Computing an FFT of Complex-Valued Data Using a Real-Only FFT Algorithm
Rick Lyons shows a compact trick to get an N-point complex FFT using only real-input FFT routines by transforming the real and imaginary parts separately and recombining their outputs. The post presents a one-line recombination formula, Xc(m) = real[Xr(m)] - imag[Xi(m)] + j{imag[Xr(m)] + real[Xi(m)]}, and an algebraic derivation based on the two-real-in-one-complex FFT identity. Useful for systems that only provide real-only FFTs.
Unit Testing for Embedded Algorithms
Unit testing is a best practice for embedded algorithm development, and Anthony Ricke shows how to apply it to DSP code so host and target behave identically. He demonstrates writing unit tests, stubbing Blackfin fixed-point functions in the workstation, and using test-driven development to safely port and optimize an average-calculation example. The SourceForge examples make the approach practical to adopt.
Using Mason's Rule to Analyze DSP Networks
When algebra gets messy, Rick Lyons shows how Mason's Rule cuts through the tedium to produce z-domain transfer functions for even nested-feedback DSP networks. The post gives a clear step-by-step procedure, definitions, and worked examples including a biquad, a DC-bias remover, and a complex multi-loop network. It also points to a public MATLAB routine to automate the bookkeeping.
The Nature of Circles
Averaging angles the usual way can produce nonsense: the mean of 0 and 359 degrees is not 179.5 when working with circular data. Peter Kootsookos shows the correct approach using vectorial or phasor averaging, converting angles to unit complex numbers and taking the argument of their sum. The short post points to directional statistics and a related IEEE paper for deeper details.
Simultaneously Computing a Forward FFT and an Inverse FFT Using a Single FFT
Rick Lyons presents a compact seven-step algorithm to compute a forward FFT and an inverse FFT at the same time using a single radix-2 complex FFT. The method builds intermediate sequences v(n) and z(n), exploits conjugate symmetry, and requires only one N-point FFT plus about 2N additions or subtractions. A clear MATLAB implementation accompanies the explanation so you can try it immediately.
Multiplierless Exponential Averaging
Rick Lyons shows how to implement exponential averaging without multiplies by exploiting a rearranged leaky-integrator form and binary shifts. He demonstrates reducing the standard two-multiply averager to a single-multiply form, then eliminating the multiply entirely when the weighting α equals reciprocals or differences of reciprocals of powers of two. The post catalogs practical α choices for fixed-point filters and flags quantization as an open issue.
Computing the Group Delay of a Filter
Rick Lyons presents a neat, practical way to get a filter's group delay directly from its impulse response using only DFTs. The method computes an N-point DFT of h(n) and of n·h(n), divides them in the frequency domain, and takes the real part to obtain group delay in samples, avoiding phase unwrapping. The post includes MATLAB code, a zero-division warning, and a caution that the method is reliable for FIR filters but not always for IIRs.
Computing Large DFTs Using Small FFTs
Rick Lyons demonstrates a practical trick for computing large N-point DFTs by combining multiple smaller radix-2 FFTs when only limited FFT sizes are available. He walks through 16-point and 24-point examples using two and three 8-point FFTs, shows how to assemble outputs with twiddle factors, and explains a symmetry that reduces twiddle storage to N/4 values. The method supports non-power-of-two DFT lengths.
Linear-phase DC Removal Filter
Rick Lyons presents a practical, multiplier-free way to remove DC while preserving linear phase by cascading D-point moving-average filters. He shows how choosing D as a power of two gives bit-shift scaling, how a dual-MA yields a narrow transition band with modest ripple, and how a quad-MA drives ripple down to near inaudible levels while noting the fixed-point accumulator sizing required.
Correlation without pre-whitening is often misleading
Correlation sounds like the obvious way to find a known pattern, but Peter Kootsookos shows why it can go badly wrong on real, nonwhite data. Using an image example with overlapping blobs, he demonstrates that pre-whitening, here done with a simple row difference, can turn a messy correlation result into a sharply localized peak.
