DR error measures: Difference between revisions

From Xenharmonic Reference
No edit summary
 
(96 intermediate revisions by the same user not shown)
Line 1: Line 1:
{{technical}}
{{technical}}
This article will describe several '''least-squares error measures for [[delta-rational chord]]s'''. They have the advantage of not fixing a particular interval in the chord when constructing the chord of best fit. However, like any other numerical measure of concordance or error, you should take them with a grain of salt.
This article will describe several '''least-squares error measures for [[delta-rational chord]]s'''. The idea behind least-squares error measures is to find the chord with the exact specified delta signature such that it deviates as little as possible from the chord that is meant to approximate that delta signature, and to measure the deviation. Least-squares error measures have the advantage of not fixing a particular interval in the chord when constructing the chord of best fit. However, like any other numerical measure of concordance or error, you should take them with a grain of salt.


== Rooted linear error ==
== Conventions and introduction ==
'''Rooted linear error''' (here ''linear'' means "in frequency space, not pitch space") measures error by optimizing how well ''cumulative'' intervals from the root real-valued harmonic match the target chord's DR signature.
The idea motivating least-squares error measures on a chord as an approximation to a given delta signature is the following:


=== Fully DR ===
We want the error of a chord 1:''r''<sub>1</sub>:''r''<sub>2</sub>:...:''r''<sub>''n''</sub> (in increasing order; we also take ''r''<sub>0</sub> = 1), with ''n'' &gt; 1, in the linear domain as an approximation to a delta-rational chord with signature +δ<sub>1</sub> +δ<sub>2</sub> ... +δ<sub>''n''</sub> (possibly with some deltas free), i.e. the target chord is
The idea motivating rooted linear error on a chord as an approximation to a given delta signature is the following (for simplicity, let’s talk about the fully DR case first):


Say we want the error of a chord 1:''r''<sub>1</sub>:''r''<sub>2</sub>:...:''r''<sub>''n''</sub> (in increasing order), with ''n'' &gt; 1, in the linear domain as an approximation to a fully delta-rational chord with signature +δ<sub>1</sub> +δ<sub>2</sub> ... +δ<sub>''n''</sub>, i.e. a chord
<math>x : x + \delta_1 : \cdots : x + \sum_{l=1}^n \delta_l</math>


<math>x : x + \delta_1 : \cdots : x + \sum_{l=1}^n \delta_l.</math>
with some root real-valued harmonic ''x''. Let <math>D_0 = 0, D_i = \sum_{k=1}^i \delta_k</math> be the delta signature +δ<sub>1</sub> +δ<sub>2</sub> ... +δ<sub>''n''</sub> written cumulatively.


Rewriting a bit, let <math>D_0 = 0, D_i = \sum_{k=1}^i \delta_k</math> be the delta signature +δ<sub>1</sub> <sub>2</sub> ... +δ<sub>''n''</sub> written cumulatively. We wish to minimize the following frequency-domain error function by optimizing ''x'':
We want to measure the error ''without having to fix any interval in the target chord'' (as one might naively fix an interval and measure errors in the other deltas in relation to the fixed interval). To do this we solve a least-squares optimization problem: use a root-sum-square objective function and optimize ''x'' (and any free deltas) to minimize that function.
 
== Domain and error model ==
We have two choices:
* to measure either the linear (frequency ratio) error or the logarithmic (cents) one (called the ''domain'');
* the collection of intervals to sum over (which we call the ''error model''):
** ''rooted'': Only intervals from the root real-valued harmonic ''x'' are chosen.
** ''pairwise'': All intervals in the chord are chosen.
** ''all-steps'': Only intervals between adjacent notes are chosen.
 
We arrive at the following general formula: Let <math>I \subseteq {\{0, 1, 2, ..., n\} \choose 2}</math> be the error model, and let <math>f_D</math> represent the domain function (identity for linear, or <math>\log</math>). Then the objective function (measuring the error) to be minimized by optimizing <math>x</math> and any free deltas is:


<math>
<math>
\displaystyle{ \sqrt{\sum_{i=1}^n \Bigg( \frac{x + D_i}{x} - r_i \Bigg)^2 }  = \sqrt{\sum_{i=1}^n \Bigg( 1 + \frac{D_i}{x} - r_i \Bigg)^2 } }.
\sqrt{\sum_{i<j, \{i,j\} \in I} \Bigg( f_D \Bigg( \frac{x + D_j}{x + D_i} \Bigg) - f_D\Bigg( \frac{r_j}{r_i} \Bigg) \Bigg)^2}.
</math>
</math>


