Utafiti huu unachunguza jinsi ya kubadilisha taarifa ya hali ya kituo yenye dimensionality kubwa (Channel State Information, CSI), ambayo inahitaji kurejeshwa kwa base station katika mifumo ya Frequency Division Duplex (FDD) massive MIMO, kuwa representation ndogo na ya haraka zaidi. Mbinu ya TAP (Tap-Assisted Parametric CSI Compression) iliyotengenezwa na waandishi hutumia lightweight 1D neural network kutoa temporal information kutoka past na current channel frequency responses, kukadiria dominant multipath propagation delays kwa single pass, kupata fractional delays kwa sub-grid interpolation, na kukokotoa complex spatial amplitudes kwa closed-form least-squares solution. Katika mazingira matano tofauti ya simulation yanayotegemea au kuendana na 3GPP, TAP inaonyesha CFR-NMSE performance yenye nguvu huku model size ikiwa chini ya megabyte moja na reported inference times zikiwa takriban 0,48–0,64 ms. Hata hivyo, matokeo yanategemea hasa ray tracing na stochastic channel simulations; utafiti hautoi real over-the-air validation na mathematical structure ya sasa imewekewa kikomo na far-field plane-wave assumption.

Wazo kuu la TAP ni kuunganisha physical interpretability ya classical compressed-sensing methods na fast inference ya deep learning. Classical Orthogonal Matching Pursuit (OMP) hutafuta dominant paths katika delay space kwa mfululizo, hivyo inaweza kuwa sahihi lakini ni iterative na computationally expensive. TAP inalenga kutumia temporal continuity ya past CSI samples kutambua likely physical delays moja kwa moja kwa neural network, kisha kufanya least-squares solution mara moja tu. Kwa hiyo network hailazimiki kuwa large encoder-decoder inayokariri channel nzima, bali inalenga kupata physical delay parameters.

Kwa mtazamo wa Uturuki, maana ya utafiti ni kwamba mbinu haihusiani moja kwa moja na network ya nchi maalum. Approach hii ya kupunguza CSI feedback burden katika FDD massive MIMO na wireless systems za baadaye zenye antenna nyingi inaweza pia kuchunguzwa katika research and development nchini Uturuki. Hata hivyo, matokeo ya UMa, InF, UMi, InH na CDL-A katika paper hayawezi kuhamishwa moja kwa moja kwenye real operator sites nchini Uturuki kisayansi. Local frequency allocations, real base-station arrays, terminal hardware, mobility, propagation properties zinazotokana na majengo na terrain, pamoja na real CSI-RS measurements, vinahitaji validation tofauti.

Tatizo kuu la utafiti ni nini?

Katika FDD massive MIMO, base station (BS) haiwezi kupata downlink channel moja kwa moja kutoka uplink channel. Kwa hiyo User Equipment (UE) lazima ikadirie downlink channel, icompress CSI yenye dimensionality kubwa sana, na kuirudisha kwa base station kupitia uplink control channel yenye capacity ndogo. Idadi ya antenna na OFDM subcarriers inapoongezeka, full Channel Frequency Response (CFR) matrix pia huongezeka.

Utafiti unaweka approaches zilizopo katika makundi matatu. Compressed-sensing methods kama OMP, LASSO na TVAL3 hutegemea physical sparsity katika delay domain lakini iterative computation huleta latency. Deep-learning methods kama familia ya CsiNet zinaweza kutoa fast inference lakini zinaweza kufungwa na antenna geometries maalum na statistics za training environments. 3GPP Type-1 na Type-2 codebooks ni lightweight computationally, lakini katika utafiti zinalinganishwa na reconstruction error kubwa zaidi.

Physical channel inamodeliwaje?

Utafiti unazingatia single-cell, FDD massive MIMO na OFDM system. Base station ina N_r × N_c uniform planar antenna array na P polarization. Total antenna port count ni:

\[ N_{\mathrm{ant}} = P N_r N_c \]

.

