12 Jul 2026
Reading the Exceptional-Point Project IV: Petermann Factor from First Principles
This self-contained post derives the Petermann factor, relates it to phase rigidity, and explains the robustness statement used in the paper.
The main relation for the reciprocal complex-symmetric arrays is
\[K=r^{-2}.\]Near an exceptional point, $r\to0$ and therefore $K\to\infty$. A large eigenvalue response is consequently accompanied by strong eigenvector nonorthogonality.
1. Minimal Background: Left and Right Eigenvectors
For a non-Hermitian matrix,
\[H\neq H^\dagger.\]Right eigenvectors satisfy
\[H\psi_k^R=E_k\psi_k^R.\]Left eigenvectors satisfy
\[(\psi_k^L)^\dagger H =E_k(\psi_k^L)^\dagger.\]Equivalently,
\[H^\dagger\psi_k^L=E_k^*\psi_k^L.\]Unlike the Hermitian case, $\psi_k^L$ and $\psi_k^R$ need not be the same vector.
2. Why Left-Right Overlap Matters
Suppose the Hamiltonian changes by a small perturbation $\Delta H$. To first order, the eigenvalue shift of a simple eigenvalue is
\[\Delta E_k \approx \frac{(\psi_k^L)^\dagger \Delta H\,\psi_k^R} {(\psi_k^L)^\dagger\psi_k^R}.\]This formula is the non-Hermitian version of ordinary first-order perturbation theory.
If the denominator
\[(\psi_k^L)^\dagger\psi_k^R\]is small, a modest perturbation can produce a large eigenvalue change. Thus left-right nonorthogonality controls eigenvalue conditioning.
3. Definition of the Petermann Factor
A normalization-independent measure is
\[K_k= \frac{ (\psi_k^R)^\dagger\psi_k^R\, (\psi_k^L)^\dagger\psi_k^L }{ |(\psi_k^L)^\dagger\psi_k^R|^2 }.\]The Cauchy-Schwarz inequality implies
\[K_k\ge1\]for an isolated simple mode.
| Value | Interpretation |
|---|---|
| $K=1$ | orthogonal, well-conditioned mode |
| $K>1$ | nonorthogonal mode |
| $K\to\infty$ | approach to a defective exceptional point |
The factor is widely associated with excess noise and linewidth enhancement in open resonant systems. Mathematically, it is closely related to the square of an eigenvalue condition number.
4. Reduction for a Complex-Symmetric Hamiltonian
The paper uses reciprocal arrays with
\[H^{\mathsf T}=H.\]If
\[H\psi_k=E_k\psi_k,\]then transposition gives
\[\psi_k^{\mathsf T}H =E_k\psi_k^{\mathsf T}.\]Thus the left row vector may be taken as $\psi_k^{\mathsf T}$, or the left column as $\psi_k^*$.
The Petermann factor becomes
\[K_k= \frac{(\psi_k^\dagger\psi_k)^2} {|\psi_k^{\mathsf T}\psi_k|^2}.\]Define phase rigidity
\[r_k= \frac{|\psi_k^{\mathsf T}\psi_k|} {\psi_k^\dagger\psi_k}.\]Therefore
\[\boxed{K_k=r_k^{-2}}.\]The relation is exact for the complex-symmetric model.
5. Two-Site Example
For
\[H_2= \frac12 \begin{pmatrix} i\gamma&\Omega\\ \Omega&-i\gamma \end{pmatrix}, \qquad q=\Omega^2-\gamma^2>0,\]the phase rigidity is
\[r_2=\frac{\sqrt q}{|\Omega|}.\]Hence
\[K_2= \frac{\Omega^2}{q}.\]As $q\to0^+$,
\[r_2\to0, \qquad K_2\to\infty.\]For example, if
\[\frac{q}{\Omega^2}=10^{-2},\]then
\[r_2=10^{-1}, \qquad K_2=10^2.\]The Petermann factor amplifies the loss of rigidity quadratically.
6. Structured Scaling for an $\mathrm{EP}_N$
For the structured spin array, define
\[\rho=\frac{q}{\Omega^2}.\]The exact near-EP phase rigidity of mode $m$ is
\[r_{jm}\sim c_{N,m}\rho^{(N-1)/2},\]where
\[c_{N,m}= \binom{N-1}{j-m}^{-1}.\]Therefore
\[K_{jm}\sim c_{N,m}^{-2} \rho^{-(N-1)}.\]The divergence becomes more severe as the exceptional-point order increases.
