1

Supplementary Figure 1: Flowchart of MUSICAL. Additional algorithmic details appear in Supplementary Note 2.
 2

Supplementary Figure 2: Window size and soft window. The effect of the window size and the soft window function is
demonstrated here using a small portion of region B of in-vitro sample 1. According to eq. (14), the value of Nw for in-vitro
sample 1 is 7. However, the result for Nw = 7 without the soft window function clearly illustrates grid-like artifacts related
to truncation. Moreover, the images obtained without the soft window function (top-row) show incorrect reconstruction since
they do not weigh the center pixel, at which the PSF is most reliable, more than the other pixels in a window. On the other
hand, use of the soft window function results in more accurate imaging result (bottom row). Further, with the use of the soft
window function, values of Nw larger than recommended in eq. (14) do not change the MUSICAL result despite using more
computational resources. But, the value of Nw less than recommended results in grid-like artifacts (bottom left).
 3

Supplementary Figure 3: Choice of σ0 in the condition M < min(N, K), example SynEx1. Synthetic example SyEx1
illustrates the choice of σ0 for the condition M < min(N, K), the effect of noise on the modified MUSIC indicator function,
and the effect of the value of α. (a) Plot of singular values. Insets show the mean image of the image stacks with and without
noise. Middle row corresponds to data without noise while bottom row corresponds to data with shot noise such that signal to
noise ratio is 128. (b,c,e,f) correspond to α = 1, where (c,f) are zoom-in of the central region of (b,e). (d,g) correspond to α = 4
and show the same region as shown in (c,f). (b) is shown in logarithmic scale for ease of visualization. The color bar of (g)
has been made slightly more red for the ease of visualization of the right most emitter. More details appear in Supplementary
Note 4. Color bar in (b) indicates logarithm of MUSICAL indicator function values. Color bars in (c-g) indicate MUSICAL
indicator function values.
 4

Supplementary Figure 4: Choice of σ0 in the condition M ≥ min(N, K), example SynEx2. Synthetic example SynEx2 in
the absence of noise is used to illustrate the effect of the value of σ0 in the condition M ≥ min(N, K). Specifically, the
sliding window at the center pixel of the image stack is considered. Plot of singular values is shown at the top. Inset shows
the emitter distribution. Different horizontal lines correspond to different values of σ0 and arrows from them point to the
corresponding modified indicator function plots (α = 1 is used). More details appear in Supplementary Note 4. Color bars
indicate MUSICAL indicator function values.
 5

Supplementary Figure 5: Illustration of the effect of imaging approximation on the mapping vector G(rn , r0m ) from a
point r0m in the sample plane to pixels rn in the image plane. Refer to Supplementary Note 1 and Supplementary Note 2
for its usage.
 6

Supplementary Figure 6: Illustration of eigenimages and projections of PSF on eigenimages. The first six eigenimages (top
row), the projections of the PSFs of the test points on the eigenimages for SynEx1 (middle row), and the logarithmic maps
of the projections (bottom row) are shown here. Actual emitter locations are shown using black dots in the top row and red
dots in the middle and the bottom rows. A detailed discussion is given in Supplementary Note 5. Color bars in the top row
indicate the pixel values of eigenimages ui . Color bars in the middle row indicate the values of kG0 (r0test ) · ui k. Color bars
in the bottom row indicate the values of log10 kG0 (r0test ) · ui k.
 7

Supplementary Figure 7: Sensitivity of MUSICAL to accurate estimation of PSF. In case that the PSF is characterized
or calibrated experimentally or does not match the actual PSF of the system, it is of interest to investigate the sensitivity of
MUSICAL to the incorrect PSF. An example of sensitivity of MUSICAL to PSF is shown here using a small portion of region
B of in-vitro sample 1. Images on the left side correspond to Airy PSFs with widths not matching with the actual Airy PSF
while the right side corresponds to images with Gaussian approximations of PSFs. When a wider PSF is used (estimated NA
smaller than actual NA), MUSICAL image appears sharper and more punctuated as seen in bottom left images. Vice versa
for a narrower PSF (estimated NA larger than actual NA) as seen in top left images. In the case that the shape of the PSF
is approximated as a Gaussian function, the MUSICAL result is almost the same for the actual and Gaussian approximated
PSFs. However, if the Gaussian approximated PSF is too narrow, then grid-like artifacts appear in the image. In conclusion,
MUSICAL is not very sensitive to the shape and width of the PSF. Nevertheless, it is significantly more sensitive to the
PSF than most single molecule localization techniques and 3B. In our observation, deconSTORM also demonstrates similar
sensitivity to PSF because of the sensitivity of deconvolution to the estimate of PSF.
 8

Supplementary Figure 8: Automatic stitching ability of the modified indicator function of MUSICAL. Synthetic example
SynEx2 with noise is used to demonstrate the automatic stitching ability of the modified MUSICAL indicator function, as
opposed to MUSIC’s original indicator function. We consider the entire image stack with Poisson noise added to the data such
that the signal to noise ratio (SNR) of the noisy image stack is 8. The singular values and the value of σ0 (gray dashed line)
are plotted in (a). σ0 is computed using eq. (28). Further, inset 1 and inset 2 show the mean image of the entire image stack
(the MUSICAL results for which are shown in (b-d)) and its central portion (the MUSICAL results for which are shown in
(e-g)), respectively. (b-d) show MUSICAL results obtained using MUSIC’s original indicator function (b), modified indicator
function of MUSICAL with α = 1 (c), and modified indicator function of MUSICAL with α = 4 (d). (e-g) are the zoom-ins of
the central portions of (b-d) respectively. The artifacts due to the pixel boundaries are clearly visible in (e) where the original
MUSIC indicator function is used. These artifacts are significantly suppressed with the modified MUSICAL indicator function
with α = 1 as seen in (f). Further, the artifacts due to pixel boundaries are completely removed using α = 4 (g). We have
observed that in general, α ≥ 2 does not have the pixel boundary artifacts if the value of σ0 is close to the optimal value. The
halo effect seen prominently in (b,c) is due to the side lobes and is discussed in Supplementary Figure 14. The role of α is
further discussed in Supplementary Note 6. Color bars in (b-g) indicate MUSICAL indicator function values.
 9

Supplementary Figure 9: FWHM and structural resolution of MUSICAL. FWHM for SynEx3 as a function of the signal
to background ratio (SBR) is shown in (a). The minimum value of FWHM is 27.11 nm for SBR 10 and the maximum value
is 158.2 nm for SBR 1.2. Quantification of structural resolution through synthetic example SynPairDelX is done in (b-e). The
−Imin
contrast used here is defined as IImax
max +Imin
where Imax is the maximum intensity among the two consecutive maxima and
Imin is the intensity of the minimum between them. (b) shows the contrast between two lines of emitter separated by distance
∆x computed using synthetic dataset SynPairDelX. The geometry of SynPairDelX is shown in the inset. The red colored star
shown on the plot in (b) corresponds to ∆x = 25 nm and has a contrast of 0.48. The mean image for the image stack and
MUSICAL result for the SynPairDelX data with ∆x = 25 nm are shown in (c,d), respectively. The intensity at section E
shown in (d) is plotted in (e). It shows that the maxima are clearly separated with a distance of 25 nm between them. No other
statistical method could resolve these two lines. STORM localized only three emitters for this data. Scale bars: (c,d) 200 nm.
 10

Supplementary Figure 10: Periodicity observed in actin filaments. MUSICAL intensity profiles of lines A1 and C1 in Fig.
1 of the main paper are shown in (a). Fourier spectra of these are shown in (b). For computing the Fourier amplitudes, the
intensity profiles were mean compensated and fast Fourier transform was computed. As noted in (b), there are two peaks in
the Fourier spectrum for the profile A1 at sampling frequencies corresponding to 62.5 nm (β1 ) and 75 nm (β2 ). The other
peaks in the Fourier spectrum of A1 are harmonics of these two sampling frequencies. Similarly, there is one clear peak in the
Fourier spectrum for the profile C1 at sampling frequency corresponding to 65.5 nm (γ1 ) and the other peak is the 4th order
harmonic of this frequency. We use SynPeriod to further validate the periodicity observed. MUSICAL result for SynPeriod is
shown in (c). (d) shows the MUSICAL intensity along the section indicated by the blue arrows in (c). (e) shows its Fourier
spectrum. The Fourier spectrum in (e) shows dominant peaks at about 50 nm or its multiples. This indicates that the peaks
observed in Fourier spectra of the in-vitro actin filaments are related to the periodicity of actin filaments. Scale bar: (c) 100
nm.
 11

Supplementary Figure 11: Comparison of density of fluorophores in in-vitro samples 1 and 2. Plots of sum of the measured
intensities in different 7 × 7 regions R1 and R2 of sample 2 (a) and R3 and R4 of sample 1 (b) are shown in (c). The plotted
intensities are an indicator of density of fluorophores in the region since all the remaining experimental factors are the same.
The intensities in sample 2 (R1 and R2) are significantly larger than in sample 1 (R3 and R4), indicating much higher density
of fluorophores in sample 2 than in sample 1. For the same distribution of on and off times, higher density of fluorophores
translates to less sparse blinking. The regions R3 and R4 correspond to the branch of actin filaments in region A of sample
1 and a single actin filament, respectively. It is seen in the inset of Supplementary Figure 11(c) that the intensity in R3 is
approximately 1.8 times the intensity in R4, indicating more densely packed fluorophores in R3. All the plots correspond to
image stacks captured with excitation laser’s intensity of 205.6 W cm−2 . Scale bars: (a) 1 µm; (b) 500 nm.
 12

