Matrix Product States and Bond-Dimension Limits

A matrix product state (MPS) takes the $2^n$ amplitudes of an $n$ -qubit wavefunction — an object too large to write down past forty qubits — and factorizes them into a chain of small tensors, one per site, contracted along a one-dimensional line. Each tensor carries one physical leg (the qubit it lives on) and up to two virtual "bond" legs stitching it to its neighbours. The dimension of those bond legs, the bond dimension $\chi$ , is the one knob that controls everything: it caps the Schmidt rank across every cut, and through it the entanglement entropy the state can hold, $S\le\log_2\chi$ . This makes the MPS the cleanest example anywhere of a representation whose memory cost and physical entanglement are the same number. The decisive fact that follows is structural: low entanglement means cheap representation. A state whose every bipartite cut carries bounded entanglement — an area law, in one dimension — can be captured to fixed accuracy by an MPS whose bond dimension grows only polynomially in system size. Hastings' theorem guarantees exactly this for the ground states of gapped local 1D Hamiltonians, which is why the density-matrix renormalization group (DMRG), read as variational optimization over MPS, is the most accurate numerical method known for 1D quantum matter. The converse is just as sharp: volume-law states — almost all of Hilbert space, and the generic output of a deep circuit — demand a bond dimension exponential in $n$ . For quantum computing the bond dimension plays a double role: it is the dial that says which circuits a laptop can secretly reproduce, and the expressive budget of a whole family of variational ansätze. This article makes $\chi$ the protagonist, following it from the SVD that builds an MPS, through the entanglement ceiling it enforces, to the area laws and circuit-depth crossovers that decide what is classically hard.

Key takeaways

A matrix product state writes the $2^n$ amplitudes of an $n$ -qubit state as a chain of contracted tensors, $c_{s_1\cdots s_n}=\mathrm{Tr}(A^{s_1}A^{s_2}\cdots A^{s_n})$ , storing $O(n\,d\,\chi^2)$ numbers instead of an exponential.
The bond dimension $\chi$ equals the largest Schmidt rank across any cut, so it sets a hard ceiling on the entanglement of every cut: $S\le\log_2\chi$ .
Ground states of gapped local 1D Hamiltonians obey an area law, $S=O(1)$ , so $\chi=\mathrm{poly}(n)$ suffices — the theorem underneath DMRG's success.
In canonical (Vidal) gauge the bond singular values are the Schmidt coefficients; dropping the smallest is the 2-norm-optimal truncation, with error equal to the discarded weight $\sum_{\alpha>\chi'}\lambda_\alpha^2$ .
Volume-law states — generic states, deep random circuits — need $\chi$ exponential in $n$ , so MPS reach only the thin, low-entanglement corner of Hilbert space.
The same object is a classical-simulation engine (TEBD), a diagnostic of quantum advantage, and a variational ansatz (sequential circuits realize an MPS exactly).
Key idea: one number, the bond dimension $\chi$ , simultaneously bounds the entanglement across every cut ( $S\le\log_2\chi$ ) and the classical cost ( $O(n\,d\,\chi^2)$ storage, $O(n\chi^3)$ per sweep); area-law states need only $\chi=\mathrm{poly}(n)$ , volume-law states need $\chi$ exponential.

The core insight

One number, the bond dimension $\chi$ , simultaneously bounds the entanglement across every cut ( $S\le\log_2\chi$ ) and the classical cost ( $O(n\,d\,\chi^2)$ storage, $O(n\chi^3)$ per sweep); area-law states need only $\chi=\mathrm{poly}(n)$ , while volume-law states need $\chi$ exponential in $n$ .

Where this lives today

The bond dimension is the working currency of two live frontiers. In quantum many-body physics, DMRG and its time-dependent cousins (TEBD, TDVP) remain the gold standard for 1D and quasi-1D systems, with modern codes pushing $\chi$ into the tens of thousands to resolve frustrated magnets, Hubbard ladders, and lattice gauge theories, and the entanglement spectrum read off a canonical MPS is now a standard probe of symmetry-protected topological order. In quantum computing the same number referees claims of quantum advantage: whenever a sampling experiment is announced beyond classical reach, tensor-network teams ask what bond dimension an MPS or 2D contraction would need to reproduce the output — and several early "supremacy" claims were narrowed exactly this way, the circuits carrying less usable entanglement than their qubit count suggested. On the hardware side, sequential (staircase) circuits prepare an exact MPS whose bond dimension is set by the number of ancillas, a variational ansatz whose expressibility is dialed by $\chi$ rather than blind depth, warm-startable from DMRG and free of some barren-plateau pathologies. The open frontier is dimensionality: the clean $S\le\log_2\chi$ story is strictly 1D, and extending controllable, contractible tensor networks to two and three dimensions (PEPS, with their contraction-hardness) is where the hardest problems sit.