Setting the derivative to 0 gives us the closed-form solution
{| class="wikitable"
|+ Objective function for various modes and error models
|-
!|Domain
!|Error model
!|Objective function
|-
!rowspan="3"|Linear
!|Rooted
||<math>\sqrt{\sum_{i=1}^n \Bigg( \frac{x + D_i}{x} - r_i \Bigg)^2 }  = \sqrt{\sum_{i=1}^n \Bigg( 1 + \frac{D_i}{x} - r_i \Bigg)^2 }</math>
|-
!|Pairwise
||<math>\sqrt{\sum_{0\leq i<j\leq n} \Bigg( \frac{x + D_j}{x + D_i} - \frac{r_j}{r_i} \Bigg)^2 }</math>
|-
!|All-steps
||<math>\sqrt{\sum_{0\leq i<n} \Bigg( \frac{x + D_{i+1}}{x + D_i} - \frac{r_{i+1}}{r_i} \Bigg)^2 }</math>
|-
!rowspan="3"|Logarithmic<br/>(nataves)
!|Rooted
||<math>\sqrt{\sum_{i=1}^n \Bigg(\log \frac{x + D_i}{r_i x} \Bigg)^2}</math>
|-
!|Pairwise
||<math>\sqrt{\sum_{0\leq i<j\leq n} \Bigg(\log \frac{x + D_j}{x + D_i} - \log \frac{r_j}{r_i} \Bigg)^2}</math>
|-
!|All-steps
||<math>\sqrt{\sum_{0\leq i<n} \Bigg(\log \frac{x + D_{i+1}}{x + D_i} - \log \frac{r_{i+1}}{r_i} \Bigg)^2}</math>
|}
To convert nataves to cents, multiply by <math>\frac{1200}{\log 2}.</math>
 
== Solution methods ==
This section gets into the depths of mathematical optimization methods used to minimize DR error. (Optimization is a whole field and there are many different methods; the reason most of those methods exist is to find solutions when finding formulas for the solutions is infeasible.)
=== Grid method (FDR case) ===
<syntaxhighlight lang="py" line="1">
class GridSolution:
    def __init__(self, x, fx):
        self.x = x
        self.fx = fx
   
    def __repr__(self):
        return f"GridSolution(x={self.x:.5f}, fx={self.fx:.5f})"


<math>x = \frac{\sum_{i=1}^n D_i }{-n + \sum_{i=1}^n r_i},</math>
def grid_method(f, window_size=100, coarse_steps=1000, fine_steps=1000):
    best_error = float("inf")
    best_x = 0
   
    coarse_step = window_size / coarse_steps
    # Fine step partitions one coarse step into smaller pieces
    fine_step = coarse_step / fine_steps
   
    # --- Phase 1: Coarse Grid Search ---
    # Search in the window (0, window_size]
    for i in range(1, coarse_steps + 1):
        x = i * coarse_step
        fx = f(x)
        if fx < best_error:
            best_error = fx
            best_x = x


which can be plugged back into
    # --- Phase 2: Fine Grid Search ---
    # Center the new search window around the best_x found in Phase 1
    # We search from (best_x - coarse_step/2) to (best_x + coarse_step/2)
   
    fine_window_lower = best_x - (coarse_step / 2)
   
    for j in range(1, fine_steps + 1):
        x = fine_window_lower + (j * fine_step)
        fx = f(x)
        if fx < best_error:
            best_error = fx
            best_x = x
           
    return GridSolution(best_x, best_error)
</syntaxhighlight>


<math>\sqrt{\sum_{1=1}^n \Bigg( 1 + \frac{D_i}{x} - r_i \Bigg)^2 }</math>
=== BFGS-B (one related delta set, arbitrary free deltas) ===
We let ''x''<sub>1</sub> = ''x'' and include additional free variables ''x''<sub>2</sub>, ..., ''x''<sub>''n''</sub>, one for every additional +?, after coalescing segments of consecutive +?'s into one +? and after trimming leading and trailing free delta segments.


to obtain the least-squares linear error.
BFGS-B is a quasi-Newton optimization method (based on BFGS) particularly suited for this problem:
* The objective function is smooth, allowing use of gradients
* Fast convergence, requiring at worst 20 iterations for accuracy
* Naturally deals with the ''x'' > 0 constraint using a log barrier and minimizing the transformed function using the unconstrained BFGS method
* Acceptable memory usage given a realistic number of parameters for practical DR chords (up to 3 interior free delta segments, thus 4 parameters).


