研究 03 · 逆问题Research 03 · Inverse problems
同基频非理想弦的谱辨识与参数重建Spectral identification and parameter reconstruction of inharmonic strings
两根弦调到同一个基频，听起来仍然不一样。差别藏在高次泛音的位置上。能不能只凭一段频谱，把张力和有效长度算回来？Two strings tuned to the same fundamental still do not sound alike. The difference hides in where the upper partials fall. Can tension and effective length be recovered from a spectrum alone?
问的是什么The question
依赖参数的振动系统，反过来问参数，是一类常见问题。谱可辨识性问得更基本：把物理参数映到一组频率，这个映射是单射吗？测量带噪声时，它稳定吗？两个问题都答不好，任何反演算法算出的数字都不值得信。Asking for the parameters of a parameter-dependent vibrating system is a familiar kind of inverse problem. Spectral identifiability asks something more basic: is the map from physical parameters to a set of frequencies injective, and is it stable when the measurement carries noise? Without answers to both, the numbers an inversion algorithm returns are not worth much.
模型The model
理想弦只有张力，实际的弦还有抗弯劲度。加上这一项，标准的劲度弦方程写成：An ideal string carries only tension; a real one also resists bending. With that term added, the standard stiff-string equation reads:
劲度把泛音推高，越高的次数推得越多，于是泛音不再是基频的整数倍。这就是非谐性，用一个参数 B 概括，标准的频率修正式为：Stiffness pushes the partials upward, and the higher the partial the more it moves, so they are no longer integer multiples of the fundamental. That is inharmonicity, summarized by a single parameter B in the standard frequency correction:
前向问题由此确定：给定物理参数，算出前若干个频率。研究里用显式有限差分离散方程，边界条件用虚拟点处理，并验证了格式的稳定性和空间二阶收敛。这套数值解算器就是谱映射的实现。That fixes the forward problem: given the physical parameters, compute the first several frequencies. The work discretizes the equation with an explicit finite difference scheme, treats the boundary with fictitious points, and verifies that the scheme is stable and second-order accurate in space. This solver is the spectral map in computable form.
高次泛音带着信息The information sits in the high partials
取两根弦，实际基频完全相同，非谐性参数不同。研究证明，归一化的谱偏差随谐波次数按 O(n2) 增长，绝对频率偏差按 O(n3) 增长。Take two strings with exactly the same actual fundamental and different inharmonicity. The study proves that their normalized spectral deviation grows as O(n2) with harmonic order, while the absolute frequency deviation grows as O(n3).
低次泛音几乎重合，难以分辨，差异要到高次才显现。频率分辨率有限时，可以推出从第几次谐波起两者才真正可分。这个次数决定了测量要做多长。The consequence is practical. Low partials nearly coincide and tell the two strings apart poorly; the gap only becomes visible higher up. Under finite frequency resolution one can derive the lowest harmonic order at which they become distinguishable, and that order sets how long the measurement has to be.
一个参数不够One parameter is not enough
只知道非谐性参数 B，张力和有效长度分不开。不同的 (T, L) 组合给出同一个 B，这是结构性的不可辨识，再精确的测量也救不回来。Knowing the inharmonicity parameter B alone does not separate tension from effective length: different pairs (T, L) give the same B. The non-identifiability is structural, and no amount of measurement precision repairs it.
补入一个绝对尺度即可，理想弦的基频正好提供它：An absolute scale fixes it, and the ideal-string fundamental provides one:
于是反演分两步：先从谱形状定出非谐性参数，再借绝对频率尺度分别确定张力与有效长度。研究证明这个两阶段框架有全局唯一性和局部 Lipschitz 稳定性，即频谱扰动不大时，参数误差被常数倍地控制住。Inversion then runs in two stages: recover the inharmonicity from the shape of the spectrum, then pin down tension and effective length separately using the absolute frequency scale. The study proves this two-stage framework has global uniqueness and local Lipschitz stability, meaning that for small spectral perturbations the parameter error stays bounded by a constant times the measurement error.
数值结果Numerical results
| 参数Parameter | 平均相对误差Mean relative error |
|---|---|
| 非谐性参数 BInharmonicity B | 1.29% |
| 有效长度 LEffective length L | 0.32% |
| 张力 TTension T | 0.66% |
B 的误差比另两项大，跟它靠高次泛音定出来有关，高次泛音的幅度小，也更容易被噪声盖住。The error on B is the largest of the three, which fits the way it is determined: from high partials, where amplitudes are small and noise covers more easily.
可以用在哪里Where it is useful
一段录音里，弦的物理参数原则上是可以读出来的。音频分析和物理建模合成都需要这类反演：前者想知道乐器的状态，后者要拿参数去驱动一个数值弦。这项工作给这类反问题补上了数学依据。In principle a recording carries the physical state of the string that produced it. Audio analysis and physical modeling synthesis both need this kind of inversion, one to learn what an instrument is doing, the other to drive a numerical string with the recovered parameters. This work supplies the mathematical footing for it.