Skip to main content

β: This English translation is in beta — the Traditional-Chinese original is the authoritative version.

Learning Path

This page expands the nine steps from the home page into a learning path you can actually follow, and extends the backbone with three advanced steps (Steps 10–12), for 12 steps in total. Each step tells you:

  • What to achieve (the learning goal of this step);
  • Which pages to read (in order);
  • Prerequisites (you will get stuck without these);
  • Expected outcome (what you should be able to do afterwards).

At the end there are two tracks: a fast track (roughly the nine-step backbone — grasp the ISF backbone in one afternoon) and a full track (all twelve steps: re-derive every formula yourself and run every lab yourself).

How to use this page: on a first pass, take the "full track" and come back to tick the checkbox after each step. For later review, the "fast track" is enough. A TODO: marker means that spot still needs manual verification against the original PDF; a "toy model" label means it is a pedagogical simplification, not transistor-level.

Course Progress
0/12 done
Progress is stored in the browser's localStorage (key isf-progress-v1), and only applies on this device and this browser.

Keep these three pages at hand as "dictionaries"​

Before you start, keep these three pages nearby; whenever you meet an unfamiliar symbol or paper reference, look it up there:

  • notation — the unified symbol table (site-wide symbols, units, per-paper cross-reference).
  • paper_summary_table — the five papers at a glance (who is responsible for what).
  • equation_index — every formula → derivation page → source.

Quick lookup (when unsure about a symbol / formula / English term, jump via the table below; when stuck, return to these four pages first):

What you want to look upWhere to go
All formulas on one page, canonical numerical examples A/B/CCheat Sheet
The meaning and units of a symbol (e.g. Γrms\Gamma_{rms}, qmaxq_{max}, c0c_0)Notation
An intuitive explanation of an English term (e.g. ISF, cyclostationary, limit cycle)Glossary
Which paper and Eq. a formula comes from, and where it is derivedEquation Index

Step 1: What an oscillator's "phase" actually is​

  • What to achieve: build the geometric picture of the limit cycle (the closed steady-state trajectory of an oscillator), and clearly separate phase (tangential along the cycle, no restoring force) from amplitude (radial away from the cycle, with a restoring force).
  • Pages to read: oscillator_phase → phase_vs_amplitude_noise.
  • Prerequisites: 2-D state space, the phase plane, the basic picture of RLC oscillation.
  • Expected outcome: you can explain "why a tangential perturbation persists forever while a radial one is pulled back" — this is exactly claim C2 (see claims_cross_reference).

Step 2: Noise is a small perturbation; the oscillator responds as LTV, not LTI​

  • What to achieve: understand that with respect to noise an oscillator is LTV (linear time-variant) — the same impulse injected at different phases produces a different phase shift, unlike an LTI system which only depends on t−τt-\tau.
  • Pages to read: lti_vs_ltv.
  • Prerequisites: Step 1; linear systems, impulse response, convolution.
  • Expected outcome: you can sketch the difference between the LTI h(t−τ)h(t-\tau) and the LTV hϕ(t,τ)h_\phi(t,\tau), corresponding to the figure lti_vs_ltv_impulse_response.png. This is claim C1.

Step 3: The operational definition of the ISF (impulse → phase)​

  • What to achieve: derive, starting from the capacitor relation q=Cvq=Cv, the operational definition of the ISF Δϕ=Γ(ω0τ) Δq/qmax\Delta\phi=\Gamma(\omega_0\tau)\,\Delta q/q_{max} ([P1] Eq.(9)–(11), p.182), and understand why Γ\Gamma is dimensionless and 2π2\pi-periodic.
  • Pages to read: impulse_to_phase_shift → isf_definition.
  • Prerequisites: Steps 1–2; the capacitor relation, unit conversions.
  • Expected outcome: you can do canonical example A in your head — with qmax=1q_{max}=1 pC, Δq=1\Delta q=1 fC, Γ=0.5\Gamma=0.5, f0=5f_0=5 GHz you get Δϕ=5×10−4\Delta\phi=5\times10^{-4} rad and Δt=15.9\Delta t=15.9 fs.

Step 4: From a single impulse to arbitrary noise (convolution)​

  • What to achieve: use superposition to generalize a single phase step into the LTV convolution ϕ(t)=1qmax∫−∞tΓ(ω0τ) in(τ) dτ\phi(t)=\frac{1}{q_{max}}\int_{-\infty}^{t}\Gamma(\omega_0\tau)\,i_n(\tau)\,d\tau ([P1] Eq.(11), p.182), and see why the upper integration limit tt (memory) makes the phase accumulate.
  • Pages to read: convolution_derivation.
  • Prerequisites: Step 3; convolution, integration.
  • Expected outcome: you can explain "phase is an integrator of noise", paving the way for the 1/f21/f^2 slope in Step 5.

