Bond-DMRG: Optimizing Virtual Bond Spaces

• Bond-DMRG is a numerical method that controls entanglement in matrix product states by regulating virtual bond spaces through adjustable dimensions and truncation.
• It employs symmetry-adapted bases and controlled bond expansion to optimize tensor operations, reducing computational cost while enhancing accuracy.
• Recent advancements extend Bond-DMRG to dynamic and two-dimensional systems with improved convergence via extrapolation and hardware-aware implementations.

Searching arXiv for recent and foundational papers on Bond-DMRG and related bond-dimension–controlled DMRG formulations. Bond-DMRG denotes a class of density-matrix renormalization group formulations in which the central numerical object is the virtual, or bond, space of a matrix product state (MPS): accuracy is controlled by bond dimensions, discarded Schmidt weight, and algorithmic rules for truncating or expanding bond spaces. In the literature considered here, the term also appears in work on bond dimerization, bond order, and molecular bond dissociation, but the common technical core is an MPS ansatz whose entanglement capacity is regulated at each bond cut (Wouters et al., 2013, 2207.14712, Ejima et al., 2017).

1. MPS bond spaces as the organizing principle

For an $L$-site problem, the MPS ansatz is written as

$$\ket{\Psi} = \sum_{\{n_k\},\{\alpha_j\}} A[1]^{n_1}_{\alpha_1} A[2]^{n_2}_{\alpha_1;\alpha_2} \cdots A[L]^{n_L}_{\alpha_{L-1}} \ket{n_1 n_2\ldots n_L},$$

where the $\alpha_j$ are the bond, or virtual, indices. In the non-symmetry-adapted picture, their dimension is denoted $D$; increasing $D$ enlarges the reachable portion of Hilbert space, subject to the trivial bound

$D_j \le \min(4^j,4^{L-j})$

in the quantum-chemistry occupation basis (Wouters et al., 2013).

In two-site DMRG, two neighboring site tensors are combined into a two-site object $B[i]$, optimized through an effective Hamiltonian eigenproblem,

$$\mathbf{H}^{\mathrm{eff}}\,\mathbf{B}[i] = \lambda\,\mathbf{B}[i],$$

and then factorized by SVD. Bond truncation enters at exactly this step: one keeps only the largest $D_i$ singular values. The discarded Schmidt weight at bond $i$ is

$$\ket{\Psi} = \sum_{\{n_k\},\{\alpha_j\}} A[1]^{n_1}_{\alpha_1} A[2]^{n_2}_{\alpha_1;\alpha_2} \cdots A[L]^{n_L}_{\alpha_{L-1}} \ket{n_1 n_2\ldots n_L},$$0

and the maximal discarded weight over a sweep,

$$\ket{\Psi} = \sum_{\{n_k\},\{\alpha_j\}} A[1]^{n_1}_{\alpha_1} A[2]^{n_2}_{\alpha_1;\alpha_2} \cdots A[L]^{n_L}_{\alpha_{L-1}} \ket{n_1 n_2\ldots n_L},$$1

is used as an entanglement-based error measure. For variational energies, a standard extrapolation is

$$\ket{\Psi} = \sum_{\{n_k\},\{\alpha_j\}} A[1]^{n_1}_{\alpha_1} A[2]^{n_2}_{\alpha_1;\alpha_2} \cdots A[L]^{n_L}_{\alpha_{L-1}} \ket{n_1 n_2\ldots n_L},$$2

so that increasing bond dimension lowers discarded weight and tightens the variational upper bound (Wouters et al., 2013).