The physical picture

Picture the full amplitude tensor of an $n$ -qubit state as one object with $2^n$ entries, and unzip it one site at a time. At each cut you split the tensor into a left and a right half and take its singular value decomposition; the singular values are the Schmidt coefficients of that cut, and the number of nonzero ones is the bond dimension there. Keep going until the whole tensor is a chain of small matrices — that chain is the MPS, exact if you keep every singular value, a controlled approximation if you keep only the largest $\chi$ . The first figure is this unzipping made literal.

Figure

A many-body amplitude tensor with 2ⁿ entries is unzipped by iterated SVD into a chain of small tensors; each bond carries at most χ Schmidt values, so the entropy of any cut obeys S ≤ log₂ χ

Think of $\chi$ as the width of an information pipe running along the chain. Every quantum correlation connecting the left half of the system to the right half must squeeze through a $\chi$ -dimensional channel at the cut, so it can carry at most $\log_2\chi$ bits of entanglement entropy. A gapped 1D ground state only needs a thin pipe of constant width; a scrambled, volume-law state needs one whose width grows exponentially. Each on-site tensor $A^{s_i}$ acts like a transfer matrix selected by the local qubit value $s_i$ , and because products of contracted transfer matrices decay with the ratio of their two dominant eigenvalues, a finite- $\chi$ MPS has strictly exponential correlations — exactly the fingerprint of a gapped phase. The bond dimension you actually need is therefore a story about entanglement scaling. The second figure is the punchline of the whole subject: how the required $\chi$ climbs with qubit number for area-law, critical, and volume-law states — three curves separated by many orders of magnitude by fifty qubits.

Figure

Bond dimension needed for the central cut versus qubit number n: an area-law state stays at χ = 4, a critical chain grows only polynomially as χ ∼ n², but a volume-law state needs χ = 2^(n/2) — about 3.4×10⁷ at n = 50, beyond any classical memory

The gauge freedom of an MPS — you can insert $GG^{-1}$ on any bond without changing the state — is physically invisible but computationally decisive. Fixing the canonical gauge makes one bond's Schmidt values manifest, so truncation and local updates become exact linear-algebra operations rather than global optimizations — the machinery under DMRG, TEBD, and every claim that a circuit can be faked on a laptop.

The core equations

An MPS is defined by giving, for each site, a set of matrices indexed by that site's physical value. For a closed (periodic) chain the amplitude is the trace of their product, which glues the two ends together.

|\psi\rangle=\sum_{s_1,\dots,s_n} \mathrm{Tr}\!\big(A^{s_1}_{[1]}A^{s_2}_{[2]}\cdots A^{s_n}_{[n]}\big)\,|s_1 s_2\cdots s_n\rangle \qquad (1)

where: $A^{s_i}_{[i]}$ — a $\chi_{i-1}\times\chi_i$ matrix for site $i$ , one for each local value $s_i\in\{0,\dots,d-1\}$ ; $d$ — the physical dimension, $d=2$ for qubits; $\chi_i$ — the bond dimension on the link between sites $i$ and $i+1$ ; $\mathrm{Tr}$ — the trace that closes a periodic chain. Reading it: to get one amplitude, pick the matrix at each site named by that qubit's value, multiply the string, and trace. For an open chain the trace is replaced by boundary vectors, $c_{s_1\cdots s_n}=v_L^{\mathsf T}A^{s_1}\cdots A^{s_n}v_R$ , with $A^{s_1}$ a row and $A^{s_n}$ a column — the form used in practice, since every bond can then be gauged independently. The bond dimension is not a free label: for the minimal MPS it is fixed by the state's entanglement. Reshape the amplitude tensor at cut $k$ into a matrix, and its rank is the Schmidt rank of the cut, which the bond must accommodate.

\chi=\max_{1\le k<n}\ \mathrm{rank}\big(\rho_{1:k}\big)=\max_k\ \chi_k \qquad (2)

where: $\rho_{1:k}$ — the reduced density matrix of the left block $\{1,\dots,k\}$ ; $\mathrm{rank}$ — the number of nonzero Schmidt coefficients across the cut $k\,|\,k{+}1$ ; $\chi_k$ — the bond dimension there. Reading it: the bond dimension of the whole MPS is the worst (largest) Schmidt rank over all its cuts. Because the eigenvalues of $\rho_{1:k}$ are the squared Schmidt coefficients — a probability distribution over at most $\chi_k$ outcomes — the cut's entropy cannot exceed the log of that count. This is the single most important inequality in the subject.