For $N=4$:
| Modes | $c_{4,m}$ | Leading $K$ |
|---|---|---|
| outer, $m=\pm3/2$ | $1$ | $\rho^{-3}$ |
| inner, $m=\pm1/2$ | $1/3$ | $9\rho^{-3}$ |
The exponent is common, while the finite prefactor distinguishes the modes.
7. Generic Structure-Breaking Scaling
A perturbation that activates the full Jordan chain gives
\[E-E_{\mathrm{EP}} \sim\delta^{1/N}.\]For a one-chain unfolding, the phase rigidity scales as
\[r_{\mathrm g} \sim|\delta|^{(N-1)/N}.\]Therefore
\[K_{\mathrm g} \sim|\delta|^{-2(N-1)/N}.\]The control-parameter exponent differs from the structured case:
| Regime | Splitting | Petermann factor |
|---|---|---|
| structured | $q^{1/2}$ | $q^{-(N-1)}$ |
| generic | $\delta^{1/N}$ | $\delta^{-2(N-1)/N}$ |
This difference is real, but it is not the end of the comparison.
8. Fixed-Splitting Robustness Law
Let $s$ be the characteristic measured eigenvalue displacement and define the dimensionless splitting
\[\sigma=\frac{s}{|\Omega|}.\]Structured regime
The structured spectrum gives
\[\sigma\sim \left|\frac{q}{\Omega^2}\right|^{1/2}.\]Therefore
\[\left|\frac{q}{\Omega^2}\right| \sim\sigma^2.\]Substitution into
\[K_{\mathrm s}\sim \left|\frac{q}{\Omega^2}\right|^{-(N-1)}\]gives
\[K_{\mathrm s}\sim \sigma^{-2(N-1)}.\]Generic regime
The generic spectrum gives
\[\sigma\sim \left|\frac{\delta}{\Omega}\right|^{1/N}\]up to an $N$-dependent constant. Hence
\[\left|\frac{\delta}{\Omega}\right| \sim\sigma^N.\]Substitution into
\[K_{\mathrm g}\sim \left|\frac{\delta}{\Omega}\right|^{-2(N-1)/N}\]again gives
\[K_{\mathrm g}\sim \sigma^{-2(N-1)}.\]Thus
\[\boxed{ K\sim\sigma^{-2(N-1)} }.\]Changing the perturbation direction changes how the experimental control parameter maps to a splitting. It does not remove the leading nonorthogonality penalty associated with an already resolved splitting.
9. Spectral Sensitivity Is Not Sensor Performance
A large $K$ is a warning that the eigenbasis is ill-conditioned. It may be associated with:
- excess spontaneous-emission noise,
- linewidth enhancement,
- enhanced response to uncontrolled perturbations,
- strong mode nonorthogonality,
- numerical sensitivity of eigenvectors.
However, $K$ alone is not a complete noise model. A quantitative sensor analysis also needs:
| Missing ingredient | Why it matters |
|---|---|
| input noise statistics | sets fluctuation strength |
| gain and loss mechanism | determines added noise |
| readout observable | determines what is measured |
| integration time | determines estimator variance |
| nonlinear saturation | limits linear amplification |
Therefore one should not claim improved signal-to-noise ratio from an eigenvalue splitting alone.
10. Common Mistakes
| Mistake | Correction | ||
|---|---|---|---|
| treating $K$ as an eigenvalue | it is an eigenvector nonorthogonality measure | ||
| using $K=r^{-1}$ | for this model, $K=r^{-2}$ | ||
| writing $K\sim s^{-2(N-1)}$ with dimensional $s$ | use $\sigma=s/ | \Omega | $ |
| assuming a large $K$ guarantees a useful sensor | a platform-specific noise model is required | ||
| ignoring modal prefactors | different modes can have the same exponent but different $K$ |
11. Final Summary
The logical chain is
\[\text{eigenvectors coalesce} \Longrightarrow r\to0 \Longrightarrow K=r^{-2}\to\infty.\]Continue with Part V: A Complete Notation and Crossover Guide.
Discussion