A quantum system can retain a visible memory of its initial arrangement while still developing complicated entanglement. Localisation does not mean that nothing happens. It means that particular forms of transport and thermalisation are restricted, under conditions that must be stated and tested.
Many-body localization extends the questions raised by Anderson localization from single-particle wavefunctions to interacting quantum systems. Its mathematical vocabulary includes local integrals of motion, or l-bits, logarithmic entanglement growth, nonthermal eigenstates, imbalance measurements, level statistics and thermal avalanches. These concepts answer different questions. A spatially concentrated wavefunction, a long-lived population pattern and a stable interacting phase are not interchangeable observations.
The central proposition of this guide is therefore: separate the mechanism of memory from the evidence that the memory survives increasing time, system size and experimental imperfections. We will derive exact two-site and two-l-bit examples, build a practical numerical test, and distinguish established calculations from model-dependent phenomenology and continuing stability questions. Nothing in the argument requires pretending that finite data already settle an infinite-system limit.
Your 50-second route
New to localisation? Start with the single-particle model and the detuned two-site calculation. Want to understand interacting memory? Read l-bits and the exact entanglement laboratory. Interpreting data? Go to imbalance and spectral diagnostics, then thermal avalanches and finite-size limits. Studying independently? Work through the practice laboratory before using its answers.
Open the reading map and conventions
Part I explains spatial localisation and the difference between weak transport and a many-body phase. Part II builds an interacting local-integral-of-motion description and derives entanglement without particle transfer. Part III tests claims using eigenstates, dynamics, disorder distributions and size scaling. Part IV examines stability, experimental boundaries and a complete numerical workflow. We use ℏ=1, Pauli matrices X,Y,Z with eigenvalues ±1, and logarithms base two for entanglement measured in bits unless stated otherwise.
The technical foundations are in Guide 45 on many-body Hamiltonians, Guide 46 on spin chains and Guide 47 on quantum chaos. This guide keeps localisation, its mechanisms and its evidentiary limits together.
A plateau is an observation over a window. A localisation length is a property of a specified state or operator. A stable phase is a statement about a family of systems and limits. Good reasoning does not substitute one for another.
1. What thermalisation would normally erase
Imagine preparing a chain with particles mainly on even sites. At the beginning, a local measurement reveals the preparation pattern. In a thermalising isolated system, sufficiently late local observations are expected to become largely insensitive to that detailed pattern once the relevant conserved quantities, such as total energy and particle number, are fixed.
The global pure state has not become a mixed thermal state under unitary evolution. Rather, small subsystems and selected observables can behave thermally because information about the preparation has spread into complicated correlations. This distinction already appeared in the eigenstate-thermalisation discussion of Guide 47.
Localisation asks whether some local information remains accessible instead. It is not enough that total particle number stays constant; that is an ordinary conservation law. The interesting question is whether a much more detailed spatial pattern continues to influence local observables beyond what the usual equilibrium variables predict.
For every proposed memory signal, therefore, first remove trivial explanations: a conserved global charge, a symmetry sector that was not matched between preparations, an exactly decoupled site or a measurement insensitive to the dynamics. Only then does the signal test nontrivial failure of thermalisation.
2. The Anderson chain
A standard one-particle lattice Hamiltonian is H=−JΣj(|j+1⟩⟨j|+|j⟩⟨j+1|)+Σj hj|j⟩⟨j|. The hopping J moves amplitude between neighbouring sites. The onsite energies hj vary from site to site.
Without disorder, translation invariance gives extended plane-wave eigenstates in a periodic chain. With suitable random onsite energies, interference among repeated scattering paths can produce eigenfunctions concentrated around particular positions. Anderson’s foundational work established the possibility of absent diffusion in random lattice models. [1]
A localised eigenfunction is often described by an envelope |ψj|∼exp(−|j−j₀|/ξ). The localisation length ξ describes the envelope, not an exact point beyond which amplitudes vanish. Exponentially small tails remain important for interactions between distant localised objects.
The model assumptions matter. Statements about one-dimensional random onsite disorder should not be transferred without checking to correlated disorder, quasiperiodic potentials, off-diagonal disorder with special symmetries, long-range hopping or higher dimensions. Those changes can alter the spectral and transport problem.
3. A transfer matrix turns a wavefunction into a recurrence
The eigenvalue equation at site j is −Jψj−1+hjψj−Jψj+1=Eψj. Rearranging gives ψj+1=[(hj−E)/J]ψj−ψj−1. In two-component form,
(ψj+1,ψj)ᵀ = [[(hj−E)/J,−1],[1,0]](ψj,ψj−1)ᵀ.
Each transfer matrix has determinant one. Products of many such matrices describe how amplitudes propagate across the sample. Their growth and decay directions connect localisation to Lyapunov-exponent methods: an inverse decay length can be extracted from the long-product behaviour under appropriate conditions.
Naively multiplying thousands of matrices is numerically unstable because one direction can grow while another shrinks. Repeated orthogonalisation or a stable factorisation is needed to measure the exponents rather than overflow the machine. A localisation calculation can therefore fail numerically even when the underlying two-by-two recurrence is correct.