S(\rho_{1:k})=-\sum_{\alpha=1}^{\chi_k}\lambda_\alpha^2\log_2\lambda_\alpha^2\ \le\ \log_2\chi_k\ \le\ \log_2\chi \qquad (3)

where: $\lambda_\alpha$ — Schmidt coefficients across cut $k$ , normalized $\sum_\alpha\lambda_\alpha^2=1$ ; $S$ — the von Neumann entanglement entropy of the block. Reading it: $\log_2\chi$ is a hard ceiling on the entanglement any cut of the state may carry. The gauge freedom is used to impose isometry conditions — $\sum_{s}A^{s\dagger}A^{s}=\mathbb{I}$ (left-canonical) and $\sum_{s}B^{s}B^{s\dagger}=\mathbb{I}$ (right-canonical) — so environment contractions collapse to the identity. Left-canonical tensors on one side of a bond and right-canonical on the other then expose the Schmidt decomposition of that bond directly: the mixed-canonical, or Vidal, form.

|\psi\rangle=\sum_{s_1\cdots s_n}A^{s_1}\!\cdots A^{s_{k}}\,\Lambda_k\,B^{s_{k+1}}\!\cdots B^{s_n}\,|s_1\cdots s_n\rangle \qquad (4)

where: $\Lambda_k$ — the diagonal matrix of Schmidt coefficients on bond $k$ , the entanglement of that cut written out explicitly; $A,B$ — left- and right-canonical tensors to the two sides. Reading it: in this gauge the bond's singular values are the Schmidt spectrum, so the entanglement is read straight off the state. Truncating a bond from $\chi_k$ to $\chi'$ then drops the smallest Schmidt values, and in the canonical gauge the resulting global error is exactly the discarded weight.

\big\||\psi\rangle-|\tilde\psi\rangle\big\|^2 \;=\; \sum_{\alpha>\chi'} \lambda_\alpha^2 \;\equiv\; \varepsilon_{\mathrm{trunc}} \qquad (5)

where: $|\tilde\psi\rangle$ — the MPS truncated to bond dimension $\chi'<\chi_k$ at one cut; $\varepsilon_{\mathrm{trunc}}$ — the discarded weight, the sum of squared dropped Schmidt values. Reading it: by Eckart–Young this keep-the-largest choice is the 2-norm-optimal rank- $\chi'$ approximation, so no other truncation does better. Both storage and the cost of contracting the network are low powers of $\chi$ — the whole point, an exponential problem traded for a polynomial one whenever $\chi$ stays modest.

\text{Storage}=O(n\,d\,\chi^2),\qquad \text{DMRG sweep}=O(n\,\chi^3),\qquad \langle\psi|O|\psi\rangle=O(n\,\chi^3) \qquad (6)

where: $\chi$ — the bond dimension, entering cubically in expectation values through transfer-matrix contraction; $d$ — physical dimension; $n$ — chain length. Reading it: doubling $\chi$ costs eight times the compute per sweep — the practical wall on how much entanglement DMRG can chase. Finally, correlations in a finite- $\chi$ MPS are set by the transfer matrix and decay geometrically with separation, which is why a fixed- $\chi$ MPS can never reproduce true criticality.

C(r)=\langle O_i O_{i+r}\rangle-\langle O_i\rangle\langle O_{i+r}\rangle\ \sim\ \Big(\tfrac{|\lambda_2|}{|\lambda_1|}\Big)^{r},\qquad \xi=\frac{1}{\ln(|\lambda_1|/|\lambda_2|)} \qquad (7)

where: $\lambda_1,\lambda_2$ — the largest and second-largest eigenvalues of the transfer matrix $E=\sum_s \bar A^{s}\otimes A^{s}$ ; $\xi$ — the correlation length. Reading it: a discrete transfer-matrix spectrum forces exponential decay with finite $\xi$ , so power-law (critical) correlations demand $\chi\to\infty$ .

Derivation AEvery quantum state is an exact matrix product state

The goal is to show constructively that the MPS is not a special ansatz but a universal factorization every state possesses, and that the bond dimensions it produces are exactly the Schmidt ranks of the cuts — so that bounding the entanglement bounds the representation. We start from the full coefficient tensor, peel off one site at a time by SVD, and track how large each bond must be; the worst case pins down the exponential wall for volume-law states.