Kwa subcarrier moja na antenna element moja, channel frequency response inamodeliwa kama jumla ya multipath components:

\[ H[n,m]=\sum_{l=1}^{L}\alpha_{l,m}e^{-j2\pi f_n\tau_l} \]

Hapa \(\alpha_{l,m}\) ni complex spatial amplitude ya path ya l kwenye antenna ya m; \(\tau_l\) ni physical propagation delay; na \(f_n\) ni frequency ya subcarrier ya n. Critical physical assumption ya utafiti ni kwamba katika far field delay ya multipath component ileile inaweza kuchukuliwa kuwa common katika antenna array, huku tofauti baina ya antennas zikiwa hasa kwenye complex spatial amplitudes.

Inverse fast Fourier transform (IFFT) ikitumika kwenye frequency axis, channel huhamishwa kwenda delay domain kama Channel Impulse Response (CIR). Kwa kuwa energy hukusanyika karibu na idadi ndogo ya dominant physical paths, CIR ina sparse structure. TAP hutumia moja kwa moja sifa hii ya kimwili.

Kwa nini classical OMP si ya haraka vya kutosha?

Katika kila iteration, OMP hulinganisha current residual signal na delay dictionary, huchagua strongest atom, huongeza active dictionary, hutatua least-squares problem, na hukokotoa residual upya. Kwa hiyo path count inapoongezeka, IFFT, correlation na matrix solving hurudiwa tena na tena.

Figure 2 ya utafiti inaonyesha kwamba path count inapoongezeka, NMSE inaboreka lakini processing time pia huongezeka kwa kiasi kikubwa. Text inataja takriban 1000 ms kwa K = 10. Hata hivyo, Table IV ya baadaye inaripoti 630,15 ms kwa OMP (K = 10). Kwa kuwa namba hizi mbili zinatofautiana ndani ya source moja, hazipaswi kuunganishwa kuwa exact OMP latency moja; zote zinaonyesha kwamba mbinu iko juu sana ya sub-millisecond target.

Kwa nini kuchagua IFFT peaks mara moja si suluhisho?

Waandishi wanaeleza sababu mbili kuu. Ya kwanza ni spectral leakage. Limited bandwidth hufanya kama rectangular window katika frequency domain na kutengeneza sinc-shaped main lobe na side lobes katika delay domain. Side lobe ya strong physical path inaweza kuchaguliwa vibaya kama weak real path.

Tatizo la pili ni grid mismatch, yaani physical delays hazidondoki moja kwa moja kwenye digital delay grid. Real delay ikiwa kati ya grid points mbili, energy husambaa kwenye neighboring atoms. Oversampling ya dictionary huongeza resolution lakini pia hufanya neighboring atoms kuwa highly correlated. Hivyo physical path moja inaweza kuonekana kama paths kadhaa tofauti.

Kwa nini delay error ni muhimu sana?

Utafiti hauishii kwenye empirical comparison; pia unachambua kihisabati jinsi delay-estimation error inavyogeuka kuwa CFR-reconstruction error. Kwa physical path moja, ikiwa true delay ni \(\tau\) na estimate ni \(\hat{\tau}=\tau+\Delta\tau\), least-squares solution inaweza kurekebisha amplitude na constant phase lakini haiwezi kuondoa frequency-dependent phase slope inayotokana na delay error.

Upper bound iliyotolewa ni:

\[ \mathrm{CFR\!-\!NMSE} \leq \frac{4\pi^2 E_f}{N_{\mathrm{sc}}}\Delta\tau^2 \]

. Hapa:

• \(N_{\mathrm{sc}}\): idadi ya OFDM subcarriers.
• \(\Delta\tau\): physical delay-estimation error.
• \(E_f=\sum_k f_k^2\): inawakilisha spread energy ya frequency points zilizotumika.

Ujumbe mkuu ni kwamba athari ya delay error hukua quadratic na katika wideband systems hata delay error ndogo inaweza kugeuka kuwa CFR error kubwa. Katika multipath case, error pia huweighted na dominant path amplitudes; kwa hiyo delay error ndogo kwenye strong line-of-sight component ni muhimu hasa.