Boundary conditions still select allowed eigenvalues. The recurrence by itself propagates any trial E; it does not prove that the resulting sequence satisfies both ends of a finite chain. This is a useful distinction between a local differential or difference equation and a complete eigenvalue problem.
4. Exact laboratory A: why a detuned pair hardly exchanges population
Before studying a random chain, solve two coupled sites. After subtracting an irrelevant mean energy, write H=δZ/2+JX. The site-energy difference is δ. Since H²=[J²+(δ/2)²]I, define Ω=√[J²+(δ/2)²].
Starting on the first site, the probability of occupying the second is
P1→2(t)=(J²/Ω²)sin²(Ωt).
The maximum transfer is 4J²/(δ²+4J²). At resonance δ=0 it reaches one. For δ=6J, it reaches only 0.1. Large energy mismatch restricts local hybridisation even though the hopping is nonzero.
This exact calculation explains one local ingredient of disorder-induced confinement. It does not prove Anderson localisation in an infinite chain: many sites create interfering paths and rare resonances that a two-site model cannot represent. A useful toy model should explain a mechanism without being promoted into a theorem about a larger system.
5. Resonant clusters are the places where a simple expansion fails
In a strong-disorder picture, a small hopping-to-detuning ratio makes perturbation theory attractive. But a random sample can contain neighbouring energies that happen to be close. For such a pair, J/δ is not small even when the typical disorder width is large.
The two-site formula shows what happens: near δ=0 the two sites mix strongly. A perturbative treatment that divides by δ must isolate that resonant pair and diagonalise it together rather than use a divergent correction.
In a larger chain, resonances can join into clusters. The mathematical challenge is to control their occurrence and their coupling across scales. It is not enough to say that every typical denominator is large. Rare denominators can govern the fate of a sufficiently large system.
This is an early version of the distinction we will later make between average disorder and rare thermal inclusions. Typical-site reasoning can be locally informative while missing the rare structures relevant to stability.
6. Inverse participation ratio: a useful but basis-dependent measure
For a normalised one-particle state, define IPR=Σj|ψj|⁴. A state on one site has IPR=1. A state uniformly spread over L sites has IPR=1/L. The reciprocal PR=1/IPR is often called a participation ratio.
Take probabilities 0.7, 0.2 and 0.1 on three sites. The IPR is 0.49+0.04+0.01=0.54, so PR≈1.852. This does not mean the particle occupies exactly 1.852 sites. It is a concentration measure that interpolates between simple limiting cases.
The basis is essential. A plane wave is extended in the site basis and concentrated in the momentum basis. A many-body computational-basis participation ratio is not automatically the same diagnostic as a real-space localisation length. Always name the basis and the normalisation before comparing IPR values.
Finite-size scaling adds information. If IPR decreases like 1/L over increasing chains, that supports extension over the available volume. If it approaches a nonzero value while the state centre remains away from boundaries, that supports localisation within that family. One value at one size cannot distinguish a truly extended state from a localisation length larger than the sample.
7. Return probability and infinite-time averaging
Prepare a particle at site j₀. Expand |j₀⟩=Σα cα|Eα⟩, with cα=⟨Eα|j₀⟩. The return amplitude is A(t)=Σα|cα|²exp(−iEαt).
If the finite-system energies are nondegenerate, averaging |A(t)|² over infinite time removes the cross terms and leaves P̄return=Σα|cα|⁴. This is an IPR in the energy-eigenstate expansion of the initial site.
A substantial return probability can therefore arise when only a limited number of eigenstates overlap the initial site. But finite isolated systems also recur, and degeneracies retain extra terms in the average. State what was averaged and which nondegeneracy assumption was used.
The calculation illustrates a broader lesson: long-time local memory is related to spectral structure, but an observed revival at one selected time is not a proof of localisation. A two-level system revives perfectly and has no thermodynamic many-body phase at all.
8. Adding interactions changes the question
Noninteracting particles can occupy localised one-particle orbitals. Interactions make the energy of one configuration depend on the occupations of others and allow many-body resonances that have no single-particle counterpart.
The question is no longer whether one wavefunction decays away from a centre. It becomes whether an interacting, isolated system can retain extensive local structure in its eigenstates and dynamics instead of serving as its own thermal bath.
A many-body basis has exponentially many configurations, but localisation in that abstract basis alone is not a complete physical definition. The useful structure must also connect to locality in real space, local observables and the stability of the description under the interactions being considered.
This is why an interacting problem cannot be settled by copying a single-particle localisation formula and replacing position with a bit string. The configuration graph, matrix-element correlations and available resonances are different.
9. A precise random-field spin-chain convention
One common model is H=(J/4)Σj[XjXj+1+YjYj+1+ΔZjZj+1]+(1/2)Σj hjZj, with open boundaries unless stated otherwise. Here X,Y,Z are Pauli matrices, so the factors of one half and one quarter would disappear or change if spin operators S=σ/2 were used instead.
