Solve the Market Split problem with Kipu Quantum's Iskay Quantum Optimizer

번역 미완료

이 페이지는 아직 한국어로 번역되지 않았습니다. 영어 원본을 보고 있습니다.

Note

Qiskit Functions are an experimental feature available only to IBM Quantum® Premium Plan, Flex Plan, and On-Prem (via IBM Quantum Platform API) Plan users. They are in preview release status and subject to change.

Usage estimate: 20 seconds on a Heron r2 processor. (NOTE: This is an estimate only. Your runtime might vary.)

Background

This tutorial demonstrates how to solve the Market Split problem using Kipu Quantum's Iskay quantum optimizer [1]. The Market Split problem represents a real-world resource allocation challenge where markets must be partitioned into balanced sales regions to meet exact demand targets.

The Market Split challenge

The Market Split problem presents a deceptively simple yet computationally formidable challenge in resource allocation. Consider a company with $m$ products being sold across $n$ different markets, where each market purchases a specific bundle of products (represented by the columns of matrix $A$). The business objective is to partition these markets into two balanced sales regions such that each region receives exactly half the total demand for every product.

Mathematical formulation:

We seek a binary assignment vector $x$, where:

• $x_j = 1$ assigns market $j$ to Region A
• $x_j = 0$ assigns market $j$ to Region B
• The constraint $Ax = b$ must be satisfied, where $b$ represents the target sales (typically half the total demand per product)

Cost function:

To solve this problem, we minimize the squared constraint violation:

$C(x) = ||Ax - b||^2 = \sum_{i=1}^{m} \left(\sum_{j=1}^{n} A_{ij}x_j - b_i\right)^2$

where:

• $A_{ij}$ represents the sales of product $i$ in market $j$
• $x_j \in \{0,1\}$ is the binary assignment of market $j$
• $b_i$ is the target sales for product $i$ in each region
• The cost equals zero precisely when all constraints are satisfied

Each term in the sum represents the squared deviation from the target sales for a particular product. When we expand this cost function, we get:

$C(x) = x^T A^T A x - 2b^T A x + b^T b$

Since $b^T b$ is a constant, minimizing $C(x)$ is equivalent to minimizing the quadratic function $x^T A^T A x - 2b^T A x$, which is exactly a QUBO (Quadratic Unconstrained Binary Optimization) problem.

Computational complexity:

Despite its straightforward business interpretation, this problem exhibits remarkable computational intractability:

• Small-scale failure: Conventional Mixed Integer Programming solvers fail on instances with as few as seven products under a timeout of one hour [4]
• Exponential growth: The solution space grows exponentially ($2^n$ possible assignments), making brute force approaches infeasible

This severe computational barrier, combined with its practical relevance to territory planning and resource allocation, makes the Market Split problem an ideal benchmark for quantum optimization algorithms [4].

What makes Iskay's approach unique?

The Iskay optimizer uses the bf-DCQO (bias-field digitized counterdiabatic quantum optimization) algorithm [1], which represents a significant advancement in quantum optimization:

Circuit efficiency: The bf-DCQO algorithm achieves remarkable gate reduction [1]:

• Up to 10 times fewer entangling gates than Digital Quantum Annealing (DQA)
• Significantly shallower circuits enable:
  • Less error accumulation during quantum execution
  • Ability to tackle larger problems on current quantum hardware
  • No need for error mitigation techniques

Non-variational design: Unlike variational algorithms requiring approximately 100 iterations, bf-DCQO typically needs only approximately 10 iterations [1]. This is achieved through:

• Intelligent bias-field calculations from measured state distributions
• Starting each iteration from an energy state near the previous solution
• Integrated classical post-processing with local search

Counterdiabatic protocols: The algorithm incorporates counterdiabatic terms that suppress unwanted quantum excitations during short evolution times, enabling the system to remain near the ground state even with rapid transitions [1].

Requirements

Before starting this tutorial, ensure that you have the following installed:

• Qiskit IBM Runtime (pip install qiskit-ibm-runtime)
• Qiskit Functions (pip install qiskit-ibm-catalog)
• NumPy (pip install numpy)
• Requests (pip install requests)
• Opt Mapper Qiskit addon (pip install qiskit-addon-opt-mapper)

You will also need to get access to the Iskay Quantum Optimizer function from the Qiskit Functions Catalog.

Setup

First, import all required packages for this tutorial.

