Skip to main content

m07 · Inverse problems and regularization (Tikhonov, L-curve)

Warning

Education only. Not for diagnosis, treatment or any clinical decision.

In 60 seconds

A forward problem predicts measurements from sources; an inverse problem recovers sources from measurements. Many inverse problems in EP are ill-posed: the forward map smooths, so very different sources give nearly the same data, and inverting it amplifies noise without limit. The simplest example is distance. Between the tissue and an electrode a height \(h\) away, in source-free blood, every spatial harmonic of wavenumber \(k\) shrinks by \(e^{-|k|h}\), so fine detail fades exponentially with distance. The singular value decomposition shows the damage: dividing by tiny singular values turns noise into nonsense. Regularization gives up the detail the data cannot support. Tikhonov’s method does this smoothly with one parameter \(\lambda\), which trades bias for noise and can be chosen by the discrepancy principle, the L-curve or generalized cross-validation. Read as Bayesian estimation, it is the best estimate under a Gaussian prior, and it comes with error bars that are only as good as that prior.

From the field

  • Unregularized solutions are useless. Naive reconstructions of heart-surface potentials from body-surface data are wildly oscillatory (Rudy and Messinger-Rapport 1988). In lab part B, 0.1% noise gives a naive reconstruction error of millions of times the signal.
  • Distance costs resolution. Noncontact mapping reconstructs endocardial potentials from a multielectrode probe in the cavity by a regularized inverse solution. In an isolated canine ventricle it located pacing sites within 10 mm and resolved sites 10–20 mm apart (Khoury et al. 1995). In the human left ventricle, timing agreement with contact electrograms worsened when the endocardium was more than 34 mm from the array (Schilling et al. 1998).
  • The parameter matters as much as the method. The L-curve corner balances residual against solution size (Hansen 1992; Hansen and O’Leary 1993), and generalized cross-validation needs no noise estimate (Golub et al. 1979). In lab part D the median error of all three rules was within 1.2 times the best possible at 1-mm height, yet the L-curve corner gave 3.5 times the best error at 0.5 mm with 0.1% noise. Report results for a range of \(\lambda\).
  • Error bars need a prior. The Bayesian reading gives a covariance for free; its intervals are right only when the prior describes the tissue (lab part E).

Derivations

Symbols used throughout. \(\mathbf x\) unknown sources (\(n\) values), \(\mathbf b\) data (\(m\) values), \(A\) the forward matrix, \(\mathbf e\) noise of standard deviation \(\sigma\) per sample. \(A = \sum_i s_i\mathbf u_i\mathbf v_i^\top\) the singular value decomposition (SVD), \(s_1 \ge s_2 \ge \dots\); \(\kappa\) condition number. \(\hat\phi(k, z)\) the Fourier coefficient of a potential at wavenumber \(k\) (rad/mm) and height \(z\) (mm); \(h\) electrode height; \(\varepsilon\) relative noise. \(f_i\) filter factors, \(k_c\) truncation index, \(\lambda\) regularization parameter, \(\tau\) safety factor. \(\sigma_x\), \(\sigma_e\) prior and noise standard deviations.

Ill-posed problems

Measurements depend linearly on sources in every model of this course (m02):

\[ \mathbf b = A\mathbf x + \mathbf e, \qquad \frac{\lVert\delta\mathbf x\rVert}{\lVert\mathbf x\rVert} \le \kappa(A)\,\frac{\lVert\delta\mathbf b\rVert}{\lVert\mathbf b\rVert}, \qquad \kappa(A) = \frac{s_1}{s_n} \tag{1}\]

The condition number \(\kappa\) bounds how much a relative error in the data can grow in the solution, and the bound is reached when the data lie along \(\mathbf u_1\) and the error along \(\mathbf u_n\). A problem is called ill-posed when a solution fails to exist, to be unique or to depend continuously on the data. The last failure is the one that matters here; after discretization it shows up as a huge \(\kappa\) that grows as the grid is refined.

