Imagine a device that turns to face a new direction and then stops turning. Its ideal gyroscope reading returns to zero. Has the sensor forgotten the turn? No: a gyroscope measures how quickly the device is turning now, not a permanent record of where it is pointing. The accumulated angle belongs to a calculation that follows the rate over time. A zero current rate means that the angle is no longer changing about the measured axis; it does not force the angle back to its starting value. Manufacturer and platform documentation identify angular rate as the gyroscope's measured quantity. [1][2]

We will follow an imaginary device through a forward turn, a pause and a reverse turn. All numbers are constructed examples, not readings collected from a phone. Most calculations use one fixed turning axis and an ideal non-rotating reference frame, leaving reference-frame motion and other axes outside that model. We will add timing and calibration errors separately so their effects remain visible. The distinction is useful whenever a screen shows a rate, an angle or an orientation: those displays may be related, but they do not contain the same information. Reading the correct one is the first step toward explaining what a movement record means.

A rate gyroscope reports angular velocity components. Its ideal zero output after a completed turn is consistent with a changed orientation because instantaneous rate and integrated angular displacement are different quantities. Analog Devices describes angular-rate sensing and integration, while Android specifies gyroscope channels in radians per second and shows incremental rotation updates. [1][2] A rate channel is therefore not an absolute heading channel, even if software later uses it to maintain an orientation estimate.

The original scalar examples assume rotation about one fixed, correctly identified axis relative to a reference treated as non-rotating. The reported values are angular rates after conversion into physical units, not raw voltages or register counts. All intervals, initial angles and errors are stipulated. We first assume an ideal sensor, then introduce timing mistakes, constant additive bias and a separate scale factor. General three-dimensional motion is addressed with an explicit rotation-order example rather than silently applying scalar angle addition to body-frame rates. No gyroscope, application, motion platform or hardware accuracy test was run for this article.

A tiny electronic gyroscope need not contain the spinning wheel that the word may bring to mind. In the vibrating-mass design described by Analog Devices, a small structure repeatedly moves back and forth. When the device rotates, the Coriolis effect produces a response across that driven motion. The structure's movement changes capacitance, an electrical property that the surrounding circuit can detect. Within the described sensing arrangement, the resulting signal is related to angular rate. [1] This is a physical mechanism for observing a turn as it happens.

The important boundary is between that rate observation and the later angle calculation. A device that has finished turning can retain its new direction without continuing to generate the same turning-rate signal. Likewise, a device turning at a steady rate can keep reporting a steady nonzero number while its direction changes continuously. Neither behavior is contradictory. The hardware mechanism does not, by itself, supply a label such as “started at ten degrees” or “now faces north.” Those labels require a reference or an initial condition. Our numerical examples will keep that reference visible instead of hiding it inside a display.

The historical ADXRS explanation describes a driven resonating mass, a transverse Coriolis response and capacitive sensing of the resulting displacement. Its stated simplified response connects angular rate with the measured differential capacitance. [1] We use that mechanism to identify the measured variable, not to import the article's historical product specifications, performance claims or application guarantees. A rate-sensitive mechanical response does not mechanically retain the integral of every previous rate sample as an absolute directional reference.

For a single ideal axis, write the converted output as ω(t), the signed angular rate in the chosen frame. A constant nonzero ω describes ongoing rotation; ω=0 describes no current rotation about that axis. A stored angle estimate θ is a separate state updated from that signal and a starting value. Distinguishing measurement from state is essential when interpreting reset behavior: resetting a software angle to zero chooses a new numerical origin, whereas a zero rate observation only contributes zero change during the corresponding interval. The two zeros refer to different quantities and are not interchangeable evidence about physical orientation.

Repeat exactly the same three intervals, but start at 200 degrees instead of 10. The changes are still +60, zero and −30 degrees, so the new final angle is 230 degrees. The gyroscope rate sequence is the same in both cases. It supplies the changes, not the missing starting direction. To express an angle relative to a chosen external reference, the calculation needs a starting value established in that reference. A label such as “heading” should therefore prompt a question about where that initial or continuing reference comes from.

There is another useful distinction in our first journey. The device turns through 60 degrees one way and 30 degrees back. Its net signed change is 30 degrees, but its total amount of turning along this one-axis path is 90 degrees. Adding the magnitudes answers a different question from adding the signed changes. A zero final rate answers neither question by itself. A single number at the end of the record cannot reconstruct the earlier journey. The time history, the starting reference and the definition of the requested angle are what make the result interpretable.

Changing θ0 from 10° to 200° while keeping the complete rate history fixed shifts every integrated angle by 190°. The final result changes from 40° to 230°, but the net displacement remains 30°. Thus integration of ideal fixed-axis rate determines angular displacement relative to an initial condition; it does not identify that initial condition from the same rate record. This is an original consequence of the integral, not a claim about a specific orientation estimator's initialization method.