Supplementary Figure 12: STORM results for in-vitro sample 2. The outline of the zoom-in regions and the section line
have the same coordinates as Fig. 2 of the main paper. The details in STORM results are significantly less than the details in
the MUSICAL results in Fig. 2 of the main paper. As inferred in Supplementary Figure 11, the blinking in sample 2 is not
sparse even for 205.6 W cm−2 and thus is not well-suited for STORM. Unsurprisingly, MUSICAL outperforms STORM for
this example. Scale bars: (a-c) 2 µm; (d-f) 1 µm.
 13

Supplementary Figure 13: Results for in-vitro sample 3 imaged at different excitation powers. Actin filaments are imaged
using a TIRF microscope and an excitation laser with power 0.93 W cm−2 (a-c, top row), 10.3 W cm−2 (d-f, second row),
and 205.6 W cm−2 (g-i, third row). The mean images are shown in (a,d,g), while the MUSICAL results are shown in
(b,e,h). Intensities in the mean image and the MUSICAL image for cross-sections D1-D3 are compared in (c,f,i), respectively.
Comparison of MUSICAL for sections D1-D3 is shown in (j). Empirical plot of power versus MUSICAL FWHM is shown
in (k). More details appear in Supplementary Note 7. Scale bars: 1 µm.
 14

Supplementary Figure 14: MUSICAL and side lobes. Mean image of in-vitro sample 1 and the region D are shown in (a).
MUSICAL result for region D is shown in (b). Normalized MUSICAL intensity profiles from sections D1-D3 are shown in
(c) and the corresponding actual intensities are shown in (d). The presence of slide lobes is clearly evident in (c). We note that
this effect is similar to the well-known effect of increased side lobes when pupil filters such as annular filters are introduced
to reduce the size of the focal spot of a focusing system [1]. Analogously, the improvement of resolution in MUSICAL as
compared to the original image stack is accompanied with side lobes. However, as noted in (d), the level of side lobes remains
almost the same for all the three curves, irrespective of their intensities. Further, we plot the side lobe levels relative to the
peak intensity of MUSICAL result (α = 4) as a function of the SBR for SynEx3 in (e). For low values of SBR, the side lobe
levels may be quite prominent. We also note that such side lobes appear in other results in electromagnetic imaging problem
as well [2], although their occurrence was not discussed. In practice, this effect of side lobes can be removed by adjusting
the minimum intensity of the color bar to a small non-zero value commensurate with the side lobe levels. More sophisticated
approaches, such as discussed in Supplementary Note 9, may be adopted. Scale bars: (a) 1 µm, (b) 500 nm. Color bar in (b)
indicates MUSICAL indicator function values.
 15

Supplementary Figure 15: MUSICAL and sensitivity to the camera noise. MUSICAL results for sample 3 acquired using
laser intensities 0.93 W cm−2 and 10.3 W cm−2 are fused as red and green channels respectively and shown here. (a) shows
the result with the samples aligned while (b) shows the fusion of MUSICAL image without sample alignment.(c,d) Standard
deviations’s map for the measured intensities using laser intensities 0.93 W cm−2 and 10.3 W cm−2 , respectively. (e,f) Mean
and standard deviations, respectively, of the measured intensities in a random image region of scMOS sensor measured without
any laser excitation. More details appear in Supplementary Note 8. Scale bars: (a-d) 1 µm; (e,f) 500 nm. Color bars in (c,d,f)
indicate standard deviation of pixel intensity values. Color bar in (e) indicate mean value of pixel intensity values.
 16

Supplementary Figure 16: Compensation for camera noise and suppression of side lobes. Results for synthetic example
SynEx4Bck are shown here. (a,b) Mean and standard deviations, respectively, of the image stack. (c) MUSICAL result. (d)
MUSICAL results with compensation of noise and suppression of side lobes using the schemes in Supplementary Note 9.
(e) scale(p) computed using eq. (30). (f) w(rtest ) computed using eq. (31). Scale bars: 500 nm. Color bars: (a) mean of
pixel intensity values; (b) standard deviation of pixel intensity values; (c,d) MUSICAL indicator function values; (e) values of
scale(p).
 17

Supplementary Figure 17: Result of side lobes suppression for in-vitro sample 2. Reduction of background using w(rtest ) of
Supplementary Note 9 in the MUSICAL result for sample 2. (a) zoom-in section of MUSICAL result for sample 2, the same
as Fig. 2(m) of the main paper. (b) the same section, but with the incorporation of w(rtest ). (c) comparison of cross-sections.
Scale bars: 1 µm.
 18

Supplementary Figure 18: Comparison between MUSICAL and STORM for in-vitro sample 1. STORM intensity profiles at
cross-sections B1-B4 in Fig. 1(f) are shown in (a). MUSICAL intensity profiles at cross-sections B1-B4 in Fig. 1(g) are shown
in (b). STORM and MUSICAL intensity profiles at cross-section C2 in Fig. 1(i,j) are shown in (c). STORM and MUSICAL
intensity profiles at cross-section C3 in Fig. 1(i,j) are shown in (d). Discussion on this figure appears in Supplementary Note
10.
 19

Supplementary Figure 19: Comparison between MUSICAL and STORM using synthetic forks similar to Fig. 1(g). Results
for synthetical experiments SynFork1 and SynFork2 which simulate a fork similar to Fig. 1(g). Top row —SynFork1;Bottom
row –SynFork2. (a,d) mean image of 1,000 frames. (b,e) STORM result (c,f) MUSICAL result. Discussion on this figure
appears in Supplementary Note 10. Scale bars: 250 nm.
 20

Supplementary Figure 20: Comparison between MUSICAL and STORM using synthetic example SynSTORM. Synthetic
example SynSTORM is used to compare MUSICAL and STORM. (a-e) correspond to fluorophores with short dark states while
(f-j) correspond to fluorophores with long dark states. (a,f) show actual emitter profiles. (b,g) show mean images of the image
stacks. (c,h) show MUSICAL result. (d,i) show STORM images and (e,j) show the points localized by STORM. Discussion
on this figure appears in Supplementary Note 10. Scale bars: 500 nm.
 21

Supplementary Figure 21: Comparison between MUSICAL and other statistical analysis method for the 3B test data
provided at [3]. Imaging results of different methods and overlay of different pairs of methods are given above. It is seen
that the shape details are better constructed by 3B (k) than any other method including MUSICAL. However, we noted that
the vertex of the polygon reconstructed by 3B highlighted using yellow arrows in (h,j-l) does not agree with the results of the
other methods. We are not sure if it is due to a local minimum in the iterative optimization of 3B or it is an accurate location
of the vertex. Scale bars: 500 nm.
 22

Supplementary Figure 22: Time lapse study for live-cell microtubules sample 1 and influence of the number of frames
used in MUSICAL. We use region B of live-cell microtubules sample 1 shown in Fig. 5 to discuss the benefit of reducing
the number of frames for live-cell studies involving dynamic molecules. MUSICAL images obtained with different numbers
of frames and in different time slots of the imaging time-line are shown above. The same lower bound criterion (0.4 times of
the 99.9% of the histogram of MUSICAL intensities) is used for all the images. MUSICAL results in the first two columns
of the top panel are significantly sharper than the remaining MUSICAL results. This is because the blurring effect due to
the local dynamics in the vicinity of the microtubule can be alleviated by using less frames. Further, the background is
relatively less cluttered in MUSICAL images with 49 frames. Higher numbers of frames imply collection of more signal from
spatially more spread traces of freely diffusing labeled molecules, and thus a more pronounced background. The bottom pane
showing MUSICAL results for increasing number of frames indicates that the structural details reconstructed by MUSICAL
start saturating at sufficiently large number of frames. Lastly, we note that the images in the first two columns of the top panel,
each using 49 frames, do not have overlapping time windows. Unsurprisingly, they differ from each other in the structural
details. This indicates the utility of MUSICAL in time lapse studies, allowing direct visualization of dynamics on small time
scales. In summary, using the smallest possible number of time frames allows for sharp, less cluttered, super-resolved images
with the potential of time lapse studies. Scale bars: 500 nm.
 23

Supplementary Figure 23: MUSICAL results for live-cell microtubules sample 2, generated using 49 frames spanning
49 ms of time. Live-cell microtubule sample 2 is imaged at a frame rate of 1,000 frames per second, i.e. 1 ms per frame. (a)
shows the mean image of 49 frames of the complete sample. (b,c) show mean and MUSICAL images of the region A. (d,e)
show mean and MUSICAL images of the region B. and (f,g) show mean and MUSICAL images of the region C. It is seen
that even in this sample, MUSICAL reconstructs sub-100 nm details. The lower bound for this example is chosen heuristically
to be 0.30 times the 99.9% of the histogram of MUSICAL intensities. The lower bound was decreased in the light of the
smaller spatial spread of diffusing molecules captured in 49 ms, as opposed to sample 1, where 49 frames correspond to 245
ms. Scale bars: (a) 1 µm; (b,c) 500 nm; (d,e) 200 nm; (f,g) 100 nm.
 24

