Recursive CSI Quantization of Time-Correlated MIMO Channels by Deep Learning Classification

Stefan Schwarz

REVIEW 3 major objections 4 minor 34 references

Recursive CSI Quantization of Time-Correlated MIMO Channels by Deep Learning Classification

T0 review · 3 major / 4 minor · reviewed 2026-08-27 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read A recursive, stage-wise Grassmannian quantizer with per-stage deep-learning classifiers can meet a 125-bit codebook's CSI distortion while sending fewer feedback bits on slowly varying MIMO channels.

desk verdict A competent extension of the author's recursive Grassmannian quantizer; the selective stage-update rule needs empirical validation before the overhead claims hold. read the letter →

arxiv 2009.13560 v1 pith:6IJ5N6ML submitted 2020-09-28 cs.IT cs.LGmath.IT

Stefan Schwarz This is my paper · ORCID

classification cs.ITcs.LGmath.IT

keywords CSIfeedbackGrassmannianquantizationrecursivemulti-stagedeeplearningclassificationtemporalchannelcorrelationselectivestageupdateMIMOchordaldistance

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

High-resolution channel-state feedback in frequency-division-duplex MIMO normally requires codebooks whose size grows exponentially with the number of antennas, making the search over codewords impractical. The paper's proposal is to replace that single large search with a recursive multi-stage Grassmannian quantizer, in which each stage handles a lower-dimensional subspace with a small codebook, and to replace each stage's search with a deep-learning classifier. It then adds a selective stage-update rule that exploits temporal correlation: when the channel has not moved much, only the later stages are refreshed. The central numerical claim is that on the Grassmann manifold G(32,1) (one-dimensional subspaces of a 32-dimensional space), the combination reaches the target average chordal-distance distortion of 0.06 with average feedback overhead below that of a single-stage 125-bit random vector quantizer at low Doppler, while staying comparable to differential and predictive Grassmannian quantization at moderate to high Doppler without on-the-fly codebook adaptation.

What carries the argument

The carrying mechanism is the recursive multi-stage Grassmannian quantizer, in which the channel subspace U[k] is quantized in R stages, each mapping an intermediate subspace of dimension d_{i-1} to a smaller subspace of dimension d_i. The stages are linked by subspace-quantization-based combining (SQBC) matrices B_i, which pass the residual subspace on to the next stage; with dimension step-size one, each stage becomes a one-dimensional Grassmannian quantization whose codebook is small enough for a DNN classifier to pick the entry. Before classification, the columns of B_i are phase-rotated so the first row is real, making the representation invariant to right-multiplication by unitary matrices. The selective stage-update rule is driven by a product-form distortion model: total squared chordal distance is written as one minus the product of per-stage terms, so for any number r of frozen stages the expected distortion can be estimated and compared with a target band set by parameters c_l and c_u. That estimate (Eq. 13) determines how few stages can be updated while keeping average distortion at the target.

What would settle it

Implement the same G(32,1) recursive quantizer with per-stage 6-bit DNN classifiers, but after freezing r stages compute the exact chordal distance of the resulting subspace rather than using the Eq. (13) estimate, and measure the average number of bits per channel use needed to hold average distortion at 0.06; if this measured overhead is materially higher than the Fig. 3 curve at low Doppler, the product-form selective-update model is the reason.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that recursive multi-stage quantization turns an intractable high-resolution CSI quantization problem into a set of small per-stage classification problems, and that temporal correlation can be exploited simply by deciding which stages to update. In the G(32,1) simulation, 31 stages with 6 bits per stage achieve the same 0.06 average chordal-distance distortion as a 125-bit single-stage RVQ. The DNN classifiers used per stage have roughly 90% classification accuracy, but the resulting distortion penalty is negligible because misclassification near codebook boundaries changes the quantized subspace very little. The selective stage-update rule then makes the average feedback overhead depend on Doppler: at normalized Doppler frequencies around 0.001, most time instants call for no update or only a few late-stage updates, while at high Doppler nearly all 31 stages are refreshed. A second simulation on a 6x2 antenna system shows recursive quantization with selective updates performing similarly to differential and predictive Grassmannian quantizers at moderate to high Doppler, but without their need to adapt the quantization codebook at every time instant.

Load-bearing premise

The results depend on the product-form distortion model of Eq. (13): total squared chordal distance is treated as one minus the product of per-stage terms, and this factorization is assumed to remain valid when the first r stages are frozen at their previous values, so the algorithm can trust its prediction of how many stages need updating.