The same logic appears in dynamical tensor-network work, where the bond dimension $$\ket{\Psi} = \sum_{\{n_k\},\{\alpha_j\}} A[1]^{n_1}_{\alpha_1} A[2]^{n_2}_{\alpha_1;\alpha_2} \cdots A[L]^{n_L}_{\alpha_{L-1}} \ket{n_1 n_2\ldots n_L},$$3 is the dimension of the virtual indices $$\ket{\Psi} = \sum_{\{n_k\},\{\alpha_j\}} A[1]^{n_1}_{\alpha_1} A[2]^{n_2}_{\alpha_1;\alpha_2} \cdots A[L]^{n_L}_{\alpha_{L-1}} \ket{n_1 n_2\ldots n_L},$$4. Compression across a bond is governed by the singular values $$\ket{\Psi} = \sum_{\{n_k\},\{\alpha_j\}} A[1]^{n_1}_{\alpha_1} A[2]^{n_2}_{\alpha_1;\alpha_2} \cdots A[L]^{n_L}_{\alpha_{L-1}} \ket{n_1 n_2\ldots n_L},$$5, with truncation error

$$\ket{\Psi} = \sum_{\{n_k\},\{\alpha_j\}} A[1]^{n_1}_{\alpha_1} A[2]^{n_2}_{\alpha_1;\alpha_2} \cdots A[L]^{n_L}_{\alpha_{L-1}} \ket{n_1 n_2\ldots n_L},$$6

and the same singular values determine the bipartite von Neumann entropy

$$\ket{\Psi} = \sum_{\{n_k\},\{\alpha_j\}} A[1]^{n_1}_{\alpha_1} A[2]^{n_2}_{\alpha_1;\alpha_2} \cdots A[L]^{n_L}_{\alpha_{L-1}} \ket{n_1 n_2\ldots n_L},$$7

At fixed bond dimension $$\ket{\Psi} = \sum_{\{n_k\},\{\alpha_j\}} A[1]^{n_1}_{\alpha_1} A[2]^{n_2}_{\alpha_1;\alpha_2} \cdots A[L]^{n_L}_{\alpha_{L-1}} \ket{n_1 n_2\ldots n_L},$$8, the maximal entropy across a cut is $$\ket{\Psi} = \sum_{\{n_k\},\{\alpha_j\}} A[1]^{n_1}_{\alpha_1} A[2]^{n_2}_{\alpha_1;\alpha_2} \cdots A[L]^{n_L}_{\alpha_{L-1}} \ket{n_1 n_2\ldots n_L},$$9, so bond-DMRG is, at base, an entanglement-budget formalism (Li et al., 30 Aug 2025).

2. Symmetry-adapted bond dimensions in ab initio quantum chemistry

In ab initio quantum chemistry, the most efficient DMRG implementations do not use a bare bond dimension $\alpha_j$0, but a symmetry-refined bond space. CheMPS2 enforces $\alpha_j$1 total spin, $\alpha_j$2 particle number, and an abelian point group $\alpha_j$3, with virtual states labeled

$\alpha_j$4

The reduced bond dimension is then

$\alpha_j$5

where $\alpha_j$6 is the multiplicity of the symmetry sector $\alpha_j$7 (Wouters et al., 2013).

This refinement matters because a single spin-$\alpha_j$8 multiplet corresponds to $\alpha_j$9 states in a non-spin-adapted MPS. CheMPS2 therefore counts retained reduced multiplets rather than raw states. In the symmetry-adapted two-site algorithm, the optimized object is a reduced tensor $D$0, and truncation keeps the $D$1 largest reduced Schmidt multiplets. Bond dimension thus becomes the number of reduced renormalized basis states rather than the number of non-adapted basis states (Wouters et al., 2013).

The underlying efficiency gain comes from Wigner–Eckart factorization. Each MPS tensor is treated as an irreducible tensor operator of $D$2, so a tensor factorizes into Clebsch–Gordan coefficients, Kronecker selection rules, and a reduced tensor. This induces sparse block structure in both the MPS and the renormalized operators, and it is precisely this block-sparse bond structure that makes large-$D$3 calculations practical (Wouters et al., 2013).

In molecular applications, the same bond-control logic is used operationally. For the carbon dimer, CheMPS2 employed a staged convergence scheme with

$D$4

alternating noise-on and noise-off instructions and $D$5. For the reported C$D$6 curves, the variational energies were converged to $D$7 from the extrapolated value, and the final extrapolated energies to better than $D$8 (Wouters et al., 2013).