Supplementary Figure 24: Result of MUSICAL for the live-cell cortical-cytoskeletal F-actin. The live-cell imaging of actin
cytoskeleton was performed in a cell line stably expressing Lifeact-GFP. The whole cytoskeleton is subject to fluorescence
labeling. However, due to the TIRF illumination geometry, mainly the cortical actin cytoskeleton was imaged. The cells were
illuminated by a 488 nm laser, excitation power of 25 W cm−2 , and imaged with 1 ms exposure. The MUSICAL result shown
was generated using 100 frames only, which corresponds to 100 ms. Mean image of 100 frames (a) has significantly less
details than the corresponding MUSICAL image (b). Zoom-ins of regions A and B are shown in (c,d), respectively. The results
are in good agreement with the known structures of the cortical F-actin mesh [4], [5]. Notably the geometry as well as the
mesh size agrees well with Fig. 3(g,h) of [4] and Fig. 6(d) of [5]. Scale bars: (a,b) 1 µm and (c,d) 100 nm.
 25

Technique Instrumentation Dye Number Frame Excitation Spatial Resolution

dependence dependence of frames rates (fps) Power view region
STED Yes Yes 1 ≈ 10-100 High Unlimited ≈ 20 nm
STORM No Yes ≥1000 ≈ 10-100 High Unlimited 20-30 nm
3B No No 100-1000 # Medium 100 to 10,000 pixels ≈ 50 nm
SOFI No No ≥100 # Medium Unlimited ≈ 100 nm
ESI No No ≥100 # Medium Unlimited ≈ 80 nm
MUSICAL No No ≥50 # Medium Unlimited 30-50 nm

Supplementary Table 1. Qualitative comparison of super-resolution techniques. # Limited by blinking statistics. More discussion
appears in Supplementary Note 11.

Sample Dye No. of washes∗ No. of Frame rate Laser Intensity

frames (fps) (W cm−2 )
In-vitro Sample 1 Phalloidin Atto 565 3 10,000 200 205.6
In-vitro Sample 2 Phalloidin Atto 565, tetraspeck 5 10,000 200 10.3, 40.2, 205.6
beads (40 nm diameter)
In-vitro Sample 3 Phalloidin Atto 565 0 20,000 100 0.93, 10.3, 205.6
Live-cell microtubules Sample 1 Lifeact GFP not applicable 1,000 200 25
Live-cell microtubules Sample 2 Lifeact GFP not applicable 49 1,000 25
Live-cell F-actin Lifeact GFP not applicable 100 1,000 25
∗
Supplementary Table 2. Details of samples are provided here. Number of washes indicate the number of times the sample
is washed before introducing the imaging buffer.
 26

S UPPLEMENTARY N OTE 1. C ONCEPT OF MUSICAL

We first begin with the mathematical model of the temporal image stack formed by blinking fluorophores. Then, the core
concept of MUSICAL is presented.
Mathematical model of the image of blinking fluorophores
For the ease of reference, we denote points in the image plane as r and the sample plane as r0 . For simplicity, we assume that
contribution from off-plane emitters can be ignored, such that the sample can be considered two-dimensional. The x and y
coordinates (i.e. lateral coordinates) of the points in the image plane are represented as (x, y). Similarly, the lateral coordinates
of the points in the sample plane are represented as (x0 , y 0 ). Necessary subscripts may be added as relevant.
Let there be M emitters (individual blinking fluorophores) located at r0m , m = 1 to M in the sample plane of a microscopy
system. We denote the temporal blinking state of the mth emitter as a binary signal bm (t), its emission strength as em , and
the emission wavelength as λ.
Let the image acquisition rate be given by ∆t−1 frames per second, where ∆t is the image acquisition time per frame. The
total emission from the emitter during the kth frame is given by
ˆ k∆t
sm (k) = em bm (t) dt. (1)
(k−1)∆t

In the imaging system, we denote the centers of the regularly arranged pixels in the image plane as rn ; n = 1 to N where N
is the number of pixels. The pixel is assumed to be square of width w. We interpret the point spread function (PSF) G(r, r0 )
of the imaging system as a mapping which maps the emission from a point r0 in the sample plane to the intensity at a point
r in the image plane. The intensity measured at the nth pixel is given by
(ynˆ+ w2 ) (xˆ
n+ 2 )
w
M
X
Ik (rn ) = G(r, r0 m )sm (k)dxdy , (2)
m=1
( yn − w
2 )( xn − w
2 )
We assume a diffraction limited imaging system in which the dimensions of pixels are significantly smaller than the extent of
the main lobe of the PSF of the system. For an imaging system whose PSF can be approximated as an Airy disk, this implies
that w  1.22λ/NA, where NA is the numerical aperture of the imaging system. Then, using the approximation
(ynˆ+ w2 ) (xˆ
n+ 2 )
w

G(r, r0 m )dxdy ≈ G(rn , r0m )w2 , (3)

(yn − w
2 )( xn − w
2 )
the intensity at the nth pixel can be simplified as
M
X
Ik (rn ) = w2 G(rn , r0m )sm (k). (4)
m=1

We consider a simple example to illustrate the nature of G(rn , r0m ), where rn are the centers of the discrete pixels in the image
space but the emitter locations r0m are not restricted to discrete grid points. The PSF of most incoherent imaging systems can
be represented using an Airy disk pattern
2πNA 0

!2
J 1 |r − Mr |
G(r, r0 ) = λM
, (5)
|r − Mr0 |
where M is the optical magnification of the imaging system and J1 () represents the Bessel function of 1st order and 1st kind.
Let us assume and imaging system with magnification 1 and NA=1.49. Emission wavelength of λ = 510 nm is assumed. Let
us consider one-dimensional case where y = 0 and y 0 = 0. In Supplementary Figure 5, we plot the values of G(rn , r0m ) for
xn = 6.5n µm, where integer n varies from −7 to 7 and four values of x0m , all of which correspond to the same pixel area
(namely n = 0). Thus, while G(r, r0 ) is a point-to-point mapping, G(rn , r0 ) is a point-to-pixel mapping. MUSICAL uses the
point-to-pixel mapping G(rn , r0 ).
Intensity measurements in the kth frame at all the pixels can be written as a matrix equation as follows:
Ik = Gsk , (6)
where,  T
Ik =Ik (r1 ) Ik (r2 ) . . . Ik (rN ) , (7)
G = w2 G0 (r01 ) G0 (r02 ) . . . G0 (r0M ) ,
 
(8)
T
G0 (r0 ) = G(r1 , r0 ) G(r2 , r0 ) . . . G(rN , r0 )

, (9)
 27

 T
sk = s1 (k) s2 (k) . . . sM (k) , (10)

and the superscript T denotes vector or matrix transpose. For the ease of further reference, we refer to G0 (r0m ), i.e. mapping
from a location where an emitter is actually present, as the emitter mapping vector (EMV). This is to differentiate the EMV
G0 (r0m ) from the more general vector G0 (r0 ).
Lastly, the intensity measurements from all the frames can be collected together as a 2-dimensional matrix, in which each row
corresponds to a pixel and each column corresponds to a frame. The complete image stack which contains images from time
frames k = 1 to K can be written as a matrix

I = [I1 I2 ... IK ] . (11)

The core concept of MUSICAL

The core concept of MUSICAL revolves around the mathematical range of the matrix I and its connection to the actual physical
system and sample that created the observed intensities in the matrix I (thus called the physical definition of the range of the
matrix). We discuss this concept below.
Physical definition of the range of the matrix I — It is evident from eqs. (6,11) that each vector in I is a linear combination
of the EMVs G0 (r0m ), m = 1 to M . Thus, the range of the matrix I is spanned by the EMVs R : {G0 (r0m ); m = 1 to M }.
Mathematical definition of the range of the matrix I — Singular value decomposition (SVD) of the matrix I gives the spatial
basis vectors uσi and temporal basis vectors vσi with associated singular values σi , such that I vσi = σi uσi . The range of the
matrix I is thus given by R : {uσi 6=0 }. The vectors uσi are referred to as eigenimages and simply written as ui .
Before proceeding, we consider two cases for the number of emitters M . These cases are M < min(N, K), which corresponds
to rank deficiency in the matrix I; and M ≥ min(N, K), which corresponds to the matrix I being full-ranked. MUSIC, the
method that inspired MUSICAL, considered only the rank-deficient case [6], [7], while MUSICAL deals with both the cases.
We note that the null space, which is important to both MUSIC and MUSICAL, is easily determinable in the rank deficient
case, whereas the full-ranked case does not have a well defined null space. Thus, we consider the two cases separately first
and then use a generalized notation that can be used for both the cases while applying MUSICAL.
Case 1: Rank deficient matrix I, M < min(N, K) — If the matrix I is rank deficient, then there exists a null space N : {uσi =0 }
which is orthogonal to R. The physical definition of the range, discussed above, implies that the EMVs are orthogonal to the
vectors in N, i.e.
G0 (r0m ) · uσi =0 = 0, (12)

where a · b denotes the vector dot product of the vectors a and b. Assuming that the fluorophores blink independent of
each other, the mapping from {G0 (r0m ); m = 1 to M } to {uσi 6=0 } is one-to-one, i.e. each {uσi 6=0 } can be represented as a
unique linear combination of {G0 (r0m ); m = 1 to M } and vice versa. Consequently, the mapping vectors at other location
r0 ∈
/ {r0m , m = 1 to M } have a non-zero projection on the null space N, i.e.

