Schmidt Decomposition Low-Rank Approximation

• Schmidt decomposition-based low-rank approximation is a method that truncates quantum state expansions to reduce circuit complexity while preserving key entanglement features.
• It enables efficient quantum image encoding in schemes like NEQR, FRQI, and QPIE by significantly lowering gate counts and circuit depth for NISQ devices.
• The approach effectively balances resource efficiency and fidelity, offering scalable techniques for state synthesis with practical trade-offs and performance guarantees.

Schmidt decomposition-based low-rank approximation is a technique for compressing and simplifying quantum state preparation by exploiting entanglement structure. This approach has recently been applied to quantum image representations, such as the Novel Enhanced Quantum Representation (NEQR), Flexible Representation of Quantum Images (FRQI), and Quantum Probability Image Encoding (QPIE), to mitigate circuit depth and gate count for practical deployment on Noisy Intermediate-Scale Quantum (NISQ) hardware. The essential idea is to perform a Schmidt (singular value) decomposition of the target quantum state, truncate the expansion at rank $r$, and define an approximate state that preserves most of the original information while being dramatically simpler to synthesize, both algorithmically and in physical resources. This article surveys the principles, methodology, performance, and trade-offs of Schmidt decomposition-based low-rank approximation, with direct reference to the current literature on quantum image encoding (Pangeva et al., 9 Jun 2026).

1. Schmidt Decomposition and Quantum State Truncation

Given a pure state $|\psi\rangle$ over $n$ qubits partitioned into subsystems $A$ and $B$ of $n_A$ and $n_B$ qubits ($n_A+n_B=n$), the Schmidt decomposition expresses $|\psi\rangle$ as

$$|\psi\rangle = \sum_{j=1}^k \lambda_j |u_j\rangle_A \otimes |v_j\rangle_B$$

where $|\psi\rangle$0, the $|\psi\rangle$1 are non-negative real Schmidt coefficients with $|\psi\rangle$2, and $|\psi\rangle$3, $|\psi\rangle$4 are orthonormal. The Schmidt rank $|\psi\rangle$5 upper bounds the amount of entanglement across the partition.

Low-rank approximation consists of truncating the sum above to the largest $|\psi\rangle$6 terms: $|\psi\rangle$7 Normalizing yields a state whose fidelity to $|\psi\rangle$8 is $|\psi\rangle$9. For many structured data sources, including natural images, the Schmidt spectrum decays rapidly, so a small $n$0 suffices for accurate approximation.

2. Application in Quantum Image Encoding

Quantum image representations (QIRs) map classical $n$1-pixel images into quantum states over $n$2 qubits, facilitating quantum parallelism in processing. Popular schemes include:

• NEQR: Pixel values stored in computational basis registers labeled by position indices.
• FRQI: Amplitude encoding with grayscale information mapped to single-qubit rotation angles.
• QPIE: Global amplitude encoding schemes.

Direct preparation of a generic image state in NEQR, for an $n$3 image with $n$4 intensity bits per pixel, requires a quantum state over $n$5 qubits: $n$6 with $n$7 ($n$8 bits) holding the intensity at location $n$9. The standard approach realizes this by $A$0 multi-controlled gates, with exponential circuit depth and gate count. On NISQ hardware this becomes impractical for modest $A$1 ($A$2).

Applying the Schmidt decomposition to $A$3 (e.g., bipartitioning the intensity vs. position register, or splitting the qubit array into two halves) reveals that, after truncating to rank $A$4, the approximate state $A$5 can be synthesized more efficiently. State-preparation algorithms for generic rank-$A$6 states (e.g., recursive QR or LRSP routines) have complexity polynomial in $A$7 and logarithmic in $A$8.

3. Compilation Complexity and Resource Scaling

A concrete performance evaluation of Schmidt decomposition-based low-rank approximation applied to NEQR, FRQI, and QPIE for $A$9 ($B$0) images is given in (Pangeva et al., 9 Jun 2026). Key metrics include:

Method	Qubits	Full-Rank Depth	LRA Depth (r=257)	Full-Rank CNOTs	LRA CNOTs (r=257)	MSE (LRA)
NEQR	20	2,737,166	735,168 (–73%)	2,072,898	751,554 (–64%)	0.2728
FRQI	13	385,025	11,256 (–97%)	221,184	7,649 (–97%)	0.277
QPIE	12	19,910	3,704 (–81%)	4,083	3,788 (–7%)	0.272

The low-rank NEQR state at $B$1 (corresponding to keeping the largest 25% of Schmidt coefficients) achieves a mean squared error (MSE) of $B$2 (on 8-bit [0,255] images) and reconstructs visually faithful images. Depth and CNOT count decrease by over $B$3. At $B$4 (50% retained), MSE vanishes, indicating perfect recovery.

