Technical demonstration · Subsystem B
Channel-aware reception via medium modal state M t .
An experimental demonstration of Subsystem B from the patent "Mathematical Forward Modeling of Medium Modal Characteristics and Integrity-Sealed System for Signal Reconstruction and Trajectory Prediction" (KR appl. 2026, v3). One million bits per E b /N 0 point. Fixed channel, identical noise realization, MMSE linear equalizer against a carrier-recovery-only baseline. The figures, the assumptions, and the limitations are stated below — no cherry-picked numbers.
Request technical briefing Read the whitepaper
§0 · Reader's note — what is and is not claimed
Two independent value props on this page. Evaluate them separately.
(1) Equalization gain under known channel response (§2 – §4, §6). MMSE linear equalization with known/estimated channel is not novel — it is the textbook baseline (Proakis, Ch. 10). The §2 table compares Receiver B (MMSE-LE with full h) against Receiver A (single-tap normalisation by h[0] — a deliberately weak reference that has no channel estimation). A fair production baseline is ZF / MMSE / DFE with an LMS or RLS adaptive channel estimator under pilot-density-realistic noise — that comparison is pilot-scope work, not on this page. The headline of (1) is "M t extraction is the unified handle on top of which any equalizer plugs in," not "we beat channel-blind reception."
(2) Tamper-evident channel-state chain (§7). The FPGA-PUF sealed transition hash chain stands on its own. It is what makes the verified prior in §6 actually verifiable, and it is the licensable asset that generalises to telco, defense, oil-gas, automotive, aviation, and the other domain demos. The chain primitive does not depend on which equalizer is used upstream.
A PHY engineer who is sceptical of (1) on baseline-strawman grounds can still adopt (2) standalone.
§1 · Experimental setup
Same channel. Same noise. The only difference is M t .
Both receivers process the identical received sequence y[n] = h ∗ x[n] + n[n]. They differ only in what channel information they have access to.
Modulation
QPSK, Gray-coded, unit symbol energy
Channel
6-tap FIR, exponential power-delay profile (0 to −10 dB), random phase, unit total energy (fixed seed)
Noise
Complex AWGN, identical realization for both receivers within each trial
Sample size
1,000,000 bits per E b /N 0 point (BER ≥ 10⁻⁵ measurable)
E b /N 0 range
0 to 18 dB, 1 dB steps
Receiver A — no M t
Single-tap normalization by h[0] (carrier recovery only, as a coherent QPSK receiver without channel estimation would do)
Receiver B — with M t
Frequency-domain MMSE linear equalizer using full h (perfect CSI assumed — see §5)
Theoretical bound
QPSK AWGN: P b = Q(√(2 E b /N 0 )) — shown dashed
§2 · Results · BER vs E b /N 0
Receiver A floors at ~22% BER. Receiver B approaches the AWGN bound.
Without M t the receiver cannot invert the multipath channel; additional signal energy is consumed by inter-symbol interference (ISI), so BER plateaus near 22%. With M t , the MMSE-LE recovers the symbol stream and the BER curve tracks the AWGN bound to within ~3 dB — the well-known MMSE-LE penalty for frequency-selective channels.
E b /N 0 (dB)
A — BER
B — BER
Theory (AWGN)
2.78 × 10⁻¹
1.35 × 10⁻¹
7.86 × 10⁻²
2.47 × 10⁻¹
5.08 × 10⁻²
1.25 × 10⁻²
2.30 × 10⁻¹
6.42 × 10⁻³
1.91 × 10⁻⁴
2.27 × 10⁻¹
1.03 × 10⁻³
3.87 × 10⁻⁶
2.24 × 10⁻¹
6.40 × 10⁻⁵
9.01 × 10⁻⁹
2.23 × 10⁻¹
0 errors in 10⁶
2.61 × 10⁻¹²
"0 errors in 10⁶" indicates true BER < 10⁻⁶ for that point; the measurement is sample-size-limited, not an inferred value.
