Source-linked AI summary
The Adjoint Is All You Need: Characterizing Barren Plateaus in Quantum Ans\"atze
Enrico Fontana, Dylan Herman, Shouvanik Chakrabarti, Niraj Kumar, Romina Yalovetzky, Jamie Heredge, Shree Hari Sureshbabu, Marco Pistoia
TL;DR
The paper addresses the unproven relationship between barren plateaus and DLA dimension in uncontrollable quantum circuits. It uses compact Lie-group representation theory to analyze LASAs and proves an inverse-DLA-dimension variance result under a 2-design assumption, while showing rapid mixing for polynomial-dimensional DLAs.
Problem
For uncontrollable circuits, the proposed relationship between trainability and DLA dimension had numerical evidence but no general proof.
Source: §1 · p. 1Method
Results
For LASAs sampled as 2-designs of the dynamical Lie group, gradient variance falls inversely with effective DLA dimension; the paper also reports rapid mixing for polynomial-dimensional DLAs.
Source: §1 · p. 1Takeaways & Limitations
Abstract
from arXiv · showhide
Using tools from the representation theory of compact Lie groups, we formulate a theory of Barren Plateaus (BPs) for parameterized quantum circuits whose observables lie in their dynamical Lie algebra (DLA), a setting that we term Lie algebra Supported Ansatz (LASA). A large variety of commonly used ans\"atze such as the Hamiltonian Variational Ansatz, Quantum Alternating Operator Ansatz, and many equivariant quantum neural networks are LASAs. In particular, our theory provides, for the first time, the ability to compute the variance of the gradient of the cost function of the quantum compound ansatz. We rigorously prove that, for LASA, the variance of the gradient of the cost function, for a 2-design of the dynamical Lie group, scales inversely with the dimension of the DLA, which agrees with existing numerical observations. In addition, to motivate the applicability of our results for 2-designs to practical settings, we show that rapid mixing occurs for LASAs with polynomial DLA. Lastly, we include potential extensions for handling cases when the observable lies outside of the DLA and the implications of our results.
1. INTRODUCTION
The paper addresses the unproven link between DLA dimension and gradient trainability in uncontrollable circuits. It develops a Lie-algebra-based account for LASAs and shows that polynomial-dimensional DLAs can mix rapidly enough to motivate the 2-design setting.
1. INTRODUCTION
- Prior symmetry-based trainability results were restricted to Schrödinger-picture subspace-controllable circuits, while the DLA-dimension connection for uncontrollable circuits had numerical support only.Some such circuits have polynomial effective DLA dimension despite an exponentially growing invariant subspace.Source: §1 · p. 1
- The paper proves that, for its considered circuit class and 2-designs of the dynamical Lie group, gradient variance falls inversely with effective DLA dimension.The authors also experimentally verify their formulas for the quantum compound ansatz.Source: §1 · p. 1
A. General Framework
- For a barren plateau, gradient variance decays exponentially with system size; under Haar invariance, the paper relates an initialization plateau to one throughout training when parameters form an approximate 2-design.The paper defines trainability as variance falling only as Ω(1/poly(n)).Source: A. General Framework · p. 2
B. DLA - BP Connection
- A Lie Algebra Supported Ansatz (LASA) restricts the measurement operator to the circuit's DLA; examples include HVA, QAOA, and some—but not all—equivariant QNNs.An equivariant operator can lie outside the DLA, so not every equivariant QNN is a LASA.Source: B. DLA - BP Connection · pp. 2–3
C. Representation Theoretic Notation
- For polynomial-dimensional DLAs, randomly selected periodic generators yield an approximate 2-design after polynomially many layers, up to a log(1/ϵ) factor.The theorem assumes an orthogonal basis of periodic skew-Hermitian generators and uniform sampling of generators and angles.Source: C. Representation Theoretic Notation · p. 4
- The 2-design analysis may not apply to other initialization schemes or correlated parameters, which the authors note may also avoid barren plateaus.Their rapid-mixing result is intended to motivate the practical relevance of the 2-design assumption.Source: C. Representation Theoretic Notation · p. 4
D. Theory of BPs for LASA
The theory gives an explicit gradient-variance expression for LASAs by decomposing the DLA into simple ideals. Its compact-group form extends the simple-group result, with the center making no contribution.
