Quantum Hamiltonian learning asks the inverse problem of quantum dynamics: from prepared probes and measured outcomes, infer the generator H that caused the observed evolution.
Forward simulation starts with H and predicts U(t)=e−iHt. System identification reverses that arrow. A model H(θ) with unknown parameter vector θ predicts a likelihood for experimental data. Statistical inference updates θ, while experiment design chooses future probe states, evolution times and measurements that make competing Hamiltonians easier to distinguish.
Guide 21 owns state tomography and Guide 22 owns process tomography. Those tasks reconstruct states or channels at a chosen time. Hamiltonian learning instead exploits the time-dependent generator structure across experiments. Guide 15 owns metrological precision bounds; this guide uses Fisher information as one diagnostic inside a broader identification problem.
Choose Hamiltonian model → design probe/time/measurement → collect outcomes → evaluate likelihood → update parameters → test residuals/model mismatch → adapt next experiment.
1. Parametric Hamiltonian model
A common finite-dimensional model is
H(θ)=Σ_{k=1}^{K} θ_k H_k
where the basis operators Hk are known and coefficients θk are unknown.
For qubits, Hk are often Pauli strings.
The identification problem is meaningful only if different θ values produce distinguishable experimental statistics under the allowed probes and measurements.
2. Single-qubit frequency learning
Take
H=(ω/2)Z.
Prepare |+⟩ and evolve for time t:
|ψ(t)⟩=e^{-iωtZ/2}|+⟩.
The Bloch vector rotates in the equatorial plane:
⟨X⟩=cos(ωt);⟨Y⟩=sin(ωt);⟨Z⟩=0.
This is Ramsey spectroscopy in its simplest mathematical form.
3. Likelihood for one X-basis measurement
Let outcome m∈{+1,−1} denote the X eigenvalue.
The Born-rule likelihood is
p(m|ω,t)=[1+m cos(ωt)]/2.
For M independent repetitions with n+ plus outcomes and n−=M−n+,
L(ω)∝p(+|ω,t)^{n_+}p(−|ω,t)^{n_-}.
Maximum likelihood chooses ω that maximises this function; Bayesian inference multiplies it by a prior.
4. Worked Ramsey likelihood
Suppose t=1 μs and 100 X measurements give 75 plus and 25 minus outcomes.
The empirical X expectation is
x̄=(75−25)/100=0.5.
An ideal maximum-likelihood phase obeys
cos(ωt)=0.5.
Within a prior window 0≤ωt≤π, this gives
ωt=π/3.
Without that prior window there are infinitely many aliases ±π/3+2πk. One evolution time cannot globally identify frequency.
5. Phase wrapping and identifiability
At fixed time t, Ramsey statistics depend on ω only through ωt modulo 2π.
Long t increases local sensitivity but shrinks the unambiguous frequency interval.
Hamiltonian learning therefore faces a precision–aliasing trade-off analogous to phase estimation:
- short times give coarse but globally unambiguous information;
- long times give fine local information but many phase aliases.
Adaptive schedules combine several times to resolve both.
6. Classical Fisher information
For the ideal X-basis Bernoulli likelihood, the Fisher information is
I(ω)=Σ_m p(m|ω,t)[∂_ω ln p(m|ω,t)]².
A direct calculation gives
I(ω)=t²
away from singular parametrisation points, with the same limiting value by continuity.
For M independent shots, Cramér–Rao gives local variance
Var(ω̂)≥1/(Mt²)
for unbiased estimators under regularity assumptions.
7. Decoherence changes the optimum time
Suppose dephasing reduces Ramsey contrast by e−t/T₂:
⟨X⟩=e^{-t/T₂}cos(ωt).
Longer t now gives more phase accumulation but less contrast.
The Fisher information no longer grows indefinitely as t². Experiment design must choose a time comparable to the coherence scale rather than “always wait longer”.
8. Bayesian Hamiltonian learning
Let prior density be p(θ).
After observing data d from experiment e, Bayes’ rule gives
p(θ|d,e)=p(d|θ,e)p(θ)/p(d|e).
Sequential experiments update the posterior recursively.