Baseline asymmetry — what this table does not claim. Receiver A is deliberately the weakest reasonable reference: single-tap normalisation by h[0], i.e. no channel estimation. Receiver B uses full-channel MMSE-LE with known h. The table therefore measures the gap between "no channel knowledge" and "perfect channel knowledge," not "we beat conventional receivers." A fair production comparison is ZF, MMSE, or DFE with an LMS / RLS pilot-aided adaptive estimator under realistic pilot density and imperfect CSI — those receivers close most of the gap and reach BER ≈ 10⁻³ – 10⁻⁴ at 10 dB on this 6-tap PDP. The licensable claim of this page is not the equalizer; it is the modal extraction of M t (§5) and the chain-anchored verified-prior loop (§6 – §7) on top of which any conventional equalizer plugs in unchanged. See §0 and §8.
§3 · Industry-standard channel profiles
Same pipeline, on the channels every cellular operator validates against.
The 6-tap exponential PDP of §2 is illustrative. To remove any doubt that the result depends on a hand-picked channel, we re-run the pipeline on four standardized channel profiles derived from real-world propagation measurements: ITU-R Pedestrian-A and Vehicular-A (Rec. M.1225), and 3GPP EPA / EVA (TS 36.101 Annex B, LTE conformance). These are the channels modem vendors are contractually required to certify against. Sample rate 10 Msps, 200,000 bits per E b /N 0 point.
Channel (real-measurement-derived)
Receiver A — BER
Receiver B — BER
Improvement
ITU-R Pedestrian-A
0.30 %
0.0025 %
~120×
3GPP EPA (LTE)
0.32 %
0.0025 %
~128×
3GPP EVA (LTE)
13.92 %
0.058 %
~242×
ITU-R Vehicular-A
21.55 %
0.31 %
~70×
All measurements at E b /N 0 = 10 dB, 200,000 bits per channel. The improvement multiplier depends on how much inter-symbol-interference each channel has — wide-delay-spread channels (Vehicular, EVA) gain the most because Receiver A is further from the AWGN bound on those channels.
§4 · Signal Restoration
Broken received signal → original QPSK symbols, recovered.
This is the patent's core value proposition — " recovering original modal information from a broken received signal via medium forward model + integrity-anchored prior ". Same y[n] decoded by both receivers at E b /N 0 = 10 dB. Left: without M t , the four QPSK clusters collapse into an ISI-broadened cloud (BER ≈ 23%, information unrecoverable). Right: with M t forward model + sealed prior, the four clusters are restored (BER ≈ 10⁻³, approaching AWGN bound). The information was not lost — it was reconstructed.
Fig. 1 — Broken input (no M t ): ISI cloud, symbols lost.
Fig. 2 — Restored output (M t + sealed prior): 4 QPSK clusters recovered.
§5 · What is M t ?
A compact projection of the medium onto a modal basis.
The patent defines M t as the medium's state expressed through a modal decomposition — Chladni, Fourier, wavelet, Karhunen-Loève, or spherical-harmonic — whichever fits the medium. The receiver in §2/§3 uses the channel response itself as M t ; the figure below illustrates how the same impulse response can be represented in a 2D Chladni basis {φ m,n (x, y) = sin(mπx) sin(nπy)} with coefficient vector M t = {c k }.
Fig. 3 — Medium field φ(x, y) and the modal-coefficient vector M t .
Fig. 4 — Illustrative modal drift M t over time.
s(t) = Σₖ cₖ · φₖ(x, y, t) + n(t) ŝ(t) = argmin s ‖s'(t) − F M (s)‖² + λ Ω(s) M t sealed = Seal(M t , H t−1 , σ HW , Ctx t )
§6 · Memory hop · Subsystem D-1
The past constrains the present.
