power (RSRP) forecasting as a representative use case. Our proposed Linear Quantum-Inspired State-Space (LiQSS) model is 10.8x-15.8x smaller and approximately 1.4x faster than prior structured state-space baselines. Relative to Transformer-based models, LiQSS achieves up to a 155x reduction in parameter count and up to 2.74x faster inference, without sacrificing forecasting accuracy.

Figures (4)

• Figure 1: System model and intelligence placement in the O-RAN architecture. Telemetry collected from O-RU/O-DU/O-CU via the E2 interface using the E2SM-KPM service model is ingested by the Near-RT RIC and processed through sliding-window preprocessing. A perceive–predict xApp operates under Near-RT constraints on memory footprint, computation complexity, and latency ($<\Delta t$), producing short-horizon KPI predictions that drive control xApps via the E2 control loop. Long-term model training and policy generation are orchestrated by the SMO and Non-RT RIC over the A1 interface, while the trained models are deployed to the Near-RT RIC on O-Cloud/edge infrastructure for real-time operation.
• Figure 2: General end-to-end architecture of the proposed quantum-inspired linear state-space tensor-network forecaster for Near-RT O-RAN telemetry. A chronologically split, train-only normalized KPI window is embedded via a TT/MPS input projection, processed by a stack of structured HiPPO–LegS mixture state-space blocks implemented as depthwise causal convolutions with squeeze–excitation gating and gated channel mixing, and summarized by a causal last-step readout. A TT/MPS prediction head then maps this causal summary to the next-step KPI estimate. During training, the model parameters (including the TT cores and SSM-mixture parameters) are optimized using a Mean Squared Error (MSE) loss with backpropagation, validation-driven early stopping, and best-checkpoint saving; during inference, the trained checkpoint is used in a no-gradient forward pass, followed by inverse normalization (and, when evaluating, metric computation).
• Figure 3: Ground truth (solid blue) versus one-step-ahead predictions (dashed orange) for key O-RAN KPIs over the last 700 test samples.
• Figure 4: Empirical evidence of linear-time behavior in LiQSS. Here, training time is the average seconds per example spent on one training step (forward pass + loss + backward pass + optimizer update) measured on a fixed batch after warm-up.