=== Partially DR (one related delta set, one free variable) ===
It is a ''quasi-Newton method'' because it uses an approximation of the Hessian (matrix of mixed second partial derivatives) of the objective function at each step.
Suppose we wish to approximate a target delta signature of the form <math>+\delta_1 +? +\delta_3</math> with the chord <math>1:f_1:f_2:f_3</math> (where the +? is free to vary). By a derivation similar to the above, the least-squares problem is


<math>
In the Python implementation below, <code>x</code> represents the vector <math>(x_1, x_2, ..., x_n),</math> <code>x0</code> is the initial guess for the solution, and <code>f</code> is the objective function.
\displaystyle {\underset{x,y}{\text{minimize}} \sqrt{\bigg(\frac{x + \delta_1}{x} - r_1 \bigg)^2 + \bigg(\frac{x+\delta_1 + y}{x} - r_2 \bigg)^2 + \bigg(\frac{x+\delta_1 + y + \delta_3}{x} - r_3 \bigg)^2 }},
</math>


where ''y'' represents the free delta +?.
<syntaxhighlight lang="py" line="1">
import numpy as np
import math


We can set the partial derivatives with respect to ''x'' and ''y'' of the inner expression equal to zero (since the derivative of sqrt() is never 0) and use SymPy to solve the system:
class BFGSSolution:
    def __init__(self, x, fx, iterations, success):
        self.x = x
        self.fx = fx
        self.iterations = iterations
        self.success = success


<syntaxhighlight lang="py">
def numerical_gradient(f, x, eps=1e-8):
import sympy
    grad = np.zeros_like(x)
x = sympy.Symbol("x", real=True)
    for i in range(len(x)):
y = sympy.Symbol("y", real=True)
        x_plus = x.copy()
d1 = sympy.Symbol("\\delta_{1}", real=True)
        x_minus = x.copy()
d2 = sympy.Symbol("\\delta_{2}", real=True)
        x_plus[i] += eps
d3 = sympy.Symbol("\\delta_{3}", real=True)
        x_minus[i] -= eps
r1 = sympy.Symbol("r_1", real=True)
        grad[i] = (f(x_plus) - f(x_minus)) / (2 * eps)
r2 = sympy.Symbol("r_2", real=True)
    return grad
r3 = sympy.Symbol("r_3", real=True)
err_squared = ((x + d1) / x - r1) ** 2 + ((x + d1 + y) / x - r2) ** 2 + ((x + d1 + y + d3) / x - r3) ** 2
err_squared.expand()
err_squared_x = sympy.diff(err_squared, x)
err_squared_y = sympy.diff(err_squared, y)
sympy.nonlinsolve([err_squared_x, err_squared_y], [x, y])
</syntaxhighlight>


The unique solution with x > 0 is
def bfgs(f, grad, x0, max_iterations=100, tolerance=1e-5):
<math>
    x = np.array(x0, dtype=float)
(x, y) = \displaystyle { \left( \frac{2 \delta_{1} + \delta_{3} + \frac{2 \left(- 2 \delta_{1}^{2} r_{1} + \delta_{1}^{2} r_{2} + \delta_{1}^{2} r_{3} - \delta_{1} \delta_{3} r_{1} + \delta_{1} \delta_{3} r_{2} - \delta_{1} \delta_{3} r_{3} + \delta_{1} \delta_{3} + \delta_{3}^{2} r_{2} - \delta_{3}^{2}\right)}{2 \delta_{1} r_{1} - 2 \delta_{1} - \delta_{3} r_{2} + \delta_{3} r_{3}}}{r_{2} + r_{3} - 2}, \  \frac{- 2 \delta_{1}^{2} r_{1} + \delta_{1}^{2} r_{2} + \delta_{1}^{2} r_{3} - \delta_{1} \delta_{3} r_{1} + \delta_{1} \delta_{3} r_{2} - \delta_{1} \delta_{3} r_{3} + \delta_{1} \delta_{3} + \delta_{3}^{2} r_{2} - \delta_{3}^{2}}{2 \delta_{1} r_{1} - 2 \delta_{1} - \delta_{3} r_{2} + \delta_{3} r_{3}}\right).}
    n = len(x)
</math>
    fx = f(x)
    g = grad(x)
   