The posterior represents both best-fit parameters and uncertainty, allowing future experiments to target the unresolved directions.
9. Particle-filter representation
For nonlinear/multimodal posteriors, represent
p(θ)≈Σ_{j=1}^{N_p} w_j δ(θ−θ_j).
After data d, update
w_j←w_j p(d|θ_j,e)
and renormalise.
When effective sample size collapses, resample/jitter particles.
This sequential Monte Carlo architecture is common in practical Bayesian quantum parameter estimation.
10. Adaptive experiment design
Choose next experiment e to maximise expected information gain
E_d[H(p(θ))−H(p(θ|d,e))]
or minimise expected posterior variance/loss.
The experiment variables can include:
- probe state;
- evolution time;
- control pulse;
- measurement basis;
- which qubits are observed.
Learning and control therefore meet directly: Guide 53 can design pulses that make unknown Hamiltonian parameters more observable.
11. Multi-parameter Pauli Hamiltonian
For n qubits, write
H=Σ_P θ_P P
over selected Pauli strings P.
A fully generic Hamiltonian has 4n Pauli coefficients, so polynomial-time learning requires structure such as locality, sparsity, translation invariance or known interaction graph.
Local k-body Hamiltonians contain only polynomially many candidate terms for fixed k when geometry is known.
12. Short-time derivative method
For observable O under unitary H,
d⟨O⟩/dt=i⟨[H,O]⟩.
If H=ΣθkHk, then
d⟨O⟩/dt=Σ_k θ_k i⟨[H_k,O]⟩.
At t≈0, derivatives measured for many prepared states/observables give a linear regression problem in θ.
This avoids long coherent evolution but can be statistically noisy because numerical derivatives divide small changes by small t.
13. Worked two-term commutator example
Let
H=(ω/2)Z+(g/2)X
and measure O=Y.
Using [Z,Y]=−2iX and [X,Y]=2iZ,
i[H,Y]=ωX−gZ.
Therefore
d⟨Y⟩/dt=ω⟨X⟩−g⟨Z⟩.
Preparing |+⟩ isolates ω because ⟨X⟩=1,⟨Z⟩=0. Preparing |0⟩ isolates −g because ⟨X⟩=0,⟨Z⟩=1.
Probe-state design can therefore turn a multivariate identification problem into nearly orthogonal equations.
14. Sparse Hamiltonian learning
Suppose H has only s significant terms in a much larger known operator dictionary.
Short-time expectation derivatives or frequency-domain measurements can produce linear observations
y=Aθ+noise.
If θ is sparse and A satisfies suitable incoherence/restricted-isometry conditions, compressed-sensing methods such as ℓ₁ minimisation can recover θ from far fewer measurements than the dictionary dimension.
Sparsity is an assumption to test, not a generic property of all many-body Hamiltonians.
15. Spectroscopy from time signals
For initial state |ψ⟩=Σjcj|Ej⟩ and observable O,
⟨O(t)⟩=Σ_{jk} c_j* c_k O_{jk} e^{i(E_j−E_k)t}.
The time signal contains frequencies equal to energy gaps Ej−Ek.
Fourier transformation reveals spectral lines when the corresponding amplitudes are nonzero.
This is ordinary spectroscopy expressed in operator language.
16. Fourier resolution
Observing a signal for total duration T gives a characteristic frequency resolution
δω≈2π/T
for simple Fourier analysis.
Closely spaced spectral lines therefore require long coherent observation, unless stronger parametric assumptions enable super-resolution methods.
Windowing reduces leakage but broadens peaks; spectral estimation has the same finite-time uncertainty familiar from classical signal processing.
17. Quantum phase estimation as spectroscopy
If controlled e−iHt is available and the input overlaps eigenstate |Ej⟩, phase estimation returns an energy estimate with probability |cj|².
This provides direct spectral sampling without classically recording a long analog time trace.
The cost is coherent controlled simulation depth scaling inversely with energy precision, plus state overlap requirements.
18. Process tomography versus Hamiltonian learning
Process tomography at one time t reconstructs channel 𝓔t.