TAP inafanya nini tofauti?

TAP inabadilisha iterative atom search ya classical OMP na lightweight 1D convolutional neural network inayoitwa DelayNet. Network haitumii current CFR pekee, bali temporal window yenye current na past CFR samples. Kwanza Hamming window na IFFT hutumika kwenye subcarrier axis ili kuhamisha data kwenda delay domain. Temporal sequence hutoa information inayosaidia kutenganisha coherent motion ya physical paths na random noise au transient side-lobe structures.

DelayNet hutoa multi-label probability heatmap katika delay axis. Architecture hutumia antenna averaging na depthwise separable convolution badala ya fully connected layers zinazotegemea moja kwa moja antenna dimensions. Hivyo lengo ni kuzuia trainable parameter count kukua pamoja na row na column count za antenna array.

Fractional delay inapatikanaje?

Katika detailed method section, parabola inafit kwenye probability values tatu karibu na top K local peaks ili kukokotoa sub-grid offset:

\[ \Delta_i = \frac{1}{2} \frac{p_{i-1}-p_{i+1}} {p_{i-1}-2p_i+p_{i+1}} \]

Kisha physical delay hupatikana kama:

\[ \hat{\tau}_k=(i+\Delta_i)\Delta t \]

. Operesheni hii inazuia delay kubaki kwenye discrete grid points pekee na, kwa kuwa gradient inaweza kurudi kwenye DelayNet, inakuwa sehemu ya end-to-end trainable structure.

Source-internal note: katika contributions list ya paper, sub-grid mechanism hii inaelezwa kama “symmetric linear power-weighted center-of-mass extractor”, huku detailed method section ikitoa parabolic interpolation equation hapo juu. Makala hii inatumia Equation 23 iliyofafanuliwa wazi kihisabati na haiunganisha kimya kimya definitions hizo mbili kama kwamba ni method moja.

Spatial amplitudes zinakokotolewaje?

Baada ya continuous delays kupatikana, Fourier dictionary inajengwa kama:

\[ A[n,k]=e^{-j2\pi f_n\hat{\tau}_k} \]

. Complex path amplitudes kwa antennas zote hukokotolewa kwa single regularized least-squares solution:

\[ \hat{X}=(A^HA+\lambda I)^{-1}A^H H_{\mathrm{tgt}} \]

Hakuna trainable parameter mpya inayotumika katika hatua hii. Katika sample calculation ya waandishi, kwa K = 20, N_sc = 1620 na N_ant = 64, algebraic recovery hii ni takriban 18 MFLOP; pipeline nzima pamoja na DelayNet imehesabiwa chini ya 153 MFLOP.

Kwa nini neural decoder haihitajiki kwenye base station?

Feedback payload ya TAP inaundwa na physical parameters: estimated delays, normalized spatial amplitudes na scale information. Base station hujenga upya continuous Fourier dictionary ileile na kutekeleza:

\[ \hat{H}=A\hat{X} \]

. Kwa hiyo hakuna separate CsiNet-like neural decoder inayohitajika kwenye base station. “Decoder-free” hapa haimaanishi channel haireconstructwi kabisa; inamaanisha reconstruction inafanywa kwa algebraic matrix operation badala ya learned neural decoder.

Training loss inafafanuliwaje?

TAP inafundishwa end-to-end kwa objective function yenye components mbili. Component ya kwanza ni MSE loss kati ya target heatmap iliyotengenezwa kutoka real physical delays na DelayNet output. Kwa kila real path, Gaussian distribution inayoweighted na amplitude yake inawekwa. Component ya pili ni CFR-NMSE loss ya final full-CFR reconstruction. Kwa hiyo network haijifunzi tu kuonyesha delay peaks, bali pia contribution ya delays hizo kwenye final channel reconstruction.

Experimental environments zinatofautianaje?