G0 (r0 ) · uσi =0 6= 0, if r0 ∈
/ {r0m , m = 1 to M }. (13)

Thus, an indicator function can be designed which tests the conditions in eqs. (12,13) at different test points r0 in the sample
plane.
Case 2: Full-rank matrix I, M ≥ min(N, K) — If M ≥ min(N, K), then the rank of the matrix is min(N, K), none of
the singular values is zero, and consequently a null space does not exist. However, the eigenimages ui now correspond to
different structural details of the arrangement of the fluorophores and their corresponding σi2 represent the energy (or strength)
of the eigenimage ui . We can choose first few eigenimages with the large eigenvalues as the representative of the structure
and assume that the presence of noise may corrupt the details in the eigenimages with smaller eigenvalues. Quantitatively, we
can choose a threshold σ0 such that σi < σ0 is considered small and the corresponding eigenimages can be designated as
belonging to the null space N : {uσi <σ0 } and the space orthogonal to it as the range space R : {uσi ≥σ0 }. We note that such
designation of the range and the null space through the threshold σ0 is not mathematically rigorous but practically useful.
As an example, if the fluorophores are arranged in a line, then the eigenimage with the largest eigenvalue indicates the structure
of a line, irrespective of the density of molecules along the line. Similarly with a circle. Thus, the major structure pattern of
simple arrangements can still be captured with very few eigenimages corresponding to the large singular values.
Thus, if a point r0 belongs to the structural details characterized by the designated range R, then G0 (r0 ) for such point has
zero projection on the designated null space N, and a non-zero projection otherwise.
Notational generalization of the two cases — The structural arrangement of fluorophores can be represented by the set of
eigenimages corresponding to the large singular values, which we loosely call the range and denote as R. The cardinality
#R of R is defined as the number of eigenimages in R. In the case M < min(N, K), #R is equal to M and the range is
 28

rigorously defined. In the case M ≥ min(N, K), #R is less than min(N, K) and is user designated. In either case, the null
space N is the space orthogonal to R.
An indicator function can be designed to test if G0 (r0 ) · {ui ∈ N} = 0. If this is so, the vector G0 (r0 ) belongs to the structure
indicated by R. Thus the design goals of MUSICAL are
1) to ensure that #R can be determined robustly,
2) to design a suitable indicator function which can provide super-resolution, and
3) to mitigate the effect of noise on the indicator function.
 29

S UPPLEMENTARY N OTE 2. A LGORITHMIC DETAILS OF MUSICAL

The flowchart of the multiple signal classification algorithm (MUSICAL) is given in Supplementary Figure 1. We discuss each
functional block of MUSICAL and its contribution in achieving the goals of MUSICAL.
Sliding window
Instead of considering the whole image stack at once, we consider only small spatial window around a pixel and slide this
window across all the pixels. The size of the spatial window is the approximate size of the main lobe of the point spread
function (in pixels). Thus, for a PSF resembling an Airy disk, the size of the spatial window can be taken as Nw = floor 1.22λ


NAw .
For computational convenience of associating the sliding window with a pixel, we use the closest odd number,
 
0.61λ
Nw = 1 + 2 floor . (14)
NAw
The use of floor (i.e., largest integer smaller than the value) implies that we consider the main lobe to an extent close to but
smaller than the span between the zeros.
The use of the sliding window contributes towards the goal 1. Choosing a small spatial window instead of the complete image
stack helps because there may be several such windows in which the number of fluorophores is small or zero and thus #R
is easily determinable. Even for the other windows, unless the number of particles in a window is large and the distribution
is random or complicated, few eigenimages can represent the distribution of the fluorophores in the window. Thus, #R can
be reasonably determined. Thus, the use of window directly helps in ensuring the applicability of the aforementioned core
concept for most scenarios.
The use of the sliding window also helps towards goal 3. Since the spatial window spans only the main lobe of the PSF, the
noise from the pixels beyond the window does not impact the indicator function computed using this window. Moreover, since
the fluorophores in center pixel of this window do not contribute significantly to the pixels beyond the main lobe of the PSF,
no significant information is lost in the context of the test points inside the center pixel.
For later reference, the window is slid one pixel at a time and the center pixel of the window is denoted as p. The image stack
corresponding to the window centered at p is denoted as Ip . In the context of Supplementary Note 1, I can be replaced by Ip
and related quantities are calculated with reference to Ip as the image stack.
Soft window function
Due to the use of sliding window, the indicator function is the most reliable for the test points in the center pixel. For the
pixels at the edge of the window, the effect of truncation due to the finite size of the window may be considered abrupt. Thus,
we apply a soft-window function, which scales the weights of the pixels according to their proximity to the center pixel. This
is elegantly achieved in the mathematical framework of Supplementary Note 1, as discussed next. Let the spatial distribution
of the soft window function be given as
h = [h(r1 ) h(r2 ) . . . h(rN )] , (15)
and the diagonal soft window transformation matrix be given by H = diag(h). We obtain the soft windowed image stack Jp
as
Jp = HIp . (16)
Further, all mapping vectors are correspondingly transformed as
G00 (r0 ) = HG0 (r0 ). (17)
After this step, in the context of Supplementary Note 1, I and G0 (r0 ) can be replaced by Jp and G00 (r0 ), respectively.
We have used 2-dimensional rotationally symmetric Gaussian function, which is centered at the center pixel rp and has a
spread (standard deviation) ρ = w Nw2−1
kr − rp k2
h(r) = exp . (18)
2ρ2
Multiple signal classification
SVD of Jp yields the eigenimages uσi and the corresponding singular values σi . In the original form of multiple signal
classification (MUSIC), the following indicator function was computed for the test points r0test in the sample plane
1
f (r0test ) = qP . (19)
0 0 2
σi <σ0 kG (rtest ) · ui k

The indicator function f (r0test ) demonstrates the following behavior:

if r0 test ∈ L

∞
f (r0test ) = (20)
finite otherwise.
where L represents the fluorophore structures represented by vectors in R. As a consequence, f (r0test ) is unbounded individually
for two closely placed emitters but bounded as test points between them. This allows two closely placed emitters to be
 30

distinguished. We note that even in ideal noiseless measurements, the numerical precision of the computation machine implies
that the values are never infinite in practice but limited by the computation precision.
We have modified the indicator function of MUSIC to achieve goal 2, contribute towards goal 3, and enable direct stitching
of the MUSIC images of different sliding windows. For the modification, we define the projection distances dPN (r0test ) and
dPR (r0test ) as follows s X
0 2
dPN (rtest ) = kG0 (r0test ) · ui k , (21)
σi <σ0
s X
2
dPR (r0test ) = kG0 (r0test ) · ui k , (22)
σi ≥σ0

where dPN (r0test ) and dPR (r0test ) are the projection of G0 (r0test ) on the null space N and the range R respectively. Then, we
define the indicator function for MUSIC as !α
0 dPR (r0test )
f (rtest ) = , (23)
dPN (r0test )
The behavior of the modified indicator function is similar to eq. (20). It has the following additional salient properties
• It incorporates the information of all the eigenimages (both R and N) whereas the information of R is completely ignored
in eq. (19).
0
• Through the use of dPR (rtest ), MUSIC image of each sliding window is automatically scaled with reference to the complete
image stack and thus, the MUSIC images positioned in the relevant pixel can simply be added for stitching all the MUSIC
images and obtaining the MUSICAL result.
0
• The value of α, typically chosen more than 1, determines the spread of f (rtest ) for a single fluorophore. With all the
parameters remaining the same, increasing the value of α makes the spread narrower and consequently the resolution better.
Thus, in the presence of significant amount of noise, resolution of MUSICAL may be improved by simply choosing larger
α. On the other hand, it increases the dynamic range of the MUSIC image. Unless mentioned, we have used α = 4 for
generating the MUSICAL results.
• The modified indicator function is a dimensionless quantity.