For a random-field ensemble, specify for example hj∼Uniform[−W,W] independently. The anisotropy is Δ and the hopping/exchange scale is J. Ratios such as W/J have meaning only with this Hamiltonian convention attached.
The total Z magnetisation is conserved. Exact diagonalisation should therefore use a fixed magnetisation block when testing level statistics. For even L, the zero-magnetisation sector has dimension binomial(L,L/2), not 2ᴸ. At L=12 that is 924; at L=20 it is 184,756.
Pal and Huse used exact diagonalisation of a random-field spin-half chain to study changes in high-energy eigenstates and finite-size indicators of the localisation transition. Such calculations established important diagnostics, but their finite sizes should not be converted into a universal critical disorder value for every spin-chain convention. [2]
10. Local integrals of motion: what makes an l-bit special
Any finite nondegenerate Hamiltonian has conserved eigenstate projectors. Calling those projectors conserved does not explain localisation, because they can be spread over the entire system. The distinctive l-bit claim is that a complete set of conserved binary operators can be chosen to be quasi-local in a suitable fully localised regime.
Write these operators as τiz, with eigenvalues ±1 and [H,τiᶻ]=0. Each is concentrated near site i but has tails extending to other sites. They are dressed versions of local physical spins rather than necessarily equal to the laboratory Zi.
Huse, Nandkishore and Oganesyan formulated a phenomenology in which fully localised many-body eigenstates are described by interacting localised pseudospins. Their description separates conserved l-bit populations from interaction-induced dephasing. [3]
That word quasi-local is the content of the theory. Removing it would make the statement almost empty: spectral decomposition always provides conserved quantities, but generally not ones a local experiment can access without controlling the whole system.
11. The l-bit Hamiltonian
A phenomenological diagonal form is H=Σi h̃iτiᶻ+Σi<j Jijτiᶻτjᶻ+Σi<j<k Kijkτiᶻτjᶻτkᶻ+…. Every term commutes with every τiz. The coefficients decay with the spatial extent of the participating operators in the localised description.
Conserved populations do not imply independent spins. The phase accumulated by a superposition of τiz states depends on the values of other l-bits through Jij and higher terms. Interactions therefore create correlations even though no τiz flips.
For a fixed background configuration of all other l-bits, spin i precesses at one frequency. A superposition of backgrounds gives several conditional frequencies. Tracing out the background makes the local coherence decrease or oscillate. This is dephasing generated internally by a closed many-body state, not necessarily decoherence caused by an external environment.
The model is useful because it turns a vague phrase such as retained memory into calculable statements. One can identify which operators are conserved, which coherences dephase, and how interaction tails set a hierarchy of times.
12. Exact laboratory B: entanglement without population transport
Take two l-bits with H=J Z_AZ_B and initial state |+⟩A|+⟩B. Both Z populations are conserved. The exact state is
|ψ(t)⟩=[e⁻ⁱᴶᵗ(|00⟩+|11⟩)+eⁱᴶᵗ(|01⟩+|10⟩)]/2.
Trace out B. The reduced state of A is ρA(t)=[I+cos(2Jt)X]/2, whose eigenvalues are p±=[1±cos(2Jt)]/2.
The entanglement entropy is S_A=h₂([1+cos(2Jt)]/2), where h₂(p)=−p log₂p−(1−p)log₂(1−p). At t=0 it is zero. At t=π/(4J), the cosine vanishes and SA=1 bit. At t=π/(2J), the subsystem is pure again and SA=0.
Throughout the calculation, neither Z population moved. Entanglement grew because the phase of A became correlated with B. This exact two-qubit example demonstrates why inhibited transport and changing quantum information are compatible. Its later revival also reminds us that two l-bits are not a thermodynamic bath.
13. Why many weak couplings produce a long hierarchy of times
Now couple one l-bit to many others through H=ZAΣj JjZj, with the other spins initially in |+⟩ states. Because all terms commute, the central-spin coherence contains a product G(t)=Πj cos(2Jjt).
A large nearby Jj changes a factor quickly. A tiny distant Jj changes it only at much later times. In a many-spin system, these different factors can make a simple local signal look relaxed long before every weak interaction has had time to build correlations.
The formula is exact for this commuting toy Hamiltonian and specified initial state. It is not a universal expression for a physical Z measurement in every random spin chain, because physical spins and dressed l-bits differ. Nevertheless, it exposes the mechanism in a form that can be checked by direct matrix evolution.
It also explains why a single fitted decay time can be misleading. A broad hierarchy of couplings can produce several stages of apparent relaxation, and a measurement window may see only one of them.
14. Derive the logarithmic distance scale
Assume an effective interaction at distance r has magnitude J(r)≈J₀exp(−r/ξ). An appreciable conditional phase needs |J(r)|t of order one. Solving that condition gives
r(t)≈ξ ln(J₀t)
for times beyond the local interaction scale. If the number of newly entangled degrees of freedom across a one-dimensional cut grows with this distance, the entropy can grow logarithmically over the regime in which the l-bit description applies.