Step 5: White noise → 1/f21/f^2, flicker → 1/f31/f^3​

  • What to achieve: derive the signature result L∝Γrms2/qmax2\mathcal{L}\propto\Gamma_{rms}^2/q_{max}^2 ([P1] Eq.(21), p.185, claim C3), and see that flicker noise upconverts into 1/f31/f^3 only through the DC term c0c_0 of the ISF ([P1] Eq.(23)(24), claims C4/C5).
  • Pages to read: white_noise_to_phase_noise → flicker_noise_upconversion.
  • Prerequisites: Step 4; PSD, Parseval.
  • Expected outcome: you can do canonical example B — the same set of numbers plugged into Eq.(21) gives L=−148.0\mathcal{L}=-148.0 dBc/Hz; and you understand the famous factor-of-2 SSB bookkeeping controversy mentioned in Section 3 of the spec.

Step 6: The Fourier view of the ISF (c0c_0, cnc_n, upconversion)​

  • What to achieve: expand the ISF into the Fourier series Γ=c02+∑cncos⁡(nω0τ+θn)\Gamma=\frac{c_0}{2}+\sum c_n\cos(n\omega_0\tau+\theta_n) ([P1] Eq.(12), p.183), understand that each cnc_n "downconverts" the noise near nω0n\omega_0 to the carrier, and that ∑cn2=2Γrms2\sum c_n^2=2\Gamma_{rms}^2 (Parseval, [P1] Eq.(20)).
  • Pages to read: fourier_series_of_isf → rms_isf.
  • Prerequisites: Step 5; Fourier series.
  • Expected outcome: you can state the mathematical basis of the design rule "symmetric waveform → c0≈0c_0\approx0 → suppressed 1/f31/f^3" (for the closed form that computes c0c_0 and the corner directly from topology parameters, see Step 10's asymmetric_isf_closed_form).

Step 7: Simulation labs — build numerical feel​

  • What to achieve: run the preceding formulas with your own hands, look at the figures, check the numbers, and turn the conversions between rad, fs, dBc/Hz and jitter into reflexes.
  • Pages to read: first numerical_feeling (three must-do mental calculations), then in order lab_01, lab_02, lab_04, lab_05, lab_06, lab_07, lab_08.
  • Prerequisites: Steps 3–6; a little Python/NumPy.
  • Expected outcome: you can verify canonical examples A/B/C in one line with the functions in simulations/common/; every figure is traceable to its script and formula in the figure_index.

Step 8: Design takeaways (symmetry, swing, slope)​

  • What to achieve: translate the formulas into design knobs — increase qmaxq_{max}, lower Γrms\Gamma_{rms}, enforce waveform symmetry to suppress c0c_0; and understand the ring result Γrms∝N−3/2\Gamma_{rms}\propto N^{-3/2} and "at fixed power and frequency, ring phase noise is almost independent of the number of stages NN" (claims C7/C8).
  • Pages to read: symmetry → lc_vs_ring.
  • Prerequisites: Steps 5–6.
  • Expected outcome: given an oscillator spec, you can say "which knob to turn first" (for the quantitative version of "how many dB from the theoretical ceiling", see Step 12's fom_limit).

Step 9: Connect to SerDes clocking (jitter, eye, PLL/CDR)​

  • What to achieve: integrate phase noise into rms jitter and connect it to eye closure and BER in a SerDes link.
  • Pages to read: psd_phase_noise_jitter → serdes_clocking_connection.
  • Prerequisites: Steps 5 and 7 (especially lab_08).
  • Expected outcome: you can do canonical example C — f0=5f_0=5 GHz, L(1MHz)=−100\mathcal{L}(1\text{MHz})=-100 dBc/Hz, 1/f21/f^2 slope, integrating 1→100 MHz ⇒σϕ=14.07\Rightarrow\sigma_\phi=14.07 mrad, σt=447.9\sigma_t=447.9 fs — and you know the integral is dominated by its lower limit (for the rigorous period/cycle-to-cycle kernels see Step 10's jitter_kernels; for the RJ/DJ decomposition and TJ@BER see Step 12's dj_dual_dirac).