D. Theory of BPs for LASA
- The compact-group result applies the simple-group variance expression to each compact simple ideal in the reductive decomposition.Compact Lie algebras decompose into simple ideals and a center.Source: D. Theory of BPs for LASA · p. 4
- The center does not contribute to the variance.The authors describe the resulting expression as a precise mathematical formula for LASAs.Source: D. Theory of BPs for LASA · p. 5
- For LASAs, the central theorem expresses gradient variance using quantities tied to the Lie algebra and its representation, with terms associated to the simple ideals.The stated setting assumes a finite-dimensional unitary representation, Lie-algebra elements for the observable and generator, and a density matrix.Source: D. Theory of BPs for LASA · pp. 4–5
E. Interpretation of Results
The variance expression interprets its numerator through operator support on simple ideals and identifies the conditions under which gradients decay exponentially. The paper also outlines a symbolic route for finding those ideals when the DLA is polynomial-dimensional.
E. Interpretation of Results
- The norms of the state, observable, and generator on each simple ideal measure their support there; the generator’s Killing norm uses the adjoint representation.The Frobenius norms of the state and observable can be interpreted as generalized measures of purity.Source: E. Interpretation of Results · p. 5
- The paper also presents an upper bound using more familiar Frobenius norms and states that its results prove and generalize the DLA-dimension conjecture for LASAs.The conjectured scaling need not hold when the initial state and observable cause other factors to dominate the decay.Source: E. Interpretation of Results · p. 5
- A barren plateau can arise from exponentially small state support on the DLA, support concentrated on an exponentially large ideal, or mutually incompatible support that makes all terms vanish.The mutually incompatible-support condition is a novel prediction and occurs only in the strict semisimple case.Source: E. Interpretation of Results · p. 5
- For polynomial-dimensional DLAs with Pauli-string generators, the authors describe an efficient procedure for discovering the simple ideals by computing and simultaneously block-diagonalizing adjoint-action matrices.The DLA basis can be found using nested commutators and linear-independence checks, with a stopping condition to avoid exponentially many iterations when the DLA is exponential.Source: E. Interpretation of Results · p. 5
F. Variance Computation for Quantum Compound Ansatz
The section applies the LASA framework to quantum compound ansätze, deriving rapid-mixing and gradient-variance results across several input states and observables.
Compound-ansatz structure
- The compound ansatz preserves Hamming weight, decomposing the representation into weight subspaces, while its DLA on the weight-n/2 subspace is su(n), not the full unitary algebra there.Thus the ansatz cannot enact all unitary transformations on that subspace; the construction uses SU(n) or SO(n) rotations on two-dimensional subspaces, including fermionic beam-splitter gates.Source: F. Variance Computation for Quantum… · pp. 5–6
Rapid mixing
- After O(tn log(1/ε)) layers, the stated SU(n) compound ansatz is an ε-approximate t-design for t ≤ n/2.The result assumes uniformly random rotation angles; t = 2 is the main case of interest for barren plateaus.Source: F. Variance Computation for Quantum… · p. 6
Gradient variance for LASA inputs
- SU-compound layers with Lie-algebra-supported measurements have no barren plateau for any fixed-Hamming-weight computational basis state.For the uniform superposition input, the variance also has no barren plateau and decays exactly with the DLA dimension n^2 − 1; numerical results agree with the predictions.Source: F. Variance Computation for Quantum… · p. 6, Figure 2 · p. 6
Single-qubit measurements
- For single-qubit measurements, the section extends the predictions to computational-basis and uniform-superposition inputs and reports numerical verification.The text says this answers an open question proposed in Ref..Source: F. Variance Computation for Quantum… · pp. 6–7
Observable outside the DLA
- With a rank-one projector outside the DLA, the compound ansatz displays a barren plateau, attributed to traceless operator components spreading across an exponentially large subset of su.The example uses an input basis state of Hamming weight n/2 and a projector onto another basis state in that space; the circuit depth exceeds O(log(n)) and the cost remains global.Source: F. Variance Computation for Quantum… · p. 7
G. Comparison with Previous Approaches
The paper contrasts state-first symmetry analysis with a Heisenberg-picture approach that uses the adjoint representation to handle subspace-uncontrollable settings. It also discusses the framework’s scope, including simulatability and observables outside the DLA.
