Sample Error Rate

Description

Sample Error Rate (SER) is a key performance indicator defined in 3GPP Release 19 specifications 38.191 and 38.769, which cover NR (New Radio) physical layer measurements and AI/ML-based CSI compression. SER measures the proportion of samples that are erroneously processed or reconstructed in a given observation window. Unlike traditional block error rate (BLER) or bit error rate (BER), SER operates at the sample level, making it particularly relevant for evaluating the fidelity of AI/ML models used for CSI feedback compression and reconstruction.

In the context of AI/ML-based CSI compression, the user equipment (UE) compresses channel state information using a neural network encoder, and the gNodeB (gNB) reconstructs it using a decoder. SER quantifies the fraction of complex-valued channel samples that deviate beyond a defined error threshold after reconstruction. This metric is crucial because traditional metrics like cosine similarity or normalized mean square error (NMSE) may not capture the perceptual or functional impact of sample-level errors on beamforming and link adaptation performance.

The calculation of SER involves comparing the original CSI samples with the reconstructed samples. A sample is considered erroneous if the error magnitude exceeds a specified threshold, which can be absolute or relative. The threshold and observation window are configurable and are typically defined in the test specifications. SER is reported as a ratio or percentage and is used in both link-level simulations and field trials to benchmark AI/ML model performance.

SER is part of a broader set of metrics introduced in Rel-19 to support the evaluation of AI/ML for NR air interface. It complements other metrics such as throughput, BLER, and beamforming gain. The metric is applicable to both frequency-division duplex (FDD) and time-division duplex (TDD) systems, and it can be used for both sub-6 GHz and millimeter-wave frequencies. Standardization of SER ensures consistent evaluation across vendors and accelerates the adoption of AI/ML in 5G-Advanced and 6G networks.

From an implementation perspective, SER requires the UE and gNB to support the necessary instrumentation to capture and compare samples. This may involve additional computational complexity, but it is essential for validating AI/ML algorithms. The metric is expected to be used in future releases for other AI/ML use cases, such as positioning and mobility optimization.

Purpose & Motivation

SER was created to address the need for a standardized, sample-level error metric for evaluating AI/ML-based physical layer processing in 5G-Advanced. Traditional metrics like BLER and BER operate at the block or bit level and may not adequately capture the performance of AI/ML models that process complex-valued samples, such as CSI compression. SER provides a more granular view of reconstruction accuracy, enabling fair comparison of different AI/ML algorithms and facilitating their integration into 3GPP networks. It also helps in setting performance requirements and test cases for AI/ML-enabled UEs and gNBs.