integral nonlinearity (INL)
/ I-N-L /
Integral nonlinearity asks a simple question: is the converter's overall response a straight line, or does it bow and bend? Ideally the relationship between input voltage and output code should be perfectly straight — double the input, double the code. INL measures how far the real curve strays from that ideal straight line, at its worst point. It is the big-picture, whole-range honesty check, the bow in the ruler seen from end to end.
Concretely, you plot the actual code-versus-voltage curve, draw the best straight line through it, and INL is the largest gap between the two, measured in LSBs. An INL of plus or minus 1 LSB means that somewhere in its range, the converter's reading is off from the ideal line by up to one whole step, even though each individual step might be fine. Example: a 12-bit ADC with 1.5 LSB INL could read a true mid-scale 2.0000 V as a code that corresponds to 2.0004 V (1.5 steps of 0.27 mV high) — a smooth, systematic bending of the transfer curve, not random noise.
Why this matters: INL is the spec that bounds the absolute accuracy of a reading after you remove simple offset and gain errors. It directly distorts signals — a sine wave passed through a converter with poor INL comes out with added harmonics, hurting SINAD and ENOB. It is distinct from its sibling DNL (which looks at the size of each individual step): you can think of INL as the running sum of the DNL errors, the accumulated bend. A converter can have small DNL everywhere yet still bow into a large INL if the little errors all lean the same way.
A 10-bit ADC over 0 to 5 V has 1 LSB = 4.88 mV. If its INL is specified as plus or minus 2 LSB, any reading can be off from the ideal straight line by up to 2 times 4.88 = 9.77 mV due to the curve's bow, separate from offset and gain.
How far the whole transfer curve bows from a straight line.
INL is the accumulated bend of the curve, roughly the running sum of all the DNL step errors. Small DNL does not guarantee small INL — if every step is just slightly large, the errors add up into a big bow even though no single step looks bad.