Stitching MUSIC images

Theoretically, the test points r0test can be any points in the sample region corresponding to the sliding window. But, for the
convenience of forming the final image by stitching the results of all the sliding windows, we decimate every pixel into a
sub-pixel grid with even number of grid points along each direction and take the centers of the sub-pixels as the test points.
Thus, the coordinates of the test points r0test in a pixel are given as (x0 + (0.5 + lx )/Lx , (y 0 + (0.5 + ly ))/Ly , where Lx , Ly
are the even numbers of sub-pixels per pixel along the x and y directions, lx , ly are integers with |lx | ≤ Lx /2, |ly | ≤ Ly /2,
and x0 , y 0 are the conjugate points of the pixel (x, y) in the sample plane.
Suppose we wish to compute the MUSICAL result for the qth pixel. When the sliding windows are centered at the pixels with
coordinates
xp = xq + wnx ; yp = yq + wny (24)
where |nx |, |ny | ≤ (Nw − 1)/2, then qth pixel is a part of these sliding windows and thus MUSIC’s indicator function is
computed for the test points in the qth pixel. Let us denote the indicator function computed for a sliding window with center
pixel (xp , yp ) by fxp ,yp (r0test ). Then, the MUSICAL result at r0test in the qth pixel can be obtained by
XX
F (r0 test ∈ qth pixel) = fxp ,yp (r0 test ) (25)
nx ny
 31

S UPPLEMENTARY N OTE 3. C ONCEPT OF RANGE AND NULL SPACE OF MATRIX AND RELATIONSHIP TO SINGULAR VALUE
DECOMPOSITION
Any matrix A of size m × n can be interpreted as a linear mapping from an n−dimensional input space to an m−dimensional
output space. Depending on the nature of A, although the output space is m− dimensional, the outputs may be restricted to a
smaller subspace of this m− dimensional space. Such nature of A is characterized by rank(A) < m. The actual output subspace
of the matrix A is called the range R of the matrix A and the matrix A is said to span the range R. The complementary
subspace to the range R, which is not spanned by the matrix A is called the null space N. Consider the following matrix as
an example:  
1 0 2 0
A= 0 1 0 2  (26)
1 1 2 2
It is evident that although the above matrix A maps inputs to a 3−dimensional space, the mapped outputs lie on a 2−dimensional
plane only. The rank of this matrix A is 2 and the aforementioned plane is the range of A. All the points in the 3−dimensional
space except the this plane belong to the null space. However, for the matrix A below, the null space is empty:
 
1 0 2 0
A= 0 1 0 2  (27)
1 1 1 1
In the situation that the input space is smaller than the output space, i.e. m > n, the null space is definitely not empty. For
example, if points on a line (a one dimensional input space) are linearly mapped to a two-dimensional or three-dimensional
output space, the mapped points would still be confined to a line.
Since the range of a matrix A is spanned by the linear mapping A, any point in the range can be expressed as a linear
combination of columns in A. On the other hand, any point in the null space cannot be expressed as a linear combination of
the columns of A. This implies that the entire null space is orthogonal to the columns of A as well as to the range of A.
Any point in an n− dimensional space can be represented as a linear combination of n independent vectors. Such a set of n
independent vectors is called the set of basis vectors or simply the basis vectors. Although there are infinitely many such sets,
it is often desirable to choose basis vectors with specific properties, such as the Euclidean norm of each vector being one, or
the basis vectors being orthogonal to each other, or one of the basis vectors being along the direction of maximum variance
for a cluster of measured points. Since MUSICAL exploits the range and the null space and their mutual orthogonality, it is
desirable to choose a set of basis vectors which allows for convenient identification of the range and the null space, as well
as inherently uses orthogonality. Singular value decomposition fulfils these requirement easily. This is because the singular
vectors are mutually orthogonal, form a set of basis vectors, and satisfy the following: Av̄i = σi ūi , where ūi , v̄i , and σi are
the ith left singular vector, right singular vector, and singular value of the matrix A, respectively. If σi 6= 0, it implies that ūi
can be represented as a linear combination of the columns in A, and thus ūi lies in the range. Similarly, if σi = 0, ūi cannot
be represented as a linear combination of the columns in A, and thus belongs to the null space.

S UPPLEMENTARY N OTE 4. T HE CHOICE OF σ0 AND THE EFFECT OF NOISE

The presence of noise affects both the range and the null spaces of the measurement matrix I (or the sliding window Jp ).
As a result of noise, not only the singular values σi , but the eigenimages ūi also change to a certain extent. The most direct
effect of noise is that the singular values may no longer be zero even in the rank-deficient case. Yet, the null space can be
0 0
determined by suitable selection of the threshold σ0 . The indirect impact of noise is that dPN (~rtest ) 6= 0 for ~rtest ∈ R. Thus,
the indicator function f (~rtest ) does not have a clear classification as indicated in eq. (20). It rather has a continuous spread as
0
we move away from a point ~rtest ∈ R. In this situation, higher value of α helps in reducing the spread.
We provide two rules of thumb to select the value of the threshold σ0 . If the signal to noise ratio SNR = kIk/kNk is known,
then σ0 is chosen as
σmax /σ0 = SNR (28)
where σmax is the maximum singular value observed for the singular values computed for all the sliding windows. If the
SNR is not known, then the value of σ0 is chosen to be slightly less than the value where a knee feature is observed in the
logarithmic plot of singular values of all the sliding windows. Here, knee feature is characterized by a point before which
singular values show a fast decaying characteristics and after which the singular values decay slowly.
We consider SynEx1 for the case M < min(N, K). It is seen in Supplementary Figure 3(a) that in the absence of noise, only
four singular values are non-zero while the remaining singular values drop to zero or numerical noise of the computer. They
are not shown for convenience of visualization. Thus, the null space is clearly defined in this case. The modified indicator
function with α = 1 computed for a grid of test points is plotted in Supplementary Figure 3(b) and the zoom-in of the center
pixel is shown in Supplementary Figure 3(c). The result clearly shows high values of the indicator function at the test points
closest to the emitter locations. Lastly, the use of α = 4 further reduces the background, as seen in Supplementary Figure 3(d).
 32

Next, we consider SynEx2 for the case M ≥ min(N, K). We consider the central region of the distribution, i.e. the sliding
window for the center most pixel in the image stack. This window of 7 × 7 pixels has 108 emitters. The singular values for
this window are shown in Supplementary Figure 4. It is seen that none of the singular values are zero, indicating that the
null space is not rigorously defined. However, the decay of singular values indicate that we may choose σ0 to allow different
definitions of the null space in the computation of the modified indicator function. For a high value of σ0 , the structure shown
in bottom-left of Supplementary Figure 4 can be identified using the modified indicator function. For subsequently smaller
value of σ0 , the modified indicator function shows increasingly more details.
Lastly, we consider the presence of noise in SynEx1. In the presence of noise, all the singular values are non-zero, as seen
in Supplementary Figure 3(a)). Thus, the null space cannot be clearly determined. Nevertheless, a knee can be observed after
4 singular values (indicated using gray dashed lines in Supplementary Figure 3(a)). Thus, using log10 σ0 = −1.2, where the
knee is approximately noted, the modified indicator functions are plotted using α = 1 in Supplementary Figure 3(e,f) and
using α = 4 in Supplementary Figure 3(g). The indicator function shows a continuous spread in the presence of noise and the
spread is significantly smaller for α = 4.

S UPPLEMENTARY N OTE 5. E XAMPLE OF EIGENIMAGES

We illustrate eigenimages and the corresponding eigenvalues for SynEx1 in Supplementary Figure 6. Only the first six
eigenimages are used for convenience. It is seen that the eigenimages with non-zero eigenvalues represent information about
the structure and the eigenimages with zero eigenvalues represent noise patterns. Other eigenimages also correspond to zero
eigenvalues and represent random noise patterns. If the emitters were not blinking, all the images in the image stack would
be the same and thus only one eigenvalue would be non-zero and would represent simply the image repeated in the entire
image stack. On the other hand, if the emitters would be emitting in synchronism with each other, the image stack could
be represented by only two images, one in which all the emitters are emitting and the other being the dark image. Even
in this case, only one eigenvalue would be non-zero and the corresponding eigenimage would be the image corresponding
to emissions from the emitters. It is the phenomenon of blinking that introduces diversity in the images and more than one
non-zero eigenvalues, each of which represents a different pattern from the image stack.
Supplementary Figure 6 also presents the projections of the PSFs of test points on the eigenimages. It is seen that the
projection at the actual emitter locations is quite high for the first eigenimage. In other eigenimages with non-zero eigenvalues,
the projections at the actual emitter locations are smaller than the projection for the first eigenimage but not equal to zero. The
projections at the actual emitter locations on the fifth and sixth eigenimages are zero. These properties are more obvious in the
log scale of the projections shown in the bottom row of Supplementary Figure 6. The projections of the PSF are similarly zero
at the emitter locations for all the eigenimages with zero eigenvalues. At other locations, the projections on the eigenimages
with zero eigenvalues is not consistently zero. These properties of the projections on the eigenimages in the null space are the
critical feature of MUSIC and MUSICAL and are exploited in their indicator functions.