Evaluation inahusisha propagation environments tano kuu: Sionna RT-based Urban Macrocell (UMa) na Indoor Factory (InF), DeepMIMO-based Urban Microcell (UMi) na Indoor Hotspot (InH), pamoja na 3GPP CDL-A inayozalishwa kupitia Sionna-PHY. Utafiti pia una scalability evaluation tofauti kwenye 1×32 ULA antenna geometry.

Katika sehemu kubwa ya main tests, 1620 subcarriers na 60 kHz subcarrier spacing hutumiwa. Carrier frequency ni 2,14 GHz kwa UMa, InF na CDL-A; 3,5 GHz kwa UMi; na 2,5 GHz kwa InH. 8×8 dual-polarized arrays hutumiwa katika UMa na InF, na 8×8 single-polarized arrays katika UMi na InH. Temporal window length imewekwa 9 kwa UMa, InF, UMi na InH, na 1 kwa CDL-A.

Channel structure inabadilikaje baina ya environments?

Figures 4 na 5 zinaonyesha kwamba indoor na outdoor delay profiles hazifanani. Katika indoor environment energy husambaa kwenye multipath components nyingi zaidi, na delay components nyingi zinaweza kuhitajika kukamata %95 ya total energy. Katika outdoor UMa structure ni sparse zaidi kwa wastani, lakini distant reflections zinaweza kurefusha tail ya delay distribution.

Kwa hiyo TAP haitegemei fixed K path count kabisa; maximum estimated path count inarekebishwa kulingana na broad environment category. Mfano, K_max = 20 kwa UMa, K_max = 90 kwa InF, na K_max = 40 kwa CDL-A.

Matokeo yanayoungwa mkono na utafiti

• TAP inaweza kufanya high-accuracy CSI compression katika evaluated simulated environments kwa physics-informed single-pass delay estimation.
• Model size hubaki chini ya 1 MB katika TAP configurations zilizotest.
• Inference times zimeripotiwa takriban 0,48–0,64 ms kwenye TAP rows.
• Transfer experiments kwa propagation environments tofauti zimefanywa bila post-training fine-tuning.
• Katika zero-shot transfer kwa antenna arrays tofauti, reported CFR-NMSE degradation imebaki chini ya 3 dB kwenye geometries zilizochunguzwa.
• Low CFR-NMSE ya TAP imegeuka kuwa spectral-efficiency results zilizo karibu na perfect-CSI curve katika CDL-A system-level experiment.
• Robustness experiments chini ya quantization na CSI-RS estimation noise pia zimefanywa.

Matokeo ambayo utafiti hauungi mkono au bado haujathibitisha

• Commercial field success katika real operator network haijaonyeshwa.
• Sim-to-real transfer haijathibitishwa kwa real over-the-air channel measurements.
• Current TAP structure haijaonyeshwa kuwa valid kwa near-field XL-MIMO.
• FR2 millimeter-wave na hybrid beamforming conditions hazijathibitishwa na current results.
• Tested zero-shot generalization si proof ya universal generalization kwa possible channel environments zote.
• Sub-1-MB model size na sub-millisecond latency si guarantee kwa UE platforms zote nje ya software/hardware na simulation conditions za paper.

Mbinu na Matokeo ya Utafiti

Metric kuu ya evaluation: CFR-NMSE

Badala ya kutumia tu “patch NMSE” inayokokotolewa kwenye cropped delay-angle patch ya CsiNet, utafiti unatumia CFR-NMSE kwenye full channel frequency response matrix:

\[ \mathrm{NMSE} = 10\log_{10} \left( \frac{\lVert H-\hat{H}\rVert_F^2} {\lVert H\rVert_F^2} \right) \]

Tofauti hii ni muhimu. CsiNet-like method inapocrop delay-angle patch, channel energy iliyo nje ya patch husetwa zero. Waandishi wanaonyesha kupitia Parseval relation kwamba real CFR error inajumuisha reconstruction error ya network ndani ya patch pamoja na energy iliyotupwa kabisa nje ya patch. Kwa hiyo comparisons katika utafiti zinafanywa kwa full CFR-NMSE.

