---
title: Every Call is Precious (ECP) Framework
url: https://www.emergentmind.com/topics/every-call-is-precious-ecp-framework
type: topic
---

# Every Call is Precious (ECP) Framework

The Every Call is Precious (ECP) framework is a theoretically principled and practically effective family of global optimization algorithms for black-box, nonconvex, Lipschitz-continuous functions with unknown Lipschitz constants. ECP and its scalable extension, ECPv2, pursue a "precious evaluation" philosophy, ensuring that each function query is potentially informative with respect to the global optimum. This approach systematically excludes wasteful or provably suboptimal evaluations, offering no-regret and minimax-optimal finite-time guarantees without requiring explicit estimation of the Lipschitz constant [2502.04290][2511.16575]. The ECP paradigm has demonstrated robustness and competitive performance on high-dimensional synthetic and real-world benchmarks.

## 1. Problem Setting and Acceptance Principle

ECP targets black-box global maximization:
\[
f : \mathcal{X} \subset \mathbb{R}^d \longrightarrow \mathbb{R}
\]
where \( f \) is unknown but assumed \( k \)-Lipschitz for some unknown \( k \ge 0 \):
\[
|f(x) - f(x')| \leq k \|x - x'\|_2 \quad \forall x, x' \in \mathcal{X}
\]
At each step, the optimizer maintains an archive \( \{(X_i, f(X_i))\}_{i=1}^t \). Rather than evaluating candidates sampled uniformly at random, ECP introduces an "acceptance region"—the set of points \( x \) that could plausibly be maximizers for some Lipschitz extension of the observed data at the current surrogate Lipschitz constant \( \varepsilon_t \). Formally:
\[
\mathcal{A}_{\text{ECP}}(\varepsilon_t, t) = \Bigl\{x \in \mathcal{X} : \min_{1 \leq i \leq t} \bigl[ f(X_i) + \varepsilon_t \|x - X_i\|_2 \bigr] \geq f^t_{\max} \Bigr\}
\]
with \( f^t_{\max} = \max_{1 \leq i \leq t} f(X_i) \).

This acceptance rule is nonparametric and adapts as additional points are evaluated. Intuitively, \( \min_i[f(X_i) + k\|x - X_i\|] \) yields the best upper bound on \( f(x) \) consistent with a \( k \)-Lipschitz function. By evaluating only if the best-case estimate exceeds the current archive maximum, ECP ensures that every accepted call is "precious" [2502.04290].

## 2. Algorithmic Workflow and Adaptation Mechanism

ECP employs a multi-scale exploration strategy where the threshold \( \varepsilon_t \) increases adaptively. The algorithm proceeds as follows:

1. **Initialization:** Draw the first sample \( X_1 \sim \mathcal{U}(\mathcal{X}) \), evaluate \( f(X_1) \), and set initial \( \varepsilon_1 \), typically \( 10^{-2} \).
2. **Sampling:** At iteration \( t \), repeatedly sample \( x \sim \mathcal{U}(\mathcal{X}) \) and check for membership in \( \mathcal{A}_{\text{ECP}}(\varepsilon_t, t) \).
3. **Acceptance and Archive Update:** Upon acceptance, evaluate \( f(x) \), add to the archive, and multiply \( \varepsilon_t \) by \( \tau_{n,d} > 1 \).
4. **Patience Control:** Maintain a "rejection counter": after \( C \) consecutive rejections, increase \( \varepsilon_t \) by \( \tau_{n,d} \).

This protocol fosters a careful balance: initially, the small \( \varepsilon_t \) yields tight regions and exploitation; as \( \varepsilon_t \) increases, the acceptance region expands, promoting exploration. The approach provably avoids indefinite rejection loops and systematically relaxes acceptance as the evaluation budget progresses [2502.04290].