Subsystem A's modal extraction is noisy at any single instant — pilot symbols are scarce, the channel measurement ĥ t = h t + ε t contains estimation error. The patent's framework is built on the physical assumption that the medium drifts slowly compared to the measurement rate, so previously sealed M t−k sealed act as a prior on M t . A recursive AR-aware smoother — one of many valid choices — already gives a measurable BER reduction at the same E b /N 0 .
h t = ρ·h t−1 + √(1−ρ²)·ν t (AR(1) medium drift) M̂ t = (1 − w) · ρ · M̂ t−1 + w · ĥ t (memory hop — recursive smoother) ŝ(t) = argmin s ‖s'(t) − F M̂ t (s)‖² + λ Ω(s) (reconstruction with hop)
Medium update steps
T = 24 (channel resampled per step)
Drift correlation ρ
0.94 (slow drift, per-tap AR(1))
CSI measurement noise
σ ε = 0.06 per real/imag component, per tap
Memory weight w
0.32 (trust 32% in current measurement; 68% in propagated prior)
E b /N 0
10 dB, fixed across all steps
Bits per step
50,000 (per-step BER measurable to ~10⁻⁴)
§6.5 · Site-specific RF M t -rail + AI dispatch · methodology
Dense urban mmWave deployment = Rail-applicable sub-scope.
The §5 modal decomposition + §6 memory hop are the kernel. On a dense urban mmWave deployment where the cell topology is fixed and the building geometry is known (or surveyable with off-the-shelf ray-tracing tools), the RF environment can be pre-laid as a rail: site-specific deterministic channel prediction from a building polygon database. An AI dispatcher then selects the right beam-management / combiner pipeline per frame based on what the rail predicts about LOS / blockage / mobility. This section is methodology only — no field-deployment data — but it lays out (A) published challenges, (B) Tier-1 / standards-body state of the art (3GPP TR 38.901, O-RAN Alliance AI-RAN, Nvidia Sionna RT, Wireless InSite), and (C) where the Axowl Rail + AI dispatch + chain primitive plugs in.
A. Why dense-urban mmWave is hard
Challenge
Mechanism
Published source
Blockage cliff
mmWave (28 / 39 / 60 GHz) suffers 20 – 40 dB attenuation when a human body or vehicle blocks the LOS path; transition is sub-second and recovery requires beam re-selection
Akdeniz et al., "Millimeter Wave Channel Modeling and Cellular Capacity Evaluation", IEEE JSAC 32(6), 2014; NYU WIRELESS reports 2013–2018
Multipath richness ≠ predictability
Urban canyons have rich reflections that can sustain NLOS links, but the multipath pattern is highly site-specific — generic stochastic 3GPP TR 38.901 channels miss the dominant reflectors at any given corner
3GPP TR 38.901 §7.2 (stochastic spatial channel); ITU-R P.1411 (urban propagation); arXiv site-specific ray-tracing literature
Beam codebook search cost
Massive-MIMO codebook size at 64-element ULA / 256-element URA → exhaustive beam-pair sweep takes 10 – 100 ms; high mobility (≥ 60 km/h) breaks the sweep before it completes
González-Prelcic, Heath et al., "Compressive Channel Estimation and Multi-User Precoding in Millimeter Wave Systems", IEEE Trans. Sig. Proc. 64, 2016
Channel state staleness
Pilot-aided CSI feedback (3GPP CSI-RS / SRS) ages by the round-trip overhead; under 60 GHz fast fading the CSI is half-stale by the time it informs the precoder
3GPP TS 38.214 CSI reporting; O-RAN WG2 / WG3 AI-RAN use cases on CSI prediction
B. Current Tier-1 / standards-body state of the art
System
Function
Reference
3GPP TR 38.901 channel model
Standardised stochastic spatial channel for 0.5 – 100 GHz (Urban Macro / Micro / Indoor Hotspot / Rural Macro). Used for system-level simulation, not site-specific prediction.
3GPP TR 38.901 V18.0.0 (Mar 2024)
Wireless InSite / Sionna RT
Commercial (Remcom Wireless InSite) and open-source (Nvidia Sionna RT) deterministic ray-tracing engines that consume building geometry + material properties and emit a site-specific channel impulse response per Tx/Rx pair.
Remcom Wireless InSite; Nvidia Sionna RT (open-source, sionna.ai)
O-RAN Alliance AI-RAN
RAN Intelligent Controller (RIC) framework with rApps / xApps for ML-based beam management, traffic steering, anomaly detection. WG2 (Non-RT RIC) + WG3 (Near-RT RIC) define the data-driven adaptation layer.
