Parker–Cao–Zaletel (Phys. Rev. B 102, 035147 (2020)) established the canonical-form, compression, and control theory of local matrix product operators (MPOs) — the operator analogue of matrix product states — but only for first-degree (extensive) operators, and left four interlocking theory gaps open: (i) a canonical form and compression with error control for 'second degree' and multi-local operators such as H^H^\widehat{H}\widehat{H}, needed for energy variances; (ii) a rigorous global error bound for infinite-MPO compression under simultaneous truncation of all bonds; (iii) a sharp necessary-and-sufficient characterization of convergence of the QR-iteration canonicalization, with cheaper preconditioning; and (iv) control of numerical precision amplification in iterative schemes such as MPO Lanczos/Krylov expansions. The direct successor line (van Damme et al. 2024; Vanthilt et al. 2026) deliberately reorganizes applications into first-degree MPOs rather than building the missing second-degree theory, and none of the four gaps has been closed. The problem is to supply this missing controlled, general compression theory for local operators represented as (infinite) MPOs.

Background

Matrix product states (MPS) encode one-dimensional quantum lattice wavefunctions as products of site-indexed matrices and underlie the density matrix renormalization group (DMRG), the standard numerical method for 1D quantum matter. Matrix product operators (MPOs) are the operator analogue — tensor-train representations acting on operator space — used to represent Hamiltonians and observables; translation-invariant versions on infinite chains (iMPOs) enable thermodynamic-limit and infinite-temperature calculations. Compressing an MPO to a smaller bond dimension (the size of the internal matrix indices) is a basic primitive of this software stack, e.g. when long-range two-body terms arising in moiré-material simulations are folded into a compact MPO.

Parker–Cao–Zaletel (PCZ) [1] gave a systematic theory for 'local' iMPOs — operators whose norm grows at most polynomially in the chain length NN — based on a degree classification that reads off the Jordan structure of the MPO transfer matrix (the matrix obtained by contracting the physical indices of neighboring operator tensors) when evaluated between injective matrix product states, i.e. states whose transfer map from physical to bond degrees of freedom is one-to-one. Zeroth-degree operators are bounded local observables; first-degree operators, including finite-range Hamiltonians H^\widehat{H} with norm ∼N\sim N, have dominant Jordan blocks of size 2; a product H^H^\widehat{H}\widehat{H} is 'second degree,' with norm ∼N2\sim N^2 and — by extrapolation of this degree pattern, which the source formalizes only through first degree — a dominant Jordan block of size 3. For first-degree iMPOs, PCZ proved an 'almost Schmidt decomposition' — a canonical form in which the Gram matrix of the left blocks is the identity up to controlled extensive corrections, analogous to the Schmidt/SVD canonical form of MPS — built in practice by a QR-iteration scheme (repeated QR decompositions driving the infinite tensor to canonical form); they derived single-bond truncation-error bounds in this canonical form, connected them to balanced truncation (a model-reduction technique from control theory that discards directions with small Hankel singular values, the control-theoretic analogues of Schmidt coefficients), and used repeated compression to build the operator Lanczos recursion O^n+1=[H^,O^n]−bnO^n−1\widehat{O}_{n+1} = [\widehat{H}, \widehat{O}_n] - b_n \widehat{O}_{n-1} starting from a local operator, as needed for operator moments and autocorrelation functions. Products H^H^\widehat{H}\widehat{H} arise when computing energy variances ⟨H^2⟩−⟨H^⟩2\langle \widehat{H}^2\rangle - \langle \widehat{H}\rangle^2 in DMRG, which is the prime motivation for second-degree objects.

Four gaps were left open. First, the degree-based canonicalization stops at first degree: no canonical form, almost-Schmidt-type decomposition, or operator-appropriate truncation with error control exists for second-degree or multi-local operators, precisely the objects needed for variances and higher moments in the thermodynamic limit. Second, the truncation-error analysis is per bond; whether simultaneously truncating every bond of an infinite MPO self-consistently (the global error) admits the same rigorous bound as in the infinite matrix product state (iMPS) case — as PCZ conjectured — is unproven. Third, the QR iteration converges only under a sufficient condition and provably fails for some first-degree iMPOs; the fallback (PCZ's Algorithm 6, a foolproof preconditioned canonicalization) is more expensive, and no necessary-and-sufficient convergence criterion or cheaper preconditioning is known. Fourth, in iterative procedures such as the Lanczos algorithm (a Krylov-subspace iteration for extremal eigenvalues and moments) implemented through repeated MPO compressions, compression errors amplify with the iteration count until the smallest above-gap singular value reaches machine precision, limiting the achievable Krylov depth; no adaptive-precision or stabilized-orthogonalization strategy tailored to the MPO algebra exists. PCZ additionally speculate about a unifying 'quantum information theory of operators' integrating algebraic structure and locality beyond 1D; that direction is motivation and context, not a deliverable of this problem.

