← Back to the chapter · Previous detail
Technical detail · IoT Mathematics: Calculus, Exponents, and Logs
Under the Hood: Models Fail at Their Assumptions
Under the hood, mathematical models fail less often because the algebra is wrong and more often because the assumptions are invisible. Free-space path loss assumes a clean propagation model, while indoor deployments add wall loss, multipath, antenna orientation, human bodies, metal racks, humidity, and channel interference. Shannon capacity is an upper bound for a channel with bandwidth and SNR; it does not include MAC overhead, duty-cycle rules, retransmissions, encryption overhead, buffering, or application latency.
Control and sampling models have similar boundaries. A derivative of a noisy signal amplifies noise unless filtering and sample time are chosen carefully. An integral of error can wind up against actuator limits unless the PID controller has anti-windup behavior. Nyquist sampling prevents aliasing only when the signal is band-limited or an anti-alias filter removes high-frequency content before the ADC. An FFT peak is meaningful only when windowing, sampling rate, leakage, and calibration are understood.
Sensor fusion models hide state assumptions. A Kalman filter needs a state vector, transition model, process noise, measurement model, measurement noise, timestamp alignment, and coordinate frame. If the GPS fix is one second old and the accelerometer frame is rotated, the matrix equations can still run while the result is physically wrong. Covariance is a promise about uncertainty; it must be updated when sensors saturate, drift, drop packets, or enter a different environment.
Cryptographic math has its own boundary. Modular exponentiation and Diffie-Hellman style exchange depend on secure parameters, randomness, authentication, freshness, key derivation, and key storage. The hard discrete-log problem does not stop a man-in-the-middle attacker unless the exchange is authenticated. Record the protocol version and library behavior when the math supports a security claim. The lesson is the same across the chapter: a formula gives leverage only when its assumptions, operating range, and validation evidence travel with the result.