wzyn20051216/matlab-agent-skills / skills/matlab-signal-vision-ai
matlab-signal-vision-ai
UnverifiedMATLAB workflow for signal processing, DSP, image processing, computer vision, lidar, medical imaging, deep learning, and AI-assisted reproducible experiments. Use for filters, spectrograms, wavelets, object detection, segmentation, classification, training, inference, generated figures, and for verifying signal or vision results — units, windows, Nyquist, filter response, BER/SNR, masks and boxes, IoU/precision/recall/F1, label alignment, transform invertibility, and split leakage.
| File | Kind | Size |
|---|---|---|
| SKILL.md | skill md | 10.5 KB |
| references/signal-measurement-contract.md | markdown | 10.3 KB |
| references/vision-verification-contract.md | markdown | 13.5 KB |
| scripts/mas_ber_check.m | other | 12.6 KB |
| scripts/mas_detect_check.m | other | 21.5 KB |
| scripts/mas_filter_check.m | other | 12.7 KB |
| scripts/mas_image_contract.m | other | 31.4 KB |
| scripts/mas_image_metrics.m | other | 4.5 KB |
| scripts/mas_measure_contract.m | other | 19.8 KB |
| scripts/mas_ofdm_ber.m | other | 6.8 KB |
| scripts/mas_sample_check.m | other | 9.6 KB |
| scripts/mas_seg_check.m | other | 23.0 KB |
| scripts/mas_spectrum_check.m | other | 7.8 KB |
| scripts/mas_split_seed.m | other | 3.6 KB |
| scripts/mas_stat_verdict.m | other | 10.0 KB |
| tests/test_signal_contract.m | other | 21.5 KB |
| tests/test_signal_vision.m | other | 15.8 KB |
| tests/test_stat_decision.m | other | 8.6 KB |
| tests/test_vision_contract.m | other | 35.3 KB |
scripts/mas_ofdm_ber.m
View on GitHubfunction r = mas_ofdm_ber(ebn0dB, opts)
%MAS_OFDM_BER QPSK-OFDM over AWGN, graded against the closed-form BER.
%
% r = mas_ofdm_ber([0 2 4 6])
% r.berSim, r.berTheory, r.relErr, r.nBits
% r.ok, r.verdict, r.tol -- verdict comes from mas_stat_verdict per point:
% PASS only where every point's count supports the closed form, FAIL where a
% powered point refutes it, UNDETERMINED where some point had no power to see
% it. r.states gives the per-point state.
%
% WHY AN OFDM TASK AND NOT "SIMULATE BPSK"
% OFDM is where a plausible-looking chain usually breaks: the IFFT/FFT pair must be
% scaled consistently, the cyclic prefix must be removed before the FFT (not after),
% and the per-subcarrier SNR must account for the energy spread across N subcarriers.
% Each of those mistakes gives a BER curve that is smooth, monotone and wrong.
%
% The reference is the textbook result for QPSK over AWGN, BER = 0.5*erfc(sqrt(Eb/N0)),
% which has nothing to do with this code. A simulation that disagrees is wrong.
%
% The check is one-sided on purpose: a simulated BER well BELOW theory is not a
% success, it means the noise scaling is too small.
if nargin < 2, opts = struct(); end
% tolMin/tolMax BRACKET the per-point counting-noise pass mark; `tol` is kept as
% the reported headline number and as the floor for well-populated points.
opts = dflt(opts, struct('nSym', 2000, 'nSub', 64, 'cpLen', 16, 'seed', 101, ...
'tol', 0.15, 'tolMin', 0.15, 'tolMax', 0.45, 'verbose', false));
nBit = opts.nSym * opts.nSub * 2; % QPSK: 2 bits per subcarrier per symbol
ebn0 = ebn0dB(:).';
berSim = zeros(size(ebn0));
berTh = zeros(size(ebn0));
rng(opts.seed);
for q = 1:numel(ebn0)
% ---- bits -> QPSK symbols (Gray coded, unit average energy) ----------
bits = randi([0 1], nBit, 1);
b = reshape(bits, 2, []).';
sym = (1 - 2*b(:, 1)) + 1j*(1 - 2*b(:, 2)); % ±1 ±1j, Es = 2
sym = sym / sqrt(2); % Es = 1
% ---- OFDM: reshape to subcarriers, IFFT, add CP ----------------------
S = reshape(sym, opts.nSub, opts.nSym);
s = ifft(S, opts.nSub, 1) * sqrt(opts.nSub); % unit-average-power time samples
sCp = [s(end-opts.cpLen+1:end, :); s];
% ---- AWGN at the sample level ---------------------------------------
% Eb/N0 -> sample noise variance.
% QPSK with unit Es: Es = 2*Eb, so the per-subcarrier SNR is 2*Eb/N0.
% The IFFT is scaled so each time sample has power Es = 1, and the FFT divides
% by sqrt(N), so the per-subcarrier signal power stays 1 and the noise power
% stays sigma2. Therefore sigma2 = 1/(2*Eb/N0), full stop.