Reported TAP performance katika environments tano

Values zifuatazo ni results za Table III. CFR-NMSE iko katika dB; value iliyo negative zaidi inaonyesha reconstruction error ndogo.

Ujumbe wa Table III si tu kwamba mean values ni ndogo. %92,71 ya UMa samples zina CFR-NMSE bora kuliko −10 dB, huku katika environments nyingine nne rate hii ikiripotiwa %99,65–%100. Hata hivyo, environments hazina same physical conditions wala same compression ratios; kwa hiyo rows hazipaswi kusomwa moja kwa moja kama ranking ya “environment ipi ni rahisi zaidi?”.

Trade-off kati ya compression quality na feedback load

Katika Figure 8, TAP hutengeneza Pareto curve kwa compression ratios tofauti. Katika UMa, TAP inaripoti −21,22 dB NMSE kwa CR = 0,0124 na bado −18,10 dB inapopunguzwa hadi CR = 0,0012. CsiNet-UPA counterpart katika paper ni −9,14 dB kwa CR = 0,0224.

Katika InF, CsiNet-UPA inatoa −7,82 dB kwa CR = 0,0105, huku TAP ikitoa −10,99 dB kwa similar CR = 0,0112 na −20,08 dB kwa CR = 0,0279. Comparison hii inaonyesha TAP inalenga trade-off ya error na feedback load kwa pamoja, si error value pekee.

Model size, latency na computational load

Table IV ni mojawapo ya tables muhimu zaidi za utafiti kwa kutathmini reconstruction accuracy pamoja na computational cost.

Table hii inaweka dai la TAP kwa usahihi zaidi: TAP haitoi absolute lowest NMSE katika kila comparison. Kwa mfano, −21,22 dB katika TAP/UMa ina maana error kubwa kuliko −31,16 dB ya OMP; lakini OMP ina reported latency ya 630,15 ms dhidi ya 0,6239 ms ya TAP. LASSO pia ina strong reconstruction accuracy ya −30,22 dB, lakini kwa CR = 0,25 na processing time ya 88648,56 ms ina computational/feedback cost tofauti kabisa. Kwa hiyo contribution kuu ya utafiti si tu “lowest error”, bali shared balance ya accuracy, latency, model size na compression load.

Model size ya CsiNet-UPA/UMa ni 657,1579 MB wakati TAP-UMa ni 0,9921 MB. Kauli ya waandishi ya “takriban mara 660 ndogo” inatokana na size difference hii.

Generalization kwa antenna geometries tofauti

Katika Figure 10, model inatumika kwenye 4×16, 16×4 na UPA sizes nyingine tofauti na training antenna geometry. Kwa mujibu wa utafiti, CFR-NMSE degradation wakati wa kuhamia mismatched geometries katika main training environments tatu imebaki chini ya 3 dB dhidi ya matched geometry. Katika 1×32 ULA experiment ya Table II, CsiNet-UPA inatoa mean −15,54 dB, huku TAP models trained katika environments tofauti zikitoa mean CFR-NMSE kati ya −25,88 na −37,01 dB.

Zero-shot generalization baina ya environments

Figure 9 ina 5×5 matrix ya training environments tano na evaluation environments tano. Kila TAP model inafundishwa kwenye source environment moja na kuendeshwa kwenye environments nyingine bila retraining. Kwa mfano, model trained kwenye InF inaripoti −47,16 dB CFR-NMSE katika UMi evaluation; model trained kwenye UMa inaripoti −26,56 dB katika InH evaluation.

Matokeo haya yanatoa strong simulation evidence kwa zero-shot environment transfer. Hata hivyo, evaluation bado imefanywa baina ya data-generation environments zilizofafanuliwa katika paper; generalization kwa real geographic field au independent operator network haijatest experimentally.