Editorial extensions

If this is right

High-resolution CSI quantization for moderate antenna counts can avoid exhaustive search over 2^b-entry codebooks, since each of R stages needs only 2^{b/R} classes and can be handled by a shallow neural network.
On time-correlated channels, average feedback overhead can be reduced by updating only a subset of stages, and at low Doppler this overhead can drop below that of a single-stage quantizer targeting the same distortion.
The same fixed quantizer structure tracks channels adequately across Doppler frequencies, matching differential and predictive Grassmannian quantizers at moderate to high Doppler without adaptive codebooks.
Online complexity is dominated by computing the combining matrices B_i, because the codebook search itself is offloaded to offline-trained DNN classifiers.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The product-form estimate in Eq. (13) assumes per-stage errors combine independently; for channels with non-isotropic residual subspaces, such as measured channels with strong line-of-sight components, the selective-update rule's bit savings are likely to change, and the exact distortion after freezing stages should be checked.
Equal bit allocation across stages is acknowledged as suboptimal, so distributing more bits to early or frequently updated stages could reduce average feedback further at the same distortion; the paper leaves this optimization open.
Because 90% classification accuracy still yields negligible distortion penalty, the per-stage codebooks could probably be made smaller or the networks shallower before distortion rises; a sweep of per-stage bit counts would locate that threshold.
The hysteresis parameters c_l and c_u could themselves be learned or adapted per channel realization, potentially replacing the hand-tuned threshold rule with a learned stage-update policy.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Request a human review

A listed scientist reviews the paper for a fee and the review publishes here regardless of verdict. See the reviewers or get listed.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

This paper takes Schwarz's own recursive multi-stage Grassmannian quantizer and does two things: it replaces exhaustive-search per-stage quantization with DNN classifiers, and it adds a selective stage-update rule that exploits temporal correlation by freezing early stages. The DNN idea is sensible because per-stage codebooks are small (6 bits, 64 classes) while single-stage codebooks blow up. The selective update is a natural extension of the author's earlier selective CSI update. The combination is new relative to the cited end-to-end DNN feedback schemes.

What is solid: the recursive quantizer's complexity reduction is real, and Table II shows the DNN classifiers track exhaustive-search distortion closely across 31 stages on G(32,1). The comparison to differential and predictive quantizers for G(6,2) is fair enough and shows comparable distortion at moderate-to-high Doppler, with the recursive quantizer avoiding on-the-fly codebook adaptation. The paper is clearly written, builds on the author's own prior work, and does not oversell: it claims a trade-off, not a breakthrough.

The soft spot is Eq. (13). The product form for total chordal distance is exact for m=1 given the per-stage distances, but the rule substitutes unconditional isotropic average distortions for stages that are conditionally not updated. When the first r stages are frozen, the residual input to the remaining stages follows the temporally correlated trajectory, so the per-stage distortion may not match the i.i.d. training average. The paper never reports the actually achieved average distortion for the Fig. 3 setup, only the bit counts. Without that, the reader cannot tell whether the advertised 0.06 distortion target is met or whether the overhead comparison to 125-bit RVQ is apples-to-apples. If the conditional distortion is higher than assumed, the rule updates too few stages and the time-averaged distortion exceeds 0.06; if lower, the savings are conservative. Either way the central overhead claim is unverified as reported. This is an addressable problem — run the simulations, report the achieved distortion, maybe characterize the conditional distribution — so it is a revision issue, not a fundamental flaw.

There are also minor gaps: no error bars anywhere, DNN architecture details are thin (layer sizes, training hyperparameters), no code or seed, and only two simulation scenarios. These are standard for the venue but worth mentioning.

Who is this for? People working in limited-feedback FDD massive MIMO. It extends an established program rather than opening a new one, but it is a competent extension. I would send it to peer review, with the request that the authors validate Eq. (13) empirically and report the achieved distortion.