3. Controlled bond expansion and the bond-space optimization debate

A more explicit use of bond-DMRG appears in controlled bond expansion (CBE). In this formulation, the point is not merely to truncate bond spaces but to expand them selectively. CBE identifies parts of the orthogonal space carrying significant weight in $D$9 and expands bonds to include only those directions. The original formulation states that CBE yields two-site accuracy and convergence per sweep at single-site costs, with the local optimization remaining fully variational (2207.14712).

The algorithm is organized around the decomposition of the two-site variational space into kept and discarded sectors. Instead of performing a full two-site optimization, CBE projects $D$0 into the relevant orthogonal bond subspace and uses this information to construct a small set of expansion vectors. The expanded one-site problem is then solved at a cost comparable to single-site DMRG, rather than the higher cost of full two-site DMRG (2207.14712).

This formulation became the subject of a technical dispute. A detailed comment argued that the projection onto the two-site tangent space is unnecessary and generally not helpful; that the sequence of five SVDs used in CBE can be replaced by a single $D$1 decomposition, optionally with one SVD, using randomized SVD; that the bond-expansion step can then scale as

$D$2

rather than

$D$3

and that the variational properties of CBE are essentially identical to those of existing schemes such as two-site DMRG and single-site subspace expansion, because truncation back to bond dimension $D$4 breaks strict hill-climbing behavior at the micro-step level (McCulloch et al., 2024).

A reply accepted randomized SVD as a promising refinement and explicitly clarified that truncation breaks strict variationality in the strong sense, but it strongly defended the two-site tangent-space projection, especially in TDVP applications. In that response, omitting the projection was argued to cause avoidable errors by spending part of the limited expansion rank on directions already in the one-site tangent space. The same reply emphasized the complementary roles of 3S mixing and CBE, and recommended a combined CBE+$D$5 strategy for efficiency and robustness (Gleis et al., 21 Jan 2025).

This exchange established a durable distinction within bond-DMRG methodology. One position emphasizes randomized linear algebra and minimal projection structure for efficient bond-space enlargement; the other emphasizes tangent-space geometry and bond-expansion directions that are explicitly orthogonal to the one-site manifold. Both agree that bond expansion is the central mechanism by which single-site-cost methods recover part of the flexibility of two-site updates (McCulloch et al., 2024, Gleis et al., 21 Jan 2025).

4. Dynamical bond-DMRG and finite-bond extrapolation

In time-dependent and frequency-domain DMRG, bond dimension controls not only ground-state accuracy but also the fidelity of response vectors and time-evolved states. In ab initio dynamics, standard DDMRG and td-DMRG employ state-averaged renormalized bases, so a fixed bond dimension $D$6 is shared among ground, operator-applied, correction-vector, or time-propagated states. The DDMRG$D$7 and td-DMRG$D$8 formulations replace that broad state averaging with more state-specific MPS representations, and the reported outcome is increased accuracy at the same bond dimension, at a nominal increase in cost (Ronca et al., 2017).

For H$D$9, DDMRG$D_j \le \min(4^j,4^{L-j})$0 at bond dimension $D_j \le \min(4^j,4^{L-j})$1 yielded an LDOS essentially indistinguishable from FCI, whereas td-DMRG$D_j \le \min(4^j,4^{L-j})$2 required $D_j \le \min(4^j,4^{L-j})$3–100 for comparable quality. The same work interpreted the improvement as an effectively better use of finite bond dimension, because the ground state and dynamical states no longer compete within a single state-averaged basis (Ronca et al., 2017).

A later benchmark on the P3HT:PCBM heterojunction model made the bond-dimension dependence explicit. That study used a unified software framework, Renormalizer, for both MPS and TTNS, and concluded that the previously reported discrepancy of up to 60% between TD-DMRG and ML-MCTDH arose primarily from insufficient bond dimensions. By increasing bond dimensions, the difference was reduced to less than 10%; with extrapolation and an optimized tensor-network structure, it was lowered to approximately 2% (Li et al., 30 Aug 2025).