System-level spectral efficiency

Katika Figure 11, reconstructed CSI katika CDL-A environment inatumika kutengeneza linear spatial precoder kwenye base station. TAP curve inafuata kwa karibu “Perfect CSI” upper bound katika SNR range iliyochunguzwa. Katika CsiNet-UPA na baadhi ya methods nyingine, CFR amplitude na phase distortions hugeuka kuwa spectral-efficiency loss kubwa zaidi.

Experiment hii inaonyesha kwamba low CFR-NMSE si abstract error metric pekee, bali pia inaonekana katika downlink spectral efficiency ndani ya simulation framework hiyo hiyo. Hata hivyo, result si real user data-rate measurement bali system-level simulation inayotokana na reconstructed CSI.

Nini kinatokea chini ya quantization?

Katika real feedback, continuous amplitude values lazima zibadilishwe kuwa finite bit sequence. Figure 12 inatumia uniform scalar quantization kwenye spatial amplitudes kwa bit widths tofauti. Kwa kuwa unquantized error floor ya TAP ni takriban −35,4 dB, impact ya quantization noise inaonekana zaidi kadri bit count inavyopungua. Hata hivyo, katika results za waandishi TAP inadumisha lower absolute CFR-NMSE kuliko compared baselines hata karibu na 4–6 bit/latent dimension.

Robustness kwa CSI-RS estimation noise

Figure 13 inachunguza method wakati initial channel estimate inayofika UE ina noise. Katika UMa na InF, TAP performance inadumishwa kwa kiasi kikubwa hata kwenye low CSI-RS SNR. Waandishi wanaeleza hili kwa mechanisms mbili: spatial averaging juu ya common physical delays za antenna elements na temporal context kutoka past CFR samples.

Ili kujaribu explanation hii, temporal window length kwenye CDL-A test imewekwa makusudi W = 1. Faster degradation ya CDL-A curve inaunga mkono, ndani ya ablation logic ya utafiti, kwamba temporal history inachangia sana noise robustness.

Mathematical limitation muhimu zaidi: near-field XL-MIMO

Current TAP model hutumia common delay moja kati ya antennas kwa kila physical path chini ya far-field assumption. Katika near-field XL-MIMO, kwa sababu ya spherical wavefront, arrival delay ya path ileile inaweza kutofautiana kwa antenna element:

\[ [h_{l,m}]_k = \alpha_{l,m} \exp\left[-j2\pi f_k(\tau_l+\Delta\tau_{l,m})\right] \]

TAP hutumia common \(\hat{\tau}_l\), wakati LS solver inaweza kurekebisha tu frequency-independent complex phase/amplitude coefficient kwa kila antenna. Residual phase error iliyotolewa katika source ni:

\[ \epsilon_{m,k} = -2\pi f_k\Delta\tau_{l,m} -\theta_m \]

. Term ya kwanza ni frequency-dependent phase slope; ya pili ni constant phase shift ambayo LS inaweza kutumia. Kwa hiyo LS solution haiwezi kuondoa phase slope. Waandishi wanakubali wazi kwamba current TAP architecture haiwezi kupanuliwa moja kwa moja kwa near-field XL-MIMO.

Tathmini ya jumla ya kisayansi

Strength kubwa zaidi ya utafiti ni kwamba neural network haijatumika kucompress CSI matrix nzima kwenye abstract latent space; badala yake inalenga estimation ya physically meaningful multipath delays. Delays zinatoka kwenye neural network, complex amplitudes kwenye analytical LS solution, na base-station reconstruction kwenye physical Fourier dictionary. Separation hii husaidia model size kubaki relatively independent na antenna geometry na kuondoa hitaji la learned decoder kwenye base station.

Strength nyingine ni kwamba utafiti hauripoti mean NMSE pekee; pia unatoa P90/P95 error levels, compression-NMSE Pareto analysis, environment transfer, antenna-geometry transfer, computational load, system-level spectral efficiency, quantization na CSI-RS noise kwa pamoja.

