Quantum criticality at strong randomness: a lesson from anomaly Yasamin Panahi, Subhayan Sahu, Naren Manjunath, Chong Wang (王翀) Abstract Quantum criticality in the presence of strong quenched randomness remains a challenging topic in modern condensed matter theory. We show that the topology and anomaly associated with average symmetry can be used to predict certain nontrivial universal properties. Our focus is on systems subject to average Lieb–Schultz–Mattis constraints, where lattice translation symmetry is preserved only on average, while on-site symmetries remain exact. We argue that in the absence of spontaneous symmetry breaking, the system must exhibit critical correlations of local operators in two distinct ways: (i) for some operator O e O_{e} charged under exact symmetries, the first absolute moment correlation | ⟨ O e ​ ( x ) ​ O e † ​ ( y ) ⟩ | ¯ \overline{|\langle O_{e}(x)O^{\dagger}_{e}(y)\rangle|} decays slowly; and (ii) for some operator O a O_{a} charged under average symmetries, the first-moment correlation ⟨ O a ​ ( x ) ​ O a † ​ ( y ) ⟩ ¯ \overline{\langle O_{a}(x)O^{\dagger}_{a}(y)\rangle} decays slowly. We verify these predictions in a few examples: the random-singlet Heisenberg spin chain in one dimension, and the disordered free-fermion critical states in symmetry class BDI in one and two dimensions. Surprisingly, even for these well-studied systems, our anomaly-based argument reveals critical correlations overlooked in previous literature. We also discuss the experimental feasibility of measuring these critical correlations. Quantum criticality in the presence of strong quenched randomness is ubiquitous in modern condensed matter physics, from superconductor-insulator transition, to quantum Hall effects and disordered quantum spin systems. Despite its long history, the subject of disordered quantum criticality has remained extremely challenging for theoretical understanding. The difficulty stems from the lack of theoretical tools that are both non-perturbative and broadly applicable. The notion of symmetry anomaly (more precisely, ’t Hooft anomaly) 63 ; 12 has become central in our non-perturbative understanding of both condensed matter physics and quantum field theory. In condensed matter systems, the most common manifestation of symmetry anomaly is the Lieb-Schultz-Mattis (LSM) constraint 35 ; 47 ; 22 and its many generalizations 59 ; 7 ; 51 ; 14 . LSM-type constraints require the ground states to be nontrivial (namely long-range entangled 5 ) for any lattice systems with translation symmetry and “fractional” symmetry quantum numbers ( e.g. spin- 1/2 under SO(3) ) per unit cell 20 ; 37 . Furthermore, they also heavily constrain what type of nontrivial low energy theory can emerge, since the anomaly of the low energy theory (IR) must match that of the microscopic lattice system (UV). A familiar example of this type is the Luttinger theorem, which demands the Fermi surface volume (an IR notion) to match with the microscopic charge density (a UV notion) 48 ; 13 . Examples of such LSM anomaly-matching with conformal field theory (CFT) in the IR have also been discussed extensively in both one and two dimensions 8 ; 27 ; 42 ; 62 ; 4 ; 6 ; 32 ; 54 ; 61 . Although ’t Hooft anomalies have played a central role in many areas of modern condensed matter and high-energy physics, they have seen relatively limited use in the study of disordered quantum criticality. An isolated example of an LSM-type constraint in a one-dimensional spin chain with exact SO(3) spin rotation symmetry and average translation symmetry was discussed in Ref. 29 . More generally, the notion of average anomaly was introduced in Ref. 39 and systematically classified in Ref. 40 . An average anomaly typically involves both an exact symmetry – one that holds for every disorder realization – and an average symmetry – one that is preserved only upon averaging over the disorder ensemble. It was argued in Ref. 39 that systems with such average anomaly must have long-range entangled ground states. Despite its early successes, the potential of the anomaly-based approach remains far from fully realized. In particular, can quantum anomalies be used to gain a deeper understanding of the universal properties of certain disordered quantum critical states? A closely related issue is that, although the notion of long-range entanglement – a consequence of anomaly – is conceptually appealing, it is not a property that can be directly and efficiently measured in experiments. It is therefore desirable to formulate predictions directly in terms of measurable quantities, such as correlation functions. In this work we uncover how anomalies constrain correlation functions for disordered quantum critical states. As an example, we consider the average LSM anomaly in general d space dimension, where lattice translation is an average symmetry, while some on-site symmetry (like SO(3) , time-reversal or fermion parity) remains an exact symmetry and is represented projectively ( e.g. a spin- 1/2 for SO(3) , Kramers’ doublet for time-reversal or unpaired Majorana for fermion parity) in the Hilbert space of a unit cell. We choose to focus on LSM anomalies as they are the most commonly encountered anomalies in condensed matter systems. Our discussion, however, holds more generally for anomalies involving both exact and average symmetry. The central result of this work is the following: Power-law rule : if a system has anomalous exact and average symmetry that are not spontaneously broken in the ground state, then for some operator O_{e} charged under exact symmetries, the first absolute moment correlation \overline{|\langle O_{e}(x)O^{\dagger}_{e}(y)\rangle|} (known as the Edwards–Anderson or EA correlator) decays critically; for some operator O_{a} charged under average symmetries, the first-moment correlation \overline{\langle O_{a}(x)O^{\dagger}_{a}(y)\rangle} decays critically. Here “critical” decay typically means power-law \sim 1/|x-y|^{2\Delta} . For typical critical phases studied in this work, the symmetry-charged operators O are local. There