## 3. Theoretical Properties and Guarantees

ECP provides the following guarantees under \( k \)-Lipschitz continuity:

- **Monotonic Acceptance Region:** For \( u \leq v \), \( \mathcal{A}_{u,t} \subseteq \mathcal{A}_{v,t} \). The region expands as \( \varepsilon_t \) increases.
- **Potential Optimality:** When \( \varepsilon_t \leq k \), \( \mathcal{A}_{\varepsilon_t,t} \subseteq \mathcal{P}_{k,t} \), the set of potential global maximizers under some completion.
- **Finite-Time and Asymptotic Regret Bounds:** For any \( \delta \), with probability \( \geq 1 - \delta \):
    \[
    \mathcal{R}_{\mathrm{ECP}}(n) \leq k\,\operatorname{diam}(\mathcal{X})\, (i^*)^{1/d} \left( \frac{\ln(1/\delta)}{n} \right)^{1/d}
    \]
    with \( i^* \leq \left\lceil \log_{\tau_{n,d}}(k/\varepsilon_1) \right\rceil \), which matches the minimax lower bound \( \Omega(k n^{-1/d}) \) for global optimization under Lipschitz continuity [2502.04290].
- **No-regret:** As \( n \to \infty \), \( \mathcal{R}_{\rm ECP}(n) \to 0 \) in probability for every \( k \)-Lipschitz \( f \).

These properties ensure strong theoretical robustness and confirm the "preciousness" principle: every function call advances, in a minimax-optimal sense, the global search.

## 4. Computational Challenges and ECPv2 Extensions

Original ECP's computational bottlenecks and conservative early-phase rejection rates are addressed in ECPv2 through three principal mechanisms [2511.16575]:

- **Adaptive Lower Bound on \( \varepsilon_t \):** At each \( t \),
    \[
    \varepsilon_t^\oslash = \frac{\max_{i \leq t} f(X_i) - \min_{i \leq t} f(X_i)}{\operatorname{diam}(\mathcal{X})}
    \]
    and \( \varepsilon_t \gets \max\{\tau_{n,d} \varepsilon_{t-1}, \varepsilon_t^\oslash\} \). This prevents vacuous acceptance regions and ensures that \( \mathcal{A}_{\text{ECPv2}} \) is nonempty.
- **Worst-\( m \) Memory Mechanism:** Only the \( m \) worst points (indexed by lowest \( f(X_i) \)) are used in the acceptance test:
    \[
    \mathcal{A}_m(\varepsilon_t, t) = \Bigl\{ x : \min_{i \in \mathcal{I}_t^m} \bigl[f(X_i) + \varepsilon_t \|x - X_i\|_2\bigr] \geq f^t_{\max} \Bigr\}
    \]
    Reduces per-iteration cost from \( O(n d) \) to \( O(m d) \) without sacrificing theoretical guarantees.
- **Fixed Random Projection:** Distances are computed in a reduced dimension \( d' \) via a Gaussian random matrix. With high probability (\( \geq 1 - 1/\beta^2 \)):
    \[
    (1 - \delta) \|x - X_i\|_2^2 \leq \| \mathbf{P} x - \mathbf{P} X_i \|_2^2 \leq (1 + \delta) \|x - X_i\|_2^2
    \]
    for all \( i \). Scale \( \varepsilon_t \) to \( \tilde{\varepsilon}_t = \varepsilon_t / \sqrt{1 - \delta} \). Computational cost per query is further reduced to \( O(\ln n) \).

**Acceptance Region Inclusion:** With probability \( \geq 1 - 1/\beta^2 \), \( \mathcal{A}_{\text{ECP}} \subseteq \mathcal{A}_{\text{ECPv2}} \), ensuring ECPv2 never rejects a point accepted by ECP.

Theoretical analysis confirms that ECPv2 preserves no-regret guarantees and optimal finite-time regret rates. Each innovation is validated through ablation and empirical studies [2511.16575].

## 5. Practical Implementation and Complexity

The core computational steps for both ECP and ECPv2 are as follows:

- **Per Candidate:** Calculate acceptance region membership by minimizing a surrogate upper bound over a subset (all, or worst-\( m \)) of previous points.
- **Per Iteration (ECPv2):**
    1. Project candidate and archive to \( \mathbb{R}^{d'} \).
    2. For each candidate, compute \( \min_{i \in \mathcal{I}_t^m} [f(X_i) + \tilde{\varepsilon}_t \|\mathbf{P}x - \mathbf{P}X_i\|] \).
    3. Accept or reject, updating the archive and predictor variables accordingly.

