Certifying Schmidt Numbers of Mixed States by Rank-Constrained Factorization

The Schmidt number of a two-party quantum state counts how many dimensions of entanglement it carries. For a noisy (mixed) state it is defined as the smallest number $k$ such that the state can be written as a probabilistic mixture of pure states, each having at most $k$ entangled terms. Computing it means searching over all such mixtures. We propose to recast that search as a smooth non-convex fitting problem: find a mixture of pure states whose coefficient matrices have rank at most $k$ and whose average reproduces the target density matrix. When the fit reaches zero error, it returns an explicit decomposition. That decomposition is a certificate anyone can check, and it proves the Schmidt number is at most $k$.

The obstacle is geometric, and the matrix-optimization literature has recently explained it. Each pure component, reshaped as an $n\times m$ matrix, must lie in the set of matrices of rank at most $k$. That set has singular "corner" points wherever rank drops below $k$. At those points, ordinary projected-gradient methods can report convergence while sitting at a point that is not stationary at all. The best decompositions of physical states often contain exactly such lower-rank components, for instance product-state pieces mixed with entangled ones, so the failure mode lands where the answer lives. We will import the remedy of Levin, Kileel and Boumal into this quantum problem. We will then pair the resulting upper bounds with existing analytic lower bounds, so that a meaningful fraction of states have their Schmidt number pinned exactly.

Core Claim

A mixed state $\rho$ on an $n\times m$ system has Schmidt number at most $k$ if and only if the following loss reaches zero for some choice of $N$ factor pairs ($N \le (nm)^2$ suffices by Carathéodory's theorem). Here $L_j \in \mathbb{C}^{n\times k}$ and $R_j\in\mathbb{C}^{m\times k}$ are unconstrained complex matrices, $\mathrm{vec}$ stacks a matrix into a vector, and component weights are absorbed into the factor norms:

$$ F_k(L,R) = \Big\| \rho - \sum_{j=1}^{N} \mathrm{vec}(L_j R_j^\top)\,\mathrm{vec}(L_j R_j^\top)^\dagger \Big\|_F^2 . $$

The claim is that minimizing $F_k$ with a second-order Riemannian trust-region method yields two things. First, it gives exact Schmidt-number upper bounds that are reliable in practice: the method reaches zero at the true $k$ far more often than first-order or projection-based alternatives. Second, it does this at dimensions where Monte Carlo search over decompositions is impractical. The method runs on the lifted space of factor pairs and adds a "hook," a rebalancing step that equalizes $L_j^\dagger L_j$ and $R_j^\dagger R_j$ (the complex analogue of the condition $L^\top L = R^\top R$ in the real setting).

The second-order method plus the hook is the key ingredient. It is the only part of the pipeline with a stationarity guarantee on the rank-bounded set, and the experiments are designed to test whether that guarantee matters in practice.

Relation to Prior Work

Two papers motivate the gap. The first is Chen's analysis of low-rank mixed states, which proves algebraic lower bounds such as $k \ge (m-\dim L_A(\rho))/r$ for a rank-$r$ state. Chen also notes that these bounds become uninformative as $r$ grows, and provides no algorithm for the matching upper bound. We use Chen's bounds as the lower half of a sandwich and supply the missing constructive upper half.

The second is Cao et al.'s Monte Carlo method for the three-tangle. It minimizes a pure-state entanglement measure plus a Frobenius penalty for mismatch with $\rho$, using parallel tempering, and reaches $10^{-7}$ agreement on up to rank-8 three-qubit states. Our loss is essentially their penalty term with the rank constraint built into the parameterization rather than scored. The difference is that we replace stochastic search with deterministic second-order optimization that has a stationarity guarantee.

Two papers supply the method. Levin et al. prove that every rank-deficient point of a bounded-rank matrix set is "apocalyptic," meaning projected-gradient stationarity measures can vanish there along a sequence while the limit is non-stationary. They show that a trust-region method on the factored space with a rebalancing hook avoids this. We apply their construction blockwise, one copy per ensemble component, coupled through a shared quartic fit term. Their analysis covers a single matrix; whether it extends cleanly to $N$ coupled blocks is a question we will check rather than assume.

Boumal, Voroninski and Bandeira show that overparameterizing a factorization removes spurious local minima for smooth semidefinite programs. Our problem is not such a program, so their theorem does not transfer. It does suggest one specific ablation: increase the number of components $N$ and measure whether spurious stationary points disappear.

Three papers constrain the design, or are the nearest neighbors we must beat. Rudra and Jivulescu compute projective tensor norms of states, including mixed ones, by plain automatic-differentiation gradient descent with adaptive rank. They report about $10^{-3}$ error against analytic values and give no certificates. The tomography works ProjFGD and Riemannian gradient descent for tomography show that factored, low-rank parameterizations scale to 13 qubits ($d=8192$ in about 20 minutes, where SDPT3 failed within 24 hours). Both reconstruct $\rho$ itself and never decompose it into rank-limited pure parts.

We also want to be candid about one point. The idea of parameterizing ensembles through the unitary freedom of a fixed square root of $\rho$ is folklore in entanglement numerics. Our contribution is not that reformulation. It is the per-component rank-$k$ factorization, the geometry-aware solver, and certification of the Schmidt number.

Research Program

We will build three solvers that share one loss and one interface.

The first is the hooked Riemannian trust-region solver on the complex factor pairs. It uses exact Hessian-vector products of the quartic loss, the rebalancing hook after each accepted step, and a continuation schedule that starts with large $N$ and prunes components whose norms vanish.