    # Approximate inverse Hessian
    # Initial approximation is the identity matrix
    H = np.eye(n)
   
    for i in range(max_iterations):
        # 0: Check convergence
        grad_norm = np.linalg.norm(g)
        if grad_norm < tolerance:
            return BFGSSolution(x, fx, i, True)
       
        # 1: Set search direction p (negative gradient direction transformed by H)
        # p = -H * g
        p = -H @ g
       
        # 2: Get alpha satisfying Wolfe conditions (Armijo rule and curvature condition)
        c1 = 1e-4
        c2 = 0.9
        max_line_search = 20
        rho_ls = 0.5
       
        gp = np.dot(g, p)
        alpha = 1.0
       
        for _ in range(max_line_search):
            x_next_guess = x + alpha * p
           
            # Check Armijo rule
            # f(x + alpha*p) <= f(x) + c1 * alpha * p^T * g
            if f(x_next_guess) <= fx + c1 * alpha * gp:
                # Check Curvature condition
                # -p^T * grad(x_next) <= -c2 * p^T * g
                g_next_guess = grad(x_next_guess)
                if -np.dot(p, g_next_guess) <= -c2 * gp:
                    break
           
            alpha *= rho_ls
       
        # 3: Set s = alpha * p and x_next = x + s
        s = alpha * p
        x_next = x + s
       
        # 4: Set y = grad(x_next) - grad(x)
        # We re-calculate grad(x_next) here to match the strict logic flow,
        # though optimization could reuse g_next_guess from the successful line search.
        g_next = grad(x_next)
        y = g_next - g
       
        # 5: BFGS Update
        # Update H += U + V
        sy = np.dot(s, y)
       
        # Prevent division by zero if step size was extremely small
        if sy == 0:
            # In a robust implementation, you might reset H to Identity here
            break
           
        Hy = H @ y
       
        # Calculate scalar for the first term: (s^T y + y^T H y) / (s^T y)^2
        scalar1 = (sy + np.dot(y, Hy)) / (sy * sy)
        U = scalar1 * np.outer(s, s)
       
        # Calculate the second term matrices
        # W = (H y) s^T + s (y^T H)
        # Note: Since H is symmetric, y^T H is equivalent to (H y)^T
        W = np.outer(Hy, s) + np.outer(s, Hy)
        V = (-1 / sy) * W
       
        H = H + U + V
       
        # Update x, fx, and g for next iteration
        x = x_next
        fx = f(x_next)
        g = g_next


=== Partially DR (one related delta set, arbitrary free deltas) ===
    return BFGSSolution(x, fx, max_iterations, False)
We similarly include a free variable to be optimized for every additional +?, after coalescing strings of consecutive +?'s and omitting the middle notes, and after trimming leading and trailing +?'s.


Todo: The L-BFGS-B algorithm is suited for five-variable (base real-valued harmonic + four free deltas; a realistic upper bound on real-world use cases of partial DR) optimization problems with bounds, so let's talk about that
def bfgs_barrier(f, bounds, x0, history_size=10, max_iterations=100, tolerance=1e-5, barrier_weight=1e-4):
    """
    Solves bounded optimization using a Log-Barrier method wrapped around BFGS.
    Note: x0 must be strictly feasible (inside bounds).
    """
    def transformed_f(x):
        penalty = 0
        for i, val in enumerate(x):
            lower, upper = bounds[i]
           
            # Hard cutoff prevents log domain errors during line search exploration
            if (lower is not None and val <= lower) or (upper is not None and val >= upper):
                return float("inf")
           
            # Log barrier penalties
            if lower is not None:
                penalty -= barrier_weight * math.log(val - lower)
            if upper is not None:
                penalty -= barrier_weight * math.log(upper - val)
       
        return f(x) + penalty
   
    # Use the transformed function for gradients as well
    grad = lambda x: numerical_gradient(transformed_f, x)
   
    result = bfgs(transformed_f, grad, x0, history_size, max_iterations, tolerance)
   
    # Restore actual function value
    result.fx = f(result.x)
    return result
</syntaxhighlight>


=== Arbitrary related delta sets ===
=== L-BFGS-B (one related delta set, arbitrary free deltas) ===
<hj>Here be dragons. No one really wants to do this, right</hj>
L-BFGS-B is an approximation to BFGS-B limiting memory usage, particularly suited for high-dimensional problems but nevertheless agreeing very well with BFGS-B for typical DR test cases. Python code for the L-BFGS method is provided below.


