arXiv subjects

Pedro Soto

Publications and source records attributed to Pedro Soto.

15 recordsLinked to original sources

arXiv2026

Verifying global identifiability of parametric linear ODE models is NP-hard

Global parameter identifiability is a property of a parametric ODE model to recover the parameter values uniquely from the input-output data. Not all parametric ODE models have this property, and checking for parameter identifiability is a prerequisite to perform numerical parameter estimation. There are many algorithms and software packages for global parameter identifiability, and frequently the runtime is large. However, the computational complexity for this problem has not been analyzed yet, though there are complexity results for local (finitely many values fit the data) parameter identifiability. In this paper, we estimate the complexity of checking global parameter identifiability over real fields for ODE models that depend linearly on the state variables and rationally on the parameters. In particular, we prove that it is equivalent to the injectivity problem.

cs.CC↗

arXiv2026

Design of MDP Convolutional Codes and Maximally Recoverable Codes Through the Lens of Matrix Completion

The matrix completion problem provides a unifying lens through which many fundamental problems in coding theory can be viewed. In this paper, we investigate Locally Recoverable Codes (LRCs) with Maximal Recoverability (MR) and Maximum Distance Profile (MDP) convolutional codes in the framework of matrix completion. In particular, we present techniques that are general enough to provide constructions for both types of codes. A common feature of our code constructions is the sparsity of their generator matrices and the property that a large number of the entries of the generator matrices are elements of a small subfield of a larger extension field.

cs.IT↗

arXiv2026

Cachemir: Fully Homomorphic Encrypted Inference of Generative Large Language Model with KV Cache

Generative large language models (LLMs) have revolutionized multiple domains. Modern LLMs predominantly rely on an autoregressive decoding strategy, which generates output tokens sequentially and employs a key-value cache (KV cache) to avoid redundant computation. However, the widespread deployment of LLMs has raised serious privacy concerns, as users are feeding all types of data into the model, motivating the development of secure inference frameworks based on fully homomorphic encryption (FHE). A major limitation of existing FHE-based frameworks is their inability to effectively integrate the KV cache, resulting in prohibitively high latency for autoregressive decoding. In this paper, we propose Cachemir, a KV Cache Accelerated Homomorphic Encrypted LLM Inference Regime to overcome this limitation. Cachemir comprises three key technical contributions: 1) a set of novel HE packing algorithms specifically designed to leverage the computational advantages of the KV cache; 2) an interleaved replicated packing algorithm to efficiently compute the vector-matrix multiplications that result from using the KV cache in Transformer linear layers; and 3) an augmented bootstrapping placement strategy that accounts for the KV cache to minimize bootstrapping cost. We demonstrate that Cachemir achieves $48.83\times$ and $67.16\times$ speedup over MOAI (ICML'25) and THOR (CCS'25) respectively on CPU and consumes less than 100 seconds on GPU to generate an output token for Llama-3-8B.

cs.CR↗

arXiv2025

Two for One, One for All: Deterministic LDC-based Robust Computation in Congested Clique

We design a deterministic compiler that makes any computation in the Congested Clique model robust to a constant fraction $\alpha<1$ of adversarial crash faults. In particular, we show how a network of $n$ nodes can compute any circuit of depth $d$, width $\omega$, and gate total fan $\Delta$, in $d\cdot\lceil\frac{\omega}{n^2}+\frac{\Delta}{n}\rceil\cdot 2^{O(\sqrt{\log{n}}\log\log{n})}$ rounds in such a faulty model. As a corollary, any $T$-round Congested Clique algorithm can be compiled into an algorithm that completes in $T^2 n^{o(1)}$ rounds in this model. Our compiler obtains resilience to node crashes by coding information across the network, where we leverage locally-decodable codes (LDCs) to maintain a low complexity overhead, as these allow recovering the information needed at each computational step by querying only small parts of the codeword. The main technical contribution is that because erasures occur in known locations, which correspond to crashed nodes, we can derandomize classical LDC constructions by deterministically selecting query sets that avoid sufficiently many erasures. Moreover, when decoding multiple codewords in parallel, our derandomization load-balances the queries per-node, thereby preventing congestion and maintaining a low round complexity. Deterministic decoding of LDCs presents a new challenge: the adversary can target precisely the (few) nodes that are queried for decoding a certain codeword. We overcome this issue via an adaptive doubling strategy: if a decoding attempt for a codeword fails, the node doubles the number of its decoding attempts. Similarly, when the adversary crashes the decoding node itself, we replace it dynamically with two other non-crashed nodes. By carefully combining these two doubling processes, we overcome the challenges posed by the combination of a deterministic LDC with a worst case pattern of crashes.