The worked example: continuation across blood. In a source-free layer the potential obeys Laplace’s equation (m02), and \(e^{-kz}\cos kx\) is a solution. Each harmonic of a potential on the tissue therefore reaches an electrode a height \(h\) above it multiplied by a gain:

\[ \hat\phi(k, h) = e^{-\lvert k\rvert h}\,\hat\phi(k, 0), \qquad \lvert k\rvert < k_{\max} = \frac{\ln(1/\varepsilon)}{h} \tag{2}\]

The m03 front shape \(s/(s^2 + a^2)\) has harmonics proportional to \(e^{-|k|a}\), so continuing a front of width \(w\) by \(h\) gives a front of width \(w + h\): the closed form of m03 (eq-m03-unipolar-closed) is this smoothing. On a grid of spacing \(\Delta x\) the gains run from 1 down to \(e^{-\pi h/\Delta x}\), so \(\kappa = e^{\pi h/\Delta x}\): about \(8\times10^{10}\) for \(h\) = 1 mm and \(\Delta x\) = 0.125 mm (lab part A). The second part of the equation is a rule of thumb. If the tissue pattern had the same amplitude in every harmonic and the noise in each harmonic were a fraction \(\varepsilon\) of it, a harmonic would be worth recovering only while its gain exceeds \(\varepsilon\). The shortest recoverable wavelength, about \(2\pi h/\ln(1/\varepsilon)\), grows in proportion to the distance: 1.4 mm at \(h\) = 1 mm and 4.1 mm at \(h\) = 3 mm for 1% noise.

Key takeaways

  • Ill-posed means small data errors cause large solution errors; \(\kappa\) measures how large.
  • Across a source-free layer, detail decays as \(e^{-|k|h}\); resolution worsens in proportion to distance.

The singular value decomposition and the naive inverse

Writing the data in the singular vectors shows what inversion does:

\[ \mathbf x_{\text{naive}} = \sum_{i} \frac{\mathbf u_i^\top\mathbf b}{s_i}\,\mathbf v_i = \mathbf x + \sum_{i} \frac{\mathbf u_i^\top\mathbf e}{s_i}\,\mathbf v_i \tag{3}\]

Noise has about the same size in every \(\mathbf u_i\), and dividing by small \(s_i\) magnifies it. A good solution needs the data coefficients \(\lvert\mathbf u_i^\top\mathbf b\rvert\) to fall faster than the \(s_i\), the discrete Picard condition (Hansen 1990a). Clean continued data meet it: their ratio to the gains is the tissue spectrum, which falls as \(e^{-|k|w}\). Noisy data level off at the noise floor, and beyond that point the ratio explodes (lab part B). For the continuation operator the singular vectors are sines and cosines and the singular values are the gains.

Key takeaways

  • The naive inverse divides noise by the singular values.
  • Plot \(\lvert\mathbf u_i^\top\mathbf b\rvert\) against \(s_i\) (a Picard plot) to see where the data stop carrying signal.

Regularization as filtering

Regularization multiplies each term by a filter factor:

\[ \mathbf x_{\text{filt}} = \sum_i f_i\,\frac{\mathbf u_i^\top\mathbf b}{s_i}\,\mathbf v_i, \qquad f_i^{\text{TSVD}} = \begin{cases} 1, & i \le k_c \\ 0, & i > k_c \end{cases}, \qquad f_i^{\text{Tikh}} = \frac{s_i^2}{s_i^2 + \lambda^2} \tag{4}\]

Truncated SVD keeps the \(k_c\) largest components. Tikhonov regularization, the minimizer of \(\lVert A\mathbf x - \mathbf b\rVert^2 + \lambda^2\lVert\mathbf x\rVert^2\), passes components with \(s_i \gg \lambda\), halves the one with \(s_i = \lambda\) and damps the rest. When the discrete Picard condition holds the two give very similar smooth solutions (Hansen 1990b); in lab part C their best errors differ by 8%. m32 writes the same method with a general penalty matrix (eq-m32-tikhonov).

Simulation

Assumptions

  • Model: three fronts with the m03 shape (Equation 2) on a periodic 32-mm line sampled at 256 points, continued to height \(h\), plus one fixed draw of white Gaussian noise (seeded generator, Box–Muller); Tikhonov reconstruction harmonic by harmonic (Equation 4). Tested against epml.inverse.
  • Parameters: front width 0.3 mm and default height 1 mm as in m03-d1 (Riccio et al. 2021); front positions (8, 18 and 21 mm), noise level and \(\lambda\) are free choices for illustration.
  • What it ignores: curved surfaces, electrode size, noise correlated between samples, and time.

Key takeaways

  • Every linear regularization method is a set of filter factors.
  • Tikhonov and truncated SVD behave alike when the Picard condition holds.