Problem Statement

Objective (one coherent goal): extend the compression theory of local matrix product operators beyond the first-degree case treated in Parker–Cao–Zaletel [1], so that compressing local operators represented as (infinite) MPOs is controlled in the same rigorous sense — canonical form, truncation-error bound, convergent canonicalization, stable iteration — that PCZ established for first degree. Concretely, deliver the following four linked results, each posed as open by the source; all four are jointly required to resolve this objective, and an answer covering only some of them is incomplete:

(1) Second-degree and multi-local operators. Define a canonical form and/or an almost-Schmidt-type decomposition for second-degree iMPOs — informally, operators with norm ∼N2\sim N^2, such as H^H^\widehat{H}\widehat{H} for a finite-range Hamiltonian, whose transfer matrices are expected to carry dominant Jordan blocks of size 3 — and for 'multi-local' operators, i.e. operators arising as products of two or more first-degree iMPOs (of degree ≥2\ge 2); and use it to compress such operators to bond dimension χ\chi with a truncation-error bound in an operator-appropriate norm or measure (one that handles the extensivity making the naive state 2-norm unsuitable). A complete answer must cover the general second-degree and multi-local case — with the degree-2 H^H^\widehat{H}\widehat{H} case as a mandatory sub-case and the construction required to extend to products of arbitrarily many first-degree iMPOs — and state what the result implies for thermodynamic-limit energy-variance computations in (i)DMRG. The source's further question 'Can it be extended beyond 1d?' may be addressed but is not required.

(2) Global error of iMPO compression. Prove (or refute) that compressing every bond of an infinite MPO self-consistently incurs a global error that combines the individual bonds' truncation errors as in the iMPS case — i.e. supply the rigorous global bound for simultaneous truncations in the infinite system that PCZ conjectured to be 'almost exactly the same as the iMPS case'.

(3) QR-iteration convergence. Find a necessary-and-sufficient condition for convergence of the simple QR-iteration canonicalization of a first-degree iMPO (PCZ give only a sufficient condition, and first-degree iMPOs exist for which the iteration fails), and a cheaper way to perform the preconditioning that makes the foolproof Algorithm 6 unnecessary in practice.

(4) Precision control in iterative compression. Devise strategies — e.g. adaptive working precision or stabilized orthogonalization tailored to the MPO algebra — that control the amplification of compression errors in iterative procedures built from repeated MPO compressions, such as the operator Lanczos recursion O^n+1=[H^,O^n]−bnO^n−1\widehat{O}_{n+1} = [\widehat{H}, \widehat{O}_n] - b_n \widehat{O}_{n-1} (with O^0\widehat{O}_0 a local operator and bnb_n the Lanczos coefficients), so that the achievable Krylov depth is no longer limited by the smallest above-gap singular value reaching machine precision; and elucidate the reason for the observed scaling χ[O^n]=O(na)\chi[\widehat{O}_n] = O(n^a), a≈2a \approx 2, of these Lanczos operators, which the source explicitly poses as an open question.

Relationship among the deliverables: (1) supplies the objects and error measures, (2) the error accounting for infinite systems, (3) the convergent canonicalization step, and (4) the iteration-level stability; together they constitute the controlled compression theory that PCZ's first-degree framework provides only partially. PCZ's speculative 'quantum information theory of operators' and higher-body balanced-truncation generalization are motivation and context, not required deliverables.

References

  1. D. E. Parker, X. Cao, and M. P. Zaletel, "Local matrix product operators: Canonical form, compression, and control theory", Phys. Rev. B 102, 035147 (2020). DOI: 10.1103/PhysRevB.102.035147, arXiv:1909.06341, https://doi.org/10.1103/PhysRevB.102.035147
  2. M. van Damme, J. Haegeman, I. McCulloch, and L. Vanderstraeten, "Efficient higher-order matrix product operators for time evolution", SciPost Phys. 17, 135 (2024). DOI: 10.21468/SciPostPhys.17.5.135, arXiv:2302.14181, https://doi.org/10.21468/SciPostPhys.17.5.135
  3. V. Vanthilt, M. van Damme, J. Haegeman, I. P. McCulloch, and L. Vanderstraeten, "Matrix Product Operator Encodings of the Magnus Expansion and Dyson Series", arXiv:2605.21597 (2026), https://arxiv.org/abs/2605.21597
  4. T. Soejima, D. E. Parker, N. Bultinck, J. Hauschild, and M. P. Zaletel, "Efficient simulation of moiré materials using the density matrix renormalization group", Phys. Rev. B 102, 205111 (2020). DOI: 10.1103/PhysRevB.102.205111, https://doi.org/10.1103/PhysRevB.102.205111
  5. M. Yang, B. Vanhecke, and N. Schuch, "Detecting Emergent Continuous Symmetries at Quantum Criticality", Phys. Rev. Lett. 131, 036505 (2023). DOI: 10.1103/PhysRevLett.131.036505, https://doi.org/10.1103/PhysRevLett.131.036505
  6. M. Fishman, S. R. White, and E. M. Stoudenmire, "The ITensor Software Library for Tensor Network Calculations", SciPost Phys. Codebases 4 (2022). DOI: 10.21468/SciPostPhysCodeb.4, https://doi.org/10.21468/SciPostPhysCodeb.4
  7. L. Balada Gaggioli and J. Mareček, "Time evolution of controlled many-body quantum systems with matrix product operators", Phys. Rev. A 112, 062612 (2025). DOI: 10.1103/9mfk-gg3x, arXiv:2509.07228, https://doi.org/10.1103/9mfk-gg3x
  8. E. Cortés Estay, N. A. Kamar, and P. Corboz, "Accurate computation of the energy variance for ground states and ⟨⟨L†L⟩⟩\langle\langle \mathcal{L}^\dagger \mathcal{L} \rangle\rangle for steady states using iPEPS", Phys. Rev. B 113, 245114 (2026). DOI: 10.1103/fkyw-c5tn, arXiv:2511.22669, https://doi.org/10.1103/fkyw-c5tn