Referee Report

3 major / 4 minor

Summary. The manuscript proposes a recursive multi-stage Grassmannian quantizer for MIMO CSI feedback, in which each stage is implemented by a DNN classifier that maps the intermediate subspace matrix to a small codebook index. To exploit temporal channel correlation, the paper introduces a selective stage-update rule: at each time instant the quantizer freezes the first r stages to their previous values and only updates the remaining stages, where r is chosen using a predicted distortion formula with two hysteresis parameters. Simulations on G(6,2) compare the approach with differential and predictive Grassmannian quantizers, and simulations on G(32,1) compare the average feedback overhead with a theoretical 125-bit single-stage RVQ baseline at target chordal distortion 0.06. The paper claims that the DNN-based recursive quantizer reduces online complexity, performs comparably to differential/predictive quantization at moderate to high Doppler, and requires less feedback than single-stage RVQ at low Doppler.

Significance. If validated, the contribution is practically relevant: it offers a hybrid model-based/DNN quantizer that keeps per-stage codebooks small enough for neural classification while retaining the performance of recursive Grassmannian quantization, and it adds a lightweight temporal-correlation mechanism that avoids on-the-fly codebook adaptation. The paper has clear strengths: the comparisons use external baselines (differential/predictive quantizers and RVQ bounds), the DNN distortion results in Table II are close to exhaustive-search values, and the central performance metric, average chordal distance distortion, is standard for the limited-feedback literature. The main claims are plausible, but the selective stage-update rule relies on an unvalidated conditional-isotropy assumption, and the reported overhead savings in Fig. 3 are not accompanied by achieved-distortion verification. The load-bearing part of the overhead claim therefore needs additional empirical support before the result can be fully accepted.

major comments (3)

[Section III-B2, Eq. (13)] Please provide an empirical validation of Eq. (13): for the Fig. 3 scenario, report the actual average distortion achieved by the selective-update algorithm as a function of Doppler, together with the predicted distortion from Eq. (12), and verify that the 0.06 target is met. This would directly address whether the substitution of unconditional averages is accurate for temporally correlated inputs.
[Section IV, Figs. 3 and 4] Please add a figure or table showing the achieved average distortion versus Doppler for the G(32,1) setup, along with the number of Monte-Carlo runs and confidence intervals, so that the reader can confirm the advertised distortion target is actually met.
[Section IV, Fig. 1] Please provide the exact stage count, bit allocation, and hysteresis parameters used in Fig. 1, and state the numerical average-feedback values corresponding to the plotted recursive-quantizer points.

minor comments (4)

[Table I] The input dimension is written as "15 · 2d_{i-1}m" which is confusing: the factor 15 is unexplained, and the intended expression is probably the concatenation of real and imaginary parts of the vectorized input, i.e., dimension 2·d_{i-1}·m. Please clarify the notation.
[Table II] The table lists only odd stage indices (1,3,...,31) even though the text says 31 stages are used. Please state explicitly that even stages are omitted from the table for space, and give the distortion values for all stages or explain why the omitted values are unnecessary.
[Section III-C] The phase-rotation preprocessing step is described only verbally. A short equation defining the phase-normalized input matrix would improve reproducibility, especially because the DNN input format is central to the classification setup.
[Section IV] The paper reports neither the number of independent channel realizations nor the number of time samples used in the simulations. Adding this information would allow the reader to judge the statistical reliability of the curves in Figs. 1-4.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation; the recursive DNN quantizer is evaluated against external baselines and its parameters are measured, not fitted to force the outcome.

full rationale

The derivation is self-contained in the relevant sense. The recursive multi-stage structure is restated in Eqs. (7)-(11), with stage-distortion constants taken from the external source [32]; the paper does not define its target result into its inputs. Table II reports measured classification accuracy and distortion of the DNN stages, obtained by training on isotropically distributed inputs with labels from exhaustive search; these values are not fitted parameters chosen to enforce the 0.06 distortion target. The selective stage-update rule in Eqs. (12)-(13) uses those measured distortions as inputs to decide how many stages to update, and Fig. 3 reports the average feedback overhead of the resulting simulated algorithm, so the overhead is a measured output rather than a back-computed prediction. The comparison with differential and predictive Grassmannian quantizers is benchmarked against external algorithms from [21]/[22] using the same channel model, providing independent reference points. The unvalidated isotropy assumption in Eq. (13) - replacing conditional updated-stage distortion with unconditional averages - is a potential validation gap that could affect whether the time-averaged distortion actually meets the 0.06 target, but it is not a circular reduction: the paper does not use the claimed overhead saving as an input to the distortion model. Self-citations such as [26] and [33] describe prior components of the method, but the equations and external evaluations are stated in the paper itself, so these citations are not load-bearing in a circular sense.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim rests on standard RVQ bounds, the author's own recursive quantizer, and the assumption that the product-form distortion model remains valid for frozen and DNN-implemented stages. No new physical entities are introduced.