can be nontrivial exceptions: for example, if the system develops nontrivial topological orders, the charge operator O may act on large strings or higher dimensional sub-manifolds and the correlation function will be non-decaying. We emphasize at the outset that we do not have a rigorous proof of our power-law rule – such results are difficult to establish even in clean systems, especially in general dimensions, although they are widely expected to hold. Instead, we present plausibility arguments and illustrate them through several nontrivial examples: the random-singlet phase in 1d spin chains 41 ; 11 ; 16 , and the disordered free-fermion critical states in symmetry class BDI, in both one and two dimensions. The one dimensional state is related to the random transverse-field Ising model 17 , and the two dimensional state, known as the Gade phase, has been extensively studied 18 ; 19 ; 38 ; 45 ; 21 ; 52 ; 30 ; 28 ; 46 ; 10 . Surprisingly, even for these relatively well-studied examples, our results predict nontrivial correlation functions that have been overlooked in the prior literature. Specifically, we show that in both examples, the first-moment dimer-dimer correlation function (see Eq. ( 6 )) exhibits a nontrivial power-law decay. Anomaly and Correlation Here we give a plausibility argument for our main claims, the “power-law rule”, about how anomaly constrains correlation functions. Let us first recall how such constraints operate in clean systems. In a clean system with global symmetry G , a nontrivial ’t Hooft anomaly guarantees that the system cannot realize a trivially gapped ground state. Equivalently, assuming a sensible notion of renormalization-group (RG) flow, a theory with a nontrivial anomaly cannot flow to a trivial infrared (IR) effective theory. For concreteness, recall two familiar examples of nontrivial ’t Hooft anomalies in condensed matter physics: (i) the boundary of a topological insulator (in two or three dimensions), where gapless edge modes are guaranteed as long as charge-conservation U(1) and time-reversal \mathbb{Z}_{2}^{T} symmetries remain unbroken; (ii) spin systems subject to LSM constraints (half-integer SO(3) spin per unit cell) in arbitrary spatial dimension d , where the ground state cannot be trivial as long as spin-rotation SO(3) and lattice translation symmetry \mathbb{Z}^{d} remain unbroken. In fact, an anomaly implies much more than merely a nontrivial IR theory. As an invariant under RG flow, a nontrivial anomaly must also be reflected in the effective IR description. This immediately implies that the symmetry involved in the anomaly must act nontrivially in the IR, in the sense that certain operators \{O\} carry charges under the anomalous symmetry and exhibit nontrivial correlation functions in the IR. Typically, \{O\} is a set of local operators, and a “nontrivial” correlation function means a power-law decay \langle O(x)O^{\dagger}(y)\rangle\sim\frac{1}{|x-y|^{2\Delta_{O}}}. (1) The rationale behind the above “power-law rule” is that if, contrary to this rule, all charged operators had uninteresting correlation functions – such as exponential decay – then the symmetry would effectively act trivially in the IR theory, in contradiction with the RG invariance of the ’t Hooft anomaly. We can again recall some familiar examples: for the gapless Dirac fermions on the boundary of a topological insulator, \{O\} includes the Dirac mass (odd under time-reversal) and the Cooper pair (charged under U(1) ); for the spin- 1/2 antiferromagnetic Heisenberg chain, \{O\} includes the spin operator \vec{S}_{i} ( i labeling the lattice site), which is charged under SO(3) , and the staggered dimer operator (-1)^{i}\vec{S}_{i}\cdot\vec{S}_{i+1} , which flips sign under translation by one unit cell T_{x} . 1 1 1 Interestingly, in the Heisenberg-chain example, all operators with power-law correlations transform trivially under translation by two unit cells T^{2}_{x} . In other words, all nontrivial IR operators carry lattice momentum either 0 or \pi . This is fully compatible with the anomaly: the \mathbb{Z}_{2} nature of the LSM anomaly implies that symmetry operations like T_{x}^{2} are not involved in the anomaly and are therefore allowed to act trivially in the IR theory. Of course, this does not forbid T_{x}^{2} from acting nontrivially on some other IR theory of the system with a very different lattice Hamiltonian. We note that not all anomalous theories satisfy the above power-law rule. Notable exceptions include spontaneous-symmetry-breaking (SSB) phases and gapped topological orders. In such cases, the anomalous symmetry can act in more subtle ways, not necessarily through local operators with power-law correlations. For example, there may exist nontrivial gapped excitations – such as domain walls in SSB phases or anyons in topological orders – that transform nontrivially under the anomalous symmetry. In this work we restrict attention to systems that do not enter SSB or topologically ordered phases – phases that are, in any case, much simpler than quantum criticality – and are therefore expected to obey the power-law rule. We emphasize again that, even for clean systems, the power-law rule has no rigorous proof in general, even though it is widely expected to be true. We now consider disordered systems with total symmetry \mathcal{G}=K\times G , where K is an exact symmetry and G is only an average symmetry (more general forms of symmetry can be treated similarly 40 ). We shall consider examples where K and G have a mixed ’t Hooft anomaly, namely the symmetry is anomalous if and only if both K and G are unbroken. A physically relevant example, which will be the main focus of this work, is LSM-type of anomalies with an average lattice translation symmetry. It was shown 39 that as long as the on-site symmetry ( SO(3) , time-reversal or fermion parity) remains exact and is represented projectively (spin- 1/2 , Kramers doublet, unpaired Majorana fermion) in each unit cell, the LSM anomaly remains nontrivial in this setting. If the on-site symmetry is broken, even only to an average symmetry by a random magnetic field, the anomaly becomes trivial, since the ground state in each disorder realization can now be a trivial product state (with the spins simply aligning with the local random field). We now discuss the power-law rule for systems with nontrivial average anomalies. A natural question immediately arises: what type of correlation function should we consider? Candidates include the first-moment correlation \overline{\langle OO^{\dagger}\rangle} and Edwards–Anderson-type correlation \overline{|\langle OO^{\dagger}\rangle|} , where \langle...