The error budget

With white noise of standard deviation \(\sigma\), the expected squared error of any filtered solution splits in two:

\[ \operatorname{E}\lVert\mathbf x_{\text{filt}} - \mathbf x\rVert^2 = \sum_i (1 - f_i)^2\,(\mathbf v_i^\top\mathbf x)^2 + \sigma^2 \sum_i \frac{f_i^2}{s_i^2} \tag{5}\]

The first sum is bias, the detail the filter removes; the second is noise passed through. Raising \(\lambda\) trades the second for the first. For a single component the best factor is \(s_i^2c_i^2/(s_i^2c_i^2 + \sigma^2)\) with \(c_i = \mathbf v_i^\top\mathbf x\), which is Tikhonov’s factor with \(\lambda = \sigma/\lvert c_i\rvert\). In lab part C the formula matches 300 simulated noise draws within 1%, and at 1-mm height with 1% noise the best relative error is 0.28. The best reconstruction separates the two fronts 3 mm apart at 1 mm, where the recording shows one merged peak, and cannot at 2 or 3 mm.

Exact

Assumptions

  • Model: Equation 5 for Tikhonov filters applied to the continuation operator (Equation 2) on a periodic 32-mm line sampled at 256 points; the truth is the three-front profile of m07-d2. Tested against epml.inverse.
  • Parameters: front width 0.3 mm and default height 1 mm as in m03-d1 (Riccio et al. 2021); noise is white, given as a fraction of the recorded peak, a free choice for illustration.
  • What it ignores: curved surfaces, electrode size, correlated noise and model error in \(A\).

Key takeaways

  • Error = bias + noise; \(\lambda\) sets the balance.
  • Tikhonov is the best per-component filter when every component has the same expected size.

Choosing λ

Without the truth, \(\lambda\) must come from the data. Two rules can be written as equations:

\[ \text{discrepancy: } \lVert A\mathbf x_\lambda - \mathbf b\rVert = \tau\lVert\mathbf e\rVert; \qquad \text{GCV: } \min_\lambda \frac{\lVert A\mathbf x_\lambda - \mathbf b\rVert^2}{\left[m - \sum_i f_i\right]^2} \tag{6}\]

The discrepancy principle asks the model to fit the data no better than the noise allows, with \(\tau \ge 1\); it needs the noise level. Generalized cross-validation (GCV) minimizes a function of the data alone; its denominator is the trace of \(I - A A_\lambda^{\#}\), with \(A_\lambda^{\#}\) the matrix that maps data to the Tikhonov solution (Golub et al. 1979). The L-curve plots \(\lVert\mathbf x_\lambda\rVert\) against \(\lVert A\mathbf x_\lambda - \mathbf b\rVert\) on log scales and picks the corner, where both are small (Hansen 1992; Hansen and O’Leary 1993); m32 gives its form (eq-m32-lambda). Hansen and O’Leary report that the L-curve is more robust than GCV when errors are correlated (Hansen and O’Leary 1993). In lab part D, at 1-mm height, the median error of each rule was within 1.2 times the best possible; at 0.5 mm with 0.1% noise the L-curve corner gave 3.5 times the best error.

Key takeaways

  • Discrepancy needs the noise level; GCV and the L-curve do not.
  • No rule is safe everywhere; show how the answer changes with \(\lambda\).

The Bayesian reading

Suppose the sources are drawn from a Gaussian prior and the noise is Gaussian. Bayes’ rule then gives a Gaussian posterior:

\[ \mathbf x \sim \mathcal N(0, \sigma_x^2 I),\ \mathbf e \sim \mathcal N(0, \sigma_e^2 I) \;\Longrightarrow\; \operatorname{E}[\mathbf x \mid \mathbf b] = (A^\top A + \lambda^2 I)^{-1}A^\top\mathbf b, \quad \operatorname{Cov}[\mathbf x \mid \mathbf b] = \sigma_e^2 (A^\top A + \lambda^2 I)^{-1}, \quad \lambda = \frac{\sigma_e}{\sigma_x} \tag{7}\]

The posterior mean is the Tikhonov solution, and \(\lambda\) becomes a ratio of noise to prior spread. The covariance gives error bars. In lab part E, 95% intervals contained the truth 95% of the time when the sources really came from the prior; they mean less when the prior is a convenience. Ridge regression in machine learning is the same estimate with a data matrix in place of \(A\) (Golub et al. 1979).