free parameters (3)

hysteresis thresholds c_u and c_l = c_u=2, c_l=1.5 in the Fig. 2 example
These tuning parameters control the trade-off between CSI update frequency and distortion; they are chosen by the author, not derived from first principles.
number of bits per stage b_i = 6 bits per stage for G(32,1); 7 bits for the G(6,2) example
Bit allocation is set to achieve a target average distortion, such as 0.06 or matching a differential quantizer. This is a design choice affecting the results.
DNN architecture and training hyperparameters = not fully specified
The network width, depth, dropout rate, training set size, and number of epochs are not reported, so the exact classifier performance is not reproducible.

assumptions (4)

standard math RVQ distortion bounds of Dai et al. [32] apply to each stage of the recursive quantizer
Eq. (11) uses the constant k_{d_{i-1},m,d_i} from [32] to predict per-stage average distortion; this is prior published theory.
domain assumption The recursive quantizer and its product distortion formula from [26] remain valid when stages are frozen
Eq. (13) extends the product form to partially frozen stages; the paper does not re-derive or validate this extension.
domain assumption Intermediate quantization inputs Bi[k] are approximately isotropically distributed
The RVQ formulas assume isotropic inputs; the paper does not measure the distribution of Bi[k] in the recursive pipeline.
domain assumption Channel follows stationary Gaussian process with known autocorrelation (Clarke's spectrum or Gauss-Markov)
The channel model in Section II is assumed; the selective update rule depends on this model only through the generation of test data.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Recursive CSI Quantization of Time-Correlated MIMO Channels by Deep Learning Classification." pith.science (2026). https://pith.science/paper/6IJ5N6ML

@misc{pith2026200913560,
  author       = {Pith},
  title        = {Pith review of: Recursive CSI Quantization of Time-Correlated MIMO Channels by Deep Learning Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6IJ5N6ML}},
  note         = {Machine review of arXiv:2009.13560}
}

read the original abstract

In frequency division duplex (FDD) multiple-input multiple-output (MIMO) wireless communications, limited channel state information (CSI) feedback is a central tool to support advanced single- and multi-user MIMO beamforming/precoding. To achieve a given CSI quality, the CSI quantization codebook size has to grow exponentially with the number of antennas, leading to quantization complexity, as well as, feedback overhead issues for larger MIMO systems. We have recently proposed a multi-stage recursive Grassmannian quantizer that enables a significant complexity reduction of CSI quantization. In this paper, we show that this recursive quantizer can effectively be combined with deep learning classification to further reduce the complexity, and that it can exploit temporal channel correlations to reduce the CSI feedback overhead.

Figures

Figures reproduced from arXiv: 2009.13560 by the authors.

**Figure 1.** Average quantization distortion on G (6, 2) versus normalized Doppler frequency. The differential and predictive quantizers provide 6 bits of feedback per time instant; the average number of feedback bits of the recursive multistage quantizer is written next to the data points. Utilizing these DNNs, the computational complexity of the quantizer is basically off-loaded to the offline training-phase of the DNNs. The … view at source ↗

**Figure 2.** Trace of the quantization error on G (6, 2) over time for the differential, predictive and recursive quantizers and a normalized Doppler of νd = 0.005. cℓ = 1.5. The black-dotted line in the figure corresponds to the maximally acceptable distortion cu d¯2 c , which determines when a stage update takes place. For predictive and differential quantization, the performance shows larger variations over time, with increas… view at source ↗

**Figure 4.** Relative frequency of the number of updated stages of [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

34 extracted references · 34 canonical work pages

[1]

An overview of limited feedback in wireless communication systems,

D. Love, R. Heath, Jr., V . Lau, D. Gesbert, B. Rao, and M. An drews, “An overview of limited feedback in wireless communication systems,” IEEE Journal on Selected Areas in Communications , vol. 26, no. 8, Oct. 2008

work page 2008
[2]

Noncoherent t rellis coded quantization: A practical limited feedback technique for m assive MIMO systems,

J. Choi, Z. Chance, D. Love, and U. Madhow, “Noncoherent t rellis coded quantization: A practical limited feedback technique for m assive MIMO systems,” IEEE Transactions on Communications , vol. 61, no. 12, pp. 5016–5029, December 2013

work page 2013
[3]

Multiple-antenna transmission with limited feedback in device-to-device networks,