\rangle denotes the quantum expectation value in a single disorder realization, and \overline{\cdot\cdot\cdot} denotes averaging over the disorder ensemble. Here we claim that for operators charged under the exact symmetry, we should consider the Edwards–Anderson (EA) correlator; while for operators charged under the average symmetry, it suffices to consider the first-moment correlator. In the Supplementary Material 53 , we provide a derivation of this result using the replica trick. Here we shall provide a physically intuitive argument. The key observation is that the power-law rule, Eq. ( 1 ), indicates that the symmetry is on the verge of spontaneous breaking. Recall that a power-law decay of a charged operator is often referred to as “quasi-long-range order” in Berezinskii–Kosterlitz–Thouless (BKT) physics, while true symmetry-breaking order corresponds to the limit \Delta_{O}=0 . Therefore we can infer which correlators should satisfy a power-law rule based on which correlator detects the limiting case of symmetry-breaking order. Let us first examine the average symmetry G . What does it mean to have G spontaneously broken? Consider a local operator O_{a} that is charged under the average symmetry G , and an explicit breaking of G would mean \overline{\langle O_{a}(x)\rangle}\neq 0 . Consequently, a spontaneous breaking of G can be detected through the two-point correlation function C_{a}(|x-y|)=\overline{\langle O_{a}(x)O^{\dagger}_{a}(y)\rangle}, (2) which approaches a nonzero constant as |x-y|\to\infty in the SSB phase. In our discussion of anomaly constraints, we assume that G is not spontaneously broken, so C_{a}(|x-y|) does not saturate to a nonzero constant and instead decays to zero at large distances. Therefore, we come to the expectation stated as the power-law rule, that there exists a G -charged operator O_{a} for which C_{a}(|x-y|) decays as a power-law. The power-law decay of the first-moment correlator can be viewed as a quasi-long-range order where the average symmetry G is on the verge of spontaneous breaking. Notably, being a first-moment quantity makes the correlation function Eq. ( 2 ) feasible for experimental observation – a point we shall return to later. Next we consider the exact symmetry K . Since K is a symmetry for each disorder realization I , K is spontaneously broken if each realization breaks K , in the sense that \langle O_{e}^{\dagger}(x)O_{e}(y)\rangle_{I}\to c_{I}\neq 0 as |x-y|\to\infty for some operator O_{e} charged under the exact symmetry K . The natural ensemble-averaged quantity that captures this sample-wise symmetry breaking order is the EA correlator C_{e}(|x-y|)=\overline{|\langle O_{e}(x)O^{\dagger}_{e}(y)\rangle|}, (3) which approaches a constant c>0 as |x-y|\to\infty in an SSB phase. As for the first-moment correlator like Eq. ( 2 ) but with O_{e} , there are two possibilities: (1) \overline{\langle O_{e}(x)O^{\dagger}_{e}(y)\rangle}\to c\neq 0 , meaning that K is also broken as an average symmetry, and (2) \overline{\langle O_{e}(x)O^{\dagger}_{e}(y)\rangle}\to 0 , meaning that K is unbroken as an average symmetry. The latter possibility is familiar in spin glass physics, where the spins order in each sample, but in random directions so that the symmetry still holds on average. Assuming the absence of SSB order we can now motivate the other part of the power-law rule: the EA correlator, Eq. ( 3 ), should decay as a power-law for some O_{e} charged under the exact symmetry K . If the anomaly remains nontrivial even after reducing K to an average symmetry, we would likewise require power-law decay for the first-moment correlator, Eq. ( 2 ). However, in this work we focus on situations where K must be an exact symmetry for the anomaly to remain nontrivial. In such cases, it is consistent for the first-moment correlator to decay exponentially, as we will see in later examples. Of course, there can be examples where the first-moment correlator decays critically even when this behavior is not required by an anomaly. In the language of quasi-long-range order, our result implies that the exact symmetry K should at least be on the verge of spontaneously breaking down to an average symmetry. We note that the relation between average anomaly and various correlation functions has been discussed from a different perspective (via mapping to mixed-state density matrices) in Ref. 60 . Examples Random Majorana lattice in d=1,2 . Our first example is the random Majorana hopping model on bipartite lattices. Consider a chain in 1d or a square lattice in 2d , with one Majorana fermion mode \gamma_{j} on each lattice site j . The Hamiltonian is simply a nearest-neighbor hopping model with random hopping amplitude: H=-i\sum_{\langle jk\rangle}t_{jk}\gamma_{j}\gamma_{k}, (4) with t_{jk}=-t_{kj} an independent random real variable on each link \langle jk\rangle . This model preserves the average lattice translation symmetry as long as the probability distribution P[t_{jk}] is translationally invariant. The fact that there is only one Majorana mode per unit cell then leads to a nontrivial average anomaly involving the exact fermion parity \mathbb{Z}_{2}^{f} and the average translation \mathbb{Z}^{d} 23 ; 1 . Additionally, there is an exact time-reversal symmetry \mathcal{T}:\gamma_{j}\to(-1)^{\sigma(j)}\gamma_{j} , where \sigma(j)=\pm 1 depending the sublattice of j . This satisfies \mathcal{T}^{2}=+1 , placing the exact symmetry in class BDI 3 . As a consequence of the anomaly, the ground state