Reading signal-chain PSD plots

Studio 2.2.4 gives each signal-chain location its own power spectral density chart. The charts start with matching dB scales and retain independent zoom, legend and enlarge controls. Curves at the same location remain together for comparison.

Chart	Signal	Comparisons kept in this chart
DPD Input	Original target x	Inputs from the selected runs
DPD Output / PA Input	Predistorted drive u = DPD(x)	Predistorted drives from the selected runs
PA Output	y = PA(u)	Linear output target g·x, PA-model predictions with and without DPD, and measured or synthetic dataset output without DPD

DPD output is the signal driving the PA. A curve called “with DPD” that was computed through a PA surrogate belongs to PA Output, not DPD Output. The linear target g·x is an output reference, not an additional captured input.

PA modeling and dataset inspection have only PA Input and PA Output charts. The Signal Generator shows only PA Input. PA Library shows the generated input and the explicitly simulated Virtual PA output separately. Missing signal locations are not fabricated. Legacy plots without enough identity information use an Unspecified signal position chart.

Legends and provenance

Starting in 2.2.13, each chart has a 14 px legend above the axes, with room for wrapped rows and separate space for the plot toolbar. Axis labels are larger too. Line color and dash distinguish curves without covering the spectral shoulders. Single-run views omit the run ID; multi-run views use R1, R2, … and list the full run-ID key underneath. Trace controls and the cursor table retain the original names, source and capture identity. Synthetic dataset outputs are labelled synthetic; mock evidence remains marked MOCK.

PA model means the trained model of the amplifier, also called a surrogate. The PA output plot includes this explanation whenever it shows DPD predictions through that model:

Legend	Meaning
Without DPD · PA model	Original input x → trained PA model: the predicted output without predistortion
With DPD · PA model	Original input x → DPD → the same trained PA model: the predicted output with predistortion
Without DPD · measured	PA output from the measured dataset, before applying DPD
Without DPD · synthetic data	PA output from a synthetic dataset, before applying DPD
Linear target g·x	The desired linear output under the result's gain reference

The two PA-model curves are predictions. A measured DPD result uses captured PA output and is labeled measured; its plot does not acquire a model-prediction explanation.

Switching a legend affects only that chart. Full input/output target and baseline curves remain available in Trace controls. Saving a review preserves independent axis ranges, selected traces, the frequency cursor, bands, reference and profiles.

Live views and exports

Training, testing, result detail, comparison, dataset inspection, Signal Generator and PA Library all use the same position labels. DPD validation probes capture u at the PA-model input during the same shadow forward pass that produces PA(u). The shadow remains in evaluation mode with RNG state preserved. Final plots use the exact saved test arrays and valid sample count.

Publication preview, PNG/SVG/PDF exports and HTML reports separate signal positions too. Existing mixed spectrum figure specifications are expanded into position panels when previewed or exported; their stored PSD arrays are not rewritten. The standalone export includes its layout helper and hash-checked replay script.

Numerical meaning

This change reorganizes plots. It does not change the Welch estimator, sample selection, amplitude scaling, PSD bins, metric definitions or integration bands. The usual spectrum units are dB re amplitude²/Hz; normalized-frequency files use cycles/sample. Generator PSD uses its documented unit-RMS dBFS reference. None of these plots implies calibrated RF watts or dBm.

Band shading follows the selected result's metric profile. Legacy and general spectral ACLR/ACPR retain their distinct definitions. Zooming does not recalculate metrics. See research review for protocol compatibility, saved provenance and frequency-cursor interpretation.

ILC runs add a waveform-specific Ideal trace to the PA Input and PA Output panels. The legend preserves that identity alongside the fitted DPD. Ideal feedback uses the current test waveform; see ILC and ILA before comparing it with a transferable model.