S UPPLEMENTARY N OTE 6. D ISCUSSION ON THE ROLE OF THE PARAMETER α

Here, we discuss further on the role of the parameter α in MUSICAL. Irrespective of the definition of the indicator function and
the value of α, resolution between two features is theoretically achieved in the case of noise-free measurements if the indicator
function is finite at a test point between the features because the contrast is perfect according to eq. (20). However, practically
the indicator function is not infinite at the location of the emitters. Further, it is not necessary that the actual emitter location
is also a test point. Nevertheless, a dip in the value of the indicator at a point between two features indicates non-zero contrast
and thus computationally resolved features. However, for visual resolution, sufficient contrast is essential. Since the indicator
function of MUSICAL is inherently non-linear through the presence of dPN in the denominator, α only tweaks the amount
of non-linearity, consequently influencing three aspects of MUSICAL result. The first consequence is obvious; α obviously
controls the contrast of the MUSICAL image. The second consequence is somewhat subtle; the value of α determines the
order of non-linearity in the stitching of the MUSICAL results of sliding windows through eq. (25). The higher the order,
lower is the contribution of the off-center sliding windows to the final MUSICAL image. A small value of α, such as 1 or 2,
implies larger contribution from the off-center sliding windows which are more susceptible to noise. A large value of α implies
less contribution from the off-center sliding windows, which do contain some information about the test points. The third
consequence is that a high value of α non-linearly scales the less prominent features and makes them visually less apparent.
The features may be less prominent due to lower number of emissions or less fluctuations. On the other hand, small value
of α emphasises the non-zero and non-uniform background, such as expected in the MUSICAL indicator function due to its
characteristic in eq. (20). We heuristically found that α = 4 provides a good trade-off for all the three aspects in general.

S UPPLEMENTARY N OTE 7. R ESULTS OF IN - VITRO SAMPLE 3 FOR DIFFERENT EXCITATION POWERS

We considered another in-vitro sample, Sample 3, for which image stacks were acquired at excitation laser powers 0.93 W
cm−2 , 10.3 W cm−2 , and 205.6 W cm−2 . The results are shown in Supplementary Figure 13. The MUSICAL results show
sharper images than the mean images for all the three powers, see Supplementary Figure 13(c,f,i). Comparison of MUSICAL
 33

results at the cross-sections D1-D3 are shown in Supplementary Figure 13(j). It is notable that as the power decreases from
205.6 W cm−2 (D3) to 10.3 W cm−2 (D2), FWHM deteriorates from 38.1 nm to 248.4 nm. However, as the power decreases
from 10.3 W cm−2 to 0.93 W cm−2 , the FWHM deteriorates by only 20 nm. The FWHM of mean intensities at the cross-
sections D1 and D2 are 320 nm and 336 nm, respectively. Thus, MUSICAL result is slightly better than the unprocessed image
stack even when the excitation power is low.
Thus, collecting the results shown for in-vitro sample 2 in Fig. 2 and the result for sample 3, we form an empirical plot of
power response of MUSICAL shown in Supplementary Figure 13(k). We note that there are three factors responsible for the
observed steep curve in Supplementary Figure 13(k). As shown in Dempsey et. al [8], reduction of excitation laser intensity has
two effects that affect the captured signal directly. First, reduced intensity reduces the number of photons emitted per switching
cycle, thus deteriorating the SNR as well as the SBR of the image stack. Notably the SBR does not change significantly in
the flat regions of the curve in Supplementary Figure 13(k). Second, reduced excitation intensity reduces the off-switching
rates and thus increases the duty cycles, which translates to less dynamic variation in blinking or flatter statistics. Further, a
third effect is the survival rate of emitters, which reduces with increasing excitation power. We believe that these three effects
compete with each other to result into the empirical plot of Supplementary Figure 13(k). We note here that since the total
acquisition time duration is significantly shorter than the time needed to reach the equilibrium on-off duty cycle (400-600
seconds [8]), the duty cycle does not reach the equilibrium value, which is almost the same for all the power values.

S UPPLEMENTARY N OTE 8. MUSICAL AND SENSITIVITY TO THE CAMERA NOISE

MUSICAL result for in-vitro sample 3 provides interesting observation for low powers. Supplementary Figure 13(b,e) show
some small particle-like artifacts, with relatively less intensity than the main structures. We discuss the cause of the appearance
of these artifacts. For this, we fuse the MUSICAL images corresponding to image stacks captured with 0.93 W cm−2 and
10.3 W cm−2 and show the fused result in Supplementary Figure 15. Supplementary Figure 15(a) shows the fused images
where the MUSICAL results have been shifted to compensate for the sample drift between the two acquisitions, whereas
Supplementary Figure 15(b) shows the fusion without compensating for the sample drift. We highlight, using white arrows,
some of the artifact in the fused images. Notably in Supplementary Figure 15(b), the artifacts are at the same location but the
samples are not aligned, as seen using two distinct non-overlapping red and green casts in the sample details. On the other
hand, Supplementary Figure 15(a) shows the samples aligned but one red and one green artifact for each artifact shown in
Supplementary Figure 15(b). In other words, if the samples are aligned the artifacts do not overlap and if the artifacts are
aligned, the samples are shifted.
This indicates that the artifacts are neither sample related nor algorithm induced, because the particles would have moved with
the sample in either of these cases. Interestingly, the standard deviation maps in Supplementary Figure 15(c,d) of the image
stacks show that some background pixels, where the artifacts were observed have larger standard deviations than the other
background pixels. Upon further analysis, we found that the dark noise characteristics of the pixels of the sCMOS camera
used for measurements are not uniform despite the in-built hot-pixel correction. Particularly, some pixels exhibit significantly
larger standard deviations than the rest as seen for a randomly selected region in Supplementary Figure 15(f), although the
mean values appear relatively uniform as seen in Supplementary Figure 15(e).
We used the semi-synthetic example SynEx4Bck, which is a fusion of synthetic data and measured dark noise, see eq. (32)
in Supplementary Methods, to validate the occurrence of such artifacts. The mean and standard deviations of the image stack
are shown in Supplementary Figure 16(a,b) and demonstrate roughly similar behaviour as the mean and standard deviation
maps of Sample 3. The SBR is about 1.5, as noted from the mean image in Supplementary Figure 16(a). The MUSICAL
result in Supplementary Figure 16(c) shows particle-like artifacts in the background in addition to the pronounced effect of
side lobes due to poor SBR. We note that the maximum MUSICAL intensity for SynEx4Bck is about 72. Thus, according
to Supplementary Figure 14(e), side lobes with intensity of about 6.5 are expected, which agrees with the observed values in
Supplementary Figure 16(c).

S UPPLEMENTARY N OTE 9. C OMPENSATING FOR KNOWN NOISE PATTERN OF THE CAMERA AND SUPPRESSING SIDE LOBES
We demonstrate here a method each for compensating for known camera noise pattern and for suppressing side lobes. While
these approaches are not needed when the SBR is sufficiently high, they may be useful for low SBR.
If an image stack of the dark noise of the same image region can be captured, its singular value decomposition can be used
to suppress these artifacts provided that they are not too close to the the features of interest. Suppose the sliding window at a
pixel p gives the small image stack Ip and noise stack Np and their respective eigenimages are denoted as umeas i and unoise
i ,
then the net projection on the dark noise space onto the measurement space is given as
P P meas noise
ui · uj
i j
d= (29)
N meas N noise
where N meas and N noise are the number of eigenimages of Ip and Np , respectively. Then, a scale factor scale(p)
scale(p) = 1 − |d| (30)
 34

can be used to scale the indicator function of MUSICAL. The scale factor for SynEx4Bck is shown in Supplementary Figure
16(e).
For compensating for the side lobes, we use non-linear weighing of the MUSICAL indicator function, where the value of the
indicator function at each test point rtest is weighed by w(rtest ), a sigmoid function defined as follows
1
w(rtest ) = 
 (31)
1 + exp −β f (rtest ) − f0

where β is a control parameter that determines the rate at which the function decays. Here, we choose β = 0.25. Further, f0
is the soft threshold for the intensity. We recommend f0 to be a fraction (we have used 0.3) of the maximum intensity of the
MUSICAL result. For example, here we have used I0 = 20 for SynEx4Bck, which is about 0.3 times the maximum MUSICAL
intensity. The weights thus calculated for SynEx4Bck are plotted in Supplementary Figure 16(f). As seen in Supplementary
Figure 16(f), the indicator function does not change much if its value is more than 40.
For SynEx4Bck, the MUSICAL result after incorporating both scale(p) and w(rtest ) is shown in Supplementary Figure 16(d).
Both the artifacts and the side lobes are significantly suppressed in this result as compared to the original MUSICAL image
shown in Supplementary Figure 16(c).
As a further example, we show the effect of w(rtest ) on the MUSICAL image of sample 2 in Fig. 2(m) of the main paper.
We use β = 1 and f0 = 6.85 which is approximately 0.3 times the maximum MUSICAL intensity. The results are shown
in Supplementary Figure 17. The MUSICAL image with w(rtest ) has a significantly lower background as evident in both
Supplementary Figure 17(b,c).