Steps 10–12 (advanced): from deep theory to system integration​

The backbone (Steps 1–9) builds the ISF machinery for a single free-running oscillator. The next three steps push the same machinery in three directions: deeper theory (Step 10), the injected oscillator (Step 11), and the whole clock system (Step 12). The former "Advanced (optional)" section has been expanded and absorbed: paper_003 and paper_004 now live in Step 11; effective_isf (the cyclostationary correction Γeff=Γ⋅α\Gamma_{eff}=\Gamma\cdot\alpha, claim C9, plus the external mathematical foundations PPV/adjoint/Floquet, claim C13, not in these 5 PDFs, supplemented from standard literature) remains optional — read it before starting Step 10 if you can.

Step 10: Advanced theory — from κ to lineshape​

  • What to achieve: take "white-noise phase diffusion has only one free parameter" all the way — the same phase-variance growth rate κ2\kappa^2 wearing five outfits (κ\kappa, DD, linewidth, ADEV, the 1/f21/f^2 coefficient); the rigorous frequency-domain kernels of the three jitters (TIE / N-period / cycle-to-cycle); when the lineshape stops being Lorentzian under flicker FM; the closed form for c0c_0 and the 1/f31/f^3 corner; and the three ways to compute Γ\Gamma directly from the waveform.
  • Pages to read (one line of "why" each):
    1. diffusion_dictionary — κ\kappa, DD, linewidth, ADEV, and the 1/f21/f^2 coefficient are five outfits of the same number; get the conversion dictionary first (canonical κ2=0.125\kappa^2=0.125 rad²/s).
    2. jitter_kernels — the three jitters are the 0th/1st/2nd-order differences of ϕ\phi; the prefactors and every single 2 are derived from first principles.
    3. beyond_lorentzian — flicker FM turns the lineshape from Lorentzian into near-Gaussian, and rigorously answers "what does the instrument actually measure".
    4. asymmetric_isf_closed_form — the [P2] App. B closed form: compute Γrms\Gamma_{rms}, c0c_0 and the 1/f31/f^3 corner directly from the stage count NN and the asymmetry AA.
    5. isf_from_waveform — the three methods of the [P1] appendix (impulse injection / closed form / first derivative), and where each extra approximation starts to fail.
  • Prerequisites: Steps 5, 6 and 9; plus lorentzian_linewidth and allan_variance (direct prerequisites of diffusion_dictionary — catch up on them first if needed).
  • Expected outcome: you can convert any jitter spec into any other representation, and you know the failure point of every approximation.
  • Self-check checkpoints:
    • With the canonical numbers (Γrms=0.5\Gamma_{rms}=0.5, qmax=1q_{max}=1 pC, Si=10−24S_i=10^{-24} A²/Hz), κ2=0.125\kappa^2=0.125 rad²/s; switching to the true ideal LC (Γrms=1/2\Gamma_{rms}=1/\sqrt2), why is it exactly 2× (0.250.25 rad²/s)?
    • Under the "single-sided SϕS_\phi, ∫0∞\int_0^\infty" convention, the period-jitter kernel's prefactor is 1/ω021/\omega_0^2; stuffing a single-sided spectrum into the 2/ω022/\omega_0^2 version common in the literature overcounts the variance by 2× (jitter by 2\sqrt2) — which bookkeeping convention does that 2 belong to?
    • Same L(10 kHz)\mathcal{L}(10\,\text{kHz}) spec: white-FM linewidth 50 Hz, flicker-FM about 3.1 kHz — why does a dBc/Hz number at a single offset not determine the linewidth at all?

