Differential Non-Linearity (DNL)

In an ideal converter, every digital code transition corresponds to an input voltage change of exactly 1 Least Significant Bit (LSB). DNL quantifies the error for each specific code.

• DNL = 0 LSB: The actual step width perfectly matches the ideal 1 LSB width.
• DNL = +0.5 LSB: The step is wider than ideal; the code persists for a larger input voltage range.
• DNL = -0.5 LSB: The step is narrower than ideal; the code is triggered over a smaller voltage window.
• DNL = -1.0 LSB: A catastrophic condition where the step width is zero, guaranteeing a missing code.

A missing code is the most severe manifestation of DNL and a definitive, permanent hardware flaw. It occurs when DNL ≤ -1.0 LSB.

• The converter completely skips a digital output code, regardless of the input signal.
• This creates a permanent, non-recoverable gap in the transfer function.
• For RF fingerprinting, a missing code acts as a high-contrast, deterministic identifier that is trivial to detect and extremely difficult to clone or mask.
• The specific pattern of missing codes across the converter's range is a direct consequence of random manufacturing mismatches in the internal resistor ladder or capacitor array.

DNL errors are not random noise; they are systematic, process-dependent artifacts of the converter's physical architecture.

• Resistor-String DACs: Mismatch in the diffused or thin-film resistors causes localized variations in the voltage drop per tap, directly creating DNL errors.
• Binary-Weighted Capacitor Arrays: Mismatch between the unit capacitors and the binary-scaled elements in a successive approximation register (SAR) ADC leads to decision errors that manifest as DNL.
• Current-Steering DACs: Static mismatch in the current source transistors creates amplitude errors that translate to non-uniform step sizes.
• These mismatches are fixed at the time of fabrication and form a static, unchangeable signature.

The DNL profile across a converter's full transfer function is a high-dimensional, device-specific vector. Unlike aggregate metrics like INL, DNL provides localized, code-by-code granularity.

• A 12-bit converter yields 4,096 distinct DNL measurements, creating a rich feature set for machine learning classifiers.
• The DNL signature is signal-independent; it is a static property of the hardware, not the waveform being digitized.
• It is robust against moderate thermal noise and can be extracted using low-frequency ramp or histogram testing methods.
• In a time-interleaved ADC, each sub-converter has its own unique DNL pattern, creating a periodic, multi-layered fingerprint.

The standard technique for extracting a DNL fingerprint is the code density histogram test.

• A high-linearity sinusoidal or ramp signal with a known probability density function is applied to the converter input.
• The frequency of occurrence of each digital code is recorded over millions of samples.
• DNL is calculated as the deviation of the actual code count from the expected count: DNL[i] = (Actual_Count[i] / Ideal_Count) - 1.
• A perfectly linear converter will have a flat histogram; peaks indicate wide codes (positive DNL), and valleys indicate narrow codes (negative DNL).
• This method provides a non-invasive way to map the entire static transfer function.

DNL and Integral Non-Linearity (INL) are mathematically linked but provide different views of the same underlying mismatch.

• INL is the running sum of DNL: INL[k] = Σ (from i=1 to k) DNL[i].
• DNL reveals local, code-to-code errors, while INL shows the cumulative deviation from the ideal straight line.
• A converter can have excellent INL but still possess a distinct DNL pattern if the local errors oscillate around zero.
• For fingerprinting, the raw DNL vector is often more discriminative than the smoothed INL curve because it preserves high-frequency spatial information about the mismatch.

A catastrophic manifestation of DNL where the error equals or exceeds -1 LSB. When DNL = -1, a step width becomes zero, meaning the converter completely skips a digital output code regardless of the input signal. These gaps are permanent, process-dependent defects that create a highly distinctive, unclonable signature. A converter with missing codes at specific transitions provides an exceptionally strong identifying feature for RF fingerprinting because the missing code locations are unique to each die.

The fundamental, irreducible difference between an analog input and its discrete digital representation, bounded by ±0.5 LSB for an ideal converter. DNL errors distort this error distribution, making it non-uniform and signal-dependent. In a perfect converter, quantization error is modeled as white noise. DNL-induced deviations create a colored noise profile that can be exploited as a device-specific signature, as the error statistics are shaped by the converter's unique transfer function imperfections.

A memoryless, amplitude-dependent distortion modeled by a polynomial transfer function. Both DNL and INL are measures of static non-linearity, meaning the error at any instant depends only on the current input amplitude, not on signal history. This time-invariant behavior makes static non-linearity a consistent, repeatable component of an RF fingerprint. The polynomial coefficients that best fit the DNL/INL profile form a compact, device-specific parameter vector for emitter identification algorithms.

A dynamic performance metric that collapses all converter imperfections—including DNL, noise, and distortion—into a single figure of merit. ENOB answers the question: 'How many bits of this converter are actually usable?' A 12-bit ADC with severe DNL errors may deliver only 9.2 ENOB. The gap between the nominal resolution and ENOB quantifies the aggregate fingerprint information available for device identification, with lower ENOB generally indicating a richer, more exploitable hardware signature.

The intentional injection of a small amount of noise—typically at the ±0.5 LSB level—prior to quantization. Dithering decorrelates quantization error from the input signal, linearizing the converter's average transfer function and eliminating the harmonic distortion caused by DNL. However, this technique modifies the native DNL fingerprint by smearing the static errors into a noise-like residual. Fingerprinting systems must account for whether a target device employs dithering, as it fundamentally alters the observable signature.