== Rooted logarithmic error ==
<syntaxhighlight lang="py" line="1">
This error measure also measures errors of rooted intervals, but measures the error in logarithmic interval distance and thus arguably has a more musically intuitive meaning.
import numpy as np
=== Fully DR ===
import math
The error function to be minimized, with units in nepers (logarithmic unit for frequency ratio of ''e''), is


<math>\sqrt{\sum_{i=1}^n \Bigg(\log \frac{x + D_i}{r_i x} \Bigg)^2}.</math>
class LBFGSSolution:
    def __init__(self, x, fx, iterations, success):
        self.x = x
        self.fx = fx
        self.iterations = iterations
        self.success = success


(To scale to cents, multiply by 1200/log 2.)
def numerical_gradient(f, x, eps=1e-8):
    grad = np.zeros_like(x)
    for i in range(len(x)):
        x_plus = x.copy()
        x_minus = x.copy()
        x_plus[i] += eps
        x_minus[i] -= eps
        grad[i] = (f(x_plus) - f(x_minus)) / (2 * eps)
    return grad


== All-interval linear error ==
def l_bfgs(f, grad, x0, history_size=10, max_iterations=100, tolerance=1e-5):
Measure all pairwise intervals, linearly
    x = np.array(x0, dtype=float) # Ensure float array
=== Fully DR ===
    fx = f(x)
<math>
    g = grad(x)
\displaystyle{ \sqrt{\sum_{0\leq i<j\leq n} \Bigg( \frac{x + D_j}{x + D_i} - \frac{r_j}{r_i} \Bigg)^2 } }.
   
</math>
    s_history = []
    y_history = []
    rho_history = []
   
    for iteration in range(max_iterations):
        grad_norm = np.linalg.norm(g)
        if grad_norm < tolerance:
            return LBFGSSolution(x, fx, iteration, True)
       
        # --- Two-loop recursion ---
        q = g.copy()
       
        # We need to store alpha to use it in the second loop
        m = len(s_history)
        alpha = [0.0] * m
       
        # First loop (Backward)
        for i in range(m - 1, -1, -1):
            alpha[i] = rho_history[i] * np.dot(s_history[i], q)
            q = q - alpha[i] * y_history[i]
       
        # Initial Hessian approximation scaling
        gamma = 1.0
        if m > 0:
            gamma = np.dot(s_history[-1], y_history[-1]) / np.dot(y_history[-1], y_history[-1])
       
        z = q * gamma


== All-interval logarithmic error ==
        # Second loop (Forward)
This error measure measures errors of ''all'' intervals, not just rooted ones.
        for i in range(m):
=== Fully DR ===
            beta = rho_history[i] * np.dot(y_history[i], z)
The error function to be minimized, with units in nepers, is
            z = z + s_history[i] * (alpha[i] - beta)
       
        p = -z # Search direction


<math>\sqrt{\sum_{0\leq i<j\leq n} \Bigg(\log \frac{x + D_j}{x + D_i} - \log \frac{r_j}{r_i} \Bigg)^2}.</math>
        # --- Line search (Backtracking) ---
        step_size = 1.0
        c1 = 1e-4
        rho_ls = 0.5
        max_line_search = 20
       
        # Check for sufficient decrease (Armijo rule)
        # f(x + alpha*p) <= f(x) + c1 * alpha * p^T * g
       
        g_dot_p = np.dot(g, p)
       
        for _ in range(max_line_search):
            x_new = x + step_size * p
            fx_new = f(x_new)
           
            if fx_new <= fx + c1 * step_size * g_dot_p:
                break
            step_size *= rho_ls
        else:
            # If line search fails to find a better point, stop or accept best attempt
            # (Here we just accept the last reduced step_size)
            pass
       
        # --- Update History ---
        s = x_new - x
        g_new = grad(x_new)
        y = g_new - g
       
        sy = np.dot(s, y)
       
        if sy > 1e-10: # Ensure curvature condition
            s_history.append(s)
            y_history.append(y)
            rho_history.append(1.0 / sy)
           