of the free-fermionic model is delocalized in any spatial dimension, even for d=1,2 . Even in the presence of local interactions, the anomaly still requires the ground state to be either long-range entangled or to spontaneously break one of the symmetries involved in the anomaly. Since fermion parity cannot be spontaneously broken, the only remaining possibility is the spontaneous breaking of the average translation symmetry. However, spontaneous breaking of an average symmetry is forbidden in d\leq 2 by the Imry-Ma constraint 26 . (a) Random Majorana lattice, d=1 (b) Random Majorana lattice, d=2 (c) Random AFM Heisenberg chain Figure 1 : Power-law finite-size scaling of correlation functions in the random Majorana lattice model and the random antiferromagnetic Heisenberg chain, for operators charged under exact symmetries ( C_{\mathrm{EA}} [Eq. 3 ] and C_{\mathrm{SS}} [Eq. 8 ] respectively) and average symmetries ( D_{\mathrm{fm}} [Eq. 6 ] for both models). We only show data with signal-to-noise ratio (SNR) and lattice separation ( r ) above the thresholds stated below. (a) 1d random Majorana lattice: exact diagonalization with 10^{4}L disorder realizations; r\geq 3 , \mathrm{SNR}\geq 10 .(b) 2d random Majorana lattice: correlations evaluated at r=r_{x} , exact diagonalization with 200L disorder realizations; r\geq 3 , \mathrm{SNR}\geq 3 . (c) Random antiferromagnetic Heisenberg chain: DMRG, with sweeps performed until the energy change between successive sweeps was below 10^{-8} , using 50 – 100 disorder realizations; r\geq 2 , \mathrm{SNR}\geq 3 . In 1d , the model Eq. ( 4 ) can be solved analytically using the strong-disorder renormalization group (SDRG) approach and is known to flow to an infinite-randomness fixed point akin to the random singlet state 17 . In 2d , the model flows to a critical delocalized state known as the Gade phase 19 ; 18 . It is also believed that the Gade phase will eventually flow to an infinite-randomness fixed point in the thermodynamic limit, although for numerically accessible system sizes its behavior is known to be more subtle 43 . The 1d random Majorana model is closely related to several other strongly interacting models. Through a Jordan–Wigner transformation, the Majorana model can be mapped to a random transverse-field Ising model, which is solved using SDRG 17 . The random-bond antiferromagnetic Heisenberg chain can likewise be solved via SDRG 16 , exhibiting universal properties similar to those of the random Majorana model, which we comment on later. We now define the relevant correlation functions. The fermion operator \gamma_{j} is charged under the exact fermion parity symmetry, so motivated by the power-law rule we study the EA correlator C_{\mathrm{EA}}(r):=\frac{1}{L^{d}}\sum_{j}\overline{|\langle i\gamma_{j}\gamma_{j+r\hat{x}}\rangle|}, (5) where we average over positions j ( L is the linear system size). We can also study the first-moment correlator G(r)=\sum_{j}\overline{\langle i\gamma_{j}\gamma_{j+r}\rangle}/L^{d} , which is not constrained by the anomaly. The numerical results are presented in Figs. 1(a) and 1(b) . We find that, in both 1d and 2d , C_{\mathrm{EA}}(r) indeed decays as a power-law, while G(r) decays exponentially for generic disorder 53 . While the behaviors are fully consistent with the power-law rule from anomaly arguments, they are hardly surprising – after all, even for an ordinary diffusive metal, the fermion Green’s function G(r) will decay exponentially beyond the mean free path, while the EA correlator Eq. ( 5 ) decays as a power-law C^{diff}_{\mathrm{EA}}\sim 1/r^{d/2} (but with a different exponent from the random Majorana model in d=1,2 ). 2 2 2 More precisely, in an ordinary diffusive metal, \overline{|\langle c^{\dagger}_{r}c_{0}\rangle|^{2}}\sim 1/r^{d} follows from the standard diffuson pole \overline{G^{R}G^{A}}\sim 1/(i\omega-Dq^{2}) . 15 The real surprise comes when we consider first-moment correlation functions for operators charged under the average translation symmetry. Let d_{j}\equiv i\gamma_{j}\gamma_{j+\hat{x}} be a dimer operator. Its k=\pi component, d_{j}^{\pi}=(-1)^{x_{j}}d_{j} , is charged under the average translation symmetry, since \overline{\langle d_{j}^{\pi}\rangle} acquires a sign under translations along \hat{x} for a translation-invariant disorder ensemble. We compute the staggered first-moment correlator D_{\mathrm{fm}}(r):=\frac{1}{L^{d}}\sum_{j}\left(\overline{\langle d_{j}d_{j+r}\rangle}-\overline{\langle d_{j}\rangle}\,\overline{\langle d_{j+r}\rangle}\right)(-1)^{r}, (6) which is equivalently a first-moment correlator for d^{\pi} . Our numerical results are shown in Figs. 1(a) and 1(b) . In d=1 , we find a clear 1/r^{2} power-law decay, while in d=2 we observe a 1/r^{3} power-law decay. Random antiferromagnetic Heisenberg chain. The power-law rule extends beyond free fermionic examples discussed so far, and holds for generic interacting systems. In the absence of analytically or numerically tractable interacting disordered systems in higher dimensions, we focus on the random antiferromagnetic (AFM) spin- \frac{1}{2} Heisenberg chain 41 ; 11 , H=\sum_{i}J_{i,i+1}~S_{i}\cdot S_{i+1}, (7) which exhibits the well-known random singlet phase 16 . This model has an exact SO(3) symmetry and an average translation symmetry, and thus realizes the average LSM anomaly 29 ; 39 . It has a power-law decaying spin-spin correlation 16 C_{\mathrm{SS}}(r):=\frac{1}{L}\sum_{i}\overline{\big\langle\vec{S}_{i}\cdot\vec{S}_{i+r}\big\rangle}\sim(-1)^{r}\,r^{-2}, (8) consistent with the ‘‘power-law rule” for the EA correlator 3 3 3 For random singlets, the standard spin-spin correlator is trivially also the EA correlator upto an overall sign, since \langle S_{i}\cdot S_{j}\rangle=-3/4 , i.e. of uniform sign, if i,j form a singlet, and zero otherwise. which we also verify in Fig. 1(c) . We further consider the dimer operator d_{i}=\vec{S}_{i}\cdot\vec{S}_{i+1} , whose staggered component is charged under the average translation symmetry, and find that the first-moment correlator defined in Eq. 6 decays as 