Using the DFT as a Filter: Correcting a Misconception
Some sources claim the DFT, when used as a filter, shifts spectral energy down to DC. Rick Lyons shows that this is not true for consecutive DFT-bin outputs and explains the cause of the confusion: the FIR interpretation requires reversing the usual twiddle-factor order. He derives the DFT-bin frequency response, shows the bandpass center at 2πm/N, and explains when decimation does produce a translation to zero Hz.
Frequency Translation by Way of Lowpass FIR Filtering
Rick Lyons shows how you can translate a signal down in frequency and lowpass filter it in a single operation by embedding cosine mixing values into FIR coefficients. The post explains how to build the translating FIR, how to choose the number of coefficient sets, and how decimation can dramatically reduce storage needs while noting practical constraints like the requirement that ft be an integer submultiple of fs.
There's No End to It -- Matlab Code Plots Frequency Response above the Unit Circle
If you want a fresh way to inspect a digital filter, this post introduces plotfil3d, a compact MATLAB function that wraps the magnitude response around the unit circle in the Z-plane so you can view it in 3D. It uses freqz to compute H(z) in dB for N points and accepts an optional azimuth to change the viewing angle; the code is provided in the appendix.
Computing an FFT of Complex-Valued Data Using a Real-Only FFT Algorithm
Rick Lyons shows a compact trick to get an N-point complex FFT using only real-input FFT routines by transforming the real and imaginary parts separately and recombining their outputs. The post presents a one-line recombination formula, Xc(m) = real[Xr(m)] - imag[Xi(m)] + j{imag[Xr(m)] + real[Xi(m)]}, and an algebraic derivation based on the two-real-in-one-complex FFT identity. Useful for systems that only provide real-only FFTs.
A poor man's Simulink
Markus Nentwig built a compact glue layer that embeds NGSPICE into Octave to cosimulate continuous-time circuits and digital control. The article walks through an RC lowpass example, the MEX-based Octave interface, and the breakpoint-driven cosimulation flow, showing how adaptive SPICE integration handles asynchronous and time-triggered events. It presents a practical, low-cost alternative to Simulink for tightly coupled analog-digital system design.
Exact Near Instantaneous Frequency Formulas Best at Zero Crossings
Cedron Dawg derives time-domain formulas that yield near-instantaneous frequency estimates optimized for zero crossings of pure tones. Complementing his earlier peak-optimized results, these difference-ratio formulas work for real and complex signals, produce four-sample estimators similar to Turners, and cancel amplitude terms, making them attractive low-latency options for clean tones while warning they degrade in noise and at peaks.
The Nature of Circles
Averaging angles the usual way can produce nonsense: the mean of 0 and 359 degrees is not 179.5 when working with circular data. Peter Kootsookos shows the correct approach using vectorial or phasor averaging, converting angles to unit complex numbers and taking the argument of their sum. The short post points to directional statistics and a related IEEE paper for deeper details.
How to Find a Fast Floating-Point atan2 Approximation
This post shows how a compact, fast atan2 can be built from a Remez-derived arctangent approximation and a matching 3rd-order polynomial. It walks through using Boost's remez_minimax to recover coefficients 0.97239411 and -0.19194795, integrating the polynomial into an atan2 with quadrant reduction, and applying branch reduction, bit tricks, and SSE2 SIMD to cut runtime while keeping max error under about 0.005 radians.
Multimedia Processing with FFMPEG
FFMPEG is a set of libraries and a command line tool for encoding and decoding audio and video in many different formats. It is a free software project for manipulating/processing multimedia data. Many open source media players are based on FFMPEG libraries.
A Complex Variable Detective Story – A Disconnect Between Theory and Implementation
A subtle phase-wrap gotcha turned a clean pencil-and-paper derivation into a software mismatch for a 5-tap FIR filter with complex coefficients. Rick Lyons shows why two algebraically equivalent-looking expressions can disagree in code, and traces the real culprit to angle limits in rectangular-form complex arithmetic. The fix is simple once you see it, but the trap is easy to miss.
