A base station with dozens of antennas can steer energy precisely toward a phone, but only if it knows the downlink channel's covariance matrix. This matrix records how strongly and how coherently signals arrive from each direction, and it changes slowly over time. In frequency-division duplex (FDD) systems, where uplink and downlink use different frequencies, the station cannot measure this matrix directly. It must rely on what the phone reports back. The standard cheap report, a 3GPP "Type I" codebook index, only names the best of a few coarse beams. The best existing trick turns many such reports into a high-resolution covariance estimate, but it needs about 33 rounds at walking speed. That makes it fragile once users move faster. Meanwhile the phone is already sending a much denser stream of other signals. It sends channel quality indicators (CQI, a few bits saying how good the link is). It also sends hybrid automatic repeat request acknowledgements (HARQ ACK/NACK, one bit per transmission saying whether a packet decoded). Current covariance estimators throw these away.
Recommender systems have faced the same kind of problem. Users leave a dense trail of weak signals such as views and clicks, and a sparse trail of strong ones such as carts and purchases. Collapsing them into one uniform signal lets the noise drown the intent. Explicitly ordering them by how much each kind reveals helps substantially. We propose to treat the base station as a recommender: each transmission "recommends" a beam to the phone. A HARQ ACK or NACK is then implicit feedback on that recommended beam. CQI is graded feedback on the same beam, and the Type I index is a sparse but explicit statement of preference among codebook beams. Covariance estimation becomes learning a low-rank preference model from heterogeneous implicit feedback that the system already pays for. It needs no new pilots and no extra uplink bits.
Core Claim
The quantity every report speaks about is the beamforming gain $g(w) = w^H C w$, where $w$ is a unit-norm beam vector and $C$ is the $N \times N$ downlink covariance at an $N$-antenna base station. Each type of report is an ordinal statement about these gains. A Type I index $p$ says $g(v_p) \ge g(v_k)$ for the other codewords $v_k$. A CQI level says $g(w_t)$ lies between two thresholds for the beam $w_t$ actually used in slot $t$. An ACK at modulation-and-coding scheme $m$ says $g(w_t)$ probably exceeded a decoding threshold $\theta_m$, and a NACK says it probably did not.
These statements differ greatly in reliability. HARQ outcomes are dense but corrupted by fast fading around the slow covariance. Type I indices are sparse and comparatively clean. This is the recommender's "flat signal problem" in physical form. We estimate $C$ by maximizing a soft ordinal likelihood with a separate noise scale per feedback type, recency weighting for mobility, and a nuclear-norm penalty that reflects the few physical propagation paths:
$$ \hat C = \arg\min_{C \succeq 0} \; \sum_{s \le t} \gamma^{\,t-s} \sum_{(a,b,\beta) \in \mathcal{P}_s} \ell\!\left(\frac{g_a(C) - g_b(C)}{\sigma_\beta}\right) + \lambda \|C\|_* . $$
The symbols are defined as follows. $\mathcal{P}_s$ is the set of ordered pairs implied by the reports received in slot $s$. Each pair compares two gains, or a gain and a threshold $\theta_m$ treated as a fixed reference. The pair says $g_a$ should exceed $g_b$, and $\beta$ labels the feedback type that produced it. $\ell(z)=\log(1+e^{-z})$ is the logistic ranking loss used in Bayesian personalized ranking. $\sigma_\beta$ is the learned noise scale of feedback type $\beta$. $\gamma \in (0,1]$ discounts stale reports, and $\lambda$ sets the low-rank pressure.
The new element is not ranking loss on its own. It is the combination of three things. First, the base station's own beam choices are treated as the probes that generate implicit feedback. Second, a hierarchy over feedback types is learned rather than handcrafted. Third, soft constraints replace hard geometric cuts. Our central empirical claim is that this estimator reaches Type II beamforming precision in clearly fewer rounds than 33 at 3 km/h, and that it keeps tracking at vehicular speeds where cut-based reconstruction stalls.