cs.DS↗

arXiv2025

A Matrix Completion Approach for the Construction of MDP Convolutional Codes

Maximum Distance Profile (MDP) convolutional codes are an important class of channel codes due to their maximal delay-constrained error correction capabilities. The design of MDP codes has attracted significant attention from the research community. However, only limited attention was given to addressing the complexity of encoding and decoding operations. This paper aims to reduce encoding complexity by constructing partial unit-memory MDP codes with structured and sparse generator matrices. In particular, we present a matrix completion framework that extends a structured superregular matrix (e.g., Cauchy) over a small field to a sparse sliding generator matrix of an MDP code. We show that the proposed construction can reduce the encoding complexity compared to the current state-of-the-art MDP code designs.

cs.IT↗

arXiv2025

Computing in a Faulty Congested Clique

We study a Faulty Congested Clique model, in which an adversary may fail nodes in the network throughout the computation. We show that any task of $O(n\log{n})$-bit input per node can be solved in roughly $n$ rounds, where $n$ is the size of the network. This nearly matches the linear upper bound on the complexity of the non-faulty Congested Clique model for such problems, by learning the entire input, and it holds in the faulty model even with a linear number of faults. Our main contribution is that we establish that one can do much better by looking more closely at the computation. Given a deterministic algorithm $\mathcal{A}$ for the non-faulty Congested Clique model, we show how to transform it into an algorithm $\mathcal{A}'$ for the faulty model, with an overhead that could be as small as some logarithmic-in-$n$ factor, by considering refined complexity measures of $\mathcal{A}$. As an exemplifying application of our approach, we show that the $O(n^{1/3})$-round complexity of semi-ring matrix multiplication [Censor-Hillel, Kaski, Korhonen, Lenzen, Paz, Suomela, PODC 2015] remains the same up to polylog factors in the faulty model, even if the adversary can fail $99\%$ of the nodes (or any other constant fraction).

cs.DS↗

arXiv2025

Optimal Secure Coded Distributed Computation over all Fields

We construct optimal secure coded distributed schemes that extend the known optimal constructions over fields of characteristic 0 to all fields. A serendipitous result is that we can encode \emph{all} functions over finite fields with a recovery threshold proportional to the complexity (tensor rank or multiplicative); this is due to the well-known result that all functions over a finite field can be represented as multivariate polynomials (or symmetric tensors). We get that a tensor of order $\ell$ (or a multivariate polynomial of degree $\ell$) can be computed in the faulty network of $N$ nodes setting within a factor of $\ell$ and an additive term depending on the genus of a code with $N$ rational points and distance covering the number of faulty servers; in particular, we present a coding scheme for general matrix multiplication of two $m \times m $ matrices with a recovery threshold of $2 m^{\omega } -1+g$ where $\omega $ is the exponent of matrix multiplication which is optimal for coding schemes using AG codes. Moreover, we give sufficient conditions for which the Hadamard-Shur product of general linear codes gives a similar recovery threshold, which we call \textit{log-additive codes}. Finally, we show that evaluation codes with a \textit{curve degree} function (first defined in [Ben-Sasson et al. (STOC '13)]) that have well-behaved zero sets are log-additive.

cs.IT↗

arXiv2024

Faster Gröbner bases for Lie derivatives of ODE systems via monomial orderings

Symbolic computation for systems of differential equations is often computationally expensive. Many practical differential models have a form of polynomial or rational ODE system with specified outputs. A basic symbolic approach to analyze these models is to compute and then symbolically process the polynomial system obtained by sufficiently many Lie derivatives of the output functions with respect to the vector field given by the ODE system. In this paper, we present a method for speeding up Gröbner basis computation for such a class of polynomial systems by using specific monomial ordering, including weights for the variables, coming from the structure of the ODE model. We provide empirical results that show improvement across different symbolic computing frameworks and apply the method to speed up structural identifiability analysis of ODE models.

cs.SC↗

arXiv2024

Algebraic Geometric Rook Codes for Coded Distributed Computing

We extend coded distributed computing over finite fields to allow the number of workers to be larger than the field size. We give codes that work for fully general matrix multiplication and show that in this case we serendipitously have that all functions can be computed in a distributed fault-tolerant fashion over finite fields. This generalizes previous results on the topic. We prove that the associated codes achieve a recovery threshold similar to the ones for characteristic zero fields but now with a factor that is proportional to the genus of the underlying function field. In particular, we have that the recovery threshold of these codes is proportional to the classical complexity of matrix multiplication by a factor of at most the genus.

cs.IT↗

arXiv2024

Algebraic identifiability of partial differential equation models

Differential equation models are crucial to scientific processes. The values of model parameters are important for analyzing the behaviour of solutions. A parameter is called globally identifiable if its value can be uniquely determined from the input and output functions. To determine if a parameter estimation problem is well-posed for a given model, one must check if the model parameters are globally identifiable. This problem has been intensively studied for ordinary differential equation models, with theory and several efficient algorithms and software packages developed. A comprehensive theory of algebraic identifiability for PDEs has hitherto not been developed due to the complexity of initial and boundary conditions. Here, we provide theory and algorithms, based on differential algebra, for testing identifiability of polynomial PDE models. We showcase this approach on PDE models arising in the sciences.