For this path, signed displacement is ∫ωdt=60°−30°=30°, while total angular travel is ∫|ω|dt=90°. These are distinct statistics of the same ideal record. A displayed orientation may additionally wrap angles modulo a full turn, so an unwrapped accumulated angle and a wrapped directional display need not show the same number after repeated revolutions. Neither convention changes the measured instantaneous rate. Specify whether a result represents net displacement, total travel or orientation relative to a reference; otherwise, a correct calculation can appear inconsistent merely because its label hides which quantity was requested.

Let the true device be still about our chosen axis, but suppose the converted rate reading contains a constant extra +0.2 degrees per second. Integrate that reading for 60 seconds and the angle estimate gains 12 degrees even though the true angle has not changed. After 300 seconds, it gains 60 degrees. The error is small when described as a rate and large when accumulated for long enough. Sensor documentation identifies bias and noise as issues that must be considered when rate is used to maintain an angle estimate. [2][3]

Collecting more samples during the same minute does not remove this particular constant offset. Each shorter interval adds a smaller error, but the intervals still add up to 60 seconds, so the total remains 0.2×60=12 degrees. A stable-looking rate is not automatically a correct rate. Real errors can also vary with time, temperature and noise; the manufacturer warns that a constant-bias assumption has limits. [3] Our calculation deliberately keeps the offset constant to isolate one mechanism. It is not a specification for any phone, nor a prediction that every real error grows in exactly the same way.

Use the scalar measurement model m(t)=ω(t)+b with constant b=0.2°/s. If the initial angle is correct and the integrator uses m without correction, the angle error is eθ(T)=∫b dt=bT. It equals 12° at 60 s and 60° at 300 s. For the stationary toy case ω=0, the whole accumulated change is error. If the initial angle already has error e0, the total becomes e0+bT: an initial-reference error persists as an additive offset while this rate bias contributes further error over time. Those two sources should not be merged into one unexplained drift label. Bias and noise are acknowledged in both the platform guidance and the manufacturer's calibration discussion. [2][3]

Partitioning the same duration into more intervals does not change this deterministic bias contribution: ΣbΔtk=bΣΔtk=bT. A higher sample count can affect other aspects of numerical reconstruction, but it does not cancel a constant additive term. Do not substitute a stochastic noise growth law for this calculation; no noise distribution, spectral density or bias-instability specification has been assumed. The manufacturer explicitly cautions that bias can vary between observations with time and temperature. [3] The example therefore demonstrates one controlled error mechanism while leaving the actual sensor's time-varying error model to separately established evidence.

Our earlier sums were deliberately limited to one fixed axis. For general three-dimensional turning, order matters. Picture an arrow initially pointing along a fixed upward z direction. A positive quarter-turn about fixed x moves it toward negative y; a following quarter-turn about fixed y leaves that direction unchanged. Reverse the order instead: the y turn moves the original arrow toward positive x, and the following x turn leaves it there. The two outcomes differ even though each sequence contains one quarter-turn about each of the same two fixed axes.

This geometric example explains why general orientation software keeps track of how rotations combine. Android's example turns a rate sample and elapsed time into a small rotation, then says to combine it with the current rotation. [2] The device's own axes can change their directions in the room during motion. Treating their three rate channels as three independent fixed-room angle counters would skip that relationship. An orientation output may use a rotation matrix or a quaternion to represent the combined result. Those are ways to describe a rotation; their presence does not eliminate timing, reference or sensor-error requirements.

Scalar fixed-axis integration does not generalize to three independent Euler-angle sums for arbitrary body-frame angular rates. A concrete rotation-order check uses active right-handed rotations about fixed world axes. Starting with ez=(0,0,1), Rx(90°) sends it to (0,−1,0), and Ry(90°) leaves that vector unchanged. In the opposite order, Ry(90°) sends ez to (1,0,0), which Rx(90°) leaves unchanged. Thus RyRx ez differs from RxRy ez. This is a geometric world-axis example, not an instruction to relabel body-frame gyro channels as world-axis commands.

Android's illustration forms an axis-angle increment from rate magnitude and elapsed time, converts it to a quaternion and a matrix, and calls for concatenation with the current rotation. Its separate rotation-vector sensor has orientation semantics rather than rad/s semantics. [2] The illustration supports the need to retain rotational state and composition order. An implementation must also choose and consistently apply frame, active/passive rotation and multiplication conventions. The example here does not prescribe one multiplication order for every library. It demonstrates why a correct one-axis sum cannot, by itself, validate an arbitrary three-dimensional orientation update.