For perfect unitary dynamics, one could take a matrix logarithm to infer H, but:
- the logarithm has branch ambiguities;
- noise makes the channel nonunitary;
- full tomography scales exponentially;
- one time point throws away known generator locality/structure.
Hamiltonian learning uses a low-parameter model directly and can be exponentially more data efficient when that model is correct.
19. Open-system generator learning
Real hardware may obey a Lindblad equation
dρ/dt=−i[H,ρ]+Σ_k γ_k𝒟[L_k]ρ.
If dissipative rates are ignored, Hamiltonian estimates can absorb decoherence into biased frequencies/couplings.
Generator learning can therefore estimate coherent θ and dissipative γ jointly.
Guide 52 owns Lindblad dynamics; here they appear as nuisance or jointly learned system-identification parameters.
20. Model mismatch
A posterior can become very narrow around the wrong model.
Examples:
- missing next-nearest-neighbour coupling;
- time-dependent drift treated as constant;
- crosstalk omitted;
- non-Markovian noise fitted as Lindblad decay;
- state-preparation or measurement errors attributed to H.
Good system identification therefore tests residuals and compares model families, not only posterior width.
21. Innovation/residual diagnostics
For predicted outcome probability pmodel and observed binary data, residuals should look statistically structureless after accounting for shot noise.
Systematic residual versus time, control amplitude or qubit location indicates a missing term.
This mirrors Guide 54’s innovation-whiteness test for continuous quantum filters.
22. SPAM errors
State-preparation and measurement (SPAM) errors can be confounded with Hamiltonian parameters.
For example, unknown X-readout contrast can mimic dephasing; a preparation phase offset can mimic an initial evolution phase.
Strategies include:
- jointly fit SPAM parameters;
- use self-consistent gate-set models;
- calibrate independent controls;
- design differential experiments cancelling nuisance offsets.
23. Gauge freedoms in identification
Some transformations of states, measurements and dynamics leave all observable probabilities unchanged.
Gate-set tomography famously has such gauge freedoms.
Hamiltonian parameters should therefore be interpreted relative to calibrated reference frames and operator conventions.
Guide 72 develops the deeper role of quantum reference frames; here the practical lesson is that “the Hamiltonian matrix” is basis/reference dependent while predicted probabilities are physical.
24. Learning interaction graphs
When the coupling graph itself is unknown, compare candidate local terms.
Short-time commutators reveal which observables begin changing under which probes; spectroscopy reveals coupled-mode splittings; sparse regression can identify a small subset of active edges.
The task combines graph discovery with parameter estimation and becomes harder when several interaction patterns produce nearly identical low-order dynamics.
25. Learning from local measurements
For geometrically local Hamiltonians, one need not always perform global tomography.
Lieb–Robinson locality limits how rapidly distant terms influence a local observable. Short-time local experiments can therefore isolate nearby couplings before information spreads across the full system.
This can make sample complexity scale with local neighbourhood size rather than total Hilbert-space dimension under suitable assumptions.
26. Classical shadows for Hamiltonian learning
Classical shadows can estimate many local Pauli observables from shared randomized measurement data.
If a learning method needs a large vector of short-time expectation derivatives, shadow measurements can reduce the cost relative to measuring every Pauli term in a separate circuit.
Guide 23 owns classical-shadow mathematics; this guide uses it as an experimental data-acquisition primitive.
27. Quantum-assisted likelihood evaluation
For a candidate θ, likelihood prediction may itself require simulating H(θ).
If classical simulation is intractable but a trusted quantum simulator can implement candidate dynamics, one quantum device can help evaluate likelihoods for another device/system.
This “quantum model of a quantum system” can reduce the classical bottleneck, but the simulator must be calibrated well enough that its own errors do not dominate inference.
28. Common misconception: one spectrum uniquely determines the Hamiltonian
Different Hamiltonians can be isospectral or produce the same observed transition frequencies under restricted probes. Identifiability depends on eigenvectors, probe preparation, observables and controls as well as energy levels.
29. Common misconception: longer evolution is always better
Longer time increases phase sensitivity but creates aliasing and suffers decoherence/model drift. Optimal interrogation time is finite and uncertainty dependent.