Serbyn, Papić and Abanin connected slow entanglement growth to interaction-induced dephasing across exponentially small energy corrections. Their argument explains a hierarchy of dephasing times within a many-body-localised description; it should not be read as an unconditional proof that such a phase is stable in every dimension. [4]
The prefactor depends on localisation structure, the cut and the initial state’s entropy density in the relevant basis. Writing S(t)=ln t without a time unit, prefactor or regime is not a complete quantitative prediction.
15. Eigenstate area laws and quench entanglement are different statements
An individual localised eigenstate can have low entanglement across a cut. A product state evolved after a quench is generally a superposition of many eigenstates. Their interaction-dependent phases can create much more entanglement than any statement about one selected eigenstate might suggest.
In finite systems, the slowly growing quench entropy can eventually saturate at a value proportional to subsystem size, often below a thermal benchmark. That does not contradict area-law eigenstate structure because the two calculations concern different states.
Bardarson, Pollmann and Moore numerically demonstrated slow, approximately logarithmic entanglement growth in strongly disordered interacting one-dimensional models, contrasting it with the noninteracting case. Their work is a primary example of interactions changing entanglement much more dramatically than visible particle transport. [5]
When reading an entropy plot, therefore, ask first: eigenstate or evolving state? Which initial state? Which subsystem? Which logarithm base? Which finite-size saturation value? Those questions are necessary before interpreting a slope as a physical phase diagnostic.
16. Echoes recover phases only when the operation is available
For the two-l-bit interaction H=JZAZB, an ideal X pulse on A changes the sign of the interaction because XAZAXA=−ZA. Evolving for t/2, applying XA, and evolving for t/2 cancels the interaction phase up to the known pulse action.
This is an exact echo identity for the toy model. It distinguishes reversible dephasing from irreversible loss of the full closed-system state. However, a laboratory pulse on the physical spin is not generally the exact X operator of a dressed l-bit.
Implementing the ideal echo may require a quasi-local operation involving several physical sites. Pulse errors, additional Hamiltonian terms and external environments further limit recovery. The existence of conserved l-bit populations does not automatically provide a practical quantum memory that corrects arbitrary errors.
This qualification is important for applications. A mechanism that preserves one classical-looking local observable can still dephase the quantum coherence required by a computation.
17. Imbalance: a direct memory observable
Prepare particles preferentially on even sites and define I(t)=[N_even(t)−N_odd(t)]/[N_even(t)+N_odd(t)]. An ideal even-only preparation has I(0)=1. Equal even and odd occupations give I=0.
A late-time nonzero imbalance shows that the measured density pattern has not fully relaxed over the observed interval. The definition makes the measurement intuitive: it compares two groups of site populations rather than reconstructing a full many-body wavefunction.
But the denominator matters if particles are lost. Preferential loss, heating, trapping inhomogeneity or unequal detection efficiency can bias the ratio. A useful analysis reports raw even and odd counts, total survival and the normalised imbalance, rather than showing the ratio alone.
Likewise, a nonzero plateau at finite time is consistent with long-lived memory but does not by itself prove an asymptotic phase. An extremely slow thermal system, an almost-integrable system or disconnected regions can also retain a pattern over a finite window. Additional diagnostics are needed.
18. Autocorrelation and the order of averaging
A spin-memory diagnostic is C(t)=(1/L)Σj⟨Zj(t)Zj(0)⟩, with the initial state or ensemble specified. At infinite temperature inside a symmetry sector, the expectation is a normalised trace over that sector rather than an arbitrary product-state average.
One can average C(t) over sites, disorder realisations, initial states and time. These operations answer different questions. Averaging over disorder can smooth oscillations that remain large in individual samples; averaging over time can hide persistent sample-specific fluctuations.
Report at least a central tendency and the spread across independent disorder realisations when that spread is important. A mean dominated by rare samples and a typical value such as a median can tell different stories. Neither should be quietly substituted for the other.
A practical study keeps the raw sample identity through the analysis. This allows a later check of whether apparent slow dynamics comes from most samples or from a small number of unusual ones.
19. Level statistics need symmetry resolution
Within one symmetry sector, sort the many-body energies and set δn=En+1−En. The adjacent-gap ratio is rn=min(δn,δn+1)/max(δn,δn+1). It compares neighbouring gaps without explicitly unfolding a smooth density of states.
For independent Poisson gaps, the mean ratio is 2ln2−1≈0.386294. A time-reversal-compatible chaotic block in the GOE class has a mean near 0.53. These benchmarks are useful comparisons, not a universal classification theorem based on one finite sample. The eigenstate and spectral approach is discussed in the random-spin-chain work of Pal and Huse. [2]
Combining unrelated symmetry sectors creates crossings that do not repel and can imitate Poisson-like statistics. Exact degeneracies can make ratios ill-defined. Restrict the energy-density window, state how degeneracies are treated, and check whether the result persists over increasing system size.
