Does IRKA merely discard state variables?

No. It projects the full dynamics into lower-dimensional spaces.

Does convergence prove global H2 optimality?

No. The targeted conditions are first-order necessary conditions.

Is it only relevant to one-input systems?

Extensions handle multiple inputs and outputs; the worked example uses the simpler single-input single-output setting.

AI Glossary / Mathematics and optimization / Intermediate

Iterative rational Krylov algorithm

Definition

The iterative rational Krylov algorithm, or IRKA, constructs a smaller linear dynamical model by repeatedly updating interpolation points using the poles of the current reduced model. It targets first-order necessary conditions for H2 approximation quality. Those conditions are useful mathematical goals, but they do not by themselves certify the best possible reduced model among all models of the chosen size.

What it means

A detailed dynamical model may contain many state variables even when users mainly care about its input-output response. Model reduction replaces those internal equations with fewer equations while trying to preserve the behavior that matters. IRKA selects projection spaces using shifted linear-system solves rather than simply deleting the least visible states.
Rational refers to the transfer functions and shifted inverse operations involved. Krylov refers to the spaces used to build the reduced representation. The method is not a generic neural-network compression algorithm, although reduced physical models can support larger data or decision workflows.

How it works

For a continuous-time single-input single-output system, write H(s)=c transpose (sI-A) inverse b. Choose a target order r and initial interpolation shifts. Construct right and left projection spaces from solves involving sigma I-A and its transpose. Form a reduced system using compatible bases, calculate its poles, and update the shifts to their reflected locations, represented by minus the reduced poles as a conjugate-closed set.
Repeat until a suitable convergence test is satisfied. At a converged solution under the relevant assumptions, interpolation values and derivatives meet the H2 first-order conditions. Check reduced-model stability and approximation quality separately. Initial shifts, numerical solves, and problem structure affect results; general convergence is not automatic.

Three practical examples

An engineering team can reduce a large linearized thermal model before repeated simulations, validating the response over the intended operating regime.
A circuit-modeling study can compare reduced input-output models with a detailed reference system and quantify approximation error.
A control research project can investigate whether an inexpensive reduced model preserves the response information needed for a subsequent design task.

Think of it this way

Think of summarizing a complicated musical response with fewer resonances, repeatedly changing where you listen according to the resonances of the current summary. The analogy is limited because IRKA uses algebraic projection and interpolation conditions. It does not select audible frequencies by human judgment or guarantee a good summary for every possible purpose.

Worked example

Consider an original two-state stable system with A=diag(-1,-3), b=(1,1), and c=(1,1). Its transfer function is H(s)=1/(s+1)+1/(s+3). We seek a one-state approximation and start with the real shift sigma=1.
The right projection vector before normalization is v=(sigma I-A) inverse b=(1/2,1/4). Because A is symmetric and b=c, the left vector can be chosen the same way. Its squared length is v transpose v=1/4+1/16=5/16. Dividing v by its length gives compatible unit bases for this one-dimensional projection.
The reduced matrix is the scalar Ar=(v transpose A v)/(v transpose v). Its numerator is -1/4-3/16=-7/16, so Ar=-7/5=-1.4. The product of the reduced input and output coefficients is (v transpose b)^2/(v transpose v)=(3/4)^2/(5/16)=9/5=1.8. Thus the reduced transfer function is Hr(s)=1.8/(s+1.4).
Check interpolation at the chosen shift. H(1)=1/2+1/4=0.75, and Hr(1)=1.8/2.4=0.75. The derivatives also match: H'(1)=-1/4-1/16=-0.3125, while Hr'(1)=-1.8/(2.4)^2=-0.3125. These matches demonstrate what the projection accomplishes at this shift.
IRKA now updates sigma to 1.4, the negative of the reduced pole -1.4. The new unnormalized vector is (1/2.4,1/4.4)=(5/12,5/22). Repeating the same scalar projection gives a reduced pole -229/157, approximately -1.458599, and residue 289/157, approximately 1.840764. The next proposed shift is therefore 229/157.
This is only the beginning of the iteration, not a declared converged solution. Also, exact matching at sigma=1 did not make the first reduced model identical to the original: H(0)=4/3, whereas Hr(0)=9/7. The difference at zero is 1/21. The example distinguishes interpolation at selected points from equality everywhere and from a global optimality certificate.

Illustrative case study

In a fictional equipment-simulation project, Elm Harbor Engineering reduces a detailed linear model to accelerate repeated analysis. It chooses several target orders, tries documented initial shifts, and compares each resulting model's stability, response error, and construction cost.
The team retains the full model as a reference. When a reduced model behaves poorly under a particular input, the analyst revisits the chosen order and validation criteria rather than relying only on a converged shift report. No measured runtime saving or accuracy result is implied by this fictional case.

Common misconceptions and mistakes

Do not interpret the word optimal as an unconditional global guarantee. Satisfying first-order necessary conditions is weaker than proving global optimality. Nor does a stable full system guarantee that every intermediate or returned reduced system from every initialization is stable.
The continuous-time shift update should not be copied unchanged into a discrete-time formulation. Also distinguish outer IRKA iterations from inner iterative linear solves: stopping those solves too loosely can alter the projection. H2 error measures a particular approximation objective, not every possible response or decision criterion.

Business use

For business applications, the central tradeoff is the cost of building and validating a reduced model versus the benefit of reusing it. Record the full model version, operating assumptions, target order, shift choices, termination conditions, and stability checks. A reduced model should be judged against the decisions it supports, with explicit limits on extrapolation beyond the conditions used to derive the original linear model.

Check your understanding

Why is sigma changed from 1 to 1.4 after the first projection?

Answer: The reduced model has pole -1.4, and the continuous-time IRKA update reflects that pole to the new shift 1.4.

Related terms

Model order reduction
Transfer function
Linear Dynamical System
Stability
Rational Krylov subspace
H2 norm
Interpolation
Reduced pole
Projection

Learn next

Recalculate the first projection using sigma=3. Compare the resulting pole and residue with those obtained at sigma=1, then verify value and derivative interpolation at the new starting shift. Explain why different starting shifts require convergence and validation checks.

Sources

Primary research: Gugercin, Antoulas, and Beattie, H2 Model Reduction for Large-Scale Linear Dynamical Systems, original paper and institutional record: https://vtechworks.lib.vt.edu/items/25885e2e-9c89-4fab-8767-c3bb6d5f2ac8 . The two-state calculations and fictional engineering case are original.

Previous: Iterative method Next: Iterative soft thresholding algorithm