For amplitude-encoded (FRQI/QPIE) schemes, the impact is even more dramatic, with over $B$5 compression in circuit resource requirements and negligible loss in visual quality at small $B$6.

4. State Preparation Circuits and Synthesis Algorithms

Full-rank NEQR synthesis requires $B$7 multi-controlled NOTs (each decomposed into $B$8 CNOTs and basis gates), with complexity $B$9. With Schmidt LRA, the preparation of the truncated rank-$n_A$0 state is achieved via recursive synthesis of the $n_A$1 and $n_A$2 blocks of the SVD, followed by controlled state assembly, at $n_A$3 to $n_A$4 complexity per standard LRSP (Low-Rank State Preparation) protocol.

The construction is as follows:

1. Reshape the amplitude vector to a $n_A$5 matrix for a balanced partition.
2. Compute the SVD to obtain the dominant $n_A$6 singular components $n_A$7.
3. Synthesize each $n_A$8 separately, followed by linear combination controlled by ancillary index qubits.
4. Normalize and assign Schmidt weights via additional multiplexers or Givens rotations.

5. Performance—Fidelity, Resource Trade-Offs, and Applicability

Low-rank truncation introduces a controllable trade-off:

• Resource efficiency: Significantly reduced depth and CNOT count; if only a few significant Schmidt coefficients exist (as for smooth or compressible data), $n_A$9 suffices.
• Accuracy: Truncation error is bounded by the discarded Schmidt weight; empirical MSE in typical images is well below 1 gray level for moderate $n_B$0.
• Determinism: NEQR and its LRA variant retain deterministic intensity retrieval: a single projective measurement yields the gray value at a given pixel in basis basis.
• NISQ compatibility: Even with aggressive LRA, NEQR requires more qubits and deeper circuits than FRQI/QPIE for a given $n_B$1, but the method makes high-fidelity synthesis feasible for images up to $n_B$2 on future NISQ machines.

Plausibly, natural images with substantial local redundancy benefit the most from this approach, since their entanglement spectrum decays rapidly, just as in classical low-rank image compression.

6. Relation to Other Quantum Image Compression and Optimization Schemes

Whereas recent circuit-level gate optimizations (such as ESOP-to-PPRM conversion (Iranmanesh et al., 2024)) reduce the number of multi-controlled gates without modifying the logical state, Schmidt decomposition-based LRA fundamentally approximates the quantum state by discarding negligible entanglement sectors. Notably, both approaches are compatible and potentially synergistic: apply PPRM circuit simplification to each kept Schmidt term for further reductions.

Compared to blockwise or hierarchical encoding strategies (e.g., SCMNEQR (Haque et al., 2022)), Schmidt LRA attacks global image structure and can adapt to arbitrary correlation patterns, not just blockwise sparsity.

7. Outlook, Limitations, and Future Research Directions

Schmidt decomposition-based low-rank approximation enables tunable resource/fidelity trade-offs for quantum state preparation in image processing and beyond (Pangeva et al., 9 Jun 2026). It is optimal in terms of entanglement retention for a fixed circuit budget. However, the classical preprocessing cost (full SVD on $n_B$3-sized amplitude vectors) remains exponential in qubit number, which may limit its use to moderate-scale images unless approximate SVD or randomized projection methods are deployed.

Potential advances include (i) hierarchical or multi-scale LRA using tensor networks; (ii) hybrid schemes combining blockwise, PPRM, and LRA methods; (iii) adaptive dynamic truncation for streaming data. In image quantum processing, Schmidt-LRA holds promise for both data upload and lossy/information-theoretic quantum image coding.

---

References

• Schmidt decomposition-based methods for efficient quantum image encoding (Pangeva et al., 9 Jun 2026)
• Gate optimization of NEQR quantum circuits via PPRM transformation (Iranmanesh et al., 2024)
• Block-wise quantum grayscale image representation and compression scheme using state connection (Haque et al., 2022)