Poisson statistics can also occur in noninteracting or integrable systems. They support a particular spectral structure but do not establish interacting localisation without additional local and dynamical evidence.
20. What the eigenstate thermalisation test actually compares
Choose a local observable O and plot Oαα=⟨Eα|O|Eα⟩ against energy density within a fixed symmetry sector. In a thermalising regime, nearby eigenstates are expected to have increasingly similar local values as the system grows, for suitable observables and windows.
A broad sample-dependent spread at similar energy density suggests retained eigenstate information. To interpret it, separate smooth energy dependence from fluctuations, and exclude obvious changes of a conserved quantum number.
Do not compare the ground state of one sample with the middle of the spectrum of another and attribute the difference to localisation. Energy density matters. A mobility-edge hypothesis concerns variation across energy, whereas a fully localised l-bit description assumes localisation across the relevant entire spectrum.
The operational question is whether two states with the same macroscopic constraints but different local preparation histories yield distinguishable local predictions at late times. An eigenstate plot and a quench experiment approach that question from complementary directions.
21. Basis participation is not a substitute for entanglement
A product state |+⟩⊗L has equal amplitudes on all 2ᴸ computational-basis strings. Its computational-basis participation is maximal, but its entanglement across every spatial cut is zero.
A Bell state on two qubits has support on only two computational strings yet has one bit of bipartite entanglement. These examples show that participation and entanglement quantify different structures.
For a localisation study, compute the quantity connected to the claim. Real-space concentration, computational-basis spreading, entanglement across a cut and local memory are not different names for one scalar measure. Agreement among several of them is informative precisely because they probe different aspects of the state.
This is a useful teaching exercise: construct two states with the same IPR but different entanglement, or the same entanglement but different site profiles. Such counterexamples prevent a diagnostic from being granted more interpretive power than it has.
22. A finite-time plateau can hide a slow decay
Consider three hypothetical curves: I₁(t)=0.2, I₂(t)=0.2t⁻⁰·⁰¹ for t≥1, and I₃(t)=0.2exp(−t/10⁶). Over a short window, all can look nearly flat within measurement uncertainty. Their asymptotic limits are different.
The example does not prescribe which fitting function is correct. It shows why a visually convincing plateau is insufficient when the distinction of interest occurs outside the observed window. A fitted asymptote can be determined more by the chosen model than by the data.
Compare alternative plausible fits, inspect residuals, extend the observation window where possible and report the range of late-time behaviour consistent with uncertainty. A clear conclusion might be that relaxation is slower than a measured bound over a stated interval. That is a useful result even when it does not establish permanent memory.
In a fictional classroom discussion, Ryan labels a flat-looking trace localised. Aisha asks what would happen if the experiment ran a hundred times longer. The right response is not to guess the future trace. It is to state which candidate behaviours the current measurements exclude and which remain compatible.
23. Finite-size drift should be measured, not explained away
Suppose a gap-ratio crossing occurs near disorder W=5 for one pair of sizes and near W=6 for a larger pair. A single claimed critical W hides that drift. It may signal corrections to scaling, a long crossover, an inappropriate ansatz or changing rare-region sensitivity.
Plot the crossing location or another finite-size threshold against size. Keep boundary conditions, disorder distribution, energy density and symmetry sector fixed. A data collapse is useful only when its parameters and corrections are constrained rather than chosen solely to make curves look similar.
The infinite-time and infinite-size limits can be delicate. A finite system eventually has a discrete spectrum and recurrences. An infinite random system can contain rare regions absent from a small sample. A conclusion should state which limit the calculation probes and which extrapolation remains an assumption.
No universal critical disorder is quoted here because changing the Hamiltonian normalisation alone changes the numerical value, and finite-size estimates do not automatically determine the asymptotic transition in every interacting model.
24. Why a rare thermal inclusion can matter
A region with unusually weak effective disorder may behave more thermally than its surroundings. If it can absorb neighbouring localised degrees of freedom, its Hilbert space grows and its level spacing decreases. That can make the enlarged region better able to absorb another neighbour.
This feedback is the thermal-avalanche scenario. De Roeck and Huveneers developed an ETH-based theory of the stability and instability of localised systems near ergodic grains, together with numerical evidence. Its assumptions and predictions should be kept distinct from an unconditional theorem for every disordered Hamiltonian. [6]
A rare inclusion is easy to miss in small samples. Its importance may increase with system size even if almost every local neighbourhood looks strongly disordered. This explains why typical-coupling arguments and rare-region stability arguments can lead to different-looking conclusions without being logically inconsistent.
25. A transparent avalanche scale estimate
Consider a hypothetical thermal region with ℓ effective spins, entropy density s and bandwidth B. Its number of relevant states scales as D≈exp(sℓ), so a representative level spacing is δbath≈Bexp(−sℓ).
Suppose the bare coupling to a spin at distance r decays as g(r)=g₀exp(−r/ξ). If the thermal-region matrix element follows an ETH-type factor D⁻¹/², a representative coupling matrix element scales as g₀exp(−r/ξ)exp(−sℓ/2).