Step 1. Reshape the amplitude tensor so the first qubit is a row index and the rest form one giant column index, giving a $d\times d^{n-1}$ matrix, and take its SVD, $c_{s_1,(s_2\cdots s_n)}=\sum_{\alpha_1}U_{s_1,\alpha_1}(\Lambda V^\dagger)_{\alpha_1,(s_2\cdots s_n)}$ . Define $A^{s_1}_{\alpha_1}=U_{s_1,\alpha_1}$ , a set of $1\times\chi_1$ objects with $\chi_1\le d$ . Why: the SVD reads off the Schmidt rank and hands you an orthonormal left basis to name the bond, so the bond dimension emerges rather than being assumed.

Step 2. Absorb $\Lambda V^\dagger$ into the remainder, merge the new bond index $\alpha_1$ with the next physical index $s_2$ into one row index, reshape and SVD again: $(\Lambda V^\dagger)_{(\alpha_1 s_2),(s_3\cdots s_n)}\to A^{s_2}(\Lambda'V'^\dagger)$ , extracting a $\chi_1\times\chi_2$ tensor $A^{s_2}$ . Why: folding the running bond together with the next site threads the chain — each SVD sees the accumulated left history plus one fresh qubit, so its bonds carry the left-right correlations.

Step 3. Iterate to the far end. After $n$ SVDs the residual is fully consumed and $A^{s_1}A^{s_2}\cdots A^{s_n}$ reproduces $c_{s_1\cdots s_n}$ exactly. Why: each step is an exact factorization, so nothing is lost — the MPS is a re-bracketing of the same numbers, agreeing to machine precision.

Step 4. Bound the bonds. At cut $k$ the bond dimension equals the rank of the reshaped matrix, $\chi_k=\mathrm{rank}(C^{(k)})\le\min(d^k,d^{n-k})$ , so the worst case is at the centre, $\chi_{\max}\sim d^{n/2}$ . If every Schmidt rank is at most $\chi$ , the MPS has bond dimension $\chi$ and stores only $O(n\,d\,\chi^2)$ numbers. Why: this single line holds the whole dichotomy — generic states hit the exponential ceiling $d^{n/2}$ , while low-entanglement states, with short Schmidt spectra, live at small $\chi$ .

Go deeper · Derivation B · 1 minWhy one bond dimension bounds the entanglement of every cut

Derivation BWhy one bond dimension bounds the entanglement of every cut

The aim is to prove the inequality that makes the bond dimension physically meaningful — a chain of $\chi$ -dimensional bonds carries at most $\log_2\chi$ bits of entanglement across any cut — and then that truncating to a smaller bond is provably optimal, with an error read off directly. We tie together three facts: the squared Schmidt values form a probability distribution, entropy on a finite support is maximized by the uniform distribution, and orthonormal Schmidt vectors collapse the truncation error to the discarded weight.

Step 1. In mixed-canonical gauge the state across bond $k$ is a Schmidt decomposition, $|\psi\rangle=\sum_{\alpha=1}^{\chi_k}\lambda_\alpha|L_\alpha\rangle|R_\alpha\rangle$ , so the reduced density matrix has eigenvalues $p_\alpha=\lambda_\alpha^2$ forming a probability distribution over at most $\chi_k$ outcomes. Why: treating the squared Schmidt values as probabilities lets the whole machinery of Shannon entropy apply to a purely quantum quantity.

Step 2. The von Neumann entropy is the Shannon entropy of that distribution, $S=-\sum_\alpha p_\alpha\log_2 p_\alpha$ , and Shannon entropy on a support of size $\chi_k$ is maximized by the flat distribution $p_\alpha=1/\chi_k$ , giving $S\le\log_2\chi_k\le\log_2\chi$ . Why: the flat spectrum is the most entangled a bond can be — equality at maximally entangled cuts — which is why a $\chi=4$ bond holds at most 2 bits.

Step 3. Now truncate: keep the largest $\chi'$ Schmidt values and drop the rest, $|\tilde\psi\rangle=\sum_{\alpha=1}^{\chi'}\lambda_\alpha|L_\alpha\rangle|R_\alpha\rangle$ . Because the $|L_\alpha\rangle$ and $|R_\alpha\rangle$ are orthonormal, the squared error telescopes to $\||\psi\rangle-|\tilde\psi\rangle\|^2=\sum_{\alpha>\chi'}\lambda_\alpha^2$ , the discarded weight. Why: orthonormality is exactly what canonical gauge buys — without it the cross terms survive and the error would not equal the spectral tail.

