Green FIR Filters with Large Ratio of Sample Rate to Bandwidth

This presentation will show you how to design and implement narrowband filters with more than an order of magnitude reduction of workload. I was recently challenged to reduce the workload for a 301 tap low pass FIR filter with sample rate 50 times the bandwidth. After my first approach in which I reduced the workload to 21 multiplies I wondered by how much could we reduce the workload? I finally stopped playing with the question when I reached 6 multiplies, which is a 50-to-1 workload reduction. The technique we present usually reduces the workload by a factor greater than 10. The only requirement to apply these techniques is that there be a large ratio of sample rate to bandwidth. Once we learn the simple trick to accomplish this reduction we then pose the next question: can we achieve similar reduction in workload when there is not a large ratio of sample rate to bandwidth? The answer surprisingly is yes? We will share the recipe for the secret sauce so you too will know how wideband filters can also be implement with more than an order of magnitude workload reduction. How about a pair of 1400 complex tap filters replaced with 100 real multiplies?

When I first started showing folks how to build FIR filters with an order of magnitude workload reduction, no-one seemed interested in clever solutions. I realized I had a marketing blind spot. I fixed that blind spot and now tell folks let me show you a green solution to your problem! There is hardly any room left on the bandwagon. 

What this presentation is about and why it matters

Fred Harris shows how to build "green" FIR filters: implementations that meet tight passband/stopband specs while using orders of magnitude fewer multiplies and adds. The trick is to exploit large ratios of sample rate to signal bandwidth (and, more generally, multirate and polyphase structures) so most heavy computation runs at a much lower rate. The result: filters that once required hundreds of taps can be realized with a few dozen operations per input, saving power, silicon area, and heat — critical in embedded systems, satellites, hearing aids, and any application where computation cost matters.

Who will benefit the most from this presentation

• DSP engineers designing channelizers, decimators/interpolators, or narrowband filters for communications, radar, or audio.
• System architects concerned with power, heat, and resource usage (satellites, mobile devices, hearing aids).
• Students and teachers who want an intuitive, practical tour of polyphase filters, multirate trade-offs, and how to reduce implementation workload.
• Anyone implementing FIR filters in hardware (ASIC/FPGA) or software who needs to shrink multiply/add counts without relaxing specs.

What you need to know

This talk assumes basic familiarity with discrete-time filtering and the Fourier transform. Key background topics that will make the talk easier to follow:

• Sinc and impulse response intuition — how a finite-length windowed sinc gives a frequency-domain main lobe and sidelobes, and how length controls the transition width.
• Relationship between filter length and transition bandwidth — the basic rule of thumb Fred uses is $N = F_s / \Delta F$, where $N$ is the number of taps, $F_s$ is the sample rate, and $\Delta F$ is the transition width (in Hz or normalized units). This sets the minimum length for a given transition.
• Polyphase decomposition and Noble identities — how moving a downsampler/upsampler through a filter produces per-phase (polyphase) filters that run at the low rate rather than computing everything at the high rate.
• Decimation/interpolation intuition — you can safely downsample after adequate antialias filtering; by merging filtering with resampling (doing the filtering in the lower-rate domain) you avoid computing transient samples that will be discarded.
• Filter-bank / FFT-based channelizers — using polyphase+FFT (or polyphase + short DFTs) lets you split many narrow bands to low-rate processing blocks and then reconstruct the desired band(s) with much less cost than a single long FIR.
• Design trade-offs —tighter stopband attenuation or smaller passband ripple increases required taps; increasing transition bandwidth or lowering sample rate reduces taps. Fred often uses a rule-of-thumb relating attenuation (in dB) to a multiplier on the base length (attenuation/22 ≈ multiplicative factor).

Glossary

• Polyphase filter — decomposition of an FIR into M sub-filters (phases) that operate at 1/M the input rate; central to efficient multirate implementations.
• Decimation — reducing the sample rate by an integer factor M (keep every Mth sample) together with the needed anti-alias filtering.
• Interpolation — increasing the sample rate by inserting zeros (or using other schemes) followed by a reconstruction filter to suppress images.
• Noble Identity — algebraic property that allows moving delays across downsamplers/upsamplers so work can be done at the lower rate.
• Transition bandwidth ($\Delta F$) — frequency interval between passband and stopband; governs required filter length together with sample rate.
• Polyphase FFT channelizer — a bank that uses polyphase decomposition plus an FFT to split the spectrum into many low-rate channels efficiently.
• Half-band filter — special FIR where every other coefficient (except center) is zero; efficient for cascaded 2:1 stages.
• Gatling-gun model — reuse one short low-rate filter with many different coefficient sets, stepped through sequentially to emulate many parallel filters with far less hardware.
• Gibbs phenomenon — ripple/overshoot resulting from truncating a bandlimited impulse response; reduced by trading transition width for sidelobe suppression.
• Equal-ripple / Remez (Parks–McClellan) — an algorithm that produces optimal minimax FIR designs (equal ripple), useful but sometimes not optimal in system-resource terms.

Suggested approach: skim the talk once to absorb the high-level strategy (downsample → cheap low-rate filter → upsample) and then rewatch the design examples (24:1, 10:1, cascaded stages) to map the trade-offs to your own specs. Expect practical tips on when to use polyphase banks, half-band cascades, or FFT-based channelizers depending on your bandwidth:sample-rate ratio.

This overview is AI-generated from the session transcript. Spot an issue? Let us know.