G. Comparison with Previous Approaches
- The Heisenberg-picture approach uses the adjoint representation to identify invariant subspaces and bypass obstacles in the subspace-uncontrollable setting.Unlike the state-first approach, these subspaces reflect symmetries in the observable’s evolution.Source: G. Comparison with Previous Approaches · pp. 7–8
3. DISCUSSION
- The framework diagnoses barren plateaus in Lie algebra supported ansätze, including HVA, QAOA, and equivariant QNNs.It also makes gradient-variance analysis possible for subspace-uncontrollable circuits such as quantum compound ansätze.Source: §3 · p. 8
- For LASAs with polynomial-dimensional DLA, the paper links the absence of barren plateaus to classical simulability of expectation values.The stated connection is limited to LASAs.Source: §3 · p. 8
- For observables with support outside the DLA, variance expressions can be obtained, but determining their scaling is generally challenging.Inverse-DLA scaling is suggested when the observable has only small support outside the DLA, as shown for the quantum compound ansatz.Source: §3 · p. 8
4. METHODS
- The method derives the DLA-dimension connection using the adjoint representation and Schur orthogonality.For an observable in the Lie algebra, its conjugation follows the adjoint representation, whose dimension clarifies the inverse-DLA scaling.Source: §4 · p. 9, A. The Adjoint Representation Connection · p. 9
A. The Adjoint Representation Connection
- The second-moment calculation simplifies because the observable enters as a tensor product, linking its conjugation to the adjoint representation when it lies in the Lie algebra.The paper treats the dynamical group as compact, ignoring its center when needed because the central component of the observable does not evolve in the Heisenberg picture.Source: A. The Adjoint Representation Connection · p. 9
B. Proof of Theorem 2.8
The proof uses compact-group representation results to simplify the second-moment integral, then relates the resulting expression to the Killing form and Casimir operator.
B. Proof of Theorem 2.8
- For a compact simple Lie group, the commutant in the lemma is the one-dimensional subspace spanned by the Casimir operator.The lemma’s setting assumes a finite-dimensional unitary representation and an element of the Lie algebra.Source: B. Proof of Theorem 2.8 · p. 10
- The representation-basis identity provides a practical way to calculate the norm of a Lie-algebra element when only the represented basis is available.The circuit setting often provides {E_i}, rather than the original basis {e_i}.Source: B. Proof of Theorem 2.8 · p. 10
- The proof expands the gradient second moment and commutators, then applies the compact-group lemma to the inner integral.The resulting expression is rearranged using the Casimir-related identity.Source: B. Proof of Theorem 2.8 · p. 10
- Antisymmetry of the commutator identifies the inner sum with the Killing form, whose negative supplies an inner product for a compact simple Lie algebra.The structure constants are given by f^j_qk = Tr([E_q, E_k]E_j).Source: B. Proof of Theorem 2.8 · pp. 10–11
C. Proof of Theorem 2.9
The proof extends the compact-group calculation across the Lie algebra’s subalgebras, where cross terms and central contributions vanish. It then applies the resulting identities to calculations for particular states.
C. Proof of Theorem 2.9
- In the compact connected Lie-group setting, the variance receives contributions from the subalgebras, while cross terms and abelian-subalgebra contributions vanish.The proof attributes the vanishing of cross terms to the fact that each subalgebra is an ideal, and the central contribution vanishes after the commutator.Source: C. Proof of Theorem 2.9 · p. 11
D. Proof of Theorem 2.11
- For the computational-basis calculation, the proof uses an orthogonal basis for the Cartan subalgebra and symmetry under simultaneous qubit permutations to infer the final form.The diagonal Casimir expression is symmetric under tensor-factor exchange and simultaneous permutations of qubit indices.Source: D. Proof of Theorem 2.11 · pp. 11–12
- The practical forms of the identities are used to evaluate the Casimir on a computational-basis state of Hamming weight k.The proof evaluates a sum using k(k−1)+(n−k)(n−k−1)−2k(n−k) = (n−2k)^2−n.Source: D. Proof of Theorem 2.11 · pp. 11–12
E. Proof of Theorem 2.12
The proofs evaluate the relevant generator contributions for a uniform-superposition state and establish a no-barren-plateau result for single-qubit σz measurements in the stated cases.
E. Proof of Theorem 2.12
- For a uniform-superposition state, only Pauli-x-type generators contribute, and only n−1 of them do not annihilate the state.The non-annihilating generators have j = i+1.Source: E. Proof of Theorem 2.12 · p. 12
F. Proof of Corollary 2.12.1
- For any k and any single-qubit σz measurement, the gradient variance vanishes only polynomially in n, implying no barren plateau.A similar result is stated for the uniform-superposition state.Source: F. Proof of Corollary 2.12.1 · p. 12
Appendix A: Introduction to Lie Groups and Representation Theory
Appendix A reviews compact Lie groups, their representations, and the representation-theoretic tools used in the paper. It introduces norms and Casimir operators, then explains weights and roots.