# Added by doQumentation — required packages for this notebook
!pip install -q numpy qiskit-addon-opt-mapper qiskit-ibm-catalog requests

import os
import tempfile
import time
from typing import Tuple, Optional

import numpy as np
import requests

from qiskit_ibm_catalog import QiskitFunctionsCatalog

from qiskit_addon_opt_mapper import OptimizationProblem
from qiskit_addon_opt_mapper.converters import OptimizationProblemToQubo

print("All required libraries imported successfully")

Configure IBM Quantum credentials

Define your IBM Quantum® Platform credentials. You will need:

• API Token: Your 44-character API key from IBM Quantum Platform
• Instance CRN: Your IBM Cloud® instance identifier

token = "<YOUR_API_KEY>"
instance = "<YOUR_INSTANCE_CRN>"

Step 1: Map classical inputs to a quantum problem

We begin by mapping our classical problem to a quantum-compatible representation. This step involves:

1. Connecting to the Iskay Quantum Optimizer
2. Loading and formulating the Market Split problem
3. Understanding the bf-DCQO algorithm that will solve it

Connect to Iskay Quantum Optimizer

We begin by establishing a connection to the Qiskit Functions Catalog and loading the Iskay Quantum Optimizer. The Iskay Optimizer is a quantum function provided by Kipu Quantum that implements the bf-DCQO algorithm for solving optimization problems on quantum hardware.

catalog = QiskitFunctionsCatalog(token=token, instance=instance)
iskay_solver = catalog.load("kipu-quantum/iskay-quantum-optimizer")

print("Iskay optimizer loaded successfully")
print("Ready to solve optimization problems using bf-DCQO algorithm")

Load and formulate the problem

Understand the problem data format

Problem instances from QOBLIB (Quantum Optimization Benchmarking Library) [2] are stored in a simple text format. Let's examine the actual content of our target instance ms_03_200_177.dat:

3 20
60   92  161   53   97    2   75   81    6  139  132   45  108  112  181   93  152  200  164   51 1002
176  196   41  143    2   88    0   79   10   71   75  148   82  135   34  187   33  155   58   46  879
68   68  179  173  127  163   48   49   99   78   44   52  173  131   73  198   84  109  180   95 1040

Format structure:

• First line: 3 20

  • 3 = number of products (constraints/rows in matrix $A$)
  • 20 = number of markets (variables/columns in matrix $A$)

• Next 3 lines: Coefficient matrix $A$ and target vector $b$

  • Each line has 21 numbers: first 20 are row coefficients, last is the target
  • Line 2: 60 92 161 ... 51 | 1002
    • First 20 numbers: How much of Product 1 each of the 20 markets sells
    • Last number (1002): Target sales for Product 1 in one region
  • Line 3: 176 196 41 ... 46 | 879
    • Product 2 sales per market and target (879)
  • Line 4: 68 68 179 ... 95 | 1040
    • Product 3 sales per market and target (1040)

Business interpretation:

• Market 0 sells: 60 units of Product 1, 176 units of Product 2, 68 units of Product 3
• Market 1 sells: 92 units of Product 1, 196 units of Product 2, 68 units of Product 3
• And so on for all 20 markets...
• Goal: Split these 20 markets into two regions where each region gets exactly 1002 units of Product 1, 879 units of Product 2, and 1040 units of Product 3

QUBO transformation

From constraints to QUBO: the mathematical transformation

The power of quantum optimization lies in transforming constrained problems into unconstrained quadratic forms [4]. For the Market Split problem, we convert the equality constraints

$Ax = b$

where $x ∈ \{0,1\}^n$, into a QUBO by penalizing constraint violations.

The penalty method: Since we need $Ax = b$ to hold exactly, we minimize the squared violation: $f(x) = ||Ax - b||^2$

This equals zero precisely when all constraints are satisfied. Expanding algebraically: $f(x) = (Ax - b)^T(Ax - b) = x^T A^T A x - 2b^T A x + b^T b$

QUBO objective: Since $b^T b$ is constant, our optimization becomes: $\text{minimize} \quad Q(x) = x^T(A^T A)x - 2(A^T b)^T x$

Key insight: This transformation is exact, not approximate. Equality constraints naturally square into quadratic form without requiring auxiliary variables or penalty parameters - making this formulation mathematically elegant and computationally efficient for quantum solvers [4]. We'll use the OptimizationProblem class to define our constrained problem, then convert it to QUBO format using OptimizationProblemToQubo, both from the qiskit_addon_opt_mapper package. This automatically handles the penalty-based transformation.

Implement data loading and QUBO conversion functions

We now define three utility functions:

1. parse_marketsplit_dat() - Parses the .dat file format and extracts matrices $A$ and $b$
2. fetch_marketsplit_data() - Downloads problem instances directly from the QOBLIB repository

def parse_marketsplit_dat(filename: str) -> Tuple[np.ndarray, np.ndarray]:
    """
    Parse a market split problem from a .dat file format.