J. Park and R. W. Heath, “Multiple-antenna transmission with limited feedback in device-to-device networks,” IEEE Wireless Communications Letters, vol. 5, no. 2, pp. 200–203, 2016

work page 2016
[4]

Limited feedback hybrid beamformin g for multi-mode transmission in wideband millimeter wave chann el,

G. Kwon and H. Park, “Limited feedback hybrid beamformin g for multi-mode transmission in wideband millimeter wave chann el,” IEEE Transactions on Wireless Communications, vol. 19, no. 6, pp. 4008–4022, 2020

work page 2020
[5]

Incremental Grassmannian f eedback schemes for multi-user MIMO systems,

A. Medra and T. N. Davidson, “Incremental Grassmannian f eedback schemes for multi-user MIMO systems,” IEEE Transactions on Signal Processing, vol. 63, no. 5, pp. 1130–1143, 2015

work page 2015
[6]

Cube-split: Structured quantizers on the grassmannian of lines,

A. Decurninge and M. Guillaud, “Cube-split: Structured quantizers on the grassmannian of lines,” in IEEE Wireless Communications and Networking Conference, pp. 1–6, March 2017

work page 2017
[7]

Hybrid beamforming with selection for multiuser massive M IMO systems,

V . V . Ratnam, A. F. Molisch, O. Y . Bursalioglu, and H. C. Papadopoulos, “Hybrid beamforming with selection for multiuser massive M IMO systems,” IEEE Transactions on Signal Processing , vol. 66, no. 15, pp. 4105–4120, 2018

work page 2018
[8]

Grassmannian prod uct codebooks for limited feedback massive MIMO with two-tier p recoding,

S. Schwarz, M. Rupp, and S. Wesemann, “Grassmannian prod uct codebooks for limited feedback massive MIMO with two-tier p recoding,” IEEE Journal of Selected Topics in Signal Processing , vol. 13, no. 5, pp. 1119–1135, Sep. 2019

work page 2019

Show all 34 references

[9]

Constructing packings in Grassmannian manifolds via alternating projec tion,

I. S. Dhillon, R. Heath, Jr., T. Strohmer, and J. A. Tropp, “Constructing packings in Grassmannian manifolds via alternating projec tion,” ArXiv e-prints, Sept. 2007

2007
[10]

A coherence-based alg orithm for optimizing rank-1 Grassmannian codebooks,

H. E. A. Laue and W. P . du Plessis, “A coherence-based alg orithm for optimizing rank-1 Grassmannian codebooks,” IEEE Signal Processing Letters, vol. 24, no. 6, pp. 823–827, 2017

2017
[11]

Constructing Grassm annian frames by an iterative collision-based packing,

B. Tahir, S. Schwarz, and M. Rupp, “Constructing Grassm annian frames by an iterative collision-based packing,” IEEE Signal Processing Letters , vol. 26, no. 7, pp. 1056–1060, July 2019

2019
[12]

A scalable fr amework for CSI feedback in FDD massive MIMO via DL path aligning,

X. Luo, P . Cai, X. Zhang, D. Hu, and C. Shen, “A scalable fr amework for CSI feedback in FDD massive MIMO via DL path aligning,” IEEE Trans. on Signal Processing , vol. 65, no. 18, pp. 4702–4716, Sep. 2017

2017
[13]

A uniﬁed transmissi on strategy for TDD/FDD massive MIMO systems with spatial basis expansion m odel,

H. Xie, F. Gao, S. Zhang, and S. Jin, “A uniﬁed transmissi on strategy for TDD/FDD massive MIMO systems with spatial basis expansion m odel,” IEEE Transactions on V ehicular Technology , vol. 66, no. 4, pp. 3170– 3184, April 2017

2017
[14]

Robust full-dimension MIMO transmission based on limited feedback angular-domain CSIT,

S. Schwarz, “Robust full-dimension MIMO transmission based on limited feedback angular-domain CSIT,” EURASIP Journal on Wireless Communications and Networking , vol. 2018, no. 1, pp. 1–20, Mar 2018

2018
[15]

Deep learning-base d CSI feedback approach for time-varying massive MIMO channels,

T. Wang, C. Wen, S. Jin, and G. Y . Li, “Deep learning-base d CSI feedback approach for time-varying massive MIMO channels, ” IEEE Wireless Communications Letters, vol. 8, no. 2, pp. 416–419, April 2019