1. Lie Groups, Lie Algebras, Representations
- Compact Lie groups admit matrix representations with Lie brackets given by commutators; their representations decompose into irreducible components and support Haar integration.The appendix also introduces equivariant maps and Schur orthogonality for representations.Source: §1 · pp. 15–16
2. Useful Representations and Norms
- Representation-dependent trace forms and norms are related by an index, providing standard, Killing, and Frobenius norm conventions.For compact simple Lie algebras, trace forms are real and definite, and any two are scalar multiples.Source: §2 · p. 16
3. Casimir Operators
- The quadratic Casimir is a central element of the universal enveloping algebra, while its tensor-power counterpart yields the split Casimir used in the main text.The appendix distinguishes these operators because the split Casimir acts proportionally to identity on irreducible components of a tensor-power representation, not generally on the original representation.Source: §3 · p. 17
4. Cartan Subalgebras, Weights, and Roots
- Cartan subalgebras simultaneously diagonalize representation operators, defining weights; in the adjoint representation, these functionals are roots.Root operators shift weight vectors, and the split Casimir is expressed in the Cartan–Weyl basis.Source: §4 · p. 18
Appendix B: Verifying Theory Reproduces Existing Results
Appendix B checks that the theory reproduces known gradient-variance results in three settings. Appendix C supplies technical lemmas, including the vanishing first Haar moment and decomposition across subalgebras.
1. SU(d)
- The fully controllable SU(d) standard-representation case reproduces the ordinary barren-plateau result and agrees exactly with the subspace-controllable result in.The Lie algebra su(d) has dimension d^2 − 1.Source: §1 · p. 19
2. SU(d/2) × SU(d/2)
- For SU(d/2) × SU(d/2), the DLA splits into two su(d/2) subalgebras, and each component of an operator projects onto its corresponding subsystem.For maximally entangled states the partial traces are proportional to identity, giving zero total variance; for an unentangled sum, the variance is the sum over subspaces.Source: §2 · p. 19
3. Spin– d−1
- In the spin-(d−1)/2 example, the DLA has dimension 3 and the adjoint representation index is 4; the calculated variance agrees with the same system’s result in.The example uses H = −iSx and O = Sx + Sy + Sz.Source: §3 · pp. 19–20
Appendix C: Proofs of Technical Lemmas
- For compact Lie groups, the first Haar moment of the gradient vanishes, so its variance equals its second moment.For reductive Lie algebras, the second-moment integral decomposes across simple ideals and the center, with cross terms vanishing by Schur orthogonality.Source: Appendix C · pp. 20–21
Appendix D: Proof of Main Text Theorem 2.13
Appendix D proves Theorem 2.13 for a quantum compound ansatz by analyzing the tensor-product representation and the gradient-variance integral. It uses the computational-basis state’s support on a particular irreducible component to simplify that integral.
Appendix D: Proof of Main Text Theorem 2.13
- The proof evaluates the second-moment operator through Schur–Weyl duality because the adjoint-representation shortcut is unavailable in this setting.The tensor-product representation is decomposed into irreducible components, with the symmetric tensor of the initial state restricting the relevant components.Source: Appendix D · p. 22
- A computational-basis state tensor itself lies in the irreducible component Vλ0, so projections onto other components do not contribute to the integral for that initial state.The proof identifies λ0 with the partition (2, . . . , 2).Source: Appendix D · pp. 22–23
- The theorem’s stated setting uses a computational-basis initial state of Hamming weight n/2 and a rank-one projector onto another basis state in that subspace.The analysis considers a subspace of dimension exponential in n, while the represented Lie algebra has dimension n^2 − 1.Source: Appendix D · pp. 22–23
- Swap invariance of the projector onto Vλ0 and its commutation with the lifted generator are used to rearrange the variance integral.The proof then uses the skew-Hermitian generator and positive semidefiniteness of −H^2 before invoking the dimension result for Vλ0.Source: Appendix D · p. 24
Appendix E: Projected Norm Lower Bound
The appendix develops a weight-space lower bound for the projected norm, then extends the variance analysis to observables outside the DLA. The extension identifies when DLA-dependent variance decay can still be bounded and where complement components complicate it.