            # Maintain history size
            if len(s_history) > history_size:
                s_history.pop(0)
                y_history.pop(0)
                rho_history.pop(0)
           
        x = x_new
        fx = fx_new
        g = g_new
   
    return LBFGSSolution(x, fx, max_iterations, False)
</syntaxhighlight>


== External links ==
== External links ==
* [https://inthar-raven.github.io/delta/ Inthar's DR chord explorer (includes fixed-root linear error calculation)]
* [https://turbofishcrow.github.io/delta/ Inthar's DR chord explorer (includes least-squares error calculation in both domains and multiple error models)]
[[Category:Atypical ratios]]
{{cat|Atypical ratios}}

Latest revision as of 04:22, 24 February 2026

This is a technical or mathematical page. While the subject may be of some relevance to music, the page treats the subject in technical language.

This article will describe several least-squares error measures for delta-rational chords. The idea behind least-squares error measures is to find the chord with the exact specified delta signature such that it deviates as little as possible from the chord that is meant to approximate that delta signature, and to measure the deviation. Least-squares error measures have the advantage of not fixing a particular interval in the chord when constructing the chord of best fit. However, like any other numerical measure of concordance or error, you should take them with a grain of salt.

Conventions and introduction

The idea motivating least-squares error measures on a chord as an approximation to a given delta signature is the following:

We want the error of a chord 1:r1:r2:...:rn (in increasing order; we also take r0 = 1), with n > 1, in the linear domain as an approximation to a delta-rational chord with signature +δ12 ... +δn (possibly with some deltas free), i.e. the target chord is

x:x+δ1::x+l=1nδl

with some root real-valued harmonic x. Let D0=0,Di=k=1iδk be the delta signature +δ12 ... +δn written cumulatively.

We want to measure the error without having to fix any interval in the target chord (as one might naively fix an interval and measure errors in the other deltas in relation to the fixed interval). To do this we solve a least-squares optimization problem: use a root-sum-square objective function and optimize x (and any free deltas) to minimize that function.

Domain and error model

We have two choices:

  • to measure either the linear (frequency ratio) error or the logarithmic (cents) one (called the domain);
  • the collection of intervals to sum over (which we call the error model):
    • rooted: Only intervals from the root real-valued harmonic x are chosen.
    • pairwise: All intervals in the chord are chosen.
    • all-steps: Only intervals between adjacent notes are chosen.

We arrive at the following general formula: Let I({0,1,2,...,n}2) be the error model, and let fD represent the domain function (identity for linear, or log). Then the objective function (measuring the error) to be minimized by optimizing x and any free deltas is:

i<j,{i,j}I(fD(x+Djx+Di)fD(rjri))2.

Objective function for various modes and error models
Domain Error model Objective function
Linear Rooted i=1n(x+Dixri)2=i=1n(1+Dixri)2
Pairwise 0i<jn(x+Djx+Dirjri)2
All-steps 0i<n(x+Di+1x+Diri+1ri)2
Logarithmic
(nataves)
Rooted i=1n(logx+Dirix)2
Pairwise 0i<jn(logx+Djx+Dilogrjri)2
All-steps 0i<n(logx+Di+1x+Dilogri+1ri)2

To convert nataves to cents, multiply by 1200log2.

Solution methods

This section gets into the depths of mathematical optimization methods used to minimize DR error. (Optimization is a whole field and there are many different methods; the reason most of those methods exist is to find solutions when finding formulas for the solutions is infeasible.)

Grid method (FDR case)

class GridSolution:
    def __init__(self, x, fx):
        self.x = x
        self.fx = fx
    
    def __repr__(self):
        return f"GridSolution(x={self.x:.5f}, fx={self.fx:.5f})"

def grid_method(f, window_size=100, coarse_steps=1000, fine_steps=1000):
    best_error = float("inf")
    best_x = 0
    
    coarse_step = window_size / coarse_steps
    # Fine step partitions one coarse step into smaller pieces
    fine_step = coarse_step / fine_steps 
    
    # --- Phase 1: Coarse Grid Search ---
    # Search in the window (0, window_size]
    for i in range(1, coarse_steps + 1):
        x = i * coarse_step
        fx = f(x)
        if fx < best_error:
            best_error = fx
            best_x = x

    # --- Phase 2: Fine Grid Search ---
    # Center the new search window around the best_x found in Phase 1
    # We search from (best_x - coarse_step/2) to (best_x + coarse_step/2)
    
    fine_window_lower = best_x - (coarse_step / 2)
    
    for j in range(1, fine_steps + 1):
        x = fine_window_lower + (j * fine_step)
        fx = f(x)
        if fx < best_error:
            best_error = fx
            best_x = x
            
    return GridSolution(best_x, best_error)

We let x1 = x and include additional free variables x2, ..., xn, one for every additional +?, after coalescing segments of consecutive +?'s into one +? and after trimming leading and trailing free delta segments.