    Parameters
    ----------
    filename : str
        Path to the .dat file containing the market split problem data.

    Returns
    -------
    A : np.ndarray
        Coefficient matrix of shape (m, n) where m is the number of products
        and n is the number of markets.
    b : np.ndarray
        Target vector of shape (m,) containing the target sales per product.
    """
    with open(filename, "r", encoding="utf-8") as f:
        lines = [
            line.strip()
            for line in f
            if line.strip() and not line.startswith("#")
        ]

    if not lines:
        raise ValueError("Empty or invalid .dat file")

    # First line: m n (number of products and markets)
    m, n = map(int, lines[0].split())

    # Next m lines: each row of A followed by corresponding element of b
    A, b = [], []
    for i in range(1, m + 1):
        values = list(map(int, lines[i].split()))
        A.append(values[:-1])  # First n values: product sales per market
        b.append(values[-1])  # Last value: target sales for this product

    return np.array(A, dtype=np.int32), np.array(b, dtype=np.int32)

def fetch_marketsplit_data(
    instance_name: str = "ms_03_200_177.dat",
) -> Tuple[Optional[np.ndarray], Optional[np.ndarray]]:
    """
    Fetch market split data directly from the QOBLIB repository.

    Parameters
    ----------
    instance_name : str
        Name of the .dat file to fetch (default: "ms_03_200_177.dat").

    Returns
    -------
    A : np.ndarray or None
        Coefficient matrix if successful, None if failed.
    b : np.ndarray or None
        Target vector if successful, None if failed.
    """
    url = f"https://git.zib.de/qopt/qoblib-quantum-optimization-benchmarking-library/-/raw/main/01-marketsplit/instances/{instance_name}"

    try:
        response = requests.get(url, timeout=30)
        response.raise_for_status()

        with tempfile.NamedTemporaryFile(
            mode="w", suffix=".dat", delete=False, encoding="utf-8"
        ) as f:
            f.write(response.text)
            temp_path = f.name

        try:
            return parse_marketsplit_dat(temp_path)
        finally:
            os.unlink(temp_path)
    except Exception as e:
        print(f"Error: {e}")
        return None, None

Load the problem instance

Now we load the specific problem instance ms_03_200_177.dat from QOBLIB [2]. This instance has:

• 3 products (constraints)
• 20 markets (binary decision variables)
• Over 1 million possible market assignments to explore ($2^{20} = 1,048,576$)

# Load the problem instance
instance_name = "ms_03_200_177.dat"
A, b = fetch_marketsplit_data(instance_name=instance_name)

if A is not None:
    print("Successfully loaded problem instance from QOBLIB")
    print("\nProblem Instance Analysis:")
    print("=" * 50)
    print(f"Coefficient Matrix A: {A.shape[0]} × {A.shape[1]}")
    print(f"   → {A.shape[0]} products (constraints)")
    print(f"   → {A.shape[1]} markets (decision variables)")
    print(f"Target Vector b: {b}")
    print("   → Target sales per product for each region")
    print(
        f"Solution Space: 2^{A.shape[1]} = {2**A.shape[1]:,} possible assignments"
    )

Convert to QUBO format

We now transform the constrained optimization problem into QUBO format:

# Create optimization problem
ms = OptimizationProblem(instance_name.replace(".dat", ""))

# Add binary variables (one for each market)
ms.binary_var_list(A.shape[1])

# Add equality constraints (one for each product)
for idx, rhs in enumerate(b):
    ms.linear_constraint(A[idx, :], sense="==", rhs=rhs)

# Convert to QUBO with penalty parameter
qubo = OptimizationProblemToQubo(penalty=1).convert(ms)

print("QUBO Conversion Complete:")
print("=" * 50)
print(f"Number of variables: {qubo.get_num_vars()}")
print(f"Constant term: {qubo.objective.constant}")
print(f"Linear terms: {len(qubo.objective.linear.to_dict())}")
print(f"Quadratic terms: {len(qubo.objective.quadratic.to_dict())}")

Convert QUBO to Iskay format

Now we need to convert the QUBO object into the dictionary format required by Kipu Quantum's Iskay Optimizer.

