Cellular Link-Quality Indicators

• Cellular link-quality indicators are quantifiable metrics that measure radio link and end-to-end performance in LTE and 5G networks.
• They combine direct physical-layer measurements (e.g., RSRP, SINR, CQI) with composite KPIs like usability, resilience, and spatial variability.
• Recent advances integrate uncertainty quantification and machine learning to enhance throughput prediction, resource allocation, and spatial coverage mapping.

Cellular link-quality indicators are quantifiable metrics that characterize the radio link and end-to-end performance of user equipment (UE) in contemporary mobile networks such as LTE and 5G NR. These indicators encompass both direct physical-layer measurements (e.g., RSRP, SINR, CQI) and higher-level composite metrics (e.g., usability, resilience), and are essential for link adaptation, resource allocation, spatial coverage mapping, machine learning-based estimation, and service quality benchmarking. Recent research further extends classical point-based metrics with uncertainty quantification and multi-dimensional Key Performance Indicators (KPIs) that jointly assess usability, stability, reliability, and spatial-temporal variability.

1. Physical-Layer and Protocol-Level Link-Quality Indicators

Cellular link quality has traditionally been quantified by standardized physical-layer metrics and protocol feedback, each of which targets distinct aspects of link reliability and capacity (Cerar et al., 2018). Principal indicators include:

• RSRP (Reference Signal Received Power): Average received power (dBm) over cell-specific reference signals, measuring large-scale pathloss and shadowing ($-140$ to $-80$ dBm typical).
• RSRQ (Reference Signal Received Quality): Ratio (dB) of RSRP to RSSI per resource block, jointly reflecting signal strength and interference/noise.
• RSSI (Received Signal Strength Indicator): Total received wideband power, including both reference and non-reference elements.
• SINR (Signal to Interference plus Noise Ratio): Linear or dB ratio of desired signal power to aggregate interference plus noise, computed as $$\mathrm{SINR} = \mathrm{RSRP}/(\mathrm{RSSI} - \mathrm{RSRP})$$.
• CQI (Channel Quality Indicator): Quantized (1–15) feedback mapping the measured SINR to the highest Modulation and Coding Scheme (MCS) ensuring BLER $\leq 10\%$ (Yin et al., 2020). CQI is central to adaptive link control.
• BLER (Block Error Rate): Fraction of unsuccessful transport blocks, primary for assessing reliability.

These metrics are measured at different temporal granularity, employ filtering or quantization, and are often pre-processed (windowed averaging, delta features, normalization) before input into higher-level analytics or learning systems.

2. Composite and Application-Oriented Link-Quality KPIs

To address the limitations of point estimates and instantaneous measures, modern frameworks such as Quality of Coverage (QoC) define a multidimensional KPI space capturing both temporal and spatial aspects of link usability (Srinivasavaradhan et al., 24 Oct 2025). Let $X(x, y, t)$ denote a performance metric at location $(x, y)$ and time $t$, and fix usability threshold $\tau$ over window $T$. The core QoC KPIs are:

• Usability ($U$): Fraction of time the link meets the usability threshold,

$$U(x, y; \tau) = \frac{1}{T} \int_0^T \mathbb{I}(\mathrm{usable}_\tau[X(x, y, t)]) \, dt$$

• Persistence ($P$): Average contiguous duration of usable intervals,

$P(x, y; \tau) = \frac{1}{N} \sum_{j=1}^N L_j$

• Performance Mean During Usable Periods ($M$): Mean median performance over all usable intervals,

$M(x, y; \tau) = \frac{1}{N} \sum_{j=1}^N M_j$

• Variability ($V$): Average normalized inter-quartile range within usable intervals,

$$V(x, y; \tau) = \frac{1}{N}\sum_{j=1}^N \frac{P_{75}(X_j) - P_{25}(X_j)}{P_{50}(X_j)}$$

• Resilience ($R$): Inverse mean duration of outages (speed of recovery),

$R(x, y; \tau) = \frac{W}{\sum_{i=1}^W D_i}$

where $N$ (resp. $W$) is the number of usable (unusable) time intervals (Srinivasavaradhan et al., 24 Oct 2025).

This five-dimensional QoC vector $(U, P, M, V, R)$ characterizes not only momentary quality but also consistency, typical performance, in-interval variation, and recovery behavior, anchoring cellular coverage assessments to application-level requirements.

3. Link-Quality Indicators in Machine Learning and Estimation

Link-quality estimation (LQE) via machine learning regularly leverages RSRP, SINR, CQI, RSRQ, and BLER as core input features; RSSI is less frequently used (Cerar et al., 2018). Empirical analyses spanning 2017–2021 found:

Indicator	Usage in ML-LQE Papers
RSRP	95%
SINR	85%
CQI	75%
RSRQ	65%
BLER	50%
RSSI	30%

Feature engineering includes smoothing, differencing, normalization, binning, and temporal stacking for time-series models (e.g., LSTM, CNN). Model types include Random Forests and DNNs for regression (throughput prediction) and SVM, LightGBM, LSTM for classification (good/poor, handover need) (Cerar et al., 2018).

Incremental inclusion of SINR and CQI to RSRP-based models confers significant predictive gains: adding SINR typically reduces throughput prediction RMSE by 25%, with CQI providing an additional 10% gain (Cerar et al., 2018). A typical pipeline predicts regression (throughput, cell edge rate) or classification (service grade, handover trigger) targets using these indicators after appropriate pre-processing.