Key takeaways

  • Tikhonov = posterior mean under a Gaussian prior, with \(\lambda = \sigma_e/\sigma_x\).
  • Error bars are a property of the prior as much as of the data.

Clinical correlate

NotePending clinical review

This section describes published findings only. It is not guidance for patient care, and it has not yet been reviewed by an EP physician.

Where this shows up in EP. Noncontact mapping reconstructs endocardial potentials from a multielectrode probe inside the cavity. In an isolated canine left ventricle, a boundary-element model with regularization reconstructed the main features of measured endocardial potentials and located pacing sites within 10 mm (Khoury et al. 1995); a follow-up study reconstructed electrograms whose correlation with measured ones exceeded 0.9 at 60% of electrodes and resolved two pacing sites 17 mm apart (Liu et al. 1997). In the left ventricles of 13 patients, the timing difference between reconstructed and contact unipolar electrograms near the array’s equator was −1.94 ± 7.12 ms (mean ± SD) where the endocardium was less than 34 mm from the array center and −14.16 ± 19.29 ms beyond it (Schilling et al. 1998). ECG imaging (m32) solves the same kind of problem from the body surface (Rudy and Messinger-Rapport 1988; Messinger-Rapport and Rudy 1990).

What the evidence is. Experimental validation in isolated canine hearts (Khoury et al. 1995; Liu et al. 1997), a comparison with contact electrograms in patients (Schilling et al. 1998), and the ECG-imaging studies discussed in m32.

What a reader should ask:

  • How far were the sources from the sensors? Resolution worsens with distance (Equation 2).
  • Which regularization, and how was its parameter chosen? Results can change with \(\lambda\) (Equation 6).
  • Was accuracy reported against direct recordings, and as a function of distance? As in (Schilling et al. 1998).

Visual summary

Schematic

forward map smooths Picard plot: data meet noise filters: TSVD and Tikhonov L-curve corner
Figure 1: Ill-posed inversion in four pictures: smoothing, the Picard plot, filtering and choosing the parameter.

Assumptions

How experts think

Before trusting a reconstruction, ask:

  1. What is the forward model, and how wrong could it be? Every error in \(A\) is inverted too (Equation 1).
  2. How far are the sources from the sensors? Distance sets the resolution limit (Equation 2).
  3. Where do the data stop carrying signal? Look at a Picard plot (Equation 3).
  4. Which filter and which \(\lambda\)? And how much does the result change across a range of \(\lambda\) (Equation 4, Equation 6)?
  5. Are error bars reported, and what prior do they assume? (Equation 7)

One sentence to keep: regularization does not recover lost detail; it decides which detail to give up.

Cheat sheet

Quantity Result
Condition number (Equation 1) \(\kappa = s_1/s_n\) bounds relative error growth
Continuation (Equation 2) gain \(e^{-\lvert k\rvert h}\); \(\kappa = e^{\pi h/\Delta x}\) on a grid; shortest wavelength \(\approx 2\pi h/\ln(1/\varepsilon)\)
Naive inverse (Equation 3) noise coefficients divided by \(s_i\); Picard: data coefficients must fall faster than \(s_i\)
Filters (Equation 4) TSVD: keep \(k_c\) terms; Tikhonov: \(s^2/(s^2 + \lambda^2)\), one half at \(s = \lambda\)
Error budget (Equation 5) bias \(\sum(1 - f)^2c^2\) + noise \(\sigma^2\sum f^2/s^2\)
Choosing \(\lambda\) (Equation 6) discrepancy (needs \(\lVert\mathbf e\rVert\)), L-curve corner, GCV
Bayesian reading (Equation 7) Tikhonov = posterior mean, \(\lambda = \sigma_e/\sigma_x\)

Values used in this module

Value Source
Canine noncontact inverse: pacing sites within 10 mm, resolution 10–20 mm Khoury et al. (1995)
Electrogram correlation > 0.9 at 60% of electrodes; sites 17 mm apart resolved Liu et al. (1997)
Human timing agreement −1.94 ± 7.12 ms within 34 mm of the array, −14.16 ± 19.29 ms beyond Schilling et al. (1998)
Front width 0.3 mm, height 1 mm (demo defaults, as in m03-d1) m03, Riccio et al. (2021)
Condition numbers, error budgets, resolution, parameter-choice scores, coverage our own code

Interview questions

Try each question before you open its model answer. In a real interview, say which assumptions you are making.