Step 4. By Eckart–Young, keeping the largest singular values is the 2-norm-optimal rank- $\chi'$ approximation of the reshaped tensor, so no other bond- $\chi'$ state approximates $|\psi\rangle$ better in Frobenius norm. Why: this promotes "drop the small ones" to a theorem, so controlling the discarded weight controls the whole simulation error.

Go deeper · Derivation C · 1 minFrom an area law to a polynomial bond dimension

Derivation CFrom an area law to a polynomial bond dimension

The purpose is to close the loop between physics and cost: if a state's entanglement obeys an area law — bounded entropy across every cut — then the bond dimension needed to represent it faithfully grows only polynomially in system size, the formal reason DMRG works so well for gapped 1D matter. The argument bounds the tail of the Schmidt spectrum with a Rényi entropy, converts it into a truncation error, and accumulates the $n-1$ cuts. Hastings' theorem supplies the area law; the derivation shows why that single input forces efficiency.

Step 1. Suppose a Rényi entropy of order $\alpha<1$ is finite and bounded across the cut, $S_\alpha=\frac{1}{1-\alpha}\log_2\sum_\beta\lambda_\beta^{2\alpha}$ . A finite $S_\alpha$ with $\alpha<1$ means $\sum_\beta\lambda_\beta^{2\alpha}$ converges, which forces the sorted Schmidt values $\lambda_1\ge\lambda_2\ge\cdots$ to decay faster than any fixed power. Why: a bounded low-order Rényi entropy is the precise way to say "the spectrum is short" — plain von Neumann entropy alone does not quite control the tail.

Step 2. Convergence of that sum implies the ordered values fall off polynomially with an exponent above one, $\lambda_\beta^2\lesssim\beta^{-1/\alpha}$ , so the discarded weight past bond $\chi$ is $\varepsilon_{\mathrm{trunc}}(\chi)=\sum_{\beta>\chi}\lambda_\beta^2\lesssim\chi^{-(1/\alpha-1)}$ . Inverting, reaching per-cut error $\varepsilon$ needs only $\chi\sim\varepsilon^{-\alpha/(1-\alpha)}$ . Why: the exponent is finite, so the required bond grows as a power of $1/\varepsilon$ , not exponentially — polynomial cost is visible already at one cut.

Step 3. Accumulate over the $n-1$ cuts. Truncation errors add in quadrature, so the global error is at most $\sqrt{\sum_{k}\varepsilon_{\mathrm{trunc}}^{(k)}}\le\sqrt{n}\,\sqrt{\varepsilon}$ , and fixed total accuracy costs only an extra factor $\sqrt{n}$ in $\varepsilon$ . Why: a chain has linearly many bonds, so errors could pile up; the quadrature bound shows only a benign $\sqrt{n}$ factor is paid.

Step 4. Together, $\chi=\mathrm{poly}(n,1/\varepsilon)$ suffices to represent any area-law state to fixed accuracy, and Hastings' theorem guarantees the constant-entropy area law for any gapped local 1D ground state. Why: this is the theorem under DMRG — why $\chi$ of a few hundred to a few thousand gives gapped-1D ground-state energies to ten significant figures.

Limiting cases

How large a bond dimension you must pay depends on how the target state's entanglement scales with system size — the four rows span the range from trivial to intractable.

Case	Result	Meaning
Product / unentangled state	$S=0\ \Rightarrow\ \chi=1$	every $A^{s_i}$ is a scalar; the MPS factorizes into single-site states.
Gapped 1D ground state (area law)	$S=O(1)\ \Rightarrow\ \chi=\mathrm{poly}(n)$	Hastings' area law guarantees efficient approximation; DMRG is exact for practical purposes.
Critical (gapless) 1D chain	$S\sim\tfrac{c}{6}\log_2 L\ \Rightarrow\ \chi\sim L^{\kappa}$	logarithmic entanglement (central charge $c$ ) forces $\chi$ to grow polynomially; usable with finite-entanglement-scaling corrections.
Volume-law / generic state	$S=\Theta(n)\ \Rightarrow\ \chi=2^{\Theta(n)}$	almost all of Hilbert space, and deep random circuits, need exponential $\chi$ — classically intractable.