Step 11: Injection locking and frequency conversion​

  • What to achieve: extend the same Γ\Gamma from phase noise into the "injected" world — the generalized Adler equation, in-lock noise shaping and out-of-lock pulling, acquisition transients and cycle slips, M:N subharmonic locking and the ILFD, mutual injection (QVCO), and finally the sub-sampling PLL that kicks the divider out of the loop.
  • Pages to read (one line of "why" each):
    1. paper_003 — the generalized Adler equation: a single first-order equation written with the ISF, giving the lock range for any topology and any injection waveform (claim C10).
    2. injection_locking_noise — a locked oscillator is a first-order PLL: its own noise is high-passed (corner ωc\omega_c), the reference low-passed; when lock fails, the one-sided pulling comb.
    3. lab_36 — the missing transient pieces: the exact closed-form acquisition trajectory, critical slowing at the lock-range edge, noise-induced cycle slips.
    4. paper_004 — the APF (the amplitude counterpart of the ISF, claim C11) and M:N subharmonic locking / ILFD.
    5. injection_locked_division — the ÷N direction of M:N locking gets its own page: the lock range rides the ISF's NN-th harmonic, and half-wave symmetry can't divide by 2.
    6. subharmonic_injection — the other half of the duality: multiplication rides the injection waveform's own NN-th harmonic, with a realignment factor β\beta and discrete-time noise shaping.
    7. lab_40 — an independent numerical check of subharmonic injection: lock range ∝1/N\propto1/N, β\beta's ODE step response, and the closed-form output jitter.
    8. quadrature_and_coupled_oscillators — mutual injection = two Adler equations: the QVCO's coupling-strength ↔ phase-error ↔ phase-noise triangular trade-off and that famous ∼3\sim3 dB.
    9. sampling_pll — sampling the zero crossing = sampling the ISF's most sensitive point: sub-sampling makes the divider noise vanish and the CP noise no longer ×N2\times N^2.
  • Prerequisites: Steps 6 and 8; catching up on effective_isf first is recommended (cyclostationary correction, see the top of this section).
  • Expected outcome: from the generalized Adler equation all the way to the noise shaping of ILFD / QVCO / sub-sampling — you can explain injection-related locking, pulling and frequency conversion along the same single Γ\Gamma.
  • Self-check checkpoints:
    • With sinusoidal injection and the ideal-LC ISF, the generalized Adler equation reduces to θ˙=Δω−ωLsin⁡θ\dot\theta=\Delta\omega-\omega_L\sin\theta with ωL=Iinj/(2qmax)\omega_L=I_{inj}/(2q_{max}); in lock, the oscillator's own noise is high-pass shaped with corner ωc=ωL2−Δω2\omega_c=\sqrt{\omega_L^2-\Delta\omega^2} (the pull-in frequency of [P3] Eq.(40)) — why does the noise suppression vanish at the edge of the lock range?
    • Out of lock, the beat frequency is ωb=Δω2−ωL2\omega_b=\sqrt{\Delta\omega^2-\omega_L^2} ([P4] Eq.(34)) and the sideband comb grows on one side only; with the canonical numbers (fL=5f_L=5 MHz, true-LC noise) at r=0.8r=0.8 the cycle-slip rate is ∼10−1.86×107\sim10^{-1.86\times10^7} — why do we say thermal slips are a "cliff", not a "slope"?
    • M:N subharmonic locking gives ωL=Iinj∣Γ~N∣/2\omega_L=I_{inj}\lvert\tilde\Gamma_N\rvert/2 ([P4] Eq.(28)–(30)): which harmonic does the ÷2 ILFD ride on? The sub-sampling PLL's in-band floor drops from −118.9-118.9 to −126.0-126.0 dBc/Hz (illustrative example) — where did the divider term go?

Step 12: System integration and measurement​

  • What to achieve: finish at the system level — book-keep a single oscillator's L(f)\mathcal{L}(f) down the clock chain to every node, compare against the FOM theoretical ceiling, understand what a reference oscillator actually buys, compute an ADC's SNR/ENOB, book-keep RJ/DJ separately into TJ@BER, read measurement plots and spurs, dodge the 12 mines, and finally use the capstone to validate the whole site end to end.
  • Pages to read (one line of "why" each):
    1. clock_chain_budget — the four bookkeeping rules (×N / ÷N / PLL / buffer) plus a complete worked chain 100 MHz→5 GHz→2.5 GHz.
    2. fom_limit — the FOM ceiling =173.8−10log⁡10(Feff)=173.8-10\log_{10}(F_{eff}) dB (300 K): know how many dB your design is from the physical limit.
    3. reference_oscillators — a crystal is just an LC tank with an outrageously high QQ: why nothing downstream can fix the reference's close-in noise.
    4. adc_aperture_jitter — sampling error = slope × timing error: clock quality directly sets the data converter's effective number of bits.
    5. dj_dual_dirac — RJ unbounded, DJ bounded: the dual-Dirac model and the industry-standard TJ@BER bookkeeping.
    6. measurement_and_spurs — the three ways to measure L(f)\mathcal{L}(f), telling spurs from random noise, and how to read a real PN plot.
    7. common_mistakes — 12 real mines: a full review of the site's factor-of-2 discipline.
    8. capstone_lc_end_to_end — the site's main spine end to end: state equations → ISF → spectrum → linewidth → jitter → BER. The finale.
  • Prerequisites: Steps 7–9; Step 10 (the jitter kernels and the diffusion dictionary are used repeatedly); Step 11 helps too (sampling_pll already appeared in Step 11).
  • Expected outcome: you can draw up — and defend — a clock noise budget: from the reference to the sampler, every stage's L(f)\mathcal{L}(f) and the final jitter have a traceable source.
  • Self-check checkpoints:
    • The worked chain (100 MHz → ×50 PLL → 5 GHz → ÷2 → 2.5 GHz → buffer) ends at an integrated jitter of 27.6 fs; under ideal ×N / ÷N, why does σt\sigma_t in seconds not change by a single fs?
    • Feeding a 5 GHz input with the canonical σt=447.9\sigma_t=447.9 fs clock: SNRjitter=37.0\text{SNR}_{jitter}=37.0 dB, ENOB =5.86=5.86 bit; 10 ENOB @ 5 GHz requires pushing σt\sigma_t down to ≤25.4\le25.4 fs — which formula is doing the conversion? (SNR=−20log⁡10(2πfinσt)\text{SNR}=-20\log_{10}(2\pi f_{in}\sigma_t))
    • TJ(BER)=DJδδ+2Q⋅σ\text{TJ}(\text{BER})=\text{DJ}_{\delta\delta}+2Q\cdot\sigma with Q(10−12)=7.03Q(10^{-12})=7.03: why is DJδδ≤DJpp\text{DJ}_{\delta\delta}\le\text{DJ}_{pp} a deliberate under-report? And why is a spur specified in dBc and not dBc/Hz?