30. Common misconception: a narrow posterior proves the model is correct
Bayesian uncertainty is conditional on the model family. A wrong model can fit confidently. Posterior predictive checks and held-out experiments are required.
31. Worked synthesis problem
A qubit obeys H=(ω/2)Z. Prepare |+⟩ and measure X after time t.
Step 1: Prediction. p(+|ω,t)=[1+cos(ωt)]/2.
Step 2: Data. 75 plus outcomes from 100 shots gives empirical p≈0.75 and ⟨X⟩≈0.5.
Step 3: Local estimate. In prior window 0≤ωt≤π, solve cos(ωt)=0.5 to get ωt≈π/3.
Step 4: Alias warning. At one time, ω could also satisfy ±π/3+2πk divided by t if the prior allows those values.
Step 5: Adaptive resolution. Use a shorter time to identify the correct coarse frequency interval, then longer times to narrow uncertainty inside that interval.
Step 6: Validation. Predict Y measurements ⟨Y⟩=sin(ωt) at held-out times and check residuals before declaring the Hamiltonian learned.
32. Practice set
- Write a linear parametric Hamiltonian model.
- For H=ωZ/2 and initial |+⟩, what are ⟨X⟩ and ⟨Y⟩?
- Write the X-measurement likelihood.
- Why does one interrogation time create frequency aliases?
- What does Fisher information quantify locally?
- Write Bayes’ update for Hamiltonian parameters.
- What is adaptive experiment design?
- How does the short-time commutator equation become a linear regression in θ?
- What frequencies appear in a time-dependent observable?
- Why is Hamiltonian learning potentially cheaper than full process tomography?
- What is model mismatch?
- Why should SPAM/reference-frame parameters sometimes be learned jointly with H?
Answers
H(θ)=Σ_kθ_kH_k.cos(ωt)andsin(ωt).p(m|ω,t)=[1+m cos(ωt)]/2.- Statistics depend on phase modulo 2π, so many frequencies yield the same phase.
- The sensitivity of an outcome distribution to small parameter changes and the local Cramér–Rao precision bound.
p(θ|d,e)∝p(d|θ,e)p(θ).- Choose future probe/time/control/measurement settings to maximise expected information or minimise expected posterior loss.
d⟨O⟩/dt=Σ_kθ_k i⟨[H_k,O]⟩, linear in θ for known probe expectations.- Energy gaps Ej−Ek with nonzero preparation/observable matrix elements.
- It exploits a low-parameter generator model across time rather than reconstructing an exponentially large channel matrix.
- The true dynamics lies outside the fitted model family or nuisance effects are omitted, biasing inferred parameters.
- Otherwise preparation/readout/reference errors can be misattributed to coherent Hamiltonian terms.
Sources and further study
[1] Nathan Wiebe, Christopher Granade, Christopher Ferrie and D. G. Cory, Hamiltonian Learning and Certification Using Quantum Resources, Physical Review Letters 112, 190501 (2014). Quantum-assisted Bayesian Hamiltonian learning.
[2] Christopher Granade and colleagues, Robust online Hamiltonian learning, New Journal of Physics 14, 103013 (2012). Sequential Bayesian/particle-filter parameter learning and adaptive experiment design.
[3] M. P. da Silva, O. Landon-Cardinal and D. Poulin, Practical characterization of quantum devices without tomography, Physical Review Letters 107, 210404 (2011). Efficient certification/learning ideas exploiting physical structure rather than full tomography.
[4] A. Shabani and colleagues, Efficient Measurement of Quantum Dynamics via Compressive Sensing, Physical Review Letters 106, 100401 (2011). Sparse/compressive reconstruction of quantum dynamical models.
Continue through Quantum Mathematics
Guide 69: Quantum State Preparation, Amplitude Encoding, QROM, Data Loading and State-Synthesis Algorithms develops input-state construction. Guide 71: Quantum Graph Algorithms, Connectivity, Triangle Finding, Shortest Paths and Quantum-Walk Search develops graph-query algorithms. Guide 72: Quantum Reference Frames, Superselection Rules, Asymmetry and Relational Quantum States clarifies reference-dependent Hamiltonian descriptions.