Least-squares magic bullets? The Moore-Penrose Pseudoinverse
Markus Nentwig walks through a practical way to remove power-line hum from measurements using the Moore-Penrose pseudoinverse. He builds a harmonic basis, computes pinv(basis) to get least-squares coefficients, and reconstructs and subtracts the hum, with a ready-to-run Matlab example. The post highlights limits and performance: basis-like signal components will be removed, and accuracy improves with the square root of sample count.
"Neat" Rectangular to Polar Conversion Algorithm
Rick Lyons revisits a clever slide-rule era trick for estimating the magnitude of a complex number without computing a square root. He highlights a neat identity, prompted by a Jerry Avins post, that converts the sqrt problem into forward and inverse trigonometric operations plus ratios. The post invites readers to derive Eq. (2) and see why a seemingly complex idea is actually simple and practical.
A Remarkable Bit of DFT Trivia
Rick Lyons highlights a surprising equality: the DFT's worst-case scalloping loss equals 2/π, the same probability that a toothpick crosses a floorboard seam in Buffon's needle problem when the toothpick equals board width. The post sketches the DFT bin-intersection derivation and connects the math to the classic probability puzzle, offering a playful insight that sharpens intuition about bin responses.
Simultaneously Computing a Forward FFT and an Inverse FFT Using a Single FFT
Rick Lyons presents a compact seven-step algorithm to compute a forward FFT and an inverse FFT at the same time using a single radix-2 complex FFT. The method builds intermediate sequences v(n) and z(n), exploits conjugate symmetry, and requires only one N-point FFT plus about 2N additions or subtractions. A clear MATLAB implementation accompanies the explanation so you can try it immediately.
There's No End to It -- Matlab Code Plots Frequency Response above the Unit Circle
If you want a fresh way to inspect a digital filter, this post introduces plotfil3d, a compact MATLAB function that wraps the magnitude response around the unit circle in the Z-plane so you can view it in 3D. It uses freqz to compute H(z) in dB for N points and accepts an optional azimuth to change the viewing angle; the code is provided in the appendix.
Fibonacci trick
Tim Wescott shares a compact, surprising trick linking Fibonacci numbers and difference equations. Start with any two consecutive Fibonacci numbers, negate the larger-magnitude one, and iterate the usual recurrence; after a few steps you'll arrive at the standard Fibonacci sequence or its negative. This behavior is specific to the Fibonacci recurrence and makes a great illustrative example for teaching linear recurrences.
Exact Near Instantaneous Frequency Formulas Best at Peaks (Part 1)
Cedron Dawg presents a new family of exact time-domain formulas to estimate the instantaneous frequency of a single pure tone. The methods generalize a known one-sample formula into k-degree neighbor-pair sums with spacing d, giving exact results in the noiseless case and tunable robustness in noise. The paper explains why real-tone estimates must be taken at peaks and shows the formulas also work for complex tones.
Correlation without pre-whitening is often misleading
Correlation sounds like the obvious way to find a known pattern, but Peter Kootsookos shows why it can go badly wrong on real, nonwhite data. Using an image example with overlapping blobs, he demonstrates that pre-whitening, here done with a simple row difference, can turn a messy correlation result into a sharply localized peak.
Exact Near Instantaneous Frequency Formulas Best at Peaks (Part 2)
Cedron Dawg derives a second family of exact time domain formulas for single-tone frequency estimation that trade a few extra calculations for improved noise robustness. Built from [1+cos]^k binomial weighting of neighbor-pair sums, the closed-form estimators are exact and are best evaluated at signal peaks for real tones, while complex tones do not share the zero-crossing limitation. Coefficients up to k=9 are provided.
A Complex Variable Detective Story – A Disconnect Between Theory and Implementation
A subtle phase-wrap gotcha turned a clean pencil-and-paper derivation into a software mismatch for a 5-tap FIR filter with complex coefficients. Rick Lyons shows why two algebraically equivalent-looking expressions can disagree in code, and traces the real culprit to angle limits in rectangular-form complex arithmetic. The fix is simple once you see it, but the trap is easy to miss.