The problem and problem_type arguments encode an optimization problem of the form

$$\begin{align} \min_{(x_1, x_2, \ldots, x_n) \in D} C(x_1, x_2, \ldots, x_n) \nonumber \end{align}$$

where

$$C(x_1, ... , x_n) = a + \sum_{i} b_i x_i + \sum_{i, j} c_{i, j} x_i x_j + ... + \sum_{k_1, ..., k_m} g_{k_1, ..., k_m} x_{k_1} ... x_{k_m}$$

• By choosing problem_type = "binary", you specify that the cost function is in binary format, which means that $D = \{0, 1\}^{n}$, as in, the cost function is written in QUBO/HUBO formulation.
• On the other hand, by choosing problem_type = "spin", the cost function is written in Ising formulation, where $D = \{-1, 1\}^{n}$.

The coefficients of the problem should be encoded in a dictionary as follows:

$$\begin{align} \nonumber &\texttt{\{} \\ \nonumber &\texttt{"()"}&: \quad &a, \\ \nonumber &\texttt{"(i,)"}&: \quad &b_i, \\ \nonumber &\texttt{"(i, j)"}&: \quad &c_{i, j}, \quad (i \neq j) \\ \nonumber &\quad \vdots \\ \nonumber &\texttt{"(} k_1, ..., k_m \texttt{)"}&: \quad &g_{k_1, ..., k_m}, \quad (k_1 \neq k_2 \neq \dots \neq k_m) \\ \nonumber &\texttt{\}} \end{align}$$

Note that the keys of the dictionary must be strings containing a valid tuple of non-repeating integers. For binary problems, we know that:

$$x_i^2 = x_i$$

for $i=j$ (since $x_i \in \{0,1\}$ means $x_i \cdot x_i = x_i$). So, in your QUBO formulation, if you have both linear contributions $b_i x_i$ and diagonal quadratic contributions $c_{i,i} x_i^2$, these terms must be combined into a single linear coefficient:

Total linear coefficient for variable $x_i$: $b_i + c_{i,i}$

This means:

• Linear terms like "(i, )" contain: original linear coefficient + diagonal quadratic coefficient
• Diagonal quadratic terms like "(i, i)" should NOT appear in the final dictionary
• Only off-diagonal quadratic terms like "(i, j)" where $i \neq j$ should be included as separate entries

Example: If your QUBO has $3x_1 + 2x_1^2 + 4x_1 x_2$, the Iskay dictionary should contain:

• "(0, )": 5.0 (combining $3 + 2 = 5$)
• "(0, 1)": 4.0 (off-diagonal term)

NOT separate entries for "(0, )": 3.0 and "(0, 0)": 2.0.

# Convert QUBO to Iskay dictionary format:

# Create empty Iskay input dictionary
iskay_input_problem = {}

# Convert QUBO to Iskay dictionary format
iskay_input_problem = {"()": qubo.objective.constant}

for i in range(qubo.get_num_vars()):
    for j in range(i, qubo.get_num_vars()):
        if i == j:
            # Add linear term (including diagonal quadratic contribution)
            iskay_input_problem[f"({i}, )"] = float(
                qubo.objective.linear.to_dict().get(i)
            ) + float(qubo.objective.quadratic.to_dict().get((i, i)))
        else:
            # Add off-diagonal quadratic term
            iskay_input_problem[f"({i}, {j})"] = float(
                qubo.objective.quadratic.to_dict().get((i, j))
            )

# Display Iskay dictionary summary
print("Iskay Dictionary Format:")
print("=" * 50)
print(f"Total coefficients: {len(iskay_input_problem)}")
print(f"  • Constant term: {iskay_input_problem['()']}")
print(
    f"  • Linear terms: {sum(1 for k in iskay_input_problem.keys() if k != '()' and ', )' in k)}"
)
print(
    f"  • Quadratic terms: {sum(1 for k in iskay_input_problem.keys() if k != '()' and ', )' not in k)}"
)
print("\nSample coefficients:")

# Get first 10 and last 5 items properly
items = list(iskay_input_problem.items())
first_10 = list(enumerate(items[:10]))
last_5 = list(enumerate(items[-5:], start=len(items) - 5))

for i, (key, value) in first_10 + last_5:
    coeff_type = (
        "constant"
        if key == "()"
        else "linear"
        if ", )" in key
        else "quadratic"
    )
    print(f"  {key}: {value} ({coeff_type})")
print("  ...")
print("\n✓ Problem ready for Iskay optimizer!")

Understand the bf-DCQO algorithm

Before we run the optimization, let's understand the sophisticated quantum algorithm that powers Iskay: bf-DCQO (bias-field digitized counterdiabatic quantum optimization) [1].

What is bf-DCQO?

bf-DCQO is based on the time evolution of a quantum system where the problem solution is encoded in the ground state (lowest energy state) of the final quantum Hamiltonian [1]. The algorithm addresses a fundamental challenge in quantum optimization:

The challenge: Traditional adiabatic quantum computing requires very slow evolution to maintain ground state conditions according to the adiabatic theorem. This demands increasingly deep quantum circuits as problem complexity grows, leading to more gate operations and accumulated errors.