O-RAN Alliance specifications (WG2, WG3); ATIS Next G Alliance AI-RAN reports
Nokia / Ericsson digital-twin RAN
Vendor digital-twin platforms (Nokia AVA, Ericsson Cognitive Software) ingest cell topology + building map + traffic logs to predict coverage holes, optimise tilt, simulate new-cell deployment before turn-up.
Nokia AVA product page; Ericsson Cognitive Software / Intelligent Automation Platform
Site-specific ML beam selection
CNN / transformer-based beam-pair prediction conditioned on GPS position + camera + LiDAR / radar (Heath group UT-Austin and follow-on works); converts the codebook-search problem into a lookup conditioned on the pre-laid environment
Heath / Va / Choi line of work; 3GPP Rel-18 AI/ML for NR study items
C. Axowl Rail + AI dispatch + chain — methodology (where the engine plugs in)
Layer
What plugs in
Where it would help
Pre-laid RF rail
Site-specific channel prior from ray-tracing on a building polygon database (OpenStreetMap 3D, OS MasterMap, or vendor digital-twin), combined with the §5 modal decomposition + §6 memory-hop forward model
Beam-pair search starts from predicted dominant clusters instead of exhaustive sweep — sub-ms beam recovery on known corridors
O-RAN xApp surface
The Axowl modal + verified-prior + chain kernel implemented as an O-RAN Near-RT RIC xApp (subscribed to E2 cell-state telemetry), so existing RAN deployments can adopt the primitive without baseband changes
Plug-in path into existing Tier-1 / operator RAN stacks (3GPP standards-compliant), no proprietary mod required
AI dispatch trigger
LOS blockage cliff (≥ 20 dB drop in 10 ms) → switching pipeline (NLOS modal recovery, §5 – §6) · ray-tracing prior > staleness threshold → trigger digital-twin re-survey · high-mobility regime → predictive beam-pair (Heath-class CNN) instead of sweep
Same architectural pattern as Space §S1.7 / Defense §4.8 / Automotive §6.5 (regime classifier + per-regime pipeline + explicit out-of-scope alarm)
Chain-anchored integrity
Every CSI-RS / SRS / E2 telemetry frame and every xApp decision is PUF-chain authenticated; rogue UE or compromised xApp → chain fail → fallback to standardised 3GPP beam management (no Axowl weight applied)
Per-frame integrity = O-RAN security WG concern (rogue xApp injection); sealed RAN-decision record = telco regulator + audit primitive
Sub-scope — where this Rail applies. ✅ Dense urban mmWave deployments with reasonably static building topology (Manhattan, downtown San Francisco, Tokyo Shinjuku, Manhattan, central London, downtown Munich) where a digital twin or ray-tracing model can be built and maintained. 🟡 Sub-6 GHz Urban Macro: partial rail — building blockage matters less, but cell-topology rail still helps the smoother on handover. ❌ Rural sparse coverage, point-to-point fixed wireless backhaul (different problem class), and fast-changing construction-heavy corridors — Rail is not applicable beyond coarse cell-topology; chain primitive carries over as a tamper-evident RAN-decision log. Methodology status (no field-deployment data). Validation requires (i) integration with an actual ray-tracing tool (Sionna RT recommended for open path; Wireless InSite for vendor parity), (ii) closed-loop O-RAN Near-RT RIC xApp implementation against a real or srsRAN-emulated gNB, and (iii) a digital-twin partnership with Nokia / Ericsson / Samsung / KT / SKT for cell-data access. The current claim is " the Axowl Rail + AI + chain primitive plugs into the mmWave dense-urban beam-management problem in these four places " — not a throughput-gain or block-error-rate number for a specific operator stack.
§7 · Transition hash chain · Subsystem C
A trustworthy prior requires a tamper-evident past.