Appendix E: Projected Norm Lower Bound
- Theorem E.2 gives a lower bound on the projected norm for a unit vector in a unitary representation of a simple Lie algebra.The bound is expressed through a norm formed from the vector’s weight-space representation.Source: Appendix E · p. 26
- A pure state supported on weight spaces whose weights differ by a root gives a positive contribution; a single weight vector gives equality in the stated condition.The text notes that the single-weight case is an example of equality, not a necessary condition.Source: Appendix E · p. 26
- For semisimple algebras, the lower-bound result applies separately to each simple ideal, with the state-dependent weight norm varying by ideal.The appendix contrasts this with the main-text upper bound, whose observable and generator norms depend on the representation.Source: Appendix E · p. 27
- Choosing an initial state with slowly decaying weight norms on subalgebras aligned between the observable and generator informs the lower bound on variance.A poor state choice can make variance fall faster than the DLA dimension; determining the relevant norms can be difficult even at small scales.Source: Appendix E · p. 27
Appendix F: Applicability of Theory beyond LASA
- For observables outside the DLA, the variance lower bound is expressed using ideal and ideal-complement components, together called the DLA component.The general observable is decomposed into DLA and orthogonal-complement parts.Source: Appendix F · pp. 27–29
- DLA-dimension decay can also appear for observables outside the DLA, while the non-ideal complement contributes no dependence on DLA size.If the complement has no support on components isomorphic to an ideal, the lower bound comes entirely from the ideal component; such isomorphic support can occur and may yield zero variance in the example given.Source: Appendix F · p. 29
1. Proof of Supplementary Theorem F.3
The proof establishes that its auxiliary quantities do not depend on the chosen isomorphism and uses Schur orthogonality to evaluate variance contributions. It then assembles the ideal-complement variance and its covariance with the ideal component.
1. Proof of Supplementary Theorem F.3
- The auxiliary quantities are well-defined because they are independent of the choice of G-isomorphism between the isomorphic components.The proof establishes this independence using Schur orthogonality for isomorphic but unequal representations.Source: §1 · pp. 29–30
- Under the stated representation hypotheses, Schur’s lemma makes the relevant composed operator a scalar multiple of the chosen G-isomorphism.The proof treats real and complex representations under their respective irreducibility conditions.Source: §1 · pp. 29–30
- The variance expression includes ideal-complement variance terms and covariance terms between the ideal and isomorphic ideal-complement components.The proof applies the main-text theorem and Lemma F.6 to obtain these contributions.Source: §1 · pp. 32–33
2. Variance of Non-ideal Complement
The section extends variance analysis to the non-ideal complement, where scaling need not depend on DLA dimension, and develops mixing bounds for LASAs, including quantum compound ansätze.
2. Variance of Non-ideal Complement
- The earlier lemmas allow exact computation of the variance of the non-ideal complement component.This component consists of irreducible representations not isomorphic to simple ideals of the DLA.Source: §2 · p. 33
- Complexifying the representation separates irreducible components, and Schur orthogonality eliminates covariance terms across the direct sum.The non-ideal terms may remain irreducible or split into two complex irreducible components.Source: §2 · p. 33
- Non-ideal variance can depend on operator spaces containing invalid quantum observables and can have no dependence on DLA dimension.With polynomially growing DLA, an exponentially large complement may contain an exponentially large irreducible subspace and potentially yield a BP; conversely, a polynomial complement may permit some observables and states to avoid one when the DLA is exponentially large.Source: §2 · pp. 33–34
Appendix G: Details of Numerical Results
- The numerical simulations use periodic quantum-compound ansätze in a brick architecture with alternating 1D-connected two-qubit layers.They use L = 12n layers and uniformly sample parameters from [0, 4π); gradient variance is estimated from 5,000 samples for n < 18 and 1,000 for n ≥ 18.Source: Appendix G · p. 34
Appendix H: Mixing time to t-designs
- The mixing analysis uses a lower bound based on minimum stable Killing rank; quantum compound ansätze saturate its lower bound, while Pauli rotations for su(2^n) saturate the upper bound.The appendix contrasts this with prior results showing efficient mixing for exponential DLAs under the su(2^n) Pauli basis.Source: Appendix H · pp. 34, 38
- For an orthogonal periodic generator basis and polynomial DLA dimension, the LASA becomes an ϵ-approximate 2-design after O(poly(n) log(1/ϵ)) layers.The spectral gap controls approximate-design formation, with additional steps exponentially suppressing the error.Source: Appendix H · pp. 35, 39
- For quantum compound ansätze, t ≤ n/2 gives an ϵ-approximate t-design for SU(n) after O(tn log(1/ϵ)) layers.The result is presented as a tighter mixing-time bound than the general LASA result.Source: Appendix H · p. 39
DISCLAIMER
The disclaimer limits the document’s purpose and disclaims representations, warranties, and liability.
DISCLAIMER
- The paper is informational material prepared by JPMorgan Chase’s Global Technology Applied Research center, not a Research Department product.Source: DISCLAIMER · p. 39
- It is not investment research or investment advice, nor a recommendation, offer, or solicitation concerning financial products or transactions.Source: DISCLAIMER · p. 39
- JPMorgan Chase and its affiliates disclaim representations, warranties, and liability regarding the document, including its completeness, accuracy, reliability, and potential legal, compliance, tax, or accounting effects.Source: DISCLAIMER · p. 39