The solution: bf-DCQO uses counterdiabatic protocols to enable rapid evolution while maintaining ground state fidelity, dramatically reducing circuit depth.

Mathematical framework

The algorithm minimizes a cost function of the form:

$\min_{(x_1,x_2,...,x_n) \in D} C(x_1,x_2,...,x_n)$

where $D = \{0,1\}^n$ for binary variables and:

$C(x) = a + \sum_i b_i x_i + \sum_{i,j} c_{ij} x_i x_j + ... + \sum g_{k_1,...,k_m} x_{k_1}...x_{k_m}$

For our Market Split problem, the cost function is:

$C(x) = ||Ax - b||^2 = x^T A^T A x - 2 b^T A x + b^T b$

The role of counterdiabatic terms

Counterdiabatic terms are additional terms introduced into the time-dependent Hamiltonian that suppress unwanted excitations during the quantum evolution. Here's why they're crucial:

In adiabatic quantum optimization, we evolve the system according to a time-dependent Hamiltonian:

$H(t) = \left(1 - \frac{t}{T}\right) H_{\text{initial}} + \frac{t}{T} H_{\text{problem}}$

where $H_{\text{problem}}$ encodes our optimization problem. To maintain the ground state during rapid evolution, we add counterdiabatic terms:

$H_{\text{CD}}(t) = H(t) + H_{\text{counter}}(t)$

These counterdiabatic terms do the following:

1. Suppress unwanted transitions: Prevent the quantum state from jumping to excited states during fast evolution
2. Enable shorter evolution times: Allow us to reach the final state much faster without violating adiabaticity
3. Reduce circuit depth: Shorter evolution leads to fewer gates and less error

The practical impact is dramatic: bf-DCQO uses up to 10 times fewer entangling gates than Digital Quantum Annealing [1], making it practical for today's noisy quantum hardware.

Bias-field iterative optimization

Unlike variational algorithms that optimize circuit parameters through many iterations, bf-DCQO uses a bias-field guided approach that converges in approximately 10 iterations [1]:

Iteration process:

1. Initial quantum evolution: Start with a quantum circuit implementing the counterdiabatic evolution protocol

2. Measurement: Measure the quantum state to obtain a probability distribution over bitstrings

3. Bias-field calculation: Analyze the measurement statistics and calculate an optimal bias-field $h_i$ for each qubit: $h_i = \text{f}(\text{measurement statistics}, \text{previous solutions})$

4. Next iteration: The bias-field modifies the Hamiltonian for the next iteration: $H_{\text{next}} = H_{\text{problem}} + \sum_i h_i \sigma_i^z$

   This allows starting near the previously found good solution, effectively performing a form of "quantum local search"

5. Convergence: Repeat until the solution quality stabilizes or a maximum number of iterations is reached

Key advantage: Each iteration provides meaningful progress toward the optimal solution by incorporating information from previous measurements, unlike variational methods that must explore the parameter space blindly.

Integrated classical post-processing

After the quantum optimization converges, Iskay performs classical local search post-processing:

• Bit-flip exploration: Systematically or randomly flip bits in the best measured solution
• Energy evaluation: Calculate $C(x)$ for each modified solution
• Greedy selection: Accept improvements that lower the cost function
• Multiple passes: Perform several passes (controlled by postprocessing_level)

This hybrid approach compensates for bit-flip errors from hardware imperfections and readout errors, ensuring high-quality solutions even on noisy quantum devices.

Why bf-DCQO excels on current hardware

The bf-DCQO algorithm is specifically designed to excel on today's noisy intermediate-scale quantum (NISQ) devices [1]:

1. Error resilience: Fewer gates (10 times reduction) means dramatically less error accumulation
2. No error mitigation required: The algorithm's inherent efficiency eliminates the need for expensive error mitigation techniques [1]
3. Scalability: Can handle problems with up to 156 qubits (156 binary variables) with direct qubit-mapping [1]
4. Proven performance: Achieves 100% approximation ratios on benchmark MaxCut and HUBO instances [1]

Now let's see this powerful algorithm in action on our Market Split problem!

Step 2: Optimize problem for quantum hardware execution

The bf-DCQO algorithm automatically handles circuit optimization, creating shallow quantum circuits with counterdiabatic terms specifically designed for the target backend.

