Engineers often ask: should I replace EMA with GSRF to reduce oscillation? The honest answer is band-dependent. GSRF Practical is a soft thermostat in log-space— a spring to a declared normal—not a linear low-pass.
What locked packs show
- ~+35% osc31 vs best EMA on phase6.2 actuator mocks (3 seeds, full series). osc31 = std of series minus rolling mean of 31 samples.
- ~−70% sweet-band spectral power vs EMA on the same packs (period roughly 50–120 samples). This is not a rolling-31-only artifact.
- Frequency-sweep sweet spot roughly P = 50–120; peak osc advantage ~+51% at P=80.
When EMA can still win on osc31
On the official frequency sweep, GSRF loses osc31 at very fast periods (~P=20) and very slow periods (~P=400–600) versus best EMA. If your “oscillation” is actually slow process drift you want tracked, EMA is often the better tool.
Why the soft thermostat helps mid-band ring
Pure-sine tests show GSRF AC gain plateaus near ~0.37 while EMA gain approaches 1.0 at long periods. Mid-band energy is clamped toward normal instead of fully followed. That is the intended identity.
Practical takeaway
- Use GSRF when the costly problem is mid-band ring / chatter around a known normal.
- Keep EMA (or k_return≈0) when the job is minimum lag tracking.
- Validate on your sampling rate: “period 80” is eighty samples—often ~80 minutes on 1‑min industrial traces.
Sources: Evidence · characterization v4 · probes v5 FULL PSD · excellence hunt v2 · phase6.2. Primary page: /evidence#frequency