2019
[16]

Exploiting bi-direction al channel reciprocity in deep learning for low rate massive MIMO CSI fe edback,

Z. Liu, L. Zhang, and Z. Ding, “Exploiting bi-direction al channel reciprocity in deep learning for low rate massive MIMO CSI fe edback,” IEEE Wireless Communications Letters, vol. 8, no. 3, pp. 889–892, 2019

2019
[17]

Deep learni ng- based limited feedback designs for MIMO systems,

J. Jang, H. Lee, S. Hwang, H. Ren, and I. Lee, “Deep learni ng- based limited feedback designs for MIMO systems,” IEEE Wireless Communications Letters , vol. 9, no. 4, pp. 558–561, 2020

2020
[18]

Grassmannian predictive co ding for delayed limited feedback MIMO systems,

T. Inoue and R. Heath, Jr., “Grassmannian predictive co ding for delayed limited feedback MIMO systems,” in 47th Annual Allerton Conference on Communication, Control, and Computing , Oct. 2009

2009
[19]

Different ial feedback of MIMO channel Gram matrices based on geodesic curves,

D. Sacristan-Murga and A. Pascual-Iserte, “Different ial feedback of MIMO channel Gram matrices based on geodesic curves,” IEEE Trans. on Wireless Communications, vol. 9, no. 12, pp. 3714–3727, Dec. 2010

2010
[20]

Grassmannian differenti al limited feedback for interference alignment,

O. El Ayach and R. Heath, Jr., “Grassmannian differenti al limited feedback for interference alignment,” IEEE Transactions on Signal Processing, vol. 60, no. 12, pp. 6481–6494, Dec 2012

2012
[21]

Adaptive quanti zation on the Grassmann-manifold for limited feedback multi-user MIMO s ystems,

S. Schwarz, R. Heath, Jr., and M. Rupp, “Adaptive quanti zation on the Grassmann-manifold for limited feedback multi-user MIMO s ystems,” in 38th International Conference on Acoustics, Speech and Sig nal Processing, pp. 5021 – 5025, V ancouver, Canada, May 2013

2013
[22]

Predictive quantization on the Stiefel manifold,

S. Schwarz and M. Rupp, “Predictive quantization on the Stiefel manifold,” IEEE Signal Processing Letters , vol. 22, no. 2, pp. 234–238, 2015

2015
[23]

Spatio-temporal co rrelated chan- nel feedback for massive MIMO systems,

Y . Ge, Z. Zeng, T. Zhang, and Y . Liu, “Spatio-temporal co rrelated chan- nel feedback for massive MIMO systems,” in IEEE/CIC International Conference on Communications in China , pp. 1–5, 2018

2018
[24]

Mimo channel in for- mation feedback using deep recurrent network,

C. Lu, W. Xu, H. Shen, J. Zhu, and K. Wang, “Mimo channel in for- mation feedback using deep recurrent network,” IEEE Communications Letters, vol. 23, no. 1, pp. 188–191, 2019

2019
[25]

Spatio-temporal representation with d eep neural re- current network in mimo csi feedback,

X. Li and H. Wu, “Spatio-temporal representation with d eep neural re- current network in mimo csi feedback,” IEEE Wireless Communications Letters, vol. 9, no. 5, pp. 653–657, 2020

2020
[26]

Reduced complexity recursive g rassmannian quantization,

S. Schwarz and M. Rupp, “Reduced complexity recursive g rassmannian quantization,” IEEE Signal Processing Letters , vol. 27, pp. 321–325, 2020

2020
[27]

A statistical theory of mobile radio rece ption,

R. H. Clarke, “A statistical theory of mobile radio rece ption,” Bell Systems Technical Journal , vol. 47, pp. 957–1000, 1968

1968
[28]

MIMO broadcast channels with ﬁnite-rate fe edback,

N. Jindal, “MIMO broadcast channels with ﬁnite-rate fe edback,” IEEE Transactions on Information Theory , vol. 52, no. 11, p. 5, Nov. 2006

2006
[29]

Limited feedback-based bl ock diagonaliza- tion for the MIMO broadcast channel,

N. Ravindran and N. Jindal, “Limited feedback-based bl ock diagonaliza- tion for the MIMO broadcast channel,” IEEE Journal on Selected Areas in Communications , vol. 26, no. 8, pp. 1473 –1482, Oct. 2008

2008
[30]