Configure the optimization

The Iskay Optimizer requires several key parameters to effectively solve your optimization problem. Let's examine each parameter and its role in the quantum optimization process:

Required parameters

Parameter	Type	Description	Example
problem	Dict[str, float]	QUBO coefficients in string-key format	{"()": -21.0, "(0,4)": 0.5, "(0,1)": 0.5}
problem_type	str	Format specification: "binary" for QUBO or "spin" for Ising	"binary"
backend_name	str	Target quantum device	"ibm_fez"

Essential concepts

• Problem format: We use "binary" since our variables are binary (0/1), representing market assignments.
• Backend selection: Choose between the available QPUs (for example, "ibm_fez") based on your needs and compute resource instance.
• QUBO structure: Our problem dictionary contains the exact coefficients from the mathematical transformation.

Advanced options (optional)

Iskay provides fine-tuning capabilities through optional parameters. While the defaults work well for most problems, you can customize the behavior for specific requirements:

Parameter	Type	Default	Description
shots	int	10000	Quantum measurements per iteration (higher = more accurate)
num_iterations	int	10	Algorithm iterations (more iterations can improve solution quality)
use_session	bool	True	Use IBM sessions for reduced queue times
seed_transpiler	int	None	Set for reproducible quantum circuit compilation
direct_qubit_mapping	bool	False	Map virtual qubits directly to physical qubits
job_tags	List[str]	None	Custom tags for job tracking
preprocessing_level	int	0	Problem preprocessing intensity (0-3) - see details below
postprocessing_level	int	2	Solution refinement level (0-2) - see details below
transpilation_level	int	0	Transpiler optimization trials (0-5) - see details below
transpile_only	bool	False	Analyze circuit optimization without running full execution

Preprocessing Levels (0-3): Specially important for larger problems that cannot currently fit on the coherence times of the hardware. Higher preprocessing levels achieve shallower circuit depths by approximations in the problem transpilation:

• Level 0: Exact, longer circuits
• Level 1: Good balance between accuracy and approximation, cutting out only the gates with angles in the lowest 10 percentile
• Level 2: Slightly higher approximation, cutting out the gates with angles in the lowest 20 percentile and using approximation_degree=0.95 in the transpilation
• Level 3: Maximum approximation level, cutting out the gates in the lowest 30 percentile and using approximation_degree=0.90 in the transpilation

Transpilation Levels (0-5): Control the advanced transpiler optimization trials for quantum circuit compilation. This can lead to an increase in classical overhead, and for some cases it might not change the circuit depth. The default value 2 in general leads to the smallest circuit and is relatively fast.

• Level 0: Optimization of the decomposed DCQO circuit (layout, routing, scheduling)
• Level 1: Optimization of PauliEvolutionGate and then the decomposed DCQO circuit (max_trials=10)
• Level 2: Optimization of PauliEvolutionGate and then the decomposed DCQO circuit (max_trials=15)
• Level 3: Optimization of PauliEvolutionGate and then the decomposed DCQO circuit (max_trials=20)
• Level 4: Optimization of PauliEvolutionGate and then the decomposed DCQO circuit (max_trials=25)
• Level 5: Optimization of PauliEvolutionGate and then the decomposed DCQO circuit (max_trials=50)

Postprocessing Levels (0-2): Control how much classical optimization, compensating for bit-flip errors with different number of greedy passes of a local search:

• Level 0: 1 pass
• Level 1: 2 passes
• Level 2: 3 passes

Transpile-only mode: Now available for users who want to analyze circuit optimization without running the full quantum algorithm execution.

Custom configuration example

Here's how you might configure Iskay with different settings:

custom_options = {
    "shots": 15_000,                                        # Higher shot count for better statistics
    "num_iterations": 12,                                   # More iterations for solution refinement
    "preprocessing_level": 1,                               # Light preprocessing for problem simplification
    "postprocessing_level": 2,                              # Maximum postprocessing for solution quality
    "transpilation_level": 3,                               # Using higher transpilation level for circuit optimization
    "seed_transpiler": 42,                                  # Fixed seed for reproducible results
    "job_tags": ["market_split"]                            # Custom tracking tags
}

For this tutorial, we will keep most of the default parameters and will only change the number of bias-field iterations:

# Specify the target backend
backend_name = "ibm_fez"

# Set the number of bias-field iterations and set a tag to identify the jobs
options = {
    "num_iterations": 3,  # Change number of bias-field iterations
    "job_tags": ["market_split_example"],  # Tag to identify jobs
}

# Configure Iskay optimizer
iskay_input = {
    "problem": iskay_input_problem,
    "problem_type": "binary",
    "backend_name": backend_name,
    "options": options,
}

print("Iskay Optimizer Configuration:")
print("=" * 40)
print(f"  Backend: {backend_name}")
print(f"  Problem: {len(iskay_input['problem'])} terms")
print("  Algorithm: bf-DCQO")