1/r^{2} (Fig. 1(c) ), exemplifying the power-law rule for average symmetries. This behavior can also be analytically explained using SDRG arguments 53 . Finally, we expect this general phenomenon to also extend to higher dimensions and beyond-LSM average anomalies, which will be interesting to study in the future. Experimental relevance A natural question regarding the power-law rule, which is especially relevant for its experimental significance, is whether these observables are self-averaging. To diagnose that, we studied the mean and standard error of each of these quantities at different r,L (the separation and the system size, respectively), and study the corresponding signal-to-noise ratios (SNR). In the Supplementary Material 53 , we provide evidence that the SNRs for these quantities satisfy an asymptotic power-law behavior, \displaystyle\mathrm{SNR}\sim L^{a}/r^{b},~~\text{with}~~b>a>0, (9) which suggests that these correlators are self-averaging in the thermodynamic setting where r is fixed and L\to\infty . We highlight that these correlators are experimentally accessible in both synthetic quantum matter (such as cold atomic gases in optical lattices), and in solid-state quantum materials. In cold atomic platforms, where snapshots of disordered ground states can be collected, higher moment correlators of Edwards–Anderson type as well as regular correlation functions can be measured to diagnose the anomaly. However, in quantum materials, EA correlators are generally inaccessible. Experimental scattering probes such as inelastic neutron scattering and X-ray scattering typically measure dynamical structure factors S_{O}(\mathbf{q},\omega) , which are space-time Fourier transforms of two-point correlation functions of suitable local operators O_{\bf{r}} . In a given sample and for an operator O_{\bf{r}} the structure factor may be written as S_{O}(\mathbf{q},\omega)=\sum_{\mathbf{r}}\int_{-\infty}^{\infty}dt\;e^{i(\omega t-\mathbf{q}\cdot\mathbf{r})}\,\langle O_{\mathbf{r}}(t)\,O_{\mathbf{0}}(0)\rangle. (10) Assuming the correlation function self averages (which we checked for all the examples we studied), the expectation value can be replaced by the sample-averaged one \langle\cdots\rangle\to\overline{\langle\cdots\rangle} . Now, the equal-time structure factor S_{O}(\mathbf{q})=\int\frac{d\omega}{2\pi}\,S_{O}(\mathbf{q},\omega) is directly related to the disorder-averaged first-moment spatial correlator, since \overline{\langle O_{\mathbf{r}}(t)\rangle}=0 for charged operators. Hence, such disorder-averaged first-moment two-point functions are experimentally accessible. For random-singlet and related critical chains at strong randomness, the spin-spin structure factor has been studied using strong-disorder RG 9 ; 44 , numerically 57 , and experimentally through indirect probes in 56 . In two dimensions, scattering studies of disordered quantum spin-liquid (QSL) candidates such as YbMgGaO 4 33 ; 49 ; 55 ; 34 reveal QSL phenomenology consistent with the possibility of higher dimensional analogues of long-range valence bond correlations 29 . Model Hamiltonians realizing random-singlet phases in 2d have also been identified numerically 36 . It will be interesting to design experiments in such materials to probe the anomaly-enforced critical correlations uncovered in our work. In a strongly disordered quantum-critical system, suitable choices of operators O_{i} that transform nontrivially under the anomalous average global symmetry (such as dimer operators in the examples studied here) may allow direct observation of the predicted nontrivial power-law decay of disorder-averaged spatial correlations via the corresponding S_{O}(\mathbf{q},\omega) or S_{O}(\mathbf{q}) . Notably, the dimer operators being spinless, should couple to charge density operators that are probed by X-ray scattering. Acknowledgements : We thank Jinmin Yi for illuminating discussions. Y.P. acknowledges support from the Natural Sciences and Engineering Research Council of Canada (NSERC) through Discovery Grants. Research at Perimeter Institute is supported in part by the Government of Canada through the Department of Innovation, Science and Economic Development; and by the Province of Ontario through the Ministry of Colleges, Universities, Research Excellence and Security. Part of this work was done at the Kavli Institute for Theoretical Physics during the “Noise-robust Phases of Quantum Matter” program, which was supported in part by grant NSF PHY-2309135 to the Kavli Institute for Theoretical Physics (KITP). Supplementary Material for “Quantum criticality at strong randomness: a lesson from anomaly” I Anomaly and correlation in replica theory We follow the standard Euclidean path integral treatment of a disordered quantum system using replica trick. The object to consider is the partition function of N identical copies of the system, all coupled with the disorder potential. The N -replicated partition function takes the form Z^{(N)}=\int Dv\prod_{a=1}^{N}D\phi_{a}\exp\left(-\sum_{a=1}^{N}S[\phi_{a},v]-S_{D}[v]\right), (11) where \phi_{a} represents the dynamical degrees of freedom in each replica and v is the random disorder potential. S[\phi_{a},v] describes the dynamics of each replica and its coupling to the disorder, while S_{D}[v] describes the probability distribution of the disorder potential. The standard practice is to calculate correlation functions from the replicated partition function, with the replica limit N\to 0 taken at the end of the calculation 2 . The exact and average symmetries act differently in the replicated theory: the exact symmetry K acts in a single replica, effectively becomes K^{\otimes N} , and the random potential v transforms trivially under K ; the average symmetry G , on the other hand, acts simultaneously in all replicas and v can transform nontrivially under G . If the original symmetry is K\times G (which is what we mainly focus on in this work), then in the replica theory the symmetry becomes K^{\otimes N}\times G . It is also useful to define the “average part” of the exact symmetry, which is