Limited feedback for interf erence align- ment in the K-user MIMO interference channel,

M. Rezaee and M. Guillaud, “Limited feedback for interf erence align- ment in the K-user MIMO interference channel,” in Proc. Information Theory W orkshop, pp. 1–5, Lausanne, Suisse, September 2012

2012
[31]

Interference alig nment under limited feedback for MIMO interference channels,

R. Krishnamachari and M. V aranasi, “Interference alig nment under limited feedback for MIMO interference channels,” IEEE Transactions on Signal Processing , vol. 61, no. 15, pp. 3908–3917, Aug 2013

2013
[32]

Quantization bounds on Gra ssmann manifolds and applications to MIMO communications,

W. Dai, Y . Liu, and B. Rider, “Quantization bounds on Gra ssmann manifolds and applications to MIMO communications,” IEEE Trans. on Information Theory , vol. 54, no. 3, pp. 1108 –1123, March 2008

2008
[33]

Subspace quantization based co mbining for limited feedback block-diagonalization,

S. Schwarz and M. Rupp, “Subspace quantization based co mbining for limited feedback block-diagonalization,” IEEE Transactions on Wireless Communications, vol. 12, no. 11, pp. 5868–5879, 2013

2013
[34]

Simulation models with correct statistical properties for Rayleigh fading channels,

Y . R. Zheng and C. Xiao, “Simulation models with correct statistical properties for Rayleigh fading channels,” IEEE Transactions on Com- munications, vol. 51, no. 6, pp. 920 – 928, June 2003

2003

Pith tools

Reviewed August 27, 2026 · model on record in the stance chip above.

[1] [1]

An overview of limited feedback in wireless communication systems,

D. Love, R. Heath, Jr., V . Lau, D. Gesbert, B. Rao, and M. An drews, “An overview of limited feedback in wireless communication systems,” IEEE Journal on Selected Areas in Communications , vol. 26, no. 8, Oct. 2008

work page 2008

[2] [2]

Noncoherent t rellis coded quantization: A practical limited feedback technique for m assive MIMO systems,

J. Choi, Z. Chance, D. Love, and U. Madhow, “Noncoherent t rellis coded quantization: A practical limited feedback technique for m assive MIMO systems,” IEEE Transactions on Communications , vol. 61, no. 12, pp. 5016–5029, December 2013

work page 2013

[3] [3]

Multiple-antenna transmission with limited feedback in device-to-device networks,

J. Park and R. W. Heath, “Multiple-antenna transmission with limited feedback in device-to-device networks,” IEEE Wireless Communications Letters, vol. 5, no. 2, pp. 200–203, 2016

work page 2016

[4] [4]

Limited feedback hybrid beamformin g for multi-mode transmission in wideband millimeter wave chann el,

G. Kwon and H. Park, “Limited feedback hybrid beamformin g for multi-mode transmission in wideband millimeter wave chann el,” IEEE Transactions on Wireless Communications, vol. 19, no. 6, pp. 4008–4022, 2020

work page 2020

[5] [5]

Incremental Grassmannian f eedback schemes for multi-user MIMO systems,

A. Medra and T. N. Davidson, “Incremental Grassmannian f eedback schemes for multi-user MIMO systems,” IEEE Transactions on Signal Processing, vol. 63, no. 5, pp. 1130–1143, 2015

work page 2015

[6] [6]

Cube-split: Structured quantizers on the grassmannian of lines,

A. Decurninge and M. Guillaud, “Cube-split: Structured quantizers on the grassmannian of lines,” in IEEE Wireless Communications and Networking Conference, pp. 1–6, March 2017

work page 2017

[7] [7]

Hybrid beamforming with selection for multiuser massive M IMO systems,

V . V . Ratnam, A. F. Molisch, O. Y . Bursalioglu, and H. C. Papadopoulos, “Hybrid beamforming with selection for multiuser massive M IMO systems,” IEEE Transactions on Signal Processing , vol. 66, no. 15, pp. 4105–4120, 2018

work page 2018

[8] [8]

Grassmannian prod uct codebooks for limited feedback massive MIMO with two-tier p recoding,

S. Schwarz, M. Rupp, and S. Wesemann, “Grassmannian prod uct codebooks for limited feedback massive MIMO with two-tier p recoding,” IEEE Journal of Selected Topics in Signal Processing , vol. 13, no. 5, pp. 1119–1135, Sep. 2019

work page 2019