Common trapS ≤ log₂ χ is a promise about contiguous cuts, and it is necessary but not sufficient

Two traps hide in the bound. First, $\chi$ bounds the entropy of a contiguous cut of the chain and nothing else directly: it makes no separate promise about a single interior qubit, nor about long-range mutual information beyond what the Schmidt rank allows. Second, and more subtly, $S\le\log_2\chi$ is necessary but not sufficient. A state can have small entropy — say $S=2$ bits — yet carry a long tail of small-but-nonzero Schmidt coefficients, so the discarded weight $\sum_{\alpha>\chi'}\lambda_\alpha^2$ stays stubbornly large until $\chi'$ is far bigger than $2^S$ . Faithfulness is decided by the full spectrum, not its entropy alone; matching the entropy is the price of admission, not proof of a cheap representation.

Worked examples

1 · The GHZ state as a bond-dimension-2 MPS (confidence builder)

Write the $n$ -qubit GHZ state $|\mathrm{GHZ}\rangle=(|00\cdots0\rangle+|11\cdots1\rangle)/\sqrt2$ as an exact MPS, find its bond dimension, and check that it saturates the entanglement bound.

Step 1. Use bulk tensors $A^0=\mathrm{diag}(1,0)$ and $A^1=\mathrm{diag}(0,1)$ (each a $2\times2$ matrix) with left and right boundary vectors $v_L=\tfrac1{\sqrt2}(1,1)$ and $v_R=(1,1)^{\mathsf T}$ . A string with any mismatch — a $0$ somewhere and a $1$ elsewhere — multiplies $A^0A^1=0$ , so its amplitude vanishes; only the all- $0$ and all- $1$ strings survive, each with amplitude $1/\sqrt2$ . This reproduces the GHZ state exactly with $2\times2$ matrices, so the bond dimension is $\chi=2$ .

Step 2. Cut the chain anywhere. The Schmidt decomposition is $\tfrac1{\sqrt2}(|0\cdots0\rangle_L|0\cdots0\rangle_R+|1\cdots1\rangle_L|1\cdots1\rangle_R)$ , with two equal Schmidt coefficients $\lambda_1=\lambda_2=1/\sqrt2$ . The Schmidt rank is 2 across every cut, matching $\chi=2$ as equation (2) requires.

Step 3. The entanglement entropy is $S=-2\cdot\tfrac12\log_2\tfrac12=\mathbf{1\ bit}$ , exactly the ceiling $\log_2\chi=\log_2 2=1$ . The GHZ state saturates $S\le\log_2\chi$ — as entangled as a $\chi=2$ bond allows — yet needs only $O(n)$ numbers, not $2^n$ . A globally entangled state can still be a constant- $\chi$ MPS, because its entanglement per cut is constant.

2 · Truncation error and the storage it buys back

One central bond of a canonical MPS has dimension $\chi=64$ . You truncate it to $\chi'=16$ , and the squares of the discarded Schmidt values sum to $10^{-6}$ . Find the resulting 2-norm error of the global state, say why the truncation is optimal, and estimate the storage saved.

Step 1. In mixed-canonical gauge the Schmidt vectors are orthonormal, so equation (5) applies directly: $\||\psi\rangle-|\tilde\psi\rangle\|^2=\sum_{\alpha>16}\lambda_\alpha^2=10^{-6}$ . Taking the square root, the global 2-norm error is $\||\psi\rangle-|\tilde\psi\rangle\|=\mathbf{10^{-3}}$ — a per-mille distance, from a millionth of discarded weight.

Step 2. This truncation is optimal by Eckart–Young: the sixteen largest singular values give the 2-norm-best rank-16 approximation, so no other bond-16 state is closer in Frobenius norm — the error is the smallest achievable at this bond dimension.

Step 3. Storage on a bond scales as $\chi^2$ (equation 6), so going from $64^2=4096$ to $16^2=256$ per physical value is a factor of 16 reduction, at the cost of a controlled $10^{-3}$ error. This is the everyday trade of TEBD and DMRG: pay a tiny, measured error to stop the bond from blowing up. Error $10^{-3}$ , optimal by Eckart–Young, and a 16× memory saving.

3 · When does a random circuit become classically hard?

Take a 1D nearest-neighbour random circuit on $n=50$ qubits. Estimate the bond dimension its output needs at depth $D=6$ and at depth $D=25$ , and say which is classically simulable.

Step 1. A 1D local random circuit builds entanglement across a central cut roughly linearly in depth until it saturates the volume law, $S\sim\min(D,\,n/2)$ bits, and the bond dimension needed is $\chi\sim2^{S}$ . The crossover is set by whether the depth has reached $n/2=25$ — the entanglement clock is the circuit depth.