the diagonal subgroup K_{\rm diag}\in K^{\otimes N} – this is the formal manifestation of the fact that an exact symmetry is automatically also an average symmetry. An anomaly between exact and average symmetry can now be understood as a mixed anomaly between K^{\otimes N} and G (see Refs. 39 ; 40 for additional subtleties). Moreover, for the anomalies considered in this work such as the average LSM, the anomaly becomes trivial if we restrict to the K_{\rm diag}\times G subgroup – namely if the exact symmetry becomes average. Treating the replicated theory as an ordinary field theory, we apply the “power-law rule” which demands some charged local operators to have nontrivial correlation functions. For the average symmetry G , we can simply consider a field \phi^{A}_{a,\gamma} in any replica index a that transforms nontrivially under G with charge \gamma . The replica-symmetrized version is simply \sum_{a}\phi^{A}_{a} . The corresponding correlation function is \displaystyle\lim_{N\to 0}\langle\sum_{a}\phi^{A}_{a,\gamma}\sum_{b}\phi^{A}_{b,-\gamma}\rangle=\overline{\langle\phi^{A}_{\gamma}\phi^{A}_{-\gamma}\rangle}, (12) which is just the ordinary first-moment correlation function. Now for the exact symmetry K , an obvious candidate of charged operator is an operator \phi^{E}_{a,\kappa} that is charged under K^{a} in any replica index a with charge \kappa . However, this operator is also charged under K_{\rm diag} , the average part of K . Since the anomaly is trivial when K^{\otimes N} is restricted to K_{\rm diag} , this additional “average charge” of \phi_{a}^{E} may cause short-range correlation. Instead, we should consider operators that are charged under K^{\otimes N} but not K_{\rm diag} . The simplest choice is \phi^{E}_{a,\kappa}\phi^{E}_{b,-\kappa} for a\neq b , or the replica-symmetrized version \sum_{a\neq b}\phi^{E}_{a,\kappa}\phi^{E}_{b,-\kappa} . The corresponding correlation function is \displaystyle\lim_{N\to 0}\langle\sum_{a\neq b,a^{\prime}\neq b^{\prime}}\phi^{E}_{a,\kappa}\phi^{E}_{b,\kappa^{\prime}}\phi^{E}_{a^{\prime},-\kappa}\phi^{E}_{b^{\prime},-\kappa^{\prime}}\rangle=\overline{\left|\langle\phi^{E}_{\kappa}\phi^{E}_{-\kappa}\rangle\right|^{2}}. (13) which is the Edwards–Anderson correlator. II Further analysis in the representative examples II.1 Anomaly constrained correlations II.1.1 Strong disorder RG for random antiferromagnetic Heisenberg chain To extend our results for the random antiferromagnetic Heisenberg chain to larger system sizes, we use techniques based on the strong-disorder renormalization group (SDRG). The SDRG approach encompasses several related schemes that we will discuss below. While the 1d and 2d Majorana models are also amenable to an SDRG analysis, these models are non-interacting, allowing us to explore sufficiently large system sizes through exact diagonalization. Below, we briefly describe the numerical SDRG protocol we used to study the random AFM spin-half Heisenberg chain. We begin with a lattice with even number of lattice sites, with a spin-half on each site. The starting Hamiltonian is H=\sum_{i}J_{i,i+1}S_{i}\cdot S_{i+1} . In the numerical implementation, we checked for both open and periodic boundary conditions, and found no qualitative difference in the correlations. The coupling on each bond is positive, J_{i,i+1}>0 . The SDRG framework proceeds as follows. Consider the strongest hopping element J_{n,n+1}\equiv\Omega . We treat \Omega/J_{ij} as a large parameter in the Hamiltonian, and decimate the strongest bond, by projecting the decimated pair of spins to their singlet state. This decimation generates a new bond between spins on site n-1 and n+1 , with an effective coupling that can be deduced from second order perturbation theory, \displaystyle H\to H+J_{n-1,n+1}^{\text{eff}}S_{n-1}\cdot S_{n+1} (14) \displaystyle J_{n-1,n+1}^{\text{eff}}=\frac{J_{n,n-1}J_{n,n+1}}{2\Omega}a>0, (21) within the fitting windows stated in the figure captions. The fitted exponent a>0 indicates self-averaging in the fixed-separation sense, since \mathrm{SNR}(r,L) increases with L when r is held fixed, while b>0 implies that at fixed L the SNR decreases with distance and large separations are statistically harder to resolve. The additional inequality b>a sharpens this statement. Along separations that scale with system size, r=\kappa L , one has \mathrm{SNR}(\kappa L,L)\sim\kappa^{-b}L^{a-b}\to 0\qquad(L\to\infty), (22) so the loss of SNR with distance outpaces the gain from increasing L . Equivalently, to maintain a fixed target SNR while probing separations that grow with L , one would need to keep r/L sufficiently small or increase the number of disorder realizations accordingly, rather than relying on larger L alone. Since \mathrm{SE}\propto N_{\mathrm{samp}}^{-1/2} , one has \mathrm{SNR}\propto\sqrt{N_{\mathrm{samp}}} at fixed (r,L) , so maintaining \mathrm{SNR}(\kappa L,L) at a fixed target value requires N_{\mathrm{samp}}\propto L^{2(b-a)} up to model-dependent prefactors. We find similar behavior for other power-law–decaying correlators of symmetry-charged operators that are not anomaly-constrained, such as the sample-connected dimer-dimer correlator also shown in Figs. 6 and 7 . (a) Spin-spin correlation, C_{\mathrm{SS}} [Eq. 8 ] (b) First-moment dimer-dimer correlation, D_{\mathrm{fm}} [Eq. 6 ] (c) Sample-connected dimer-dimer correlation, D_{\mathrm{sc}} [Eq. 18 ] Figure 6 : Signal-to-noise ratio (SNR) analysis for power-law–decaying correlators in the random antiferromagnetic Heisenberg chain, computed using strong-disorder RG with 10^{5} disorder realizations. Left panels show \mathrm{SNR} versus L at fixed separation r ; right panels show \mathrm{SNR} versus r at fixed system size L . The fit uses 20\leq L\leq 100 and 3\leq r\leq 40 , with \mathrm{SNR}\geq 5 . (a) Edwards–Anderson correlation, C_{\mathrm{EA}} [Eq. 3 ] (b) First-moment dimer-dimer correlation, D_{\mathrm{fm}} [Eq. 6 ] (c) Sample-connected dimer-dimer correlation, D_{\mathrm{sc}} [Eq. 18 ] Figure 7 : Signal-to-noise ratio (SNR) analysis for power-law–decaying correlators in the 1d