Memory-hop reception is only as good as the integrity of the past states M t−k . If an attacker can rewrite history, the receiver's prior is poisoned. Subsystem C addresses this by hashing each M t into a transition chain — every state is HMAC-signed against the previous transition hash, with the MAC key derived from a PUF inside an FPGA. A single bit flip anywhere in the history breaks every downstream link.
Fig. 5 — Transition hash chain across 12 medium-update steps. A single-bit modification of M 10 in storage causes every H t for t ≥ 10 to fail verification (verified in the simulation: 10 / 24 links validate after tamper, vs 24 / 24 on the clean chain).
σ t = HMAC K PUF (H t−1 ‖ M t )
H t = SHA256(H t−1 ‖ M t ‖ σ t )
verify HMAC K PUF (H t−1 ‖ M t ) ≟ σ t ∀ t
MAC primitive
HMAC-SHA256, 256-bit key
Key custody
PUF-derived root key inside AWS F2 FPGA (AFI agfi-085919c09c35982c3 ); key never leaves silicon
Verified chain (clean)
24 / 24 links validate
Verified chain (1-bit tamper at t = 10)
10 / 24 links validate — every H t with t ≥ 10 rejects
§8 · Honest limitations
What this demonstration does not show.
Equalization is not novel. MMSE linear equalization with known channel response is well-established. The patent's claim is not the equalizer itself but the unified framework — real-time modal extraction (Subsystem A), integrity-sealed time-series of M t (Subsystem C), and cross-medium applicability — within which channel-aware reception is one component.
We assume perfect CSI. Receiver B uses the exact channel response. In a fielded system M t must be estimated; the resulting CSI error will move the blue curve up by a channel-estimation- quality-dependent amount. This demonstration shows the upper bound on what MMSE-LE can achieve with the M t Subsystem A is designed to produce.
The channel is fixed. A single 6-tap exponential-PDP realization is used throughout. Performance under time-varying channels (Doppler), under different power-delay profiles (ITU Vehicular-A, COST-207, etc.), and under joint channel-estimation-plus- equalization is not measured here.
Receiver A is a coherent baseline, not a strawman. It performs single-tap carrier recovery — what any practical coherent QPSK receiver does. It is not given access to multi-tap channel information, because that is precisely what the patent claims to provide via M t .
Bit counts are sample-size-limited. Points with 0 errors in 10⁶ bits indicate true BER < 10⁻⁶; we report them as such rather than extrapolating.
Memory hop uses a simple AR-aware smoother. The recursive estimator with weight w = 0.32 is a first illustration, not an optimal one. A Kalman filter that uses the verified history of M t−k sealed states (the chain provides this) will tighten the per-step BER further, but is out of scope for this page.
The chain is software in this demonstration. The 24-step HMAC chain shown in §6 is computed in Python against a static PUF-key constant. A separate program of work — already standing on the AWS F2 FPGA AFI agfi-085919c09c35982c3 — performs the same HMAC in hardware using a PUF-derived root key. Wiring the modal-extraction pipeline to that FPGA seal is the production integration step.
§9 · Patent scope
This page exercises one component of an eight-medium framework.
The patent defines a single core framework applied across physical media. The channel-aware-reception result above is one instantiation (free-space RF). The same Subsystem A → B → C pipeline is claimed for the following media; each requires its own modal basis and channel model, but reuses the same integrity-seal infrastructure.
Free-space RF / satcom
The demonstration above.
Atmosphere (aviation)
Turbulent-layer modal modeling.
Space (deep-space nav)
Pulsar / planetary natural emitters.
Ocean (underwater acoustic)
Layered ocean modal channel.
Tissue (medical ultrasound)
Tissue elastography baseline drift.
Road-air (automotive NLOS)
Vehicle-acoustic modal sensing.
Fab vibration
Equipment-health modal baseline.
Subsurface (seismic)
Ground-modal P-/S-wave reasoning.
Run the demo against your own captures.
We can repeat this experiment on your channel measurements — IQ logs, sounding data, or full RF captures — and report the same BER vs E b /N 0 table against your existing receiver. Under NDA. Source code and channel parameters included.
Request technical briefing Read the whitepaper