Channel State Information (CSI)

Unlike coarse Received Signal Strength Indicator (RSSI) measurements, CSI decomposes the channel response across individual Orthogonal Frequency-Division Multiplexing (OFDM) subcarriers. This provides a high-resolution spectral fingerprint of the environment.

• Captures amplitude and phase for each subcarrier.
• Reveals frequency-selective fading patterns invisible to RSSI.
• Enables detection of subtle environmental changes, such as human movement or device location shifts.

CSI is a complex matrix representing the channel's state across both time and frequency domains. This duality allows for the extraction of distinct feature types.

• Static CSI: Represents stable multipath structures, useful for device fingerprinting and passive localization.
• Dynamic CSI: Captures time-varying fluctuations caused by moving objects, enabling device-free activity recognition and gesture detection.
• The Channel Impulse Response (CIR) is derived via an Inverse Fast Fourier Transform (IFFT), revealing multipath delay profiles.

CSI acts as a passive radar, where every object in the environment becomes a virtual sensor. Minute perturbations in the propagation paths alter the CSI matrix.

• Multipath propagation causes signals to reflect, diffract, and scatter, creating unique interference patterns.
• A person walking through a room alters these paths, creating a detectable Doppler shift and amplitude variance.
• This sensitivity is the foundation for Wi-Fi sensing applications, including fall detection and occupancy monitoring.

CSI measurements inherently contain the unique hardware impairments of the transmitter's analog front-end. These imperfections are embedded in the channel estimate.

• I/Q imbalance and phase noise from the local oscillator create distinctive constellation distortions.
• Power amplifier non-linearity introduces spectral regrowth patterns.
• These artifacts are stable over time and serve as robust, unclonable identifiers for Radio Frequency Fingerprinting (RFF).

In Time Division Duplex (TDD) systems, the uplink and downlink channels are considered reciprocal. CSI measured on one link can be used to infer the state of the reverse link.

• This principle is critical for Massive MIMO beamforming, where the base station uses uplink CSI to steer downlink energy toward the user.
• Reciprocity holds for the physical propagation path but requires calibration to compensate for mismatched transmit/receive chains at the base station.

CSI is often characterized statistically to reduce feedback overhead and enable robust system design. Key statistical parameters define the channel's behavior.

• Delay spread quantifies the time dispersion of multipath components.
• Coherence bandwidth defines the frequency range over which the channel response remains correlated.
• Coherence time specifies the duration over which the channel impulse response is invariant, determined by the maximum Doppler shift.

The process of decomposing a learned signal representation into independent, interpretable factors of variation. The goal is to separate the latent space into:

• Device-specific factors: Stable hardware impairments unique to the transmitter.
• Channel-specific factors: Variable multipath, fading, and noise. By discarding the channel factors at inference time, the model achieves robust identification.

A kernel-based statistical measure used to quantify the distance between probability distributions of features from different domains. As a regularization term during training, minimizing MMD between source and target domain feature distributions forces the network to learn domain-invariant representations, aligning features from diverse channel conditions into a common space.

A technique that trains fingerprinting models on a massive variety of simulated channel conditions rather than attempting to adapt to a specific target domain. By randomizing parameters like delay spread, Doppler shift, and noise floor during training, the real-world deployment environment appears as just another variation, dramatically improving sim-to-real transfer robustness.

A metric learning loss function that structures the embedding space by enforcing a margin of separation. For each training sample:

• Anchor: A signal from a specific device.
• Positive: Another signal from the same device under different channel conditions.
• Negative: A signal from a different device. The loss minimizes anchor-positive distance while maximizing anchor-negative distance, learning channel-robust clusters.