Relation to Prior Work
The method's closest relative, and the work it most directly builds on, is Pushing The Limit of Type I Codebook For FDD Massive MIMO Beamforming. That paper reads each Type I report as a linear cut that removes part of a feasible set of covariances. It adds nuclear-norm regularization and designs informative probing matrices, and on QuaDRiGa it reaches Type II precision after about 33 rounds with a 32-antenna array and a 3 km/h user. We keep its low-rank prior and its insight that probing can be designed. We change two things. We admit the denser HARQ and CQI streams, and we replace hard cuts with soft, reliability-weighted ordinal terms. Hard cuts built from noisy one-bit HARQ outcomes would quickly make the feasible set empty.
Massive MIMO with Imperfect Channel Covariance Information motivates working with covariance rather than instantaneous channels. It shows that sample covariance fails when observations are fewer than antennas and that shrinkage and indirect estimation cut pilot needs sharply. That work lives in a reciprocity setting with uplink pilots, not FDD feedback.
The recommender side supplies the method. Accurate multi-behavior sequence-aware recommendation via graph convolution networks enforces ranking margins along a Buy > Cart > View hierarchy with non-uniform pairwise sampling. It also uses dense auxiliary interactions to warm up sparse target behaviors. We borrow both ideas: pairwise ordinal objectives across behavior tiers, and dense HARQ outcomes helping the sparse Type I reports. That paper flags its handcrafted hierarchy and linear cascade as limitations, which is why we learn $\sigma_\beta$ instead of fixing the order. Micro-Behavior Encoding for Session-based Recommendation shows that the meaning of an action depends on its pairing with other actions. Our analogue is that an ACK carries different information depending on the modulation scheme and beam it was sent with, which our pairs encode explicitly. RBPR separates weak "browsing" signals from strong "purchasing" signals within one ranking model, a direct precedent for mixing feedback tiers in a single pairwise loss.
Two papers constrain the design. Deep CNN-Based Channel Estimation for mmWave Massive MIMO Systems reports that learned estimators degrade when channel statistics shift, for example between urban micro and urban macro scenarios. This pushes us toward a model-based estimator with few learned parameters (the $\sigma_\beta$, $\gamma$, $\lambda$) rather than a deep network trained per environment. Joint Visibility Region and Channel Estimation for Extremely Large-Scale MIMO Systems shows how hard intermediate decisions propagate errors in two-stage estimators. This is another reason we keep every report as a soft term in one objective rather than thresholding it first.
Research Program
The work has four parts.
The first part is a feedback simulator layered on QuaDRiGa. It emits Type I indices, CQI, rank indicators, and HARQ outcomes with realistic periodicities, so that HARQ appears every scheduled slot and PMI every reporting period. Errors come from fast fading around the slow covariance and from link adaptation choices. The scheduler's beam and modulation choices are logged as the "recommendations" that generated each outcome.
The second part is the estimator. We solve the objective above online by projected gradient steps on a low-rank factorization $C = UU^H$. The noise scales $\sigma_\beta$ are fitted by maximizing marginal likelihood on a held-out window, and $\gamma$ is chosen by a speed-aware rule that we will derive and test.
The third part is a beam-selection policy. Because HARQ feedback only speaks about beams the station actually used, we compare purely greedy beam use against light exploration, such as occasional perturbed beams in low-priority slots. This lets us measure how much exploration the ordinal estimator needs. It mirrors exposure bias in recommenders and the probing design of the cutting-plane paper.
The fourth part is analysis. We will characterize when HARQ pairs are identifiable at all. A NACK on the chosen beam says little about directions orthogonal to it, so we expect HARQ to sharpen the dominant eigenvector and leave secondary paths to Type I reports. We will test this by measuring per-eigenvector error.
Baselines
The naive reference point is plain Type I beamforming, which uses the latest reported codeword. Next to it sits a sample-covariance estimate built by averaging outer products of reported codewords with diagonal shrinkage, as in Massive MIMO with Imperfect Channel Covariance Information. Any method that does not beat these on beamforming gain is not worth its complexity.
The strongest existing competitor is the cutting-plane reconstruction of Pushing The Limit of Type I Codebook. It is the only method in the file that recovers high-resolution FDD covariance from legacy feedback. Its reported result is convergence to Type II beamforming precision after about 33 communication rounds on QuaDRiGa with a 32-antenna base station and a 3 km/h user. We will reimplement it on our simulator in that same configuration and first confirm that we reproduce roughly 33 rounds before comparing. It will receive the same probing freedom and the same per-update solver time cap, with its regularization weight tuned on the same validation grid and trial budget as ours. Beating it means reaching the same Type II precision in clearly fewer rounds at 3 km/h, and holding a beamforming-gain advantage at 30 and 60 km/h, where its round count may exceed the channel's stationarity time.