Q1. What makes an inverse problem ill-posed?

Model answer

Its solution does not depend continuously on the data: small data errors cause large solution errors. After discretization this shows as a huge condition number that grows with refinement (Equation 1).

Q2. Why does recording farther from the tissue lose detail?

Model answer

Across source-free blood each harmonic shrinks by \(e^{-|k|h}\), so short wavelengths fall below the noise first; the shortest recoverable wavelength grows in proportion to \(h\) (Equation 2).

Q3. Condition number \(10^{10}\), data accurate to \(10^{-4}\). What can you say about the solution?

Model answer

Only that its relative error could be as large as \(10^{6}\): without regularization the bound guarantees nothing (Equation 1).

Q4. Explain the naive inverse in terms of the SVD.

Model answer

Each data coefficient is divided by its singular value; noise coefficients are divided too, so small singular values amplify noise (Equation 3).

Q5. What is a Picard plot and what do you look for?

Model answer

Data coefficients \(\lvert\mathbf u_i^\top\mathbf b\rvert\) against singular values. They should fall faster than the \(s_i\); where they level off at the noise floor, the data stop carrying signal (Equation 3) (Hansen 1990a).

Q6. Write the Tikhonov filter factor and interpret it.

Model answer

\(f = s^2/(s^2 + \lambda^2)\): about 1 for \(s \gg \lambda\), one half at \(s = \lambda\), about \(s^2/\lambda^2\) below (Equation 4).

Q7. Tikhonov or truncated SVD?

Model answer

When the discrete Picard condition holds they give very similar solutions (Hansen 1990b); Tikhonov is smooth in \(\lambda\) and needs no SVD, TSVD is simple to explain. In lab part C their best errors differ by 8%.

Q8. Split the expected error of a regularized solution.

Model answer

Bias \(\sum(1 - f_i)^2(\mathbf v_i^\top\mathbf x)^2\) from removed detail plus noise \(\sigma^2\sum f_i^2/s_i^2\) passed through (Equation 5).

Q9. Compare the discrepancy principle, the L-curve and GCV.

Model answer

Discrepancy needs the noise level and fits the data only to that level; the L-curve picks the corner of solution size against residual; GCV minimizes a predicted-error function of the data. Hansen and O’Leary report the L-curve more robust than GCV to correlated errors (Equation 6) (Hansen and O’Leary 1993; Golub et al. 1979).

Q10. What is the Bayesian meaning of λ?

Model answer

With a Gaussian prior of spread \(\sigma_x\) and noise \(\sigma_e\), Tikhonov is the posterior mean with \(\lambda = \sigma_e/\sigma_x\) (Equation 7).

Q11. Why did noncontact reconstructions lose timing accuracy far from the array?

Model answer

Detail decays with distance from the sources (Equation 2), so far sites are reconstructed from fewer usable components. In patients, timing agreement worsened beyond 34 mm from the array (Schilling et al. 1998).

Q12. A colleague shows one beautiful reconstruction. What do you ask for?

Model answer

The forward model and its errors, the regularization and how \(\lambda\) was chosen, results over a range of \(\lambda\), validation against direct recordings, and accuracy as a function of distance (Equation 6).

Code lab

The lab builds the continuation problem, breaks the naive inverse and repairs it.

Part What you do Data
A The forward map: continuation, the m03 front shape and the condition number (Equation 1, Equation 2) synthetic
B The naive inverse and the Picard plot (Equation 3) synthetic
C Filters, the error budget and resolution (Equation 4, Equation 5) synthetic
D Choosing λ: discrepancy, L-curve, GCV (Equation 6) synthetic
E The Bayesian reading and credible intervals (Equation 7) synthetic

Each part ends with checks: lines that pass with the code as given. If you change something and a check fails, that is information, not an error to hide.

Self-check

Short checks you can do in your head. Open each answer after you have tried it.

S1. Singular values \(10^{0}\) to \(10^{-8}\). Condition number?

Answer

\(10^{8}\) (Equation 1).

S2. A harmonic with wavelength 2 mm, electrode 1 mm away. Gain?

Answer

\(k = \pi\) rad/mm, so \(e^{-\pi} \approx 0.04\) (Equation 2).

S3. Noise 1% of each harmonic, height 2 mm. Shortest recoverable wavelength by the rule of thumb?