BFGS-B is a quasi-Newton optimization method (based on BFGS) particularly suited for this problem:

  • The objective function is smooth, allowing use of gradients
  • Fast convergence, requiring at worst 20 iterations for accuracy
  • Naturally deals with the x > 0 constraint using a log barrier and minimizing the transformed function using the unconstrained BFGS method
  • Acceptable memory usage given a realistic number of parameters for practical DR chords (up to 3 interior free delta segments, thus 4 parameters).

It is a quasi-Newton method because it uses an approximation of the Hessian (matrix of mixed second partial derivatives) of the objective function at each step.

In the Python implementation below, x represents the vector (x1,x2,...,xn), x0 is the initial guess for the solution, and f is the objective function.

import numpy as np
import math

class BFGSSolution:
    def __init__(self, x, fx, iterations, success):
        self.x = x
        self.fx = fx
        self.iterations = iterations
        self.success = success

def numerical_gradient(f, x, eps=1e-8):
    grad = np.zeros_like(x)
    for i in range(len(x)):
        x_plus = x.copy()
        x_minus = x.copy()
        x_plus[i] += eps
        x_minus[i] -= eps
        grad[i] = (f(x_plus) - f(x_minus)) / (2 * eps)
    return grad

def bfgs(f, grad, x0, max_iterations=100, tolerance=1e-5):
    x = np.array(x0, dtype=float)
    n = len(x)
    fx = f(x)
    g = grad(x)
    
    # Approximate inverse Hessian
    # Initial approximation is the identity matrix
    H = np.eye(n)
    
    for i in range(max_iterations):
        # 0: Check convergence
        grad_norm = np.linalg.norm(g)
        if grad_norm < tolerance:
            return BFGSSolution(x, fx, i, True)
        
        # 1: Set search direction p (negative gradient direction transformed by H)
        # p = -H * g
        p = -H @ g
        
        # 2: Get alpha satisfying Wolfe conditions (Armijo rule and curvature condition)
        c1 = 1e-4
        c2 = 0.9
        max_line_search = 20
        rho_ls = 0.5
        
        gp = np.dot(g, p)
        alpha = 1.0
        
        for _ in range(max_line_search):
            x_next_guess = x + alpha * p
            
            # Check Armijo rule
            # f(x + alpha*p) <= f(x) + c1 * alpha * p^T * g
            if f(x_next_guess) <= fx + c1 * alpha * gp:
                # Check Curvature condition
                # -p^T * grad(x_next) <= -c2 * p^T * g
                g_next_guess = grad(x_next_guess)
                if -np.dot(p, g_next_guess) <= -c2 * gp:
                    break
            
            alpha *= rho_ls
        
        # 3: Set s = alpha * p and x_next = x + s
        s = alpha * p
        x_next = x + s
        
        # 4: Set y = grad(x_next) - grad(x)
        # We re-calculate grad(x_next) here to match the strict logic flow, 
        # though optimization could reuse g_next_guess from the successful line search.
        g_next = grad(x_next)
        y = g_next - g
        
        # 5: BFGS Update
        # Update H += U + V
        sy = np.dot(s, y)
        
        # Prevent division by zero if step size was extremely small
        if sy == 0: 
            # In a robust implementation, you might reset H to Identity here
            break 
            
        Hy = H @ y
        
        # Calculate scalar for the first term: (s^T y + y^T H y) / (s^T y)^2
        scalar1 = (sy + np.dot(y, Hy)) / (sy * sy)
        U = scalar1 * np.outer(s, s)
        
        # Calculate the second term matrices
        # W = (H y) s^T + s (y^T H)
        # Note: Since H is symmetric, y^T H is equivalent to (H y)^T
        W = np.outer(Hy, s) + np.outer(s, Hy)
        V = (-1 / sy) * W
        
