The question: What does Channel State Information actually capture about human presence, and why does it outperform RSSI for fine-grained wireless sensing?
Your Wi-Fi router is already measuring your breathing — it just doesn't know it yet. Every OFDM (Orthogonal Frequency-Division Multiplexing — the technique that packs Wi-Fi data across dozens of parallel frequency channels at once) packet that bounces off your chest carries a tiny phase shift back to the access point. At 2.4 GHz, a chest displacement of just 6.25 cm — half a wavelength — rotates that phase a full 180°. The signal-strength number your phone calls RSSI averages all of that away into a single scalar. CSI (Channel State Information — the per-subcarrier amplitude and phase matrix the physical layer computes for every received packet) keeps every subcarrier's reading intact, turning a commodity Wi-Fi link into something closer to a respiration-grade radar.
RSSI is a lossy projection, not a measurement
A single 802.11n CSI snapshot on a 3-antenna, 30-subcarrier link returns 90 complex values — one amplitude and one phase angle per subcarrier per antenna pair. RSSI returns one real number: the amplitude average across all those subcarriers, collapsed to the MAC layer with no frequency resolution whatsoever shahbazian2023_1172. That is not a simplification of CSI. It is a lossy projection that discards 89 of those 90 complex values before any sensing algorithm ever sees the data.
The resolution gap runs deeper than byte count. RSSI lives at the MAC layer and updates once per packet — coarse time resolution, no ability to distinguish which frequencies were perturbed. CSI is a physical-layer measurement: it carries subcarrier-level frequency resolution and resolves individual multipath clusters — the separate echoes that arrive after bouncing off walls, furniture, and bodies via different paths fallani2026_04be ↗. Vital-sign monitoring needs to detect chest displacement at the scale of millimetres. RSSI cannot see it — not because it lacks precision in some engineering sense, but because the information was never there to begin with.
The 180° phase swing that makes breathing visible
When a signal travels a path of length d to a reflector and back, it accumulates a phase shift of 2π·(2d)/λ, where λ is the wavelength. At 2.4 GHz, λ ≈ 12.5 cm. A chest wall moving through half that distance — 6.25 cm, a realistic tidal breathing amplitude — changes the round-trip path length by a full wavelength. The detectable swing across one half-cycle of breathing is 180° jiang2018_5aae.
That 180° swing is not a noise-floor curiosity. It dominates the CSI phase trace when a person sits within a few metres of the link. Because each OFDM subcarrier carries an independent phase measurement, even a single breathing episode appears as a coherent, frequency-correlated modulation across dozens of subcarriers simultaneously — a pattern that no single-number RSSI reading could ever encode ahmad2024_8639.
The practical consequence: CSI can resolve the chest-wall micro-motion of a seated person breathing quietly. RSSI cannot resolve it at all. This is not a claim about algorithm sophistication; it is a consequence of what each metric discards at the hardware layer.
Why a body modulates the channel at all: Fresnel zones
The phase-rotation argument explains how displacement shows up in CSI. Fresnel-zone geometry explains why a stationary body two metres from the link affects the channel in the first place.
The First Fresnel Zone is an ellipsoidal region around the direct Tx–Rx path. Any object inside this ellipsoid diffracts the signal — bends energy around itself — producing a measurable drop in received power when it crosses the line between transmitter and receiver. Outside the First Fresnel Zone, reflection dominates: the body scatters energy toward the receiver rather than blocking it wu2022_75d3.
This matters for deployment. A person standing directly between router and receiver sits inside the First Fresnel Zone and causes strong diffraction-based attenuation that shows up cleanly in CSI amplitude. The same person standing to the side of the link sits in the reflection regime: their contribution to the channel is weaker and phase-dependent rather than amplitude-dominant. Link geometry — where you mount the nodes, not just what hardware you use — determines what CSI can resolve. Our October library deployment with the ESP32-C5 sensing nodes has to account for this from the floor plan up.
The channel encodes the room, not just the human
Here is where the picture gets uncomfortable.
Meneghello et al. collected CSI amplitude traces in the same empty bedroom on different days and computed the Pearson correlation between them. The result was near zero meneghello2023_0a93 ↗. No human present. Same room. Same hardware. Different day. The channel fingerprint had changed enough that the two traces were statistically uncorrelated.
That finding reframes what CSI actually measures. CSI does not measure human presence. It measures the full multipath environment — furniture, humidity, temperature gradients, the position of a door, the presence or absence of a person — and a model trained on today's channel fingerprint is silently learning the room as it was today, not the room as it will be tomorrow.
This is the root cause of every model-drift problem the series will tackle in later episodes. EXP-002 (CSI Model Drift Measurement Over Time) is our operational attempt to quantify how fast a trained model decays in the same physical space — turning Meneghello's headline statistic into a concrete maintenance schedule for the library install. EXP-001 (BLE-Assisted CSI Ground Truth Collection) pairs CSI with BLE-derived occupancy ground truth precisely because CSI carries no scalar people-count channel: the amplitude-and-phase wiggles need a second modality to tell the model what they mean.
The phase channel most pipelines quietly discard
Most CSI sensing pipelines in the active literature use amplitude only. Phase measurements from commodity hardware arrive with a hardware-induced bias — a random offset that varies per packet, per antenna, per boot — and correcting it requires either a calibrated reference or careful sanitisation that many implementations skip. The result is that the 180° phase swing, the single most information-rich feature CSI offers over RSSI, gets thrown away before training begins.
The consensus view treats this as a reasonable engineering trade-off. We are less sure. If phase is what makes breathing-resolution sensing physically possible, and if our own EXP-001 pipeline is currently amplitude-only, then the accuracy ceiling we are measuring in EXP-001 may not be the ceiling of CSI sensing — it may be the ceiling of CSI sensing with one hand tied behind its back.
Which raises the question we will have to answer before the series moves on to crowd counting or domain transfer: if the phase channel is the mechanism that separates CSI from RSSI in the first place, and most lab pipelines silently drop it — what accuracy are we leaving on the table, and will we ever instrument ourselves well enough to find out?