Step 3: Execute using Qiskit primitives

We now submit our problem to run on IBM Quantum hardware. The bf-DCQO algorithm will:

1. Construct shallow quantum circuits with counterdiabatic terms
2. Execute approximately 10 iterations with bias-field optimization
3. Perform classical post-processing with local search
4. Return the optimal market assignment

# Submit the optimization job
print("Submitting optimization job to Kipu Quantum...")
print(
    f"Problem size: {A.shape[1]} variables, {len(iskay_input['problem'])} terms"
)
print(
    "Algorithm: bf-DCQO (bias-field digitized counterdiabatic quantum optimization)"
)

job = iskay_solver.run(**iskay_input)

print("\nJob successfully submitted!")
print(f"Job ID: {job.job_id}")
print("Optimization in progress...")
print(
    f"The bf-DCQO algorithm will efficiently explore {2**A.shape[1]:,} possible assignments"
)

Monitor job status

You can check the current status of your optimization job. The possible statuses are:

• QUEUED: Job is waiting in the queue
• RUNNING: Job is currently executing on quantum hardware
• DONE: Job completed successfully
• CANCELED: Job was canceled
• ERROR: Job encountered an error

# Check job status
print(f"Job status: {job.status()}")

Wait for completion

This cell will block until the job completes. The optimization process includes:

• Queue time (waiting for quantum hardware access)
• Execution time (running the bf-DCQO algorithm with approximately 10 iterations)
• Post-processing time (classical local search)

Typical completion times range from a few minutes to tens of minutes depending on queue conditions.

# Wait for job completion
while True:
    status = job.status()
    print(
        f"Waiting for job {job.job_id} to complete... (status: {status})",
        end="\r",
        flush=True,
    )
    if status in ["DONE", "CANCELED", "ERROR"]:
        print(
            f"\nJob {job.job_id} completed with status: {status}" + " " * 20
        )
        break
    time.sleep(30)

# Retrieve the optimization results
result = job.result()
print("\nOptimization complete!")

Step 4: Post-process and return result in desired classical format

We now post-process the quantum execution results. This includes:

• Analyzing the solution structure
• Validating constraint satisfaction
• Benchmarking against classical approaches

Analyze results

Understand the result structure

Iskay returns a comprehensive result dictionary containing:

• solution: A dictionary mapping variable indices to their optimal values (0 or 1)
• solution_info: Detailed information including:
  • bitstring: The optimal assignment as a binary string
  • cost: The objective function value (should be 0 for perfect constraint satisfaction)
  • mapping: How bitstring positions map to problem variables
  • seed_transpiler: Seed used for reproducibility
• prob_type: Whether the solution is in binary or spin format

Let's examine the solution returned by the quantum optimizer.

# Display the optimization results
print("Optimization Results")
print("=" * 50)
print(f"Problem Type: {result['prob_type']}")
print("\nSolution Info:")
print(f"  Bitstring: {result['solution_info']['bitstring']}")
print(f"  Cost: {result['solution_info']['cost']}")
print("\nSolution (first 10 variables):")
for i, (var, val) in enumerate(list(result["solution"].items())[:10]):
    print(f"  {var}: {val}")
print("  ...")

Solution validation

Now we validate whether the quantum solution satisfies the Market Split constraints. The validation process checks:

What is a constraint violation?

• For each product $i$, we calculate the actual sales in Region A: $(Ax)_i$
• We compare this to the target sales $b_i$
• The violation is the absolute difference: $|(Ax)_i - b_i|$
• A feasible solution has zero violations for all products

What we expect:

• Ideal case: Total violation = 0 (all constraints perfectly satisfied)
  • Region A gets exactly 1002 units of Product 1, 879 units of Product 2, and 1040 units of Product 3
  • Region B gets the remaining units (also 1002, 879, and 1040 respectively)
• Good case: Total violation is small (near-optimal solution)
• Poor case: Large violations indicate the solution doesn't satisfy the business requirements

The validation function will compute:

1. Actual sales per product in each region
2. Constraint violations for each product
3. Market distribution between regions

def validate_solution(A, b, solution):
    """Validate market split solution."""
    x = np.array(solution)
    region_a = A @ x
    region_b = A @ (1 - x)
    violations = np.abs(region_a - b)

    return {
        "target": b,
        "region_a": region_a,
        "region_b": region_b,
        "violations": violations,
        "total_violation": np.sum(violations),
        "is_feasible": np.sum(violations) == 0,
        "region_a_markets": int(np.sum(x)),
        "region_b_markets": len(x) - int(np.sum(x)),
    }