S UPPLEMENTARY N OTE 10. C OMPARISON BETWEEN MUSICAL AND SMLM TECHNIQUES

The intensity plots of STORM and MUSICAL results at the cross-sections B1-B4 of region B in in-vitro sample 1 shown
in Fig. 1(f,g) are given in Supplementary Figure 18(a,b). The intensity plots of STORM and MUSICAL at the cross-sections
C2-C3 of region C of sample 1 shown in Fig. 1(i,j) are given in Supplementary Figure 18(c,d).
The cross-sections B1-B4 represent a situation of branching of a filament into two, where B1 shows clearly separated branches,
B2 shows a line very close to the junction, B3 is approximately on the junction, and B4 is the unbranched filament. In
Supplementary Figure 18(a,b), the STORM and MUSICAL results at sections B1 and B4 point to similar inferences, i.e. two
clearly separated branches for B1 and an unbranched filament for B4. However, MUSICAL and STORM results at B2 show
quite different results. STORM shows approximately 5 peaks at about 193 nm, 233 nm, 320 nm, 380 nm, and 493 nm, while
MUSICAL shows three peaks at 161 nm, 228 nm, and 309 nm. It is indeed difficult to conclude the structure at B2. The
results of MUSICAL and STORM for B3 are also interesting. MUSICAL shows two separable but closely located maxima
(separated by about 33 nm) with the minimum between them located at about 250 nm. On the other hand, STORM shows a
non-smooth plateau spread from about 220 nm to 320 nm.
STORM and MUSICAL results for region C are shown in Fig. 1(i,j). MUSICAL clearly indicates a branching but STORM
does not. In the cross-section plots for C2 (just before branching, Supplementary Figure 18(c)) and C3 (after branching,
Supplementary Figure 18(d)), it is seen that STORM gives only one maximum while MUSICAL gives two maxima for C3.
Also, the maximum of STORM for C3 is closer to the right side maximum generated by MUSICAL.
For further analysis, we use synthetic examples SynFork1 and SynFork2 to emulate forks such as seen in regions B and
C of sample 1. SynFork1 and SynFork2 have similar geometry but different blinking rates. Details about these appear in
Supplementary Methods. It suffices to mention here that SynFork2 has significantly slow blinking than SynFork1. The mean
images of the image stack, STORM result, and MUSICAL result are respectively shown in Supplementary Figure 19(a-c) for
SynFork1 and Supplementary Figure 19(d-f) for SynFork2. It is seen that STORM can reconstruct the fork with good fidelity
only when the blinking is extremely sparse. The nature of STORM result for SynFork1 with speckles between the prongs
matches with the STORM result for the fork in Fig. 1(f).
To illustrate the effect of presence of emitters placed closely in a small volume on STORM and MUSICAL, we consider
synthetic example SynSTORM, the details of which are provided in Supplementary Methods and the layout is shown in
Supplementary Figure 20(a,f). The width of each of the five lines is about 7 nm such that each line mimics an actin filament.
The two closest and parallel lines are separated by 50 nm. The top row corresponds to short dark states (τon = 0.05s, τoff = 50s)
while the bottom row corresponds to long dark states (τon = 0.05s, τoff = 250s), where τon and τoff represent average time
spent by molecules in the bright (on) and dark (off) states respectively.
In addition to the imaging result of STORM shown in Supplementary Figure 20(d,i), we also provide localization results of
STORM in Supplementary Figure 20(e,j). In the case of long dark states, the number of bright fluorophores at a given time
is small even at the junction. Thus, in comparison to MUSICAL result (Supplementary Figure 20(h)), the performance of
STORM (Supplementary Figure 20(i)) is superior in this case. On the other hand, in the case of shorter dark states, the number
of bright fluorophores in a small volume along the parallel lines is sufficiently high to render STORM ineffective in resolving
the lines (Supplementary Figure 20(d)). MUSICAL can, however, still resolve the lines. Further, the MUSICAL result and the
details reconstructed at the lines and junction are similar in both the cases.
 35

S UPPLEMENTARY N OTE 11. C OMPARISON OF THE COMPUTATIONAL ASPECTS OF SUPER - RESOLUTION TECHNIQUES
We present a succinct juxtaposition of the computational aspects of MUSICAL with the other methods in Supplementary Table
1. While the number of frames required by MUSICAL is very small, SOFI and the single iteration ESI have similarly low
requirement; the basic condition for all three is capturing a statistically significant sample of blinking. On the other hand,
MUSICAL allows any scale of sub-pixelation without additional iterations unlike those in ESI or the cross-correlations in
SOFI. MUSICAL scores better than 3B in computation time but 3B outperforms MUSICAL in image sharpness as the signal
to noise ratio becomes poorer, due to its iterative rejection of less likely localization candidates. Further, as compared to SOFI,
ESI, as well as 3B, MUSICAL requires the PSF to be known with sufficient accuracy although its sensitivity to the error in
PSF estimation is reduced by choosing a sufficiently high value of α. The only aspect where MUSICAL is somewhat heuristic
is in the choice of its control parameter σ0 . However, the rules of thumb noted in Supplementary Note 4 are sufficient in most
cases. Further, the appearance of grid artifacts despite using the modified indicator function and α ≥ 2 is a good indication of
an inappropriate choice of the value of σ0 .
MUSICAL and Spatial Covariance Reconstructive (SCORE) Super-Resolution Fluorescence Microscopy
Spatial Covariance Reconstructive (SCORE) Super-Resolution Fluorescence Microscopy [9] is a recent method that also uses
eigenimages, though quite differently from MUSICAL. It provides resolution comparable to SOFI and ESI. Here, we highlight
some important conceptual differences between SCORE and MUSICAL:
• Score uses only those eigenimages that belong to the range of the measurement matrix. On the other hand, the use of
eigenimage in the null space is the core of MUSICAL and is even more important than the use of eigenimages in the
range.
• MUSICAL uses the information of PSF not only for computing the projections on the range and the null space but also
to incorporate sliding window function, which is critical in reducing the impact of noise and other emitters distant from
the emitters in the sliding window. SCORE does not use any sliding window; its use of PSF is limited to the computation
of the projections onto the eigenimages in the range.
• SCORE casts the reconstruction as a minimization problem of minimizing the difference between the measured and
estimated covariances after estimating an initial covariance matrix using an exponential indicator function of the projection.
MUSICAL does not need to solve an iterative optimization and is expected to be more computation efficient for this reason.
 36

S UPPLEMENTARY M ETHODS .
Implementation and values of control parameters of MUSICAL and other methods
Here, we discuss the details of the implementations and values of control parameters for all the methods used in this paper
for reporting results.
MUSICAL — MUSICAL has been implemented on Matlab 2012b. Parallelization has not been implemented. For generating
the results in the papers, camera’s offset value of 80 is subtracted for experimental images and synthetic examples for which the
signal to background ratio is defined. Gaussian soft window of size 7×7 pixels is used. The standard deviation of the Gaussian
soft window is 3 pixels. α = 4 is used. The signal to background ratio of the mean image after subtraction of the offset was
used as the SNR estimate for selecting σ0 as discussed in Supplementary Note 4. If the result appeared unsatisfactory, the
knee criterion was used for generating the MUSICAL result. For synthetic examples without noise, the knee criterion is used.
The point spread function used in MUSICAL is assumed to be the standard Airy disk computed using the specified emission
wavelength, numerical aperture, magnification, and pixel size.
STORM [10] — The rainSTORM Matlab code provided by the authors of [11] was used for generating the STORM images.
Least squares based fitting of Gaussian function along x and y axes is done independently on the pixels which show non-zero
difference from all immediate neighbors. We chose a region of interest of 3 pixels, which translates to using 7 pixels along either
direction for fitting Gaussian function. We tried values 4 and 5 for region of interest as well, but found that using 3 as region
of interest gave sharper images. Further, we chose an estimate of the standard deviation of the Gaussian function as 2.9 pixels,
which corresponds to 188.5 nm. However, as noted in the user guide, the choice of this parameters is almost inconsequential
since the range of standard deviation of Gaussian function is quite large. The tolerance is set to 0.2, as recommended in the
user guide. The number of iterations for fitting the Gaussian function is set to 10, which is more than the recommended value
6. For Fig. 3(b) of the main paper, the STORM and PALM results provided with the dataset are used. Suitable colorbar has
been used for better and brighter visualization of these results.
DeconSTORM [12] — Matlab implementation of deconSTROM provided by the authors of [12] on the project webpage [13]
is used. Default values were used for all the control parameters except the parameter β, i.e. the probability of an inactive
emitter will become active in the next frame. Relevant optical parameters of the experiment were provided. Different values
in the range β ∈ [6.5 × 10−5 , 0.65] were attempted and the best result was reported. β ∼ 6.5 × 10−5 suited the tubulin long
sequence data in Data-SMLM because of very sparse blinking. β ∼ 0.65 suited the tubulin high density data in Data-SMLM
and SynEx2 due to non-sparse blinking. β ∼ 0.0065 was used for the remaining data. Sub-pixelation by a factor of 8 was
used.
3B [14] — The 32-bit ImageJ plugin for 3B provided by the authors of [14] on their project page [3] was used for 3B
computation. The default hardcoded values of the control parameters were used. Relevant optical parameters of the experiment
were provided. For generating the 3B image using the localization results of 3B, reconstructed pixel size of 3.25 nm (sup-
pixelation by a factor of 20) and reconstruction blur PSF of 30 nm was used. Results in Fig. 4(a,b) and Supplementary Figure
21 are obtained after 50 iterations and results in Fig. 4(c,d) of the main paper are obtained after 65 iterations.
ESI [15] — Fiji plugin provided by the authors of [15] was used for ESI computation. For the results generated in Fig. 4 of
the main paper, two iterations of order √ 2 each were run. The number of images binned in the first iteration was a factor of
the number of frames K closest to the K. Number of bins for entropy was set as 100. Only 1 iteration of ESI was used to
generate the result. Logarithmic scale was used for Fig. 4(c,d) of the main paper.
SOFI [16] — Matlab code provided by Dr. Joerg Enderlein for the article [16] was used for SOFI. Cumulants of order 5 were
used for Fig. 4(a,b) and Supplementary Figure 21 and order 6 were used for Fig. 4(c,d) of the main paper. Sub-pixelation by
a factor of 5 was used.
Color scale used for all super-resolution methods — The non-linearity of all the super-resolution methods mandates the use
of a heuristic range of intensities for better visualization of the results. We have clipped the maximum values of the intensity
in the following manner. For a result, the cumulative histogram is computed using 10000 uniformly sized bins. The bin at
which the cumulative histogram reaches 99.9% is identified. The colorbar of the result is clipped at the center of this bin.
Thus, only 0.1% pixels in the result have intensity higher than the maximum intensity of the colorbar used for visualization.
For MUSICAL, typically 99.96% is used instead of 99.9%.
Machine used for computation — The results of all the methods were generated using a Dell Precision T5610 workstation
with 64-bit Intel Xeon CPU E5-2630 v2 2.60 GHz and 64 GB RAM. The 32-bit ImageJ plugin of 3B [14] was used, however
on the same system. The multicore ImageJ implementation of ESI [15] was used. No parallelization was incorporated for
SOFI, deconSTORM, and MUSICAL.
 37

