⚛️ quantum physics

Reducing C-NOT Counts for State Preparation and Block Encoding via Diagonal Matrix Migration

This paper introduces a diagonal matrix migration technique to reduce C-NOT gate counts for quantum state preparation and block encoding, achieving significant improvements over existing algorithms like Plesch-Brukner, including a leading term of $(11/12)2^n$ for general states and $(11/48)4^n$ for block encoding, which notably surpasses the theoretical lower bound for $n$ -qubit unitary synthesis.

Original authors: Zexian Li, Guofeng Zhang, Xiao-Ming Zhang

Published 2026-03-18

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Zexian Li, Guofeng Zhang, Xiao-Ming Zhang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to build a complex machine out of Lego bricks. In the world of quantum computing, these "bricks" are called gates, and the most expensive, difficult-to-manufacture brick is the C-NOT gate. It's like the heavy, reinforced steel beam of your Lego castle: you need it to hold things together, but using too many of them makes your castle slow, unstable, and expensive to build.

The paper you provided is about a new set of blueprints (algorithms) that allows engineers to build two specific, crucial parts of a quantum computer—State Preparation and Block Encoding—using significantly fewer of these heavy steel beams.

Here is the breakdown of their breakthrough, explained simply:

1. The Problem: The "Heavy Lifting" of Quantum Computers

Before a quantum computer can solve a problem (like simulating a new drug or optimizing a stock portfolio), it first needs to get the data into the machine.

State Preparation: This is like loading a specific pattern of colored marbles into a box. You need to arrange them perfectly to represent your data.
Block Encoding: This is like putting a complex 2D map into a 3D hologram projector so the computer can read it.

Currently, the standard methods for doing this are like trying to carry a piano up a staircase one brick at a time. They work, but they use way too many C-NOT gates (the heavy bricks), making the process inefficient.

2. The Solution: "Diagonal Matrix Migration"

The authors, Zexian Li and his team, invented a clever trick they call Diagonal Matrix Migration.

The Analogy: The Moving Truck
Imagine you are moving houses. You have a bunch of heavy furniture (C-NOT gates) and a bunch of light, flat boxes (Diagonal Matrices).

The Old Way: You try to carry the heavy furniture and the boxes separately. You make two trips. It takes a long time, and you use a lot of fuel (C-NOTs).
The New Way (Migration): The authors realized that the "flat boxes" (diagonal matrices) are special. They can slide under or through the heavy furniture without getting in the way. Because of this, you can move the heavy furniture and the boxes in a single trip, or even rearrange the furniture so you don't need to move it as much.

In technical terms, they found a way to slide these "diagonal" operations through the circuit so they cancel out or merge with other steps, effectively deleting the need for several heavy C-NOT gates.

3. The Results: Cutting the Weight in Half

By using this "migration" technique, they achieved two major victories:

For State Preparation (Loading the Data):
- Before: The best method (from 2011) required a certain number of heavy bricks.
- Now: They reduced the number of bricks needed by about 10-15%.
- Analogy: If you were building a wall that needed 100 heavy bricks, they found a way to build it with only 91. It sounds small, but in quantum computing, every single brick counts because the machine is so fragile.
For Block Encoding (The 3D Hologram):
- Before: The methods were very heavy, using a lot of bricks.
- Now: They reduced the brick count by nearly 50% for the main part of the calculation.
- The Shocking Part: They did this so efficiently that their method actually uses fewer bricks than the theoretical minimum required to build a perfect, generic 3D shape (a "unitary"). It's like building a house using fewer bricks than the laws of physics say are necessary for a generic house, because they found a specific shortcut for this type of house.

4. The "Low-Rank" Bonus

The paper also tackles a special case: Low-Rank Matrices.

Analogy: Imagine you are organizing a library. Most libraries have unique books on every shelf (Full-Rank). But some libraries have many copies of the same few books (Low-Rank).
The Trick: If you know the books are repetitive, you don't need to build a unique shelf for every single copy. You can stack them.
The Result: For these "repetitive" data sets, their method is even more efficient, scaling down the cost based on how repetitive the data is.

Why Does This Matter?

Think of a quantum algorithm as a race car.

The Engine: The math that solves the problem.
The Fuel: The C-NOT gates.

If your car uses too much fuel just to get to the starting line (State Preparation) or to load the map (Block Encoding), you run out of fuel before you can even race. By cutting the fuel consumption of these setup steps, the authors are giving quantum computers a much longer range. This means we can solve bigger, more complex problems in science and engineering before the computer "runs out of gas" (decoheres or makes errors).

In summary: This paper is a masterclass in efficiency. It takes the heavy, clumsy steps of setting up a quantum computer and replaces them with a sleek, sliding-door technique that saves massive amounts of resources, bringing us one step closer to practical, real-world quantum computing.

1. Problem Statement

Quantum algorithms for scientific computing rely heavily on two fundamental input models: Quantum State Preparation (encoding a vector into a quantum state) and Block Encoding (embedding a matrix into a larger unitary operator).

The Bottleneck: The circuit complexity, specifically the count of C-NOT gates (the primary source of error and latency in current quantum hardware), often dominates the end-to-end complexity of these algorithms.
The Gap: Existing methods, such as the Plesch–Brukner (PB) algorithm for state preparation and various block encoding protocols, have not yet reached the theoretical lower bounds for C-NOT counts. There is a significant gap between current leading constants and the theoretical minimums required by circuit topology.