        H = H + U + V
        
        # Update x, fx, and g for next iteration
        x = x_next
        fx = f(x_next)
        g = g_next

    return BFGSSolution(x, fx, max_iterations, False)

def bfgs_barrier(f, bounds, x0, history_size=10, max_iterations=100, tolerance=1e-5, barrier_weight=1e-4):
    """
    Solves bounded optimization using a Log-Barrier method wrapped around BFGS.
    Note: x0 must be strictly feasible (inside bounds).
    """
    def transformed_f(x):
        penalty = 0
        for i, val in enumerate(x):
            lower, upper = bounds[i]
            
            # Hard cutoff prevents log domain errors during line search exploration
            if (lower is not None and val <= lower) or (upper is not None and val >= upper):
                return float("inf")
            
            # Log barrier penalties
            if lower is not None:
                penalty -= barrier_weight * math.log(val - lower)
            if upper is not None:
                penalty -= barrier_weight * math.log(upper - val)
        
        return f(x) + penalty
    
    # Use the transformed function for gradients as well
    grad = lambda x: numerical_gradient(transformed_f, x)
    
    result = bfgs(transformed_f, grad, x0, history_size, max_iterations, tolerance)
    
    # Restore actual function value
    result.fx = f(result.x)
    return result

L-BFGS-B is an approximation to BFGS-B limiting memory usage, particularly suited for high-dimensional problems but nevertheless agreeing very well with BFGS-B for typical DR test cases. Python code for the L-BFGS method is provided below.

import numpy as np
import math

class LBFGSSolution:
    def __init__(self, x, fx, iterations, success):
        self.x = x
        self.fx = fx
        self.iterations = iterations
        self.success = success

def numerical_gradient(f, x, eps=1e-8):
    grad = np.zeros_like(x)
    for i in range(len(x)):
        x_plus = x.copy()
        x_minus = x.copy()
        x_plus[i] += eps
        x_minus[i] -= eps
        grad[i] = (f(x_plus) - f(x_minus)) / (2 * eps)
    return grad

def l_bfgs(f, grad, x0, history_size=10, max_iterations=100, tolerance=1e-5):
    x = np.array(x0, dtype=float) # Ensure float array
    fx = f(x)
    g = grad(x)
    
    s_history = []
    y_history = []
    rho_history = []
    
    for iteration in range(max_iterations):
        grad_norm = np.linalg.norm(g)
        if grad_norm < tolerance:
            return LBFGSSolution(x, fx, iteration, True)
        
        # --- Two-loop recursion ---
        q = g.copy()
        
        # We need to store alpha to use it in the second loop
        m = len(s_history)
        alpha = [0.0] * m 
        
        # First loop (Backward)
        for i in range(m - 1, -1, -1):
            alpha[i] = rho_history[i] * np.dot(s_history[i], q)
            q = q - alpha[i] * y_history[i]
        
        # Initial Hessian approximation scaling
        gamma = 1.0
        if m > 0:
            gamma = np.dot(s_history[-1], y_history[-1]) / np.dot(y_history[-1], y_history[-1])
        
        z = q * gamma

        # Second loop (Forward)
        for i in range(m):
            beta = rho_history[i] * np.dot(y_history[i], z)
            z = z + s_history[i] * (alpha[i] - beta)
        
        p = -z # Search direction

        # --- Line search (Backtracking) ---
        step_size = 1.0
        c1 = 1e-4
        rho_ls = 0.5
        max_line_search = 20
        
        # Check for sufficient decrease (Armijo rule)
        # f(x + alpha*p) <= f(x) + c1 * alpha * p^T * g
        
        g_dot_p = np.dot(g, p)
        
        for _ in range(max_line_search):
            x_new = x + step_size * p
            fx_new = f(x_new)
            
            if fx_new <= fx + c1 * step_size * g_dot_p:
                break
            step_size *= rho_ls
        else:
            # If line search fails to find a better point, stop or accept best attempt
            # (Here we just accept the last reduced step_size)
            pass 
        
        # --- Update History ---
        s = x_new - x
        g_new = grad(x_new)
        y = g_new - g
        
        sy = np.dot(s, y)
        
        if sy > 1e-10: # Ensure curvature condition
            s_history.append(s)
            y_history.append(y)
            rho_history.append(1.0 / sy)
            
            # Maintain history size
            if len(s_history) > history_size:
                s_history.pop(0)
                y_history.pop(0)
                rho_history.pop(0)
            
        x = x_new
        fx = fx_new
        g = g_new
    
    return LBFGSSolution(x, fx, max_iterations, False)