Synthetic examples
SynEx1: This example consists of four emitters arranged at the corners of a square of size 30 nm whose corners lie on the
x and y axes. For this example, we assume that the image is captured on 7 × 7 pixel array only, which corresponds to Nw in
eq. (14). Its image stack contains K = 49 frames. Since the pixel array is of size 7 × 7 only, there is only one sliding window.
This example is used to illustrate the effect of noise and the value of α on the modified indicator function.
SynEx2: This example consists of four circles, each of radius 500 nm. Each circle has 180 emitters uniformly located along
the circumference of the circle. The centers of the circles are located at the four corners of a square such that the distance
between the diagonal corners is 1100 nm and the corners lie on the axes. All the other parameters of the measurement system
are the same as SynEx1, except that each image in the temporal image stack is of size 51 × 51 pixels. This example is used
to illustrate MUSICAL imaging and the effect of the threshold σ0 for the case M ≥ min(N, K).
SynEx3: SynEx3 has 500 emitters arranged along a line of 200 nm. The image stack has K = 1000 frames. It is used to
quantify full width at half maximum (FWHM) and side lobe level as a function of signal to background ratio.
SynEx4: SynEx4 has exactly the same details as SynEx2, except that the image stack of SynEx4 has 1000 frames. Then, an
image stack, BckEx4, captured for a random region in the actual sCMOS sensor without any laser excitation is fused with the
SynEx4 to form a new image stack SynEx4Bck as below
SynEx4(p, k)
SynEx4Bck(p, k) = 100 + BckEx4(p, k) (32)
max
0 0
SynEx4(p0 , k 0 )
p ,k

where p, k denote a pixel and a frame, respectively. Here, 100 is used since the average intensity value of the image stack
BckEx4 is 100. Thus, SynEx4Bck has SBR 2. Only SynEx4Bck is used. Results for this example appear in Supplementary
Note 8.
SynPairDelX: This set of synthetic examples consist of a sample containing two lines of 500 nm each. Each line contains 51
emitters placed uniformly at a separation of 10 nm. The distance between the two lines is characterized by ∆x, as shown in
Fig. 1(e) of the main paper. One image stack of 1000 frames is generated for each value of ∆x, which varies between 20 nm
and 60 nm. The file with name SynPairDelX500Angstrom represents the image stack with ∆x = 500 Å= 50nm, i.e. . The
image stacks are noise free.
SynPeriod: SynPeriod which has 33 emitters placed along 3 lines which are 100 nm apart. Each line has 11 emitters uniformly
placed along a length of 500 nm, consecutive emitters being 50 nm apart. Example contains 1,000 frames captured at a frame
rate of 200 frames per second and the SBR is 100.
SynFork1 and SynFork2: SynFork1 and SynFork2 have a geometry similar to the fork seen in Fig. 1(b3) of the main paper.
In each of them, the fork has two prongs with an angle of 30◦ between them attached to a stem; each prong and the stem are
500 nm long and have uniformly distributed 200 emitters. Image stack contains K = 1000 frames and has SBR 10.
SynSTORM: Image stacks for the SynSTORM are simulated using Matlab code testSTORM, provided by the authors of
[11]. Photon emission rate of 100,000 photons sec−1 , average time of 50,000 seconds before bleaching occurs, and emission
wavelength of 510 nm are used. The parameters of the imaging system match the experimental system. No background is
assumed (background set to 0). The image stacks contain 3000 frames each, taken at 20 frames per second.
Simulation of blinking: SynPairDelX image stacks, SynEx1, SynEx2, SynEx3, SynEx4, SynPeriod, SynFork1 and SynFork2
are generated assuming Poisson blinking model, where the length of switching times ton and toff are computed using a Poisson
distribution with the average on time τon and the average off time τoff , respectively. We have used τon = 0.3 ms, τoff = 0.7 ms
for SynEx1 and SynPeriod; τon = 3 ms, τoff = 7 ms for SynEx2, SynEx3, and SynEx4; τon = 2 ms, τoff = 998 ms for
SynFork1; τon = 0.2 ms, τoff = 999.8 ms for SynFork2; and τon = 0.3 ms, τoff = 999.7 ms for SynPair stacks. The photon
emission rate is assumed to be 100,000 photons sec−1 .
Simulation of images: All the examples assume an emission wavelength λ = 510nm, a TIRF system of NA =1.49,
magnification 100X, and sCMOS camera with pixel size 65 nm.
Simulation of noise: For the example SynEx2 used in Supplementary Note 11, we add camera’s offset value of 100 and add
Poisson noise. In order to achieve a desired SBR, we scale the intensities in SynEx2 between 0 and 100(SBR-1) and add the
imaging system’s offset value of 100 to the scaled intensities. Then Poisson noise is added to each intensity value. Similarly,
the offset and Poisson noise is added to SynEx3. Offset value of 100, SBR value of 10, and Poisson noise is used for SynFork1
and SynFork2.

S UPPLEMENTARY R EFERENCES
[1] Born, M. & Wolf, E. Principles of optics: electromagnetic theory of propagation, interference and diffraction of light (Cambridge university press,
1999).
[2] Hou, S., Solna, K. & Zhao, H. A direct imaging algorithm for extended targets. Inverse Problems 22, 1151 (2006).
[3] Cox, S. 3b microscopy analysis software (2013). URL http://www.coxphysics.com/3b/#download.
[4] Bovellan, M. et al. Cellular control of cortical actin nucleation. Current Biology 24, 1628–1635 (2014).
[5] Charras, G. T., Hu, C.-K., Coughlin, M. & Mitchison, T. J. Reassembly of contractile actin cortex in cell blebs. The Journal of cell biology 175, 477–490
(2006).
 38

[6] Chen, X. & Agarwal, K. MUSIC algorithm for two-dimensional inverse problems with special characteristics of cylinders. IEEE Transactions on
Antennas and Propagation 56, 1808–1812 (2008).
[7] Agarwal, K. & Chen, X. Applicability of MUSIC-type imaging in two-dimensional electromagnetic inverse problems. IEEE Transactions on Antennas
and Propagation 56, 3217–3223 (2008).
[8] Dempsey, G. T., Vaughan, J. C., Chen, K. H., Bates, M. & Zhuang, X. Evaluation of fluorophores for optimal performance in localization-based
super-resolution imaging. Nature Methods 8, 1027–1036 (2011).
[9] Deng, Y., Sun, M., Lin, P.-H., Ma, J. & Shaevitz, J. W. Spatial covariance reconstructive (score) super-resolution fluorescence microscopy. PloS one 9,
e94807 (2014).
[10] Rust, M. J., Bates, M. & Zhuang, X. Sub-diffraction-limit imaging by stochastic optical reconstruction microscopy (STORM). Nature Methods 3,
793–796 (2006).
[11] Rees, E. J. et al. Blind assessment of localisation microscope image resolution. Optical Nanoscopy 1, 1–10 (2012).
[12] Mukamel, E. A., Babcock, H. & Zhuang, X. Statistical deconvolution for superresolution fluorescence microscopy. Biophysical journal 102, 2391–2400
(2012).
[13] Zhuang, X. Deconstorm: software for analyzing storm data (2012). URL http://zhuang.harvard.edu/decon_storm.html.
[14] Cox, S. et al. Bayesian localization microscopy reveals nanoscale podosome dynamics. Nature Methods 9, 195–200 (2012).
[15] Yahiatene, I., Hennig, S., Müller, M. & Huser, T. Entropy-based super-resolution imaging (ESI): From disorder to fine detail. ACS Photonics 2,
1049–1056 (2015).
[16] Dertinger, T., Colyer, R., Iyer, G., Weiss, S. & Enderlein, J. Fast, background-free, 3d super-resolution optical fluctuation imaging (SOFI). Proceedings
of the National Academy of Sciences 106, 22287–22292 (2009).