Answer

\(2\pi \cdot 2/\ln 100 \approx 2.7\) mm (Equation 2).

S4. \(s = \lambda\). Tikhonov factor?

Answer

One half (Equation 4).

S5. You double λ. Which part of the error grows and which shrinks?

Answer

Bias grows, noise shrinks (Equation 5).

S6. Prior spread 1, noise 0.01. Tikhonov λ?

Answer

0.01 (Equation 7).

Literature figures

These figures are worth looking at alongside this module. We link to them instead of copying them: some licenses do not allow reuse here, and others we could not confirm. Each entry says what to look for and where this module explains it.

  • Noncontact inverse solutions. Khoury and colleagues compare measured and inversely computed endocardial potential maps and isochrones (Khoury et al. 1995). This module: ill-posed problems and regularization.
  • Accuracy against distance. Schilling and colleagues plot timing differences between reconstructed and contact electrograms against distance from the array (Schilling et al. 1998). This module: continuation.
  • The L-curve. Hansen shows the L-curve of a discrete ill-posed problem and its corner (Hansen 1992). This module: choosing λ.

Further reading

References

Golub, Gene H., Michael Heath, and Grace Wahba. 1979. “Generalized Cross-Validation as a Method for Choosing a Good Ridge Parameter.” Technometrics 21 (2): 215–23. https://doi.org/10.1080/00401706.1979.10489751.
Hansen, Per Christian. 1990a. “The Discrete Picard Condition for Discrete Ill-Posed Problems.” BIT Numerical Mathematics 30 (4): 658–72. https://doi.org/10.1007/BF01933214.
Hansen, Per Christian. 1990b. “Truncated Singular Value Decomposition Solutions to Discrete Ill-Posed Problems with Ill-Determined Numerical Rank.” SIAM Journal on Scientific and Statistical Computing 11 (3): 503–18. https://doi.org/10.1137/0911028.
Hansen, Per Christian. 1992. “Analysis of Discrete Ill-Posed Problems by Means of the L-Curve.” SIAM Review 34 (4): 561–80. https://doi.org/10.1137/1034115.
Hansen, Per Christian, and Dianne P. O’Leary. 1993. “The Use of the L-Curve in the Regularization of Discrete Ill-Posed Problems.” SIAM Journal on Scientific Computing 14 (6): 1487–503. https://doi.org/10.1137/0914086.
Khoury, D S, B Taccardi, R L Lux, P R Ershler, and Y Rudy. 1995. “Reconstruction of Endocardial Potentials and Activation Sequences from Intracavitary Probe Measurements. Localization of Pacing Sites and Effects of Myocardial Structure.” Circulation 91 (3): 845–63. https://doi.org/10.1161/01.cir.91.3.845.
Liu, Z W, P Jia, P R Ershler, et al. 1997. “Noncontact Endocardial Mapping: Reconstruction of Electrograms and Isochrones from Intracavitary Probe Potentials.” Journal of Cardiovascular Electrophysiology 8 (4): 415–31. https://doi.org/10.1111/j.1540-8167.1997.tb00807.x.
Messinger-Rapport, B J, and Y Rudy. 1990. “Noninvasive Recovery of Epicardial Potentials in a Realistic Heart-Torso Geometry. Normal Sinus Rhythm.” Circulation Research 66 (4): 1023–39. https://doi.org/10.1161/01.res.66.4.1023.
Riccio, Jennifer, Alejandro Alcaine, Sara Rocher, Laura Martinez-Mateu, et al. 2021. “Characterization of Atrial Propagation Patterns and Fibrotic Substrate with a Modified Omnipolar Electrogram Strategy in Multi-Electrode Arrays.” Frontiers in Physiology 12: 674223. https://doi.org/10.3389/fphys.2021.674223.
Rudy, Y, and B J Messinger-Rapport. 1988. “The Inverse Problem in Electrocardiography: Solutions in Terms of Epicardial Potentials.” Critical Reviews in Biomedical Engineering 16 (3): 215–68.
Schilling, R J, N S Peters, and D W Davies. 1998. “Simultaneous Endocardial Mapping in the Human Left Ventricle Using a Noncontact Catheter: Comparison of Contact and Reconstructed Electrograms During Sinus Rhythm.” Circulation 98 (9): 887–98. https://doi.org/10.1161/01.cir.98.9.887.