Dividing by the level spacing gives a dimensionless hybridisation estimate (g₀/B)exp(−r/ξ)exp(sℓ/2). Distance suppresses coupling, while the region’s growing Hilbert space enhances access to closely spaced levels.
This estimate explains the competition; it is not a universal transition criterion. Geometry, conservation laws, the growth rule for ℓ, matrix-element statistics and the validity of ETH inside the inclusion all matter. Its value is that every assumption is visible and can be challenged rather than hidden in the word avalanche.
26. What a conditional mathematical proof establishes
Imbrie’s work on one-dimensional random quantum spin chains proves localisation under a stated assumption limiting level attraction, using a sequence of local unitary transformations. It supplies a mathematically controlled example of the quasi-local diagonalisation programme. [7]
The conclusion is neither that interacting localisation has no rigorous basis nor that every experimentally used random Heisenberg chain is covered by the proof. A theorem applies to its model, hypotheses and parameter regime.
For a reader, the productive task is to map each assumption onto the proposed system: dimensionality, interaction locality, disorder distribution, strength of perturbations and the required spectral regularity. A missing match does not disprove the physical phenomenon; it marks the boundary of what that theorem certifies.
This is a general mathematical skill. Respecting a theorem’s hypotheses makes the result more useful, because it tells us exactly what additional argument or measurement is needed to extend it.
27. Random and quasiperiodic potentials are not the same disorder ensemble
Independent random onsite fields allow rare spatial fluctuations governed by a probability distribution. A quasiperiodic potential such as hj=Wcos(2παj+φ), with irrational α in an ideal infinite construction, has deterministic spatial correlations.
Changing the phase φ samples translations of a structured potential, not an ensemble of independent site energies. Rare-region arguments based on independent randomness therefore cannot be transferred unchanged.
On a finite lattice, irrational α is represented numerically and boundary commensurability can matter. State the chosen frequency, phase ensemble and boundary conditions. Two calculations both labelled disordered can have materially different spatial statistics.
The same caution applies to experimental speckle and other correlated potentials. The word disorder does not specify a universal probability model.
28. A dated experimental example of the distinction
In the preprint Stability of many-body localization in two dimensions, first submitted on 28 August 2025, Hur and colleagues studied ultracold atoms in optical lattices with variable sizes up to 24 by 24 sites. Using imbalance, they compared random and quasiperiodic potentials. For random disorder, the crossover shifted toward stronger disorder as size increased; for the quasiperiodic case, they reported no clear size dependence over their studied range. [8]
The authors interpret the random-disorder trend as consistent with an avalanche scenario and the quasiperiodic result as suggesting possible stability. Those are appropriately different claims. Absence of a clear drift over available sizes is not itself a proof of stability at arbitrarily large size and time.
This example is useful because it makes the experimental comparison explicit. The system size, potential class and measured observable are part of the conclusion. It should not be abbreviated to a universal statement that two-dimensional localisation has either been proved or ruled out.
29. Long-range interactions change the dephasing and stability scales
The logarithmic entangling-distance estimate used exponentially decaying couplings. If instead a model has J(r)≈J₀/rᵅ, the simple condition J(r)t≈1 gives r(t)≈(J₀t)^(1/α), not a logarithm.
That algebra does not establish an entanglement theorem or a localisation transition for every long-range model. It shows that changing the interaction tail changes the timescale argument at its starting point.
Long-range exchange can also create additional transport channels, while long-range diagonal interactions can mainly dephase in a particular effective model. These should not be collapsed into one category. Specify the operator structure as well as the spatial decay.
Comparisons between trapped-ion chains, short-range spin models and optical-lattice systems require this distinction. Similar-looking slow signals can arise from different Hamiltonians and therefore need different extrapolations.
30. An external bath changes the isolation assumption
A system coupled to uncontrolled environmental degrees of freedom is not the isolated Hamiltonian problem described by an exact set of l-bits. Environmental transitions can flip or dephase the physical and dressed variables, and a sufficiently effective thermal reservoir can destroy the retained memory at late times.
A finite-time experiment may still display localisation-like dynamics before bath-induced relaxation dominates. That makes the bath coupling and observation window quantitative parts of the model, not reasons to discard the experiment.
Use Guide 52 when a Markovian description is justified and Guide 76 when the environment retains relevant memory. A slow decay caused by weak environmental coupling and a slow decay caused by rare internal resonances are different mechanisms.
31. A driven system needs a Floquet version of the question
Periodic driving removes ordinary energy conservation. The natural spectral object becomes the one-period unitary and its quasienergies, as developed in Guide 73.
A driven localised description would need quasi-local structure compatible with that periodic evolution. A high-frequency prethermal plateau is another possible reason for slow heating and can exist without asserting permanent localisation.
Therefore nonheating over a measured interval does not uniquely identify Floquet localisation. Compare dependence on drive frequency, disorder, interactions and time. A prethermal explanation and a localised explanation can agree over an initial window but make different predictions at longer times or under changed driving conditions.