2. Methodology

The authors propose a unified approach based on Diagonal Matrix Migration (DMM) combined with recursive Block-ZXZ decomposition.

Core Techniques:

Diagonal Matrix Migration (DMM):
- The authors exploit the commutativity properties of diagonal matrices ( $\Delta$ ) and uniformly controlled rotations about the z-axis ($UCRZ$).
- In standard synthesis, a diagonal matrix is often "trapped" between C-NOT gates. The DMM technique allows these diagonal matrices to be "migrated" (commuted) through C-NOT gates and $UCRZ$ layers.
- Once migrated, these diagonal matrices can be merged into the next recursive step or absorbed into subsequent unitary/isometry operations, effectively reducing the number of C-NOT gates required for the overall synthesis.
Recursive Block-ZXZ Decomposition:
- Building on the work of Krol et al., the authors use a recursive strategy to decompose $n$ -qubit unitaries and isometries.
- State Preparation: The algorithm reshapes the state vector, performs Singular Value Decomposition (SVD), and recursively prepares the Schmidt coefficients while synthesizing the basis vectors ( $U$ and $V$ ) only up to a diagonal matrix.
- Block Encoding: The matrix is decomposed into singular vectors and singular values. The synthesis of the singular vectors is optimized by treating them as isometries (partial unitaries) and applying the DMM technique to merge diagonal terms.
Isometry Synthesis Optimization:
- Unlike full unitary synthesis which requires three multiplexor operations, the authors demonstrate that synthesizing an isometry (mapping $2^{n-1} \to 2^n$ ) requires only two multiplexor operations when utilizing the diagonal migration technique.

3. Key Contributions

A. Optimized State Preparation (SPDMM)

Result: The authors present the State Preparation via Diagonal Matrix Migration (SPDMM) algorithm.
Improvement: They reduce the leading constant of the C-NOT count from $23/24$ (Plesch–Brukner, 2011) to $11/12$ .
Theorem 1: An arbitrary $n$ -qubit state can be prepared using at most $\frac{11}{12} 2^n$ C-NOT gates.
Lower Bound: They also establish a theoretical lower bound of $\frac{1}{2} 2^n$ for general state preparation.

B. Single Ancilla Block Encoding (SIABLE)

Result: A new protocol for Single Ancilla Block Encoding (SIABLE) that uses the spectral norm ( $\|A\|_2$ ) as the optimal scaling factor.
Improvement: For a full-rank $2^{n-1} \times 2^{n-1}$ matrix, the leading term of the C-NOT count is reduced to $11/48 \times 4^n$ .
Significance: This result is remarkable because the leading constant ( $11/48 \approx 0.229$ ) is lower than the established lower bound for general $n$ -qubit unitary synthesis ( $1/4 = 0.25$ ). This is possible because block encoding has weaker constraints than full unitary synthesis.
Theorem 2: A single-ancilla block encoding for a full-rank matrix uses at most $\frac{11}{48} 4^n - 2^n + \frac{7}{3}$ C-NOT gates.

C. Low-Rank Matrix Encoding

Result: A specialized protocol for rank- $K$ matrices, which are common in practical applications (e.g., recommendation systems, LLMs).
Improvement: The C-NOT count leading term is refined to $(K + 11/12) 2^n$ .
Theorem 5: For a rank- $K$ matrix, the encoding requires at most $(K + \frac{11}{12}) 2^n + O(Kn^2)$ C-NOT gates.

4. Results and Experimental Validation

The authors validated their algorithms using the QCLAB MATLAB library.

State Preparation (Table I):
- For $n=15$ qubits, the proposed SPDMM method requires 29,627 C-NOT gates.
- Comparison: Plesch–Brukner (PB) requires 38,813; Isometry synthesis requires 32,752.
- SPDMM achieves the lowest count, approaching the theoretical lower bound of 16,373.
Block Encoding (Table III):
- For $n=7$ (encoding a $64 \times 64$ matrix), SIABLE requires 3,629 C-NOT gates.
- Comparison: FABLE requires 8,192; BITBLE requires 8,190; QSD requires 7,660.
- SIABLE significantly outperforms existing methods and even surpasses the theoretical lower bound for unitary synthesis (which is 4,091 for $n=7$ ), confirming that block encoding is inherently more efficient than full unitary synthesis.
Low-Rank Performance (Table IV):
- The protocol shows substantial savings for low-rank matrices. For example, encoding a rank-1 matrix for $n=9$ requires only 654 gates, compared to 59,565 for a full-rank matrix.

5. Significance

Hardware Efficiency: By reducing the C-NOT count, the proposed methods directly lower the circuit depth and error accumulation, making quantum algorithms for linear algebra more feasible on Near-Term Intermediate-Scale Quantum (NISQ) devices and early fault-tolerant machines.
Theoretical Advancement: The work bridges the gap between practical circuit synthesis and theoretical limits. Specifically, achieving a constant lower than the unitary synthesis lower bound for block encoding highlights the unique efficiency gains possible when relaxing the "full unitary" constraint.
Scalability: The recursive nature of the algorithm and the specific optimization for low-rank matrices make these methods highly scalable for large-scale scientific computing applications where matrices are often sparse or low-rank.

In conclusion, this paper introduces a novel "Diagonal Matrix Migration" technique that fundamentally improves the efficiency of quantum circuit synthesis for state preparation and block encoding, setting new state-of-the-art benchmarks for C-NOT gate counts.