The second is a projection solver. It takes a gradient step on full pure-state vectors and then truncates each reshaped component back to rank $k$ by a singular value decomposition. This is exactly the projected-gradient scheme that Levin et al. predict can stall at corners.

The third is a penalized first-order solver in the style of Rudra and Jivulescu. It runs Adam on the same factors with no hook and no curvature information.

Certification works as follows. A run "succeeds at $k$" if the loss falls below $10^{-12}$. We then re-verify the reconstructed decomposition in extended precision. We round the factors and check positivity and the trace-distance residual. We also check that each component's $(k{+}1)$-th singular value is exactly zero by construction. Because each component is a product of $n\times k$ and $k\times m$ matrices, this holds structurally, so the only quantity to verify is the residual.

For each state we then report the gap between Chen's lower bound and the smallest $k$ certified from above. A zero gap is an exact Schmidt number.

On the analysis side, we will study the landscape on small instances: $3\times3$, $3\times 9$, and $4\times 4$ systems. Many random restarts will let us catalogue stationary points with positive loss at the true $k$ and record whether they contain rank-deficient components. That tests directly whether apocalyptic corners, rather than ordinary local minima, cause projection methods to fail.

Baselines

The naive reference point needs no optimization. We take the eigendecomposition of $\rho$, which is itself a valid ensemble, and report the maximal Schmidt rank among its eigenvectors as the upper bound, with Chen's analytic bound as the lower bound. Any method that does not close this gap more often than doing nothing is useless. The fraction of states where the eigendecomposition already certifies the answer is our floor.

The strongest existing method for decomposition search is Cao et al.'s replica-exchange Monte Carlo, adapted to minimize $F_k$ over rank-$k$ factors instead of the three-tangle energy. It is the strongest competitor because parallel tempering is designed to tunnel through exactly the barriers we worry about. The cited paper reports agreement with analytic three-tangle values to errors as low as $10^{-7}$ on up to rank-8 states. There is no reported result on Lacuna for Monte Carlo or any other numerical Schmidt-number computation, so we will establish that number ourselves.

For fairness, every method gets the same states, the same $N$, the same wall-clock budget on the same hardware, and the same number of random initializations. We will tune temperatures and swap rates on a held-out set of states with equal tuning effort for each method. Beating this baseline means a higher success rate at the true $k$ under equal time, or equal success at least ten times faster, on systems beyond $4\times4$.

The ablations isolate the key ingredient. The projection solver removes the lifted geometry. The Adam solver removes curvature and the hook. A trust-region variant without the hook removes only the rebalancing step. If the full method does not outperform the no-hook variant specifically on states whose optimal decomposition contains rank-deficient components, then the apocalyptic-point story is not what drives performance, and we will say so.

Convex semidefinite relaxations are not a baseline for the upper bound. No convex program in the file computes Schmidt-number upper bounds. The positivity-of-partial-transpose hierarchies, such as the $PPT_k$ construction, bound other quantities from the opposite side.

Why not just…?

Why not just run Adam on the factors, as in the projective-norm work? This is the most serious threat. On small, well-conditioned instances it may work as well as the full method. Rudra and Jivulescu get roughly $10^{-3}$ accuracy this way, which is two orders of magnitude short of what exact certification needs. That gap could, however, reflect their problem rather than the optimizer. The Adam ablation, run under equal budget on the same states, settles the question. If Adam matches our success rate at the true $k$ across sizes, the contribution shrinks to the certification pipeline and the loss formulation, and we will report it that way.

Why not just alternate between truncating each component's singular values and refitting the mixture? This scheme is cheap and intuitive, and it is precisely projected-gradient descent onto the bounded-rank set. Levin et al. prove that this method's stationarity measure can go to zero at non-stationary corner points. Our landscape study tests whether that pathology occurs in Schmidt-number instances. The prediction is that projection runs will stall with positive loss at the true $k$, with rank-deficient components, more often than hooked runs.

Why not just use Monte Carlo, which already handles rough landscapes? Cao et al. show it is accurate on three qubits, but its cost grows with the number of ensemble partitions, and it gives no stationarity guarantee. The equal-time comparison on growing $n\times m$ systems will show whether it remains competitive. If it does, a hybrid that uses Monte Carlo for initialization and trust-region refinement is the natural fallback.

Proposed Experiments

The first test uses Chen's rank-3 state on a $3\times9$ system, whose Schmidt number is proven to be exactly 3. A sound method must drive the loss to zero at $k=3$ and must not reach zero at $k=2$ for any $N$. We will report success rates per method over at least 100 restarts.

The second test uses random low-rank states on $n\times m$ systems with $n,m$ up to 16 and ranks $r$ from 2 up to and beyond $n$. In these regimes Chen's generic bound gives a lower bound that is tight in some cases and loose in others. We will report the fraction of instances with a zero lower-to-upper gap, along with runtime as dimension grows.

The third test uses families with closed-form Schmidt numbers under noise, such as maximally entangled states mixed with white noise. Here we check that the certified $k$ jumps at the known noise thresholds.

The decisive outcome comes from the stratified comparison on states whose optimal decomposition is known to contain lower-rank components. If the hooked trust-region solver reaches the true $k$ substantially more often than both the projection solver and the no-hook variant at equal compute, the geometric remedy is doing real work in quantum entanglement quantification. If it does not, the proposal reduces to a useful certification tool built on simpler optimizers.