32. A numerical workflow that preserves the question
Start by fixing the Hamiltonian convention, boundary conditions, disorder ensemble, symmetry sector and energy-density window. Store the random seed or actual field list for every sample. Construct H and verify Hermiticity and the commutator with the claimed conserved charge.
For small L, compare two independent constructions: a full Pauli-tensor matrix and a matrix built directly in the conserved bit-string sector. Their matched eigenvalues and observable predictions provide a useful implementation check.
Calculate at least one eigenstate quantity and one dynamical quantity. For example, combine gap ratios and eigenstate entanglement with imbalance and quench entanglement. Repeat over increasing L and independent disorder samples, keeping the physical conventions unchanged.
Separate numerical error from physical uncertainty. An eigensolver residual tests a matrix calculation. A time-step convergence test checks propagation. Disorder-sample uncertainty describes an ensemble. Finite-size drift concerns extrapolation. None is a substitute for the others.
33. A research result should come with a claim boundary
A useful result might read: in this fixed-sector random-field chain, over these sizes and times, the imbalance decays more slowly than this bound, the gap-ratio distribution approaches this benchmark, and the entanglement growth is compatible with a logarithm over this interval.
That statement is stronger than an unqualified label because another researcher can reproduce it and identify which additional tests would distinguish competing explanations. The uncertainty is attached to a specific extrapolation rather than spread vaguely over everything.
The goal is not to avoid conclusions. It is to draw the conclusion that the evidence actually supports and then make the remaining gap explicit. A well-controlled finite-size result can guide new theory and experiments even when the ultimate thermodynamic interpretation remains under investigation.
34. Practice laboratory with worked answers
Problem 1: local detuning. For H=δZ/2+JX, let δ=4J. What is the maximum probability of transfer between the two sites?
Answer. The maximum is 4J²/(16J²+4J²)=1/5. Detuning suppresses transfer, but the system still oscillates. This two-site result explains off-resonant mixing, not a complete many-body localisation phase.
Problem 2: participation. A state has probabilities 1/2, 1/4 and 1/4 on three sites. Find its IPR and participation ratio.
Answer. IPR=1/4+1/16+1/16=3/8, so PR=8/3. The result is a concentration measure, not a literal noninteger count of occupied sites.
Problem 3: conserved-sector dimension. How many computational-basis states occur in an eight-spin zero-magnetisation sector?
Answer. Four spins are up and four down, so the number is binomial(8,4)=70. Using all 256 states in one level-statistics calculation would mix magnetisation sectors unless the Hamiltonian breaks that conservation law.
Problem 4: l-bit meaning. Every eigenstate projector commutes with H. Why is that not already an explanation of local memory?
Answer. Generic eigenstate projectors are nonlocal. The l-bit description requires quasi-local conserved operators related to physical positions. Without that locality, conservation alone does not tell a local experiment what information it can retain.
Problem 5: entanglement without hopping. For H=JZAZB and initial |++⟩, find the reduced state of A at t=π/(4J).
Answer. cos(2Jt)=0, so ρA=I/2 and the two pure-state subsystems share one bit of entanglement. Neither Z population changed. Conditional phases created entanglement without population transport.
Problem 6: interaction range. If J(r)=J₀exp(−r/ξ), compare the distances that can acquire order-one phase at times t and 100t.
Answer. The difference is ξln100≈4.605ξ. A hundredfold increase in time only adds a constant number of localisation lengths in this logarithmic estimate. This is the origin of the broad hierarchy, not a guarantee that every entropy curve has the same prefactor.
Problem 7: imbalance. At a measurement time there are 70 detected particles on even sites and 30 on odd sites. Find I. Then explain why total particle loss must still be tracked.
Answer. I=(70−30)/100=0.4. The ratio shows an uneven surviving population. Without initial counts and loss information, it does not distinguish retained density memory from a bias introduced by selective loss or detection.
Problem 8: gap ratios. Three consecutive gaps are 0.2, 0.5 and 0.1. Compute the two adjacent ratios. Can their mean establish a phase?
Answer. The ratios are 0.4 and 0.2, with mean 0.3. Two values are far too few. A phase interpretation needs a controlled symmetry sector, a suitable energy window, many samples and size scaling, plus other diagnostics.
Problem 9: eigenstate versus quench entropy. Why can area-law eigenstates coexist with extensive late-time entanglement after a product-state quench?
Answer. The quench state is a coherent superposition of many eigenstates. Their relative interaction phases can entangle subsystems. A property of each eigenstate separately does not determine the entanglement of every superposition.
Problem 10: rare regions. In the illustrative avalanche estimate, doubling the thermal region size lowers its many-body level spacing exponentially. Why might that compete with decaying spatial coupling?
Answer. A larger bath has more closely spaced states with which a neighbour can hybridise. Under the stated ETH estimate, the coupling-to-spacing ratio contains exp(sℓ/2−r/ξ). The competition depends on how the bath grows and on the model assumptions; the formula is not a universal threshold.