Step 2. At $D=6$ (about $\log_2 50\approx5.6$ , logarithmically shallow), $S\approx6$ and $\chi\approx2^6=\mathbf{64}$ , so storage $O(n\,d\,\chi^2)=50\cdot2\cdot64^2\approx4\times10^{5}$ numbers is trivial on a laptop. Shallow 1D circuits are efficiently simulable precisely because their bond dimension stays polynomial.

Step 3. At $D=25=n/2$ the cut is volume-saturated, $S\approx25$ and $\chi\approx2^{25}\approx\mathbf{3.4\times10^7}$ ; storage balloons to $\approx1\times10^{17}$ numbers, worse than the $2^{50}\approx1.1\times10^{15}$ dense amplitudes, so the MPS has stopped helping. The shallow-to-deep crossover at $D\sim n/2$ is exactly the boundary between "fakeable on a laptop" and quantum-hard.

Physics of the portrait

The matrix product state was discovered several times before anyone saw the discoveries were the same — an exact solvable-model ansatz, a numerical fixed point, and a simulation data structure, all one object seen from three directions.

Origins · 1 minmatrix product states and how this was discovered

matrix product states

the χ-bounded tensor-network ansatz cₛ₁…ₛₙ = Tr(Aˢ¹ ⋯ Aˢⁿ) · many-body entanglement and quantum simulation

The matrix product state has a tangled genealogy, invented independently in three communities that only later saw they described one object. In condensed matter it appeared in 1987 as the exact ground state Affleck, Kennedy, Lieb, and Tasaki (AKLT) wrote for a spin-1 chain — a valence-bond wavefunction that, read modern, is a bond-dimension-2 MPS. Numerical physics arrived from the opposite end: Steven White's density-matrix renormalization group of 1992 was for years simply the most accurate method for 1D systems, and Stellan Östlund and Stefan Rommer recognized in 1995 that its fixed point is a matrix product state. The third face came when Guifré Vidal showed in 2003 that a truncated- $\chi$ MPS efficiently simulates "slightly entangled" computations, making the bond dimension the explicit dial of classical hardness. Frank Verstraete and Ignacio Cirac's generalization to higher-dimensional tensor networks and Matthew Hastings' 2007 area-law proof for gapped 1D ground states sealed the synthesis — finally explaining why the MPS is so good at exactly the states that matter. A measured empirical success became a theorem about entanglement.

History

1987

— Affleck, Kennedy, Lieb, and Tasaki write the AKLT state, an exact matrix product ground state

1992

— White introduces the density-matrix renormalization group

1995

— Östlund and Rommer recast the DMRG fixed point as a matrix product state

2003

— Vidal simulates slightly entangled circuits with truncated MPS

2007

— Hastings proves the area law for gapped one-dimensional ground states

Going further

Texts. Nielsen & Chuang, Quantum Computation and Quantum Information, §2.5 (Schmidt decomposition) and Ch. 4 · Preskill, Lecture Notes on Quantum Computation, Ch. 4 (entanglement, area laws) · Watrous, The Theory of Quantum Information, Ch. 6 · Wilde, Quantum Information Theory, 2nd ed., Ch. 11 (entropies) · Schollwöck, "The density-matrix renormalization group in the age of matrix product states", Ann. Phys. 326, 96 (2011) · Verstraete, Cirac & Murg, "Matrix product states, projected entangled pair states, and variational renormalization group methods", Adv. Phys. 57, 143 (2008) · Hastings, "An area law for one-dimensional quantum systems", J. Stat. Mech. (2007) P08024 · Vidal, "Efficient classical simulation of slightly entangled quantum computations", Phys. Rev. Lett. 91, 147902 (2003).

Provenance. The MPS and canonical forms, the iterated-SVD construction, and the $O(n\chi^3)$ costs follow Schollwöck and Verstraete–Cirac–Murg; the Schmidt decomposition and the $S\le\log_2\chi$ bound follow Nielsen & Chuang §2.5 and Wilde Ch. 11; truncation optimality is Eckart–Young; the area law is Hastings (J. Stat. Mech. 2007 P08024); the circuit-simulation reading is Vidal (PRL 91, 147902); and the critical scaling $S\sim(c/6)\log_2 L$ is Calabrese–Cardy. A verifier should: (1) reconstruct a random 4-qubit state by left-to-right SVD and confirm the contracted MPS matches to machine precision, with $\chi_k=\mathrm{rank}(\rho_{1:k})$ ; (2) confirm the GHZ tensors give $(|0\rangle^{\otimes n}+|1\rangle^{\otimes n})/\sqrt2$ with $\chi=2$ , $S=1$ ; (3) check that discarded weight $10^{-6}$ yields a $10^{-3}$ 2-norm error; (4) verify $(\lambda_2/\lambda_1)^r$ correlation decay via the transfer-matrix eigenvalues. All bond-dimension and storage figures are illustrative arithmetic, not empirical claims.