random Majorana lattice, computed using exact diagonalization with 5000 disorder realizations. Left panels show \mathrm{SNR} versus L at fixed separation r ; right panels show \mathrm{SNR} versus r at fixed system size L . The fit uses 40\leq L\leq 500 and 3\leq r\leq 40 , with \mathrm{SNR}\geq 10 . For the 2d random Majorana lattice, Eq. ( 21 ) provides a reasonable baseline fit, but the residuals retain a systematic dependence on the aspect ratio r/L . This indicates that, at the system sizes accessible here, the approach to the asymptotic regime depends not only on r and L separately, but also on the finite-size ratio r/L . Rather than imposing an explicit r/L cutoff, which would introduce a direct bias towards a specific power-law scaling regime, we incorporate the leading finite-size correction as a multiplicative scaling function f(r/L) and fit \mathrm{SNR}(r,L)\sim\frac{L^{a}}{r^{b}}\,f(r/L),\qquad f(r/L)\approx\exp\!\big(\alpha\,r/L\big). (23) This modification captures the curvature seen in \log\mathrm{SNR} versus \log r at fixed L while preserving the expected large- L limit of the d=1 ansatz. At fixed r and L\to\infty , the ratio r/L\to 0 so f(r/L)\to 1 , and the scaling reduces to a pure power-law. Figure 8 shows that Eq. ( 23 ) provides a consistent description for the anomaly-constrained Edwards–Anderson and first-moment dimer-dimer correlators in the 2d Majorana lattice model, and it also fits the sample-connected dimer-dimer correlator, with fit parameters listed in the figure. We note that in our 2d results, the first-moment dimer-dimer correlator exhibits b>a in Fig. 8 (b), so along separations that scale with system size, such as r=\kappa L , one expects \mathrm{SNR}(\kappa L,L)\sim\kappa^{-b}L^{a-b} and hence a decreasing SNR with increasing L along r\propto L . By contrast, the Edwards–Anderson correlator in Fig. 8 (a) yields ba is not a uniform feature of all d=2 observables at the sizes accessible here. These behaviors of the SNR motivates keeping L_{\max}\approx 50 in the d=2 analysis, since pushing to substantially larger L while maintaining comparable statistical control at large separations would require a much larger number of disorder realizations. The sign of \alpha is informative about the leading finite-size drift with c=r/L . For the Edwards–Anderson correlator and the sample-connected dimer-dimer correlator we find \alpha<0 [Fig. 8 (a,c)], indicating an additional suppression of the SNR as one probes separations that are a larger fraction of the system size, beyond what is captured by the baseline L^{a}/r^{b} scaling. This effect is especially pronounced for D_{\mathrm{sc}} , where the large negative \alpha shows that the SNR deteriorates rapidly as r/L increases. For the first-moment dimer-dimer correlator we find \alpha>0 [Fig. 8 (b)], so the r/L -dependent correction bends in the opposite direction and partially compensates that loss over the fitting window. Overall, the SNR analysis serves as a quantitative sanity check on the regime in which disorder-averaged correlators can be reliably interpreted as exhibiting power-law scaling, given the system sizes and disorder ensembles accessible in our numerics. (a) Edwards–Anderson correlation, C_{\mathrm{EA}} [Eq. 3 ] (b) First-moment dimer-dimer correlation, {D}_{\mathrm{fm}} [Eq. 16 ] (c) Sample-connected dimer-dimer correlations, D_{\mathrm{sc}} [Eq. 18 ] Figure 8 : Signal-to-noise ratio (SNR) analysis for power-law–decaying correlators in the 2d random Majorana lattice, computed using exact diagonalization with 4000 disorder realizations. Left panels show \mathrm{SNR} versus L at fixed separation r ; right panels show \mathrm{SNR} versus r at fixed system size L . The fit uses 12\leq L\leq 52 and 4\leq r\leq 24 , with \mathrm{SNR}\geq 3 . IV Strong disorder RG argument for the dimer-dimer correlation functions In this section we present an SDRG argument that supports our numerical results for the dimer-dimer correlations in the 1 d random-singlet phase of the Majorana or spin models we study. As discussed in the main text, we construct dimer operators that are charged (i.e. their expectation values transform by a sign) under the average translation symmetry. It is natural to parametrize the dimer operator as d_{i}\;\approx\;(-1)^{i}\,\varepsilon_{i}, (24) where the expectation value \overline{\langle d_{i}\rangle} vanishes, and \varepsilon_{i} is translation-invariant. This makes it clear that any long-distance connected dimer-dimer correlation function has the asymptotic form \bigl\langle d_{i}d_{i+r}\bigr\rangle_{\rm conn}\;\sim\;(-1)^{r}F(r), (25) with a real function F(r) that (in the examples we study) decays with r . In what follows we use SDRG to determine this power-law decay for systems of random-singlet type. In the SDRG procedure, at each step we identify the currently strongest bond and denote its coupling by \Omega . We decimate this bond by forming a singlet and removing the two spins from the low-energy Hilbert space. As the RG proceeds, more spins are eliminated and the remaining active sites become increasingly sparse. If the strongest bond at the n th decimation step is \Omega_{n} , we parametrize the RG flow by the scale \Gamma=-\frac{\log\Omega}{\log\Omega_{n}} . If we start at the random-singlet fixed point, the arguments of Fisher show that the density of surviving (active) spins at RG scale \Gamma behaves as n(\Gamma)\sim\Gamma^{-2} . 