Problem 11: a flat trace. An imbalance is unchanged within error bars from t=10 to t=100. What is the strongest immediate conclusion?
Answer. No decay larger than the experimental resolution was detected over that window. A nonzero infinite-time limit requires additional evidence or assumptions. The bounded observation remains useful and should be reported with its time interval and uncertainty.
Problem 12: a changing model. A short-range chain is replaced by one with power-law couplings. Why must the logarithmic entanglement argument be revisited?
Answer. The argument began with exponentially decaying interactions. A power-law tail gives a different relation between distance and interaction time. It can also change resonant transport and stability. Reusing the old conclusion without its original premise is invalid.
35. A four-stage study plan
Stage one: solve a small matrix. Derive the detuned two-site propagator, check the resonant and strongly detuned limits, and compute an IPR. Make the distinction between a calculation and a localisation claim explicit in your written conclusion.
Stage two: separate population and phase. Derive the two-l-bit reduced density matrix. Plot or tabulate its Z populations, X coherence and entropy at several times. The populations remain fixed while the other quantities change. This is the cleanest route past the misconception that localisation freezes all quantum information.
Stage three: compare diagnostics. Build a small random-field chain in one magnetisation sector. Calculate level ratios, eigenstate entanglement and an imbalance quench. Repeat over several field samples. Explain what each diagnostic measures and why agreement is more informative than any one label.
Stage four: audit an extrapolation. Read a primary paper and identify its largest size, longest time, disorder ensemble, physical model and claimed limit. Distinguish measured quantities from theoretical interpretations. Then propose one change of size, time, potential or observable that would most clearly separate two plausible explanations.
Frequently asked questions
Is many-body localisation just Anderson localisation with more particles? No. Interactions generate conditional energy shifts, dephasing and many-body resonances. The relevant conserved structure is quasi-local in an interacting many-body Hilbert space, not simply a collection of fixed one-particle orbitals.
Does a nonzero imbalance prove stable localisation? It proves retained imbalance over the measured window after experimental biases are controlled. Stability under longer times, larger sizes and rare regions is an additional question.
Can entanglement grow while particles stay localised? Yes. The exact H=JZAZB example demonstrates phase-generated entanglement without population transfer. In a many-body localised description, a hierarchy of weak interactions can extend this mechanism over long times.
Is a Poisson-like spectrum enough? No. Integrability, noninteracting structure and unresolved symmetries can also suppress level repulsion. Combine spectral information with locality, eigenstate and dynamical tests.
What can be stated confidently? Exact toy models establish particular mechanisms; conditional theorems establish results within their hypotheses; finite experiments establish behaviour over specified sizes and times. A useful account keeps all three, rather than weakening them into one vague claim or exaggerating them into a universal verdict.
Sources and further study
[1] P. W. Anderson, Absence of Diffusion in Certain Random Lattices, Physical Review 109, 1492 (1958). Foundational single-particle disorder localisation.
[2] Arijeet Pal and David A. Huse, The many-body localization phase transition, Physical Review B 82, 174411 (2010). Exact-diagonalisation eigenstate and finite-size diagnostics.
[3] David A. Huse, Rahul Nandkishore and Vadim Oganesyan, Phenomenology of fully many-body-localized systems, Physical Review B 90, 174202 (2014). Local conserved pseudospins and interacting dephasing.
[4] Maksym Serbyn, Z. Papić and Dmitry A. Abanin, Universal slow growth of entanglement in interacting strongly disordered systems, Physical Review Letters 110, 260601 (2013). Dephasing mechanisms and logarithmic entanglement within the localised description.
[5] Jens H. Bardarson, Frank Pollmann and Joel E. Moore, Unbounded growth of entanglement in models of many-body localization, Physical Review Letters 109, 017202 (2012). Numerical quench dynamics contrasting interacting and noninteracting systems.
[6] Wojciech De Roeck and François Huveneers, Stability and instability towards delocalization in MBL systems, Physical Review B 95, 155129 (2017). An ETH-based theory of thermal inclusions, with stated assumptions and stability predictions.
[7] John Z. Imbrie, On Many-Body Localization for Quantum Spin Chains, Journal of Statistical Physics 163, 998–1048 (2016). A one-dimensional proof under an assumption limiting level attraction.
[8] Junhyeok Hur and colleagues, Stability of many-body localization in two dimensions, arXiv preprint first submitted 28 August 2025. Variable-size optical-lattice measurements comparing random and quasiperiodic potentials; the finite observations and the proposed stability interpretation are kept distinct here.
Continue through Quantum Mathematics — Batch 19
Guide 73: Quantum Floquet Systems, Periodic Driving, Quasienergies, Magnus Expansion and Floquet Engineering studies repeated driving and prethermal windows. Guide 74: Non-Hermitian Quantum Mechanics, PT Symmetry, Exceptional Points and Biorthogonal Spectra distinguishes skin accumulation from disorder localisation. Guide 76: Non-Markovian Quantum Dynamics, Memory Kernels, Nakajima–Zwanzig Equations and Information Backflow studies memory in reduced open-system evolution.