4. Channel Quality Indicator (CQI): Protocol Function and Predictive Modeling

CQI holds a privileged role as a protocol feedback mechanism in LTE/5G NR. It is computed by each UE as an integer $c \in \{1,\dots,15\}$, encoding the highest MCS sustainable under present SINR for a target BLER. The mapping is defined by standardized SINR thresholds $\gamma_1,\dots,\gamma_{14}$ such that

$$c = \begin{cases} 1, & \gamma < \gamma_1 \ 2, & \gamma_1 \leq \gamma < \gamma_2 \ \vdots\ k, & \gamma_{k-1} \leq \gamma < \gamma_k \ 15, & \gamma \geq \gamma_{14} \end{cases}$$

(Yin et al., 2020). CQI drives downlink resource allocation via MCS selection in the MAC scheduler, directly influencing the instantaneous throughput: $R_t = \sum_{i=1}^n r(M_{i,t}) \cdot (1 - B_{i,t})$ with $M_{i,t} = m(\mathrm{CQI}_i(t))$ and $B_{i,t}$ the error rate. Feedback delay $\tau$ degrades resource allocation efficacy, with 0.5 ms of delay resulting in up to 15–20% throughput loss at high UE mobility. Deep learning time-series models (LSTM) trained on recent CQI vectors, and updated online to track channel dynamics, significantly reduce this penalty, producing up to 12% throughput gain at 70 m/s versus delayed-CQI baseline (Yin et al., 2020).

5. Stochastic Models and Temporal/Spatial Correlation of Link Quality

The temporal and spatial correlation structure of cellular link quality is nontrivial, particularly due to interference persistence across time slots and spatial locations. In heterogeneous cellular networks (HCNs), the aggregate interference experienced by a user is correlated—quantified by the interference correlation coefficient

$$\rho_{t} = \frac{1}{\mathbb{E}[h^2]} = 0.5 \quad \text{(for Rayleigh fading)}$$

which is invariant to network tiers, BS density, and SIR threshold (Sheng et al., 2014). The single-slot success probability for SIR threshold $\beta_k$ is

$$P_s = \sum_{k=1}^K \frac{\lambda_k P_k^\delta \beta_k^{-\delta}}{(\sum_\ell \lambda_\ell P_\ell^\delta)C(\delta)}$$

with $\delta = d/\alpha$. The $n$-slot joint success probability,

$p^{(n)} = \frac{p^{(1)}}{D_n(\delta)}$

is governed by the diversity polynomial $D_n$, capturing temporal dependence driven entirely by path-loss exponent and number of time slots. Ignoring such correlation underestimates outage risk by 30–50%. The conditional $n$-th slot success given $n-1$ previous successes is determined by these parameters alone, independent of other network features (Sheng et al., 2014).

6. Uncertainty Quantification and Data-Driven Coverage Mapping

Recent approaches systematically quantify the uncertainty in predicted link-quality scores, delivering spatially-gridded confidence intervals to better guide data collection and coverage mapping (Jiang et al., 2024). Ensemble Spatial Conformal Prediction (ESCP) constructs rigorous prediction intervals by:

• Fitting a point-wise regression model $\hat{f}$ for each spatial location,
• Computing nonconformity scores as absolute residuals $\epsilon_i = |Y_i - \hat{f}(X_i)|$,
• Building prediction intervals via split conformal and ESCP methods, including localized calibration, bootstrapping, and quantile regression,
• Generating uncertainty maps by evaluating interval width $w(x)$ over dense grids,
• Deploying active sampling by targeting locations with highest uncertainty $w(x)$ to improve training-data quality in critical regions.

Coverage and interval width are assessed empirically; valid conformal methods achieve coverage near the target $1-\alpha$, with ESCP yielding spatially-adaptive, statistically efficient intervals (Jiang et al., 2024).

7. Passive Probing and Network Usage Indicators

Beyond protocol-layer indicators, passive client-side probing architectures enable inference of link quality with minimal network participation. The C³ACE framework leverages downlink control channel (PDCCH) analysis to determine the number of active UEs ($N$) and estimate expected throughput for fair schedulers: $$T_\mathrm{UE}^\mathrm{expected} = \frac{1}{N+1} N_\mathrm{RB}^\mathrm{total} r_\mathrm{RB}(\mathrm{MCS})$$ Passive histogram-based RNTI validation achieves sub-Mbit/s throughput estimation error and low false-positive rates. Such metrics inform local offloading, network selection, and real-time control decisions on resource-constrained UEs (Falkenberg et al., 2017).

---

Cellular link-quality indicators collectively span low-level physical metrics, dynamically-evolving protocol feedback, and high-order spatiotemporal performance dimensions. Modern research advances the paradigm from single-metric coverage to multidimensional, application-driven, and uncertainty-informed quantification, incorporating both the physics of radio propagation and the realities of noisy, sparse, or heterogeneous empirical measurement [(Jiang et al., 2024); (Srinivasavaradhan et al., 24 Oct 2025); (Cerar et al., 2018); (Yin et al., 2020); (Falkenberg et al., 2017); (Sheng et al., 2014)].