Memory cost is \( O(n d) \); distance computations scale as \( O(m d') \), with \( d' = O(\ln n) \).

ECPv2 pseudocode is as follows:

```python
# ECPv2 Pseudocode Sketch
# Input: budget n, domain X, function f, parameters eps1, tau_nd, C, m, delta, beta
Construct random projection P of size d'
Sample X1 ~ U(X); observe f(X1)
t = 1; eps = eps1; h_reject = 0
while t < n:
    Propose Y ~ U(X); hatY = P Y
    if h_reject >= C:
        eps = tau_nd * eps
        h_reject = 0
    eps_oslash = (f_max^t - f_min^t) / diam(X)
    eps = max(eps, eps_oslash)
    Identify worst-m: I_t^m
    if min_{i in I_t^m} [f(X_i) + eps/sqrt(1-delta) * ||P Y - P X_i||] >= f_max^t:
        Evaluate f(Y), append, t += 1, h_reject=0
    else:
        h_reject +=1
Return argmax_{i<=n} f(X_i)
```
(ECPv2's actual implementation should refer to details in [2511.16575].)

## 6. Experimental Benchmarks and Performance

Comprehensive benchmark studies compare ECP/ECPv2 to established methods including AdaLIPO, DIRECT, SMAC3, Dual-Annealing, CMA-ES, and Bayesian techniques. Key settings include:

- **Benchmarks:** High-dimensional synthetic (Rosenbrock \( d=3,100,200,300,500 \), Powell \( d=1000 \), etc.) and low-dimensional testbeds.
- **Metrics:** Simple regret versus evaluation budget, wall-clock time.
- **Default hyperparameters:** \( \varepsilon_1 = 0.01, \tau_{n,d} = 1 + 1/(nd), C = 1000, m = 8, \delta = 2/3, \beta = 5 \).

Results confirm that ECPv2 matches or outperforms all competing methods in final regret, with particularly notable acceleration on high-dimensional tasks (up to \( 2\times \) faster than ECP in wall-clock time, achieving equal or better regret). Ablation experiments demonstrate the independent and combined impacts of lower bounding, worst-\( m \), and projection mechanisms in reducing computational burden [2511.16575].

A summary of empirical findings is provided below:

| Method      | Benchmark Coverage      | Regret Performance         | Wall-clock Speed        |
|-------------|------------------------|---------------------------|------------------------|
| ECPv2       | Broad (2–1000D)        | Optimal/near-optimal      | Best/far-above ECP     |
| ECP         | Broad (2–1000D)        | Optimal/near-optimal      | Slower than ECPv2      |
| SOTA others | Broad (varied)         | Sometimes close           | Variable               |

On \( d=500 \) Rosenbrock, ECPv2 achieves optimal regret in roughly half the wall-clock time required by ECP, with other optimizers typically slower and/or less robust for unknown Lipschitz constants [2511.16575].

## 7. Limitations, Extensions, and Applications

ECP's foundational assumption is global Lipschitz continuity of the objective. If \( f \) is highly non-Lipschitz or exhibits severe local ruggedness, the core acceptance-rejection logic may become less effective. In extremely high dimensions or for very large evaluation budgets, surrogate-based Bayesian optimization or evolutionary algorithms may achieve superior sample efficiency, although ECPv2 narrows this gap through algorithmic acceleration.

Extensions proposed include:

- Alternative metrics beyond Euclidean for structured domains.
- Integrating lightweight surrogate models once \( \varepsilon_t \) substantially exceeds true \( k \).
- Online adaptation of growth rule \( \tau_{n,d} \) and patience parameter \( C \).
- Plug-and-play integration in black-box optimization pipelines.

ECP code is available at https://github.com/fouratifares/ECP [2502.04290].

ECP and ECPv2 offer minimax-optimal global optimization for Lipschitz-continuous functions with unknown smoothness, distinguished by their rigorous acceptance rule, adaptive behavior, and scalable implementation [2502.04290][2511.16575].

Source: https://www.emergentmind.com/topics/every-call-is-precious-ecp-framework