That benchmark also formalized a practical convergence protocol. A uniform fixed bond dimension $D_j \le \min(4^j,4^{L-j})$4 was used on every bond; observables were examined as functions of $D_j \le \min(4^j,4^{L-j})$5; and linear extrapolation in $D_j \le \min(4^j,4^{L-j})$6 was applied only in the asymptotic large-$D_j \le \min(4^j,4^{L-j})$7 regime. The same study reported that a tailored tree topology, TreeX, reduced the maximum entanglement entropy across bonds relative to both a simple MPS and a reference tree, thereby increasing effective entanglement capacity per unit bond dimension (Li et al., 30 Aug 2025).

5. Large bond dimensions in two dimensions

In two spatial dimensions, bond-DMRG becomes an exercise in managing the exponential growth of required MPS entanglement capacity. For square lattices of size $D_j \le \min(4^j,4^{L-j})$8, the required bond dimension for an area-law state scales at least as

$D_j \le \min(4^j,4^{L-j})$9

and the leading DMRG cost scales as $B[i]$0. This scaling was demonstrated explicitly in large-scale TPU implementations, which reached bond dimension

$B[i]$1

on $B[i]$2 free fermions and $B[i]$3 transverse-field Ising systems; optimizing a single MPS tensor at that $B[i]$4 took about 2 minutes on 1,024 TPU v3 cores (Ganahl et al., 2022).

The same bond-dimension logic governs high-accuracy studies of competing two-dimensional ordered phases. On the triangular-lattice Heisenberg antiferromagnet, large-scale DMRG with bond dimension up to $B[i]$5 and width up to $B[i]$6 produced the benchmark estimates

$B[i]$7

after linear extrapolation in truncation error and finite-size scaling (Huang et al., 2023).

In the square-lattice $B[i]$8–$B[i]$9 Heisenberg model, conventional DMRG was used up to $$\mathbf{H}^{\mathrm{eff}}\,\mathbf{B}[i] = \lambda\,\mathbf{B}[i],$$0, while Fully Augmented Matrix Product States (FAMPS) reached $$\mathbf{H}^{\mathrm{eff}}\,\mathbf{B}[i] = \lambda\,\mathbf{B}[i],$$1 on $$\mathbf{H}^{\mathrm{eff}}\,\mathbf{B}[i] = \lambda\,\mathbf{B}[i],$$2 systems. FAMPS introduces an additional layer of disentanglers and retains cost scaling

$$\mathbf{H}^{\mathrm{eff}}\,\mathbf{B}[i] = \lambda\,\mathbf{B}[i],$$3

while empirically giving energies comparable to ordinary DMRG at approximately three times larger bond dimension. With truncation-error extrapolation, finite-size scaling, and finite-$$\mathbf{H}^{\mathrm{eff}}\,\mathbf{B}[i] = \lambda\,\mathbf{B}[i],$$4 scaling, that study identified the VBS phase as plaquette rather than columnar (Huang et al., 2024).

These examples show that, in two dimensions, bond-DMRG is not a narrow algorithmic variant but a regime of practice defined by extremely large $$\mathbf{H}^{\mathrm{eff}}\,\mathbf{B}[i] = \lambda\,\mathbf{B}[i],$$5, careful control of truncation error, and often hardware-aware implementations or augmented MPS ansätze (Ganahl et al., 2022, Huang et al., 2024, Huang et al., 2023).