Prerequisites. Schmidt Decomposition and Bipartite Entanglement · Singular Value Decomposition · Von Neumann Entropy · Parameterized Quantum Circuits and Ansatz Design. Leads to. Variational Quantum Eigensolver · Expressibility, Entangling Capability, and Trainability Trade-offs · Barren Plateaus in Variational Landscapes · Hardware-Efficient and Chemistry-Inspired Ansätze.

Ask the tutor: Ask me to walk you through building the exact MPS of any small state by hand — or to show where a given circuit's bond dimension crosses from laptop-easy to quantum-hard.

Is an MPS an approximation or an exact representation?

Both, and which one is up to you. The iterated-SVD construction shows every state has an exact MPS if you let the bond dimension grow to whatever the Schmidt ranks demand, up to $2^{n/2}$ at the centre. The approximation enters only when you cap $\chi$ and drop the smallest Schmidt values: for an area-law state that costs almost nothing, because the spectrum is short; for a volume-law state it is ruinous, because you delete most of the state. So "MPS" names a family from exact-but-huge to cheap-but-approximate, with $\chi$ as the slider.

Why does the bound involve log₂ χ rather than χ itself?

Because entanglement entropy counts bits, and a $\chi$ -dimensional bond is a channel that carries at most $\log_2\chi$ qubits of entanglement — the same reason a classical variable with $\chi$ outcomes holds at most $\log_2\chi$ bits of information. The maximum is reached when the Schmidt spectrum is flat, all $\chi$ values equal to $1/\sqrt\chi$ , giving $S=\log_2\chi$ exactly; any peaking lowers the entropy below that ceiling, which is why real states usually sit well under $\log_2\chi$ even at fixed bond dimension.

If DMRG is just variational MPS, why was it invented as something else?

History ran backwards. White introduced DMRG in 1992 as a real-space renormalization procedure — keep the block states carrying the most weight in a density matrix, discard the rest — and it worked stunningly well before anyone fully understood why. Only in 1995 did Östlund and Rommer show its fixed point is a matrix product state, and only later was the whole method reframed as variational energy minimization over MPS of fixed bond dimension. The "keep the largest density-matrix eigenvalues" rule is, in retrospect, exactly the optimal Schmidt truncation of a bond — DMRG was a correct algorithm in search of the language the MPS supplied.

Can an MPS ever represent a critical (gapless) state?

Only approximately, but systematically. A finite- $\chi$ MPS has a discrete transfer-matrix spectrum, so its correlations always decay exponentially with a finite length $\xi\sim1/\ln(\lambda_1/\lambda_2)$ — never the true power law of a critical point. Increasing $\chi$ grows $\xi$ , so the MPS mimics criticality out to longer distances; studying how observables drift with $\chi$ and extrapolating to $\chi\to\infty$ (finite-entanglement scaling) recovers the critical exponents. Critical states are thus reachable in the limit, never at any finite bond dimension.

What makes the canonical gauge worth the trouble?

Without it almost every operation is global; with it, they become local. Fixing the gauge makes the tensors partial isometries — left-canonical on one side of a bond, right-canonical on the other — and then three things become free: the norm is just $\sum_\alpha\lambda_\alpha^2$ on any bond, a single-site expectation value contracts to one small local tensor sandwich, and a bond truncation has an error equal to the discarded Schmidt weight exactly. Every efficient MPS algorithm — DMRG, TEBD, expectation values — is really moving the canonical centre along the chain and doing local linear algebra at each stop.

Does a small entanglement entropy guarantee a cheap MPS?

No, and this is a genuine trap. The entropy is a single number summarizing the whole spectrum, and $S\le\log_2\chi$ is only necessary — the floor you cannot go below. A state can have modest entropy yet a long tail of slowly decaying Schmidt values, so the discarded weight $\sum_{\alpha>\chi'}\lambda_\alpha^2$ stays non-negligible until $\chi'$ far exceeds $2^S$ . What controls the true cost is how fast the sorted spectrum decays, not the entropy it integrates to — which is why area laws, by bounding a low-order Rényi entropy, do the real work.