q-bio.QM↗

arXiv2023

Near-Optimal Fault Tolerance for Efficient Batch Matrix Multiplication via an Additive Combinatorics Lens

Fault tolerance is a major concern in distributed computational settings. In the classic master-worker setting, a server (the master) needs to perform some heavy computation which it may distribute to $m$ other machines (workers) in order to speed up the time complexity. In this setting, it is crucial that the computation is made robust to failed workers, in order for the master to be able to retrieve the result of the joint computation despite failures. A prime complexity measure is thus the \emph{recovery threshold}, which is the number of workers that the master needs to wait for in order to derive the output. This is the counterpart to the number of failed workers that it can tolerate. In this paper, we address the fundamental and well-studied task of matrix multiplication. Specifically, our focus is on when the master needs to multiply a batch of $n$ pairs of matrices. Several coding techniques have been proven successful in reducing the recovery threshold for this task, and one approach that is also very efficient in terms of computation time is called \emph{Rook Codes}. The previously best known recovery threshold for batch matrix multiplication using Rook Codes is $O(n^{\log_2{3}})=O(n^{1.585})$. Our main contribution is a lower bound proof that says that any Rook Code for batch matrix multiplication must have a recovery threshold that is at least $\omega(n)$. Notably, we employ techniques from Additive Combinatorics in order to prove this, which may be of further interest. Moreover, we show a Rook Code that achieves a recovery threshold of $n^{1+o(1)}$, establishing a near-optimal answer to the fault tolerance of this coding scheme.

cs.DC↗

arXiv2023

Robust Parameter Estimation for Rational Ordinary Differential Equations

We present a new approach for estimating parameters in rational ODE models from given (measured) time series data. In typical existing approaches, an initial guess for the parameter values is made from a given search interval. Then, in a loop, the corresponding outputs are computed by solving the ODE numerically, followed by computing the error from the given time series data. If the error is small, the loop terminates and the parameter values are returned. Otherwise, heuristics/theories are used to possibly improve the guess and continue the loop. These approaches tend to be non-robust in the sense that their accuracy depend on the search interval and the true parameter values; furthermore, they cannot handle the case where the parameters are locally identifiable. In this paper, we propose a new approach, which does not suffer from the above non-robustness. In particular, it does not require making good initial guesses for the parameter values or specifying search intervals. Instead, it uses differential algebra, interpolation of the data using rational functions, and multivariate polynomial system solving. We also compare the performance of the resulting software with several other estimation software packages.

cs.MS↗

arXiv2022

Lightweight Projective Derivative Codes for Compressed Asynchronous Gradient Descent

Coded distributed computation has become common practice for performing gradient descent on large datasets to mitigate stragglers and other faults. This paper proposes a novel algorithm that encodes the partial derivatives themselves and furthermore optimizes the codes by performing lossy compression on the derivative codewords by maximizing the information contained in the codewords while minimizing the information between the codewords. The utility of this application of coding theory is a geometrical consequence of the observed fact in optimization research that noise is tolerable, sometimes even helpful, in gradient descent based learning algorithms since it helps avoid overfitting and local minima. This stands in contrast with much current conventional work on distributed coded computation which focuses on recovering all of the data from the workers. A second further contribution is that the low-weight nature of the coding scheme allows for asynchronous gradient updates since the code can be iteratively decoded; i.e., a worker's task can immediately be updated into the larger gradient. The directional derivative is always a linear function of the direction vectors; thus, our framework is robust since it can apply linear coding techniques to general machine learning frameworks such as deep neural networks.

cs.LG↗

arXiv2022

More Efficient Identifiability Verification in ODE Models by Reducing Non-Identifiability

Structural global parameter identifiability indicates whether one can determine a parameter's value from given inputs and outputs in the absence of noise. If a given model has parameters for which there may be infinitely many values, such parameters are called non-identifiable. We present a procedure for accelerating a global identifiability query by eliminating algebraically independent non-identifiable parameters. Our proposed approach significantly improves performance across different computer algebra frameworks.

cs.SC↗

arXiv2022

Random Alloy Codes and the Fundamental Limits of Coded Distributed Tensors

Tensors are a fundamental operation in distributed computing, \emph{e.g.,} machine learning, that are commonly distributed into multiple parallel tasks for large datasets. Stragglers and other failures can severely impact the overall completion time. Recent works in coded computing provide a novel strategy to mitigate stragglers with coded tasks, with an objective of minimizing the number of tasks needed to recover the overall result, known as the recovery threshold. However, we demonstrate that this strict combinatorial definition does not directly optimize the probability of failure. In this paper, we focus on the most likely event and measure the optimality of a coding scheme more directly by its probability of decoding. Our probabilistic approach leads us to a practical construction of random codes for matrix multiplication, i.e., locally random alloy codes, which are optimal with respect to the measures. Furthermore, the probabilistic approach allows us to discover a surprising impossibility theorem about both random and deterministic coded distributed tensors.

cs.IT↗