Limitation muhimu zaidi ni kwamba matokeo yote yanabaki ndani ya simulation na simulation-based datasets. Waandishi pia wanaona future work ya kujaribu sim-to-real gap kwa real over-the-air channel measurements na kupanua kwa near-field MIMO na FR2 millimeter-wave systems kuwa muhimu.

Dokezo la Chanzo na Mbinu

Jina kamili la kazi asilia: Temporal Channel Estimation for Generalized CSI Feedback

Waandishi na mpangilio: Minwoo Kim; Hyeonsu Lyu; Sehyun Ryu; Sojeong Park; Hyun Jong Yang.

Equal contribution / co-first author: Hakuna tamko kama hilo katika version iliyopakiwa.

Corresponding author: Hyun Jong Yang.

Taasisi: Source inahusisha Minwoo Kim, Sehyun Ryu na Sojeong Park na Department of Electrical Engineering, Pohang University of Science and Technology (POSTECH), Pohang 37673, South Korea. Kwa Hyeonsu Lyu na Hyun Jong Yang, source ina “Institute of New Media and Communications, Seoul 08826, South Korea”. Hyun Jong Yang pia anahusishwa na Department of Electrical and Computer Engineering, Seoul National University, Seoul 08826, South Korea. Kwa kuwa affiliation line kwenye source ina closure/punctuation issue, text haijajengwa upya kimya kimya.

Aina ya source na publication status: Kazi hii ni preprint ambayo haijapitia peer review; matokeo yanapaswa kutathminiwa kwa kuzingatia publication stage hii.

Platform: arXiv, Electrical Engineering and Systems Science / Signal Processing (eess.SP).

arXiv ID: arXiv:2608.01713v1.

Submission date: 3 Agosti 2026.

DOI: 10.48550/arXiv.2608.01713. DOI hii imetolewa na arXiv kupitia DataCite kwa arXiv record na haipaswi kutafsiriwa kama peer-reviewed journal DOI.

Official link:https://arxiv.org/abs/2608.01713

Journal / conference / publisher: Hakuna peer-reviewed journal au conference publication iliyotajwa katika version iliyochunguzwa.

License: arXiv.org non-exclusive distribution license. License hii haijachukuliwa kama Creative Commons reuse license; kwa hiyo original figures za source paper hazijanakiliwa moja kwa moja.

Funding: Hakuna separate funding statement katika version iliyochunguzwa.

Data availability: Hakuna separate data-availability statement katika version iliyochunguzwa.

Conflict of interest: Hakuna separate conflict-of-interest statement katika version iliyochunguzwa.

CRediT/author contributions: Hakuna CRediT au detailed author-contribution statement katika version iliyochunguzwa.

Source-internal editorial inconsistencies: Ingawa ukurasa wa kwanza una text “Manuscript received April 19, 2021; revised August 16, 2021”, kazi imewasilishwa arXiv mwaka 2026 na bibliography yake inajumuisha publications za 2026. Kwa hiyo dates za 2021 hazikutumika kama real current submission/revision dates. Pia, sub-grid extractor definition katika contributions section haioani kikamilifu na explicit parabolic interpolation ya Equation 23. Abstract, contributions na conclusion pia hutumia dB ranges tofauti kwa performance improvement; kwa hiyo Verianla text imeegemea direct numeric values za comparative tables.

Methodological boundary: Matokeo yanategemea Sionna RT, DeepMIMO, Sionna-PHY/CDL-A na related simulation setups. Utafiti hautoi real field/over-the-air validation. Architecture inategemea far-field common-delay assumption na structural limitation kwa near-field XL-MIMO imeonyeshwa wazi.

Content method: Maelezo haya ya Verianla yanategemea tu kazi iliyochunguzwa kwa scientific findings. External sources zilitumika tu kwa bibliographic verification ya title, author order, arXiv ID, DOI, publication status na license; hakuna external scientific experiment, performance result au mechanism mpya iliyoongezwa.