The upper references are full Type II feedback and genie-known covariance. Type II reaches up to 30% better beamforming precision than Type I in the baseline configurations of the cutting-plane paper. Neither upper reference is a competitor, since both cost far more uplink bits. They tell us how much of the gap we close.
Three ablations isolate the key ingredients. The first gives the cutting-plane method the same HARQ and CQI information as hard cuts. The second uses our soft ordinal loss with all $\sigma_\beta$ fixed equal, which is the "flat signal" version. The third uses our loss with Type I reports only. Together they separate the value of extra feedback streams, of soft constraints, and of learned per-type reliability.
Why not just…?
Why not just add CQI and HARQ to the cutting-plane method as extra cuts? This is the most serious threat, and it is our first ablation. Our expectation is that HARQ outcomes, which flip with fast fading even when the covariance is fixed, will generate mutually inconsistent cuts within a few dozen slots. The feasible set would then collapse or require ad hoc cut deletion. Only CQI, being averaged, is likely to work as a hard cut. If hard cuts with all feedback types match our soft estimator, the contribution shrinks to "use the extra feedback." We will report it that way and drop the ranking machinery.
Why not just upgrade to Type II feedback? Type II gives better precision, but at a substantial uplink cost. The cutting-plane authors also stress that Type I is the deployed standard across 4G and 5G. Our proposal adds zero uplink bits, so Type II serves as a ceiling rather than an alternative.
Why not just train a neural network that maps the feedback stream to covariance? This is feasible and would likely be fast at inference. However, Deep CNN-Based Channel Estimation documents degradation under mismatched channel statistics, and the base station's scheduler makes the feedback distribution depend on its own policy. We will train a sequence-model baseline on urban micro scenarios and test it on urban macro scenarios and at unseen speeds. We will count it as a real competitor if it holds up under that shift.
Why not just weight report types by hand? A fixed hierarchy, such as trusting Type I fully and HARQ half as much, is the recommender default, and the MBA paper names its tuning burden as a limitation. Reliability here depends on speed and signal-to-noise ratio, so we compare learned $\sigma_\beta$ against the best hand-tuned fixed weights chosen with the same tuning budget.
Proposed Experiments
All experiments run on QuaDRiGa at the cutting-plane paper's setting of 32 antennas and 3 km/h. We then sweep speeds of 3, 30, and 60 km/h and signal-to-noise ratios, and cover urban micro and urban macro, line-of-sight and non-line-of-sight scenarios. The metrics are rounds to Type II beamforming precision, normalized beamforming gain $g(\hat w)/g(w^\star)$ over time, downlink spectral efficiency, and covariance error split by eigenvector. We report solver time per update for every method.
The decisive checks are as follows. First, at 3 km/h, does our estimator reach Type II precision in fewer than the reported 33 rounds? Second, at 30 km/h and above, does it keep a gain advantage while cutting-plane tracking lags? Third, does the soft ordinal loss with learned reliability beat hard cuts fed the same reports? Fourth, does the advantage vanish when HARQ outcomes are randomly shuffled across slots? This is a sanity check that the gain truly comes from the information in HARQ. Fifth, how much exploration does the beam policy need before HARQ helps beyond the dominant direction?
There is no reported result on Lacuna for any covariance estimator that uses HARQ or CQI feedback, so the comparison with the cutting-plane method's 33 rounds is the anchor for this work.
References
- Pushing The Limit of Type I Codebook For Fdd Massive Mimo Beamforming: A Channel Covariance Reconstruction Approach
- Massive MIMO with imperfect channel covariance information
- Accurate multi-behavior sequence-aware recommendation via graph convolution networks
- Micro-Behavior Encoding for Session-based Recommendation
- RBPR: Role-based Bayesian Personalized Ranking for Heterogeneous One-Class Collaborative Filtering
- Deep CNN based Channel Estimation for mmWave Massive MIMO Systems
- Joint Visibility Region and Channel Estimation for Extremely Large-Scale MIMO Systems