# Convert bitstring to list of integers and validate
optimal_assignment = [
    int(bit) for bit in result["solution_info"]["bitstring"]
]
validation = validate_solution(A, b, optimal_assignment)

Interpret the validation results

The validation results show whether the Quantum Optimizer found a feasible solution. Let's examine the following:

Feasibility check:

• is_feasible = True means the solution perfectly satisfies all constraints (total violation = 0)
• is_feasible = False means some constraints are violated

Sales analysis:

• Compare Target versus Actual sales for each product
• For a perfect solution: Actual = Target for all products in both regions
• The difference indicates how close we are to the desired market split

Market distribution:

• Shows how many markets are assigned to each region
• There's no requirement for equal numbers of markets, only that sales targets are met

print("Solution Validation")
print("=" * 50)
print(f"Feasible solution: {validation['is_feasible']}")
print(f"Total constraint violation: {validation['total_violation']}")

print("\nSales Analysis (Target vs Actual):")
for i, (target, actual_a, actual_b) in enumerate(
    zip(validation["target"], validation["region_a"], validation["region_b"])
):
    violation_a = abs(actual_a - target)
    violation_b = abs(actual_b - target)
    print(f"  Product {i+1}:")
    print(f"    Target: {target}")
    print(f"    Region A: {actual_a} (violation: {violation_a})")
    print(f"    Region B: {actual_b} (violation: {violation_b})")

print("\nMarket Distribution:")
print(f"  Region A: {validation['region_a_markets']} markets")
print(f"  Region B: {validation['region_b_markets']} markets")

Solution quality assessment

Based on the validation results above, we can assess the quality of the quantum solution:

If is_feasible = True (Total violation = 0):

• The Quantum Optimizer successfully found an optimal solution
• All business constraints are perfectly satisfied
• This demonstrates quantum advantage on a problem where classical solvers struggle [4]

If is_feasible = False (Total violation > 0):

• The solution is near-optimal but not perfect
• Small violations may be acceptable in practice
• Consider adjusting optimizer parameters:
  • Increase num_iterations for more optimization passes
  • Increase postprocessing_level for more classical refinement
  • Increase shots for better measurement statistics

Cost function interpretation:

• The cost value from solution_info equals $||Ax - b||^2$
• Cost = 0 indicates perfect constraint satisfaction
• Higher cost values indicate larger constraint violations

Conclusion

What we accomplished

In this tutorial, we successfully:

1. Loaded a real optimization problem: Obtained a challenging Market Split instance from the QOBLIB benchmark library [2]
2. Transformed to QUBO format: Converted the constrained problem into an unconstrained quadratic formulation [3]
3. Leveraged advanced quantum algorithms: Used Kipu Quantum's bf-DCQO algorithm with counterdiabatic terms [1]
4. Obtained optimal solutions: Found feasible solutions satisfying all constraints

Key takeaways

Algorithm innovation: The bf-DCQO algorithm represents a significant advancement [1]:

• 10 times fewer gates than digital quantum annealing
• Approximately 10 iterations instead of approximately 100 for variational methods
• Built-in error resilience through circuit efficiency

Counterdiabatic terms: Enable rapid quantum evolution while maintaining ground state fidelity, making quantum optimization practical on today's noisy hardware [1].

Bias-field guidance: The iterative bias-field approach allows each iteration to start near previously found good solutions, providing a form of quantum-enhanced local search [1].

Next steps

To deepen your understanding and explore further:

1. Try different instances: Experiment with other QOBLIB instances of varying sizes
2. Tune parameters: Adjust num_iterations, preprocessing_level, postprocessing_level
3. Compare with classical: Benchmark against classical optimization solvers
4. Try different strategies: Try to find a better encoding for the problem or formulate it as HUBO (if possible)
5. Apply to your domain: Adapt the QUBO/HUBO formulation techniques to your own optimization problems

References

[1] IBM Quantum. "Kipu Quantum Optimization." IBM Quantum Documentation.

[2] QOBLIB - Quantum Optimization Benchmarking Library. Zuse Institute Berlin (ZIB). https://git.zib.de/qopt/qoblib-quantum-optimization-benchmarking-library

[3] Glover, F., Kochenberger, G., & Du, Y. (2019). "Quantum bridge analytics I: a tutorial on formulating and using QUBO models." 4OR: A Quarterly Journal of Operations Research, 17(4), 335-371.

[4] Lodi, A., Tramontani, A., & Weninger, K. (2023). "The Intractable Decathlon: Benchmarking Hard Combinatorial Problems." INFORMS Journal on Computing.