16 . So if the chain initially has N_{0} sites, the number of active spins is N(\Gamma)\sim N_{0}/\Gamma^{2} . The probability that a site is active at scale \Gamma is therefore p_{\text{surv}}(\Gamma)\;=\;\frac{N(\Gamma)}{N_{0}}\;\sim\;\Gamma^{-2}, (26) The same fixed-point solution implies that the typical spacing between active spins grows as \ell(\Gamma)\sim\Gamma^{2} . To probe correlations at separation r we should thus consider the RG scale \Gamma_{r} where \ell(\Gamma_{r})\sim r , i.e. \Gamma_{r}^{2}\sim r\quad\Rightarrow\quad\Gamma_{r}\sim\sqrt{r}, (27) Inserting this into Eq. ( 26 ) gives p_{\text{surv}}(\Gamma_{r})\sim\Gamma_{r}^{-2}\sim r^{-1}, (28) so each site has probability of order 1/r to remain active up to the stage at which singlets of length \sim r are being formed. From this perspective there are three qualitatively distinct ways the SDRG flow can treat the two dimers at (i,i+1) and (i+r,i+r+1) : (a) the most common configuration is that all four spins in (i,i+1) and (i+r,i+r+1) are decimated separately at early RG scales, forming short singlets with nearby neighbours; by the time singlets of length \sim r appear, there is no active degree of freedom left in either dimer region, and the two bonds are effectively independent. Since each site survives up to scale \Gamma_{r} only with probability p_{\text{surv}}(\Gamma_{r})\sim 1/r , the probability that any of the four spins is still active at that scale is of order 1/r . Thus, the probability that all four have already been decimated is P_{(a)}(r)=1-\mathcal{O}\!\left(\frac{1}{r}\right), so this configuration occurs with probability of order 1, P_{(a)}(r)\approx 1 , for large r . (b) More rarely, one effective spin from the vicinity of (i,i+1) and one from (i+r,i+r+1) both survive until they become nearest neighbours among the active sites and form a single long singlet of length \sim r ; this requires two specific sites to survive and therefore occurs with probability P_{(b)}(r)\sim p_{\text{surv}}(\Gamma_{r})^{2}\sim r^{-2} . (c) Even more rarely, all four spins associated with the two dimers (or their RG descendants) survive to scale \Gamma_{r} and form two distinct long singlets spanning the two dimer regions; this requires four sites to survive and occurs with probability P_{(c)}(r)\sim p_{\text{surv}}(\Gamma_{r})^{4}\sim r^{-4} . We first consider the disorder-averaged dimer-dimer correlation \overline{\langle d_{i}d_{i+r}\rangle} . In the SDRG picture this average can be expressed as a weighted sum over the three configurations (a)–(c), \overline{\langle d_{i}d_{i+r}\rangle}=\sum_{k=a,b,c}P_{(k)}(r)\,\mathcal{D}_{(k)}, (29) where P_{(k)}(r) is the probability of case (k) and \mathcal{D}_{(k)} is the corresponding (order–one) value of \langle d_{i}d_{i+r}\rangle in that configuration. In the typical configuration (a) both dimers are short-range bonds, so \langle d_{i}d_{i+r}\rangle is controlled by local physics in the two dimer regions and is independent of r once it is large. The rarer configurations (b) and (c) shift this local value only by an amount of order one, but their contributions are suppressed by their small probabilities, P_{(b)}(r)\sim r^{-2} and P_{(c)}(r)\sim r^{-4} . As a result, \overline{\langle d_{i}d_{i+r}\rangle}=\text{const.}+\mathcal{O}\!\left(\frac{1}{r^{2}}\right), (30) and the unconnected disorder average is dominated by an r -independent background; the staggered, (-1)^{r} –oscillating piece only appears in the subleading corrections. The behavior changes once we consider the first-moment dimer-dimer correlator D_{\mathrm{fm}}(r) defined in Eq. ( 6 ). For the SDRG estimate it is convenient to focus on a representative pair of dimers at separation r , for which Eq. ( 6 ) reduces to D_{\rm fm}(r)=\left(\overline{\langle d_{i}d_{i+r}\rangle}-\overline{\langle d_{i}\rangle}\,\overline{\langle d_{i+r}\rangle}\right)(-1)^{r}, (31) By translation invariance the disorder-averaged one-point functions are site independent, \overline{\langle d_{i}\rangle}\approx\overline{\langle d_{i+r}\rangle}\equiv m , so D_{\mathrm{fm}}(r)=\sum_{k=a,b,c}P_{(k)}(r)\,\mathcal{D}_{(k)}-m^{2}=\sum_{k=a,b,c}P_{(k)}(r)\,\Delta_{(k)}, (32) with \Delta_{(k)}\equiv\mathcal{D}_{(k)}-m^{2} . Thus D_{\text{fm}}(r) is still a first-moment quantity over the three SDRG configurations, but with shifted weights \Delta_{(k)} . For the typical case (a) the two dimer regions are effectively independent, so \mathcal{D}_{(a)}\approx m^{2} and hence \Delta_{(a)}\approx 0 ; meaning that the large r -independent contribution that dominated \overline{\langle d_{i}d_{i+r}\rangle} is removed by the subtraction. The remaining contribution is governed by the rare cases (b) and (c), for which \Delta_{(b)},\Delta_{(c)}=\mathcal{O}(1) while P_{(b)}(r)\sim r^{-2} and P_{(c)}(r)\sim r^{-4} . In case (b) the long singlet that connects the two dimer regions couples to the staggered piece (-1)^{i}\varepsilon_{i} , so the surviving contribution is staggered and decays with the probability P_{(b)}(r) . Combining this with the general structure discussed above, we obtain D_{\mathrm{fm}}(r)\sim\frac{A}{r^{2}}+\mathcal{O}\!\left(\frac{1}{r^{4}}\right), (33) Finally, we consider the sample-connected dimer-dimer correlator D_{\mathrm{sc}}(r) defined in Eq. ( 18 ). Again focusing on a representative pair of dimers at separation r , Eq. ( 18 ) can be written as D_{\mathrm{sc}}(r)=\overline{\Bigl(\langle d_{j}d_{j+r}\rangle-\langle d_{j}\rangle\langle d_{j+r}\rangle\Bigr)}(-1)^{r}, (34) In case (a) the two dimer regions are independent within each realization, so \langle d_{i}d_{i+r}\rangle\approx\langle d_{i}\rangle\langle d_{i+r}\rangle and the sample-connected correlation is essentially zero. In case (b) a single long singlet connects the two regions; both d_{i} and d_{i+r} are controlled by the same two-spin state, and to leading order \langle d_{i}d_{i+r}\rangle again factorizes into \langle d_{i}\rangle\langle d_{i+r}\rangle , so this case also does not contribute at leading order. The minimal configuration that produces a nonzero sample-connected contribution is case (c), with two distinct long singlets of length \sim r spanning the two dimers. In this situation d_{i}d_{i+r} contains cross-terms in which one factor comes from singlet #1 and the other from singlet #2; these are genuine four-spin contributions that cannot be written as \langle d_{i}\rangle\langle d_{i+r}\rangle , and hence survive in the sample-connected combination with an \mathcal{O}(1) value. Since such configurations occur with probability P_{(c)}(r)\sim r^{-4} and inherit the momentum- \pi charge of the dimer operator, the sample-connected correlation function decays as D_{\text{sc}}(r)\sim\frac{B}{r^{4}}. (35) in agreement with the numerical results.