%
% The first version also divided by (N+cp)/N "because the CP carries noise". That
% is wrong: the CP samples are DISCARDED at the receiver, so the noise in them
% goes with them. The CP penalty is a TRANSMIT power penalty (you must radiate
% more to keep the same received Es), not a receive-SNR penalty. Applying it here
% inflated sigma2 by 1.25x and produced BER 0.102 where theory says 0.0786.
ebn0lin = 10^(ebn0(q)/10);
sigma2 = 1/(2*ebn0lin);
rx = sCp + sqrt(sigma2/2)*(randn(size(sCp)) + 1j*randn(size(sCp)));
% ---- receiver: drop CP, FFT, slice ----------------------------------
rx = rx(opts.cpLen+1:end, :);
Y = fft(rx, opts.nSub, 1) / sqrt(opts.nSub);
y = Y(:);
bHat = [real(y) < 0, imag(y) < 0];
bHat = reshape(bHat.', [], 1);
berSim(q) = mean(bHat ~= bits);
berTh(q) = 0.5*erfc(sqrt(ebn0lin));
end
relErr = (berSim - berTh) ./ max(berTh, eps);
% TOLERANCE FROM COUNTING NOISE, NOT FROM A ROUNDED CONSTANT
% Under Gaussian noise the bit errors are (very nearly) a Poisson count, so the
% relative standard error of the measured BER is ~1/sqrt(nErr). A single flat
% tolerance cannot be right for every point: at 0 dB a 2^18-bit run counts
% ~20000 errors (0.7 % noise) while at 8 dB it counts ~49 (14 % noise). A flat
% 15 % therefore sits BELOW the noise floor at 8 dB and rejects correct code
% about half the time -- a flaky test is worse than no test, because the first
% response to a flaky test is to stop reading it.
% So the per-point pass mark is 3 sigma of that point's own counting noise,
% never below opts.tolMin and never above opts.tolMax.
nErr = berSim * nBit;
sigmaRel = 1 ./ sqrt(max(nErr, 1));
tolPt = max(opts.tolMin, min(opts.tolMax, 3 * sigmaRel));
r = struct();
r.ebn0dB = ebn0; r.berSim = berSim; r.berTheory = berTh;
r.relErr = relErr; r.nBits = nBit; r.nSym = opts.nSym; r.nSub = opts.nSub;
r.cpLen = opts.cpLen; r.seed = opts.seed;
r.tol = opts.tol;
r.tolMin = opts.tolMin;
r.tolMax = opts.tolMax;
r.nErrors = nErr;
r.sigmaRel = sigmaRel;
r.tolPerPoint = tolPt;
% ---- verdict from the five-state evidence flow, not from a flat tolerance --
% Each point is a binomial count graded against the closed form. The tolerance
% above stays as a reported diagnostic, but it no longer decides anything,
% because a two-state PASS/FAIL cannot express the case this ladder actually
% hits: a short run at high Eb/N0 counts ZERO errors when theory predicts 0.06,
% and that observation neither supports nor refutes the simulation -- it simply
% has no power. Reporting it as FAIL was the defect.
ev = cell(1, numel(ebn0));
states = cell(1, numel(ebn0));
for q = 1:numel(ebn0)
ev{q} = mas_stat_verdict(round(berSim(q)*nBit), nBit, struct( ...
'criterion', 'matches-theory', 'theory', max(berTh(q), realmin)));
states{q} = ev{q}.state;
end
if any(strcmp(states, 'sufficient-refutation'))
r.verdict = 'FAIL';
agg = states{find(strcmp(states, 'sufficient-refutation'), 1)};
elseif any(~strcmp(states, 'sufficient-support'))
r.verdict = 'UNDETERMINED';
agg = states{find(~strcmp(states, 'sufficient-support'), 1)};
else
r.verdict = 'PASS';
agg = 'sufficient-support';
end
r.states = states;
r.state = agg;
r.evidence = ev;
r.nPointRefuted = nnz(strcmp(states, 'sufficient-refutation'));
r.nPointUnpowered = nnz(strcmp(states, 'insufficient-sample') | ...
strcmp(states, 'interval-straddles-target'));
% One-sided would be wrong here, so it is not one-sided: a BER well BELOW theory
% means the noise variance is too small, which is the same defect as too much noise
% and just as fatal. The pass mark is wide enough to absorb counting noise but far
% narrower than the defect this check exists to catch -- applying the CP penalty at
% the receiver shifts BER by ~30 %.
r.ok = strcmp(r.verdict, 'PASS');
r.detail = sprintf('%d carriers, CP %d, %d bits: BER %s vs theory %s (rel %s, tol %s, %d refuted / %d unpowered of %d points, %s)', ...
opts.nSub, opts.cpLen, nBit, mat2str(berSim, 3), mat2str(berTh, 3), ...
mat2str(relErr, 2), mat2str(tolPt, 2), r.nPointRefuted, r.nPointUnpowered, ...
numel(ebn0), r.verdict);
end
% ---------------------------------------------------------------------------
function o = dflt(o, d)
f = fieldnames(d);
for k = 1:numel(f)
if ~isfield(o, f{k}), o.(f{k}) = d.(f{k}); end
end
end
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| planning-and-task-breakdownaddyosmani/agent-skillsBreaks work into ordered tasks. Use when you have a spec or clear requirements and need to break work into implementable tasks. Use when a task feels too large to start, when you need to estimate scope, or when parallel work is possible. | Stars: 119k | Updated | Claude CodeCodex | Verified | Install |
| screenpipethedotmack/claude-memSearch screen and audio history recorded by Screenpipe and optionally save selected context into claude-mem. Use when the user asks to recall desktop activity, a meeting, or something seen in another app through Screenpipe. | Stars: 118k | Updated | Claude CodeCodex | Verified | Install |
| contact-widgetnexu-io/open-designSelf-contained floating chat widget with welcome screen, social links, meeting button, and message input. Single HTML file, zero dependencies. | Stars: 108k | Updated | Claude CodeCodex | Verified | Install |
| meeting-notesnexu-io/open-designMeeting notes page — title bar with attendees, agenda checklist, decisions block, action items table with owners + dates, and a "next meeting" footer. Use when the brief mentions "meeting notes", "minutes", "1:1 notes", "all-hands recap", or "会议纪要". | Stars: 108k | Updated | Claude CodeCodex | Verified | Install |