6. Virtual bonds and physical bonds

A persistent source of ambiguity is the difference between MPS bonds and physical bonds. In the quantum-chemistry formulation of CheMPS2, this distinction is explicit: MPS bonds are virtual indices of dimension $$\mathbf{H}^{\mathrm{eff}}\,\mathbf{B}[i] = \lambda\,\mathbf{B}[i],$$6 or $$\mathbf{H}^{\mathrm{eff}}\,\mathbf{B}[i] = \lambda\,\mathbf{B}[i],$$7, whereas physical chemical bonds, such as the C–C bond in C$$\mathbf{H}^{\mathrm{eff}}\,\mathbf{B}[i] = \lambda\,\mathbf{B}[i],$$8, are unrelated except indirectly through correlation and entanglement. Near bond breaking, entanglement grows, and larger MPS bond dimensions are needed; but the two notions of bond are not identical (Wouters et al., 2013).

In one-dimensional correlated-electron models with explicit bond dimerization, by contrast, physical bond modulation is built directly into the Hamiltonian. For the half-filled extended Hubbard model with dimerized hopping,

$$\mathbf{H}^{\mathrm{eff}}\,\mathbf{B}[i] = \lambda\,\mathbf{B}[i],$$9

DMRG and iDMRG use a natural two-site unit cell. There, bond dimerization replaces the SDW/BOW region of the pure EHM by a symmetry-protected-topological Peierls insulator, and the direct SPT–CDW transition is Ising-like with $D_i$0 at weak coupling, reaches a tricritical Ising point with $D_i$1, and becomes first order beyond that (Ejima et al., 2017).

A related but distinct usage occurs in frustrated spin chains. For the zigzag spin-$D_i$2 chain with first- and second-neighbor antiferromagnetic exchange, a modified DMRG algorithm that adds four sites per iteration was introduced because the dominant $D_i$3 bonds are long-range in the usual site ordering. That modified algorithm yielded accurate results up to $D_i$4 for the magnetic gap $D_i$5, the BOW amplitude $D_i$6, the spiral wavelength $D_i$7, and the spin correlation length $D_i$8 (Kumar et al., 2010).

In non-Born–Oppenheimer molecular models, bond-DMRG may refer primarily to the physical bonding problem while still being numerically organized around MPS bond control. A one-dimensional diatomic molecule with quantum electrons and quantum nuclei was mapped to a two-leg ladder and treated with a three-site DMRG algorithm plus compressed long-range interactions. In that setting, DMRG resolved binding and unbinding as functions of nuclear mass ratio and spin sector, while the computational challenge remained the management of entanglement and truncation on the MPS bonds (Yang et al., 2018).

7. Limits, exact low-rank cases, and hybrid extensions

Bond-DMRG also has rigorous low-rank and hybrid-algorithm extensions. For two-electron wavefunctions, it was proved that there exists an orbital basis in which any state is an MPS with maximal bond dimension 3, and that 3 is optimal in the generic case. A direct consequence is that QC-DMRG with bond dimension 3, combined with fermionic mode optimization, exactly recovers the FCI energy for two-electron systems (Friesecke et al., 2021). This is an exceptional case, but it shows that bond dimension can sometimes be characterized exactly rather than only empirically.

At the opposite extreme, recent hybrid quantum-classical work treats the central bond itself as the key diagnostic. In a DMRG-guided MPS encoding protocol for quantum ground-state preparation, the central-bond Schmidt rank of intermediate states was found to grow logistically with the number of encoding layers. Its inflection point $D_i$9 marks the boundary of the efficient encoding regime, and beyond that point the additional number of encoding layers required to reach target infidelity $i$0 empirically scales as

$i$1

That observation motivates terminating encoding at $i$2 and using probabilistic imaginary-time evolution to remove the remaining excited-state weight (Watanabe et al., 28 May 2026).

Taken together, these developments delimit the scope of bond-DMRG. In ordinary many-body problems, bond dimension is a tunable numerical resource whose adequacy must be assessed by discarded weight, truncation-error extrapolation, or convergence across multiple ansätze. In special sectors, such as two-electron quantum chemistry, the required bond rank can be determined exactly. And in hybrid settings, central-bond diagnostics can be elevated from a convergence measure to a design principle for quantum algorithms (Friesecke et al., 2021, Watanabe et al., 28 May 2026).