Two tracks​

Fast track (about one afternoon, backbone only)​

This corresponds to the Steps 1–9 backbone (Steps 10–12 are not part of the fast track). The goal is "understand what the ISF is and how it determines phase noise". You do not re-derive anything and do not run the simulations yourself:

  1. oscillator_phase (only the limit cycle and the phase/amplitude picture)
  2. lti_vs_ltv (the core LTV conclusion and figure)
  3. impulse_to_phase_shift (memorize the operational definition + example A)
  4. white_noise_to_phase_noise (memorize Eq.(21) and example B)
  5. fourier_series_of_isf (the intuition of c0c_0 suppressing 1/f31/f^3)
  6. numerical_feeling (the three mental calculations)
  7. serdes_clocking_connection (the jitter and eye conclusions)

After this you should be able to answer: what the ISF is, why LTV, L∝Γrms2/qmax2\mathcal{L}\propto\Gamma_{rms}^2/q_{max}^2, why symmetry matters, and how phase noise becomes jitter.

Full track (re-derive every step + run every lab)​

Follow Steps 1 through 12 above in order, cover the answer and re-derive every formula yourself, run every lab yourself with python scripts/run_all_sims.py and check figures against numbers (the advanced labs referenced in Steps 10–12 included). Suggested pace:

StageStepsKey deliverable (self-check)
Geometric intuition1–2Can draw the limit cycle; can state LTV vs. LTI
ISF core3–6Can re-derive Eq.(9)→(11)→(12)→(20)→(21)→(24)
Hands-on7Can reproduce examples A/B/C; orders of magnitude match
Design8–9Can list the design knobs; can integrate L\mathcal{L} into σt\sigma_t
Advanced theory10Can convert freely among κ\kappa, DD, linewidth, ADEV and the 1/f21/f^2 coefficient; can name where every 2 comes from
Injection & conversion11Can derive ωc\omega_c, ωb\omega_b and the M:N lock condition from the generalized Adler equation; can say where sub-sampling wins
System integration12Can book-keep L(f)\mathcal{L}(f) down a full clock chain; can draw up and defend a jitter budget

Key takeaways​

  • Twelve steps = the nine-step backbone (phase geometry → LTV → ISF definition → convolution → white/flicker noise → Fourier → labs → design → SerDes) plus three advanced steps (Step 10 κ and lineshape → Step 11 injection locking and frequency conversion → Step 12 system integration and measurement).
  • The four quick-lookup pages (return there first when stuck): cheat_sheet, notation, glossary, equation_index (see the "Quick lookup" table at the top of this page).
  • The fast track covers the nine-step backbone; the full track walks all twelve steps: re-derive every formula, run every lab.
  • Steps 10–12 are advanced: injection/APF now live in Step 11; PPV/adjoint and much of the instrument/architecture knowledge in the system-integration pages come